linux_dsm_epyc7002/drivers/gpu/drm/i915/gt/intel_breadcrumbs.c

393 lines
11 KiB
C
Raw Normal View History

drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
/*
* Copyright © 2015 Intel Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*
*/
#include <linux/kthread.h>
drm/i915: Seal races between async GPU cancellation, retirement and signaling Currently there is an underlying assumption that i915_request_unsubmit() is synchronous wrt the GPU -- that is the request is no longer in flight as we remove it. In the near future that may change, and this may upset our signaling as we can process an interrupt for that request while it is no longer in flight. CPU0 CPU1 intel_engine_breadcrumbs_irq (queue request completion) i915_request_cancel_signaling ... ... i915_request_enable_signaling dma_fence_signal Hence in the time it took us to drop the lock to signal the request, a preemption event may have occurred and re-queued the request. In the process, that request would have seen I915_FENCE_FLAG_SIGNAL clear and so reused the rq->signal_link that was in use on CPU0, leading to bad pointer chasing in intel_engine_breadcrumbs_irq. A related issue was that if someone started listening for a signal on a completed but no longer in-flight request, we missed the opportunity to immediately signal that request. Furthermore, as intel_contexts may be immediately released during request retirement, in order to be entirely sure that intel_engine_breadcrumbs_irq may no longer dereference the intel_context (ce->signals and ce->signal_link), we must wait for irq spinlock. In order to prevent the race, we use a bit in the fence.flags to signal the transfer onto the signal list inside intel_engine_breadcrumbs_irq. For simplicity, we use the DMA_FENCE_FLAG_SIGNALED_BIT as it then quickly signals to any outside observer that the fence is indeed signaled. v2: Sketch out potential dma-fence API for manual signaling v3: And the test_and_set_bit() Fixes: 52c0fdb25c7c ("drm/i915: Replace global breadcrumbs with per-context interrupt tracking") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190508112452.18942-1-chris@chris-wilson.co.uk
2019-05-08 18:24:52 +07:00
#include <trace/events/dma_fence.h>
#include <uapi/linux/sched/types.h>
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
#include "i915_drv.h"
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
static void irq_enable(struct intel_engine_cs *engine)
{
if (!engine->irq_enable)
return;
/* Caller disables interrupts */
spin_lock(&engine->i915->irq_lock);
engine->irq_enable(engine);
spin_unlock(&engine->i915->irq_lock);
}
static void irq_disable(struct intel_engine_cs *engine)
{
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
if (!engine->irq_disable)
return;
/* Caller disables interrupts */
spin_lock(&engine->i915->irq_lock);
engine->irq_disable(engine);
spin_unlock(&engine->i915->irq_lock);
}
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
static void __intel_breadcrumbs_disarm_irq(struct intel_breadcrumbs *b)
{
lockdep_assert_held(&b->irq_lock);
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
GEM_BUG_ON(!b->irq_enabled);
if (!--b->irq_enabled)
irq_disable(container_of(b,
struct intel_engine_cs,
breadcrumbs));
b->irq_armed = false;
}
void intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine)
{
struct intel_breadcrumbs *b = &engine->breadcrumbs;
if (!b->irq_armed)
return;
spin_lock_irq(&b->irq_lock);
if (b->irq_armed)
__intel_breadcrumbs_disarm_irq(b);
spin_unlock_irq(&b->irq_lock);
}
static inline bool __request_completed(const struct i915_request *rq)
{
return i915_seqno_passed(__hwsp_seqno(rq), rq->fence.seqno);
}
__maybe_unused static bool
check_signal_order(struct intel_context *ce, struct i915_request *rq)
{
if (!list_is_last(&rq->signal_link, &ce->signals) &&
i915_seqno_passed(rq->fence.seqno,
list_next_entry(rq, signal_link)->fence.seqno))
return false;
if (!list_is_first(&rq->signal_link, &ce->signals) &&
i915_seqno_passed(list_prev_entry(rq, signal_link)->fence.seqno,
rq->fence.seqno))
return false;
return true;
}
drm/i915: Seal races between async GPU cancellation, retirement and signaling Currently there is an underlying assumption that i915_request_unsubmit() is synchronous wrt the GPU -- that is the request is no longer in flight as we remove it. In the near future that may change, and this may upset our signaling as we can process an interrupt for that request while it is no longer in flight. CPU0 CPU1 intel_engine_breadcrumbs_irq (queue request completion) i915_request_cancel_signaling ... ... i915_request_enable_signaling dma_fence_signal Hence in the time it took us to drop the lock to signal the request, a preemption event may have occurred and re-queued the request. In the process, that request would have seen I915_FENCE_FLAG_SIGNAL clear and so reused the rq->signal_link that was in use on CPU0, leading to bad pointer chasing in intel_engine_breadcrumbs_irq. A related issue was that if someone started listening for a signal on a completed but no longer in-flight request, we missed the opportunity to immediately signal that request. Furthermore, as intel_contexts may be immediately released during request retirement, in order to be entirely sure that intel_engine_breadcrumbs_irq may no longer dereference the intel_context (ce->signals and ce->signal_link), we must wait for irq spinlock. In order to prevent the race, we use a bit in the fence.flags to signal the transfer onto the signal list inside intel_engine_breadcrumbs_irq. For simplicity, we use the DMA_FENCE_FLAG_SIGNALED_BIT as it then quickly signals to any outside observer that the fence is indeed signaled. v2: Sketch out potential dma-fence API for manual signaling v3: And the test_and_set_bit() Fixes: 52c0fdb25c7c ("drm/i915: Replace global breadcrumbs with per-context interrupt tracking") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190508112452.18942-1-chris@chris-wilson.co.uk
2019-05-08 18:24:52 +07:00
static bool
__dma_fence_signal(struct dma_fence *fence)
{
return !test_and_set_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags);
}
static void
__dma_fence_signal__timestamp(struct dma_fence *fence, ktime_t timestamp)
{
fence->timestamp = timestamp;
set_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &fence->flags);
trace_dma_fence_signaled(fence);
}
static void
__dma_fence_signal__notify(struct dma_fence *fence)
{
struct dma_fence_cb *cur, *tmp;
lockdep_assert_held(fence->lock);
lockdep_assert_irqs_disabled();
list_for_each_entry_safe(cur, tmp, &fence->cb_list, node) {
INIT_LIST_HEAD(&cur->node);
cur->func(fence, cur);
}
INIT_LIST_HEAD(&fence->cb_list);
}
void intel_engine_breadcrumbs_irq(struct intel_engine_cs *engine)
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
{
struct intel_breadcrumbs *b = &engine->breadcrumbs;
drm/i915: Seal races between async GPU cancellation, retirement and signaling Currently there is an underlying assumption that i915_request_unsubmit() is synchronous wrt the GPU -- that is the request is no longer in flight as we remove it. In the near future that may change, and this may upset our signaling as we can process an interrupt for that request while it is no longer in flight. CPU0 CPU1 intel_engine_breadcrumbs_irq (queue request completion) i915_request_cancel_signaling ... ... i915_request_enable_signaling dma_fence_signal Hence in the time it took us to drop the lock to signal the request, a preemption event may have occurred and re-queued the request. In the process, that request would have seen I915_FENCE_FLAG_SIGNAL clear and so reused the rq->signal_link that was in use on CPU0, leading to bad pointer chasing in intel_engine_breadcrumbs_irq. A related issue was that if someone started listening for a signal on a completed but no longer in-flight request, we missed the opportunity to immediately signal that request. Furthermore, as intel_contexts may be immediately released during request retirement, in order to be entirely sure that intel_engine_breadcrumbs_irq may no longer dereference the intel_context (ce->signals and ce->signal_link), we must wait for irq spinlock. In order to prevent the race, we use a bit in the fence.flags to signal the transfer onto the signal list inside intel_engine_breadcrumbs_irq. For simplicity, we use the DMA_FENCE_FLAG_SIGNALED_BIT as it then quickly signals to any outside observer that the fence is indeed signaled. v2: Sketch out potential dma-fence API for manual signaling v3: And the test_and_set_bit() Fixes: 52c0fdb25c7c ("drm/i915: Replace global breadcrumbs with per-context interrupt tracking") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190508112452.18942-1-chris@chris-wilson.co.uk
2019-05-08 18:24:52 +07:00
const ktime_t timestamp = ktime_get();
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
struct intel_context *ce, *cn;
struct list_head *pos, *next;
LIST_HEAD(signal);
spin_lock(&b->irq_lock);
if (b->irq_armed && list_empty(&b->signalers))
__intel_breadcrumbs_disarm_irq(b);
list_for_each_entry_safe(ce, cn, &b->signalers, signal_link) {
GEM_BUG_ON(list_empty(&ce->signals));
list_for_each_safe(pos, next, &ce->signals) {
struct i915_request *rq =
list_entry(pos, typeof(*rq), signal_link);
GEM_BUG_ON(!check_signal_order(ce, rq));
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
if (!__request_completed(rq))
break;
GEM_BUG_ON(!test_bit(I915_FENCE_FLAG_SIGNAL,
&rq->fence.flags));
drm/i915: Seal races between async GPU cancellation, retirement and signaling Currently there is an underlying assumption that i915_request_unsubmit() is synchronous wrt the GPU -- that is the request is no longer in flight as we remove it. In the near future that may change, and this may upset our signaling as we can process an interrupt for that request while it is no longer in flight. CPU0 CPU1 intel_engine_breadcrumbs_irq (queue request completion) i915_request_cancel_signaling ... ... i915_request_enable_signaling dma_fence_signal Hence in the time it took us to drop the lock to signal the request, a preemption event may have occurred and re-queued the request. In the process, that request would have seen I915_FENCE_FLAG_SIGNAL clear and so reused the rq->signal_link that was in use on CPU0, leading to bad pointer chasing in intel_engine_breadcrumbs_irq. A related issue was that if someone started listening for a signal on a completed but no longer in-flight request, we missed the opportunity to immediately signal that request. Furthermore, as intel_contexts may be immediately released during request retirement, in order to be entirely sure that intel_engine_breadcrumbs_irq may no longer dereference the intel_context (ce->signals and ce->signal_link), we must wait for irq spinlock. In order to prevent the race, we use a bit in the fence.flags to signal the transfer onto the signal list inside intel_engine_breadcrumbs_irq. For simplicity, we use the DMA_FENCE_FLAG_SIGNALED_BIT as it then quickly signals to any outside observer that the fence is indeed signaled. v2: Sketch out potential dma-fence API for manual signaling v3: And the test_and_set_bit() Fixes: 52c0fdb25c7c ("drm/i915: Replace global breadcrumbs with per-context interrupt tracking") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190508112452.18942-1-chris@chris-wilson.co.uk
2019-05-08 18:24:52 +07:00
clear_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
if (!__dma_fence_signal(&rq->fence))
continue;
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
/*
* Queue for execution after dropping the signaling
* spinlock as the callback chain may end up adding
* more signalers to the same context or engine.
*/
i915_request_get(rq);
list_add_tail(&rq->signal_link, &signal);
}
/*
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
* We process the list deletion in bulk, only using a list_add
* (not list_move) above but keeping the status of
* rq->signal_link known with the I915_FENCE_FLAG_SIGNAL bit.
*/
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
if (!list_is_first(pos, &ce->signals)) {
/* Advance the list to the first incomplete request */
__list_del_many(&ce->signals, pos);
if (&ce->signals == pos) /* now empty */
list_del_init(&ce->signal_link);
}
}
spin_unlock(&b->irq_lock);
list_for_each_safe(pos, next, &signal) {
struct i915_request *rq =
list_entry(pos, typeof(*rq), signal_link);
drm/i915: Seal races between async GPU cancellation, retirement and signaling Currently there is an underlying assumption that i915_request_unsubmit() is synchronous wrt the GPU -- that is the request is no longer in flight as we remove it. In the near future that may change, and this may upset our signaling as we can process an interrupt for that request while it is no longer in flight. CPU0 CPU1 intel_engine_breadcrumbs_irq (queue request completion) i915_request_cancel_signaling ... ... i915_request_enable_signaling dma_fence_signal Hence in the time it took us to drop the lock to signal the request, a preemption event may have occurred and re-queued the request. In the process, that request would have seen I915_FENCE_FLAG_SIGNAL clear and so reused the rq->signal_link that was in use on CPU0, leading to bad pointer chasing in intel_engine_breadcrumbs_irq. A related issue was that if someone started listening for a signal on a completed but no longer in-flight request, we missed the opportunity to immediately signal that request. Furthermore, as intel_contexts may be immediately released during request retirement, in order to be entirely sure that intel_engine_breadcrumbs_irq may no longer dereference the intel_context (ce->signals and ce->signal_link), we must wait for irq spinlock. In order to prevent the race, we use a bit in the fence.flags to signal the transfer onto the signal list inside intel_engine_breadcrumbs_irq. For simplicity, we use the DMA_FENCE_FLAG_SIGNALED_BIT as it then quickly signals to any outside observer that the fence is indeed signaled. v2: Sketch out potential dma-fence API for manual signaling v3: And the test_and_set_bit() Fixes: 52c0fdb25c7c ("drm/i915: Replace global breadcrumbs with per-context interrupt tracking") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190508112452.18942-1-chris@chris-wilson.co.uk
2019-05-08 18:24:52 +07:00
__dma_fence_signal__timestamp(&rq->fence, timestamp);
spin_lock(&rq->lock);
__dma_fence_signal__notify(&rq->fence);
spin_unlock(&rq->lock);
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
i915_request_put(rq);
}
}
void intel_engine_signal_breadcrumbs(struct intel_engine_cs *engine)
{
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
local_irq_disable();
intel_engine_breadcrumbs_irq(engine);
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
local_irq_enable();
}
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
static void signal_irq_work(struct irq_work *work)
{
struct intel_engine_cs *engine =
container_of(work, typeof(*engine), breadcrumbs.irq_work);
intel_engine_breadcrumbs_irq(engine);
}
drm/i915/guc: Always enable the breadcrumbs irq The execlists emulation on top of the GuC (used for scheduling and preemption) depends on the MI_USER_INTERRUPT for its notifications and tasklet action. As we always employ the irq, there is no advantage in ever disabling it while we are using the GuC, so allow us to arm the breadcrumb irq when enabling GuC submission and disarm upon disabling. The impact should be lessened by the delayed irq disabling we do (we only disable after receiving an interrupt for which no one was wanting), but allowing guc to explicitly manage the irq in relation to itself is simpler and prevents an issue with losing an interrupt for preemption as it is not coupled to an active request. Internally, we add a reference counter (breadcrumbs.irq_enabled) as a simple mechanism to allow GuC to keep the breadcrumb irq enabled. To improve upon always enabling the irq while guc is selected, we need to hook into the parking facility of intel_engines so that we only enable the breadcrumbs while the GT is active (one step better would be to individually park/unpark each engine). In effect, this means that we keep the breadcrumb irq always enabled for the entire duration the guc is busy, whereas before we would try to switch it off whenever we idled for more than interrupt with no associated waiters. The difference *should* be negligible in practice! v2: Stop abusing fence signaling (and its auxiliary data structures) to enable the breadcrumbs irqs. Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Cc: Michał Winiarski <michal.winiarski@intel.com>, Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> Reviewed-by: Michał Winiarski <michal.winiarski@intel.com>, Link: https://patchwork.freedesktop.org/patch/msgid/20171025143943.7661-3-chris@chris-wilson.co.uk
2017-10-25 21:39:42 +07:00
void intel_engine_pin_breadcrumbs_irq(struct intel_engine_cs *engine)
{
struct intel_breadcrumbs *b = &engine->breadcrumbs;
spin_lock_irq(&b->irq_lock);
if (!b->irq_enabled++)
irq_enable(engine);
GEM_BUG_ON(!b->irq_enabled); /* no overflow! */
spin_unlock_irq(&b->irq_lock);
}
void intel_engine_unpin_breadcrumbs_irq(struct intel_engine_cs *engine)
{
struct intel_breadcrumbs *b = &engine->breadcrumbs;
spin_lock_irq(&b->irq_lock);
GEM_BUG_ON(!b->irq_enabled); /* no underflow! */
if (!--b->irq_enabled)
irq_disable(engine);
spin_unlock_irq(&b->irq_lock);
}
static void __intel_breadcrumbs_arm_irq(struct intel_breadcrumbs *b)
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
{
struct intel_engine_cs *engine =
container_of(b, struct intel_engine_cs, breadcrumbs);
drm/i915/guc: Always enable the breadcrumbs irq The execlists emulation on top of the GuC (used for scheduling and preemption) depends on the MI_USER_INTERRUPT for its notifications and tasklet action. As we always employ the irq, there is no advantage in ever disabling it while we are using the GuC, so allow us to arm the breadcrumb irq when enabling GuC submission and disarm upon disabling. The impact should be lessened by the delayed irq disabling we do (we only disable after receiving an interrupt for which no one was wanting), but allowing guc to explicitly manage the irq in relation to itself is simpler and prevents an issue with losing an interrupt for preemption as it is not coupled to an active request. Internally, we add a reference counter (breadcrumbs.irq_enabled) as a simple mechanism to allow GuC to keep the breadcrumb irq enabled. To improve upon always enabling the irq while guc is selected, we need to hook into the parking facility of intel_engines so that we only enable the breadcrumbs while the GT is active (one step better would be to individually park/unpark each engine). In effect, this means that we keep the breadcrumb irq always enabled for the entire duration the guc is busy, whereas before we would try to switch it off whenever we idled for more than interrupt with no associated waiters. The difference *should* be negligible in practice! v2: Stop abusing fence signaling (and its auxiliary data structures) to enable the breadcrumbs irqs. Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Cc: Michał Winiarski <michal.winiarski@intel.com>, Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> Reviewed-by: Michał Winiarski <michal.winiarski@intel.com>, Link: https://patchwork.freedesktop.org/patch/msgid/20171025143943.7661-3-chris@chris-wilson.co.uk
2017-10-25 21:39:42 +07:00
lockdep_assert_held(&b->irq_lock);
if (b->irq_armed)
return;
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
/*
* The breadcrumb irq will be disarmed on the interrupt after the
* waiters are signaled. This gives us a single interrupt window in
* which we can add a new waiter and avoid the cost of re-enabling
* the irq.
*/
b->irq_armed = true;
/*
* Since we are waiting on a request, the GPU should be busy
* and should have its own rpm reference. This is tracked
* by i915->gt.awake, we can forgo holding our own wakref
* for the interrupt as before i915->gt.awake is released (when
* the driver is idle) we disarm the breadcrumbs.
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
*/
if (!b->irq_enabled++)
drm/i915/guc: Always enable the breadcrumbs irq The execlists emulation on top of the GuC (used for scheduling and preemption) depends on the MI_USER_INTERRUPT for its notifications and tasklet action. As we always employ the irq, there is no advantage in ever disabling it while we are using the GuC, so allow us to arm the breadcrumb irq when enabling GuC submission and disarm upon disabling. The impact should be lessened by the delayed irq disabling we do (we only disable after receiving an interrupt for which no one was wanting), but allowing guc to explicitly manage the irq in relation to itself is simpler and prevents an issue with losing an interrupt for preemption as it is not coupled to an active request. Internally, we add a reference counter (breadcrumbs.irq_enabled) as a simple mechanism to allow GuC to keep the breadcrumb irq enabled. To improve upon always enabling the irq while guc is selected, we need to hook into the parking facility of intel_engines so that we only enable the breadcrumbs while the GT is active (one step better would be to individually park/unpark each engine). In effect, this means that we keep the breadcrumb irq always enabled for the entire duration the guc is busy, whereas before we would try to switch it off whenever we idled for more than interrupt with no associated waiters. The difference *should* be negligible in practice! v2: Stop abusing fence signaling (and its auxiliary data structures) to enable the breadcrumbs irqs. Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Cc: Michał Winiarski <michal.winiarski@intel.com>, Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> Reviewed-by: Michał Winiarski <michal.winiarski@intel.com>, Link: https://patchwork.freedesktop.org/patch/msgid/20171025143943.7661-3-chris@chris-wilson.co.uk
2017-10-25 21:39:42 +07:00
irq_enable(engine);
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
}
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
void intel_engine_init_breadcrumbs(struct intel_engine_cs *engine)
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
{
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
struct intel_breadcrumbs *b = &engine->breadcrumbs;
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
spin_lock_init(&b->irq_lock);
INIT_LIST_HEAD(&b->signalers);
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
init_irq_work(&b->irq_work, signal_irq_work);
}
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
void intel_engine_reset_breadcrumbs(struct intel_engine_cs *engine)
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
{
struct intel_breadcrumbs *b = &engine->breadcrumbs;
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
unsigned long flags;
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
spin_lock_irqsave(&b->irq_lock, flags);
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
if (b->irq_enabled)
irq_enable(engine);
else
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
irq_disable(engine);
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
spin_unlock_irqrestore(&b->irq_lock, flags);
}
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine)
{
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
}
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
bool i915_request_enable_breadcrumb(struct i915_request *rq)
{
drm/i915: Seal races between async GPU cancellation, retirement and signaling Currently there is an underlying assumption that i915_request_unsubmit() is synchronous wrt the GPU -- that is the request is no longer in flight as we remove it. In the near future that may change, and this may upset our signaling as we can process an interrupt for that request while it is no longer in flight. CPU0 CPU1 intel_engine_breadcrumbs_irq (queue request completion) i915_request_cancel_signaling ... ... i915_request_enable_signaling dma_fence_signal Hence in the time it took us to drop the lock to signal the request, a preemption event may have occurred and re-queued the request. In the process, that request would have seen I915_FENCE_FLAG_SIGNAL clear and so reused the rq->signal_link that was in use on CPU0, leading to bad pointer chasing in intel_engine_breadcrumbs_irq. A related issue was that if someone started listening for a signal on a completed but no longer in-flight request, we missed the opportunity to immediately signal that request. Furthermore, as intel_contexts may be immediately released during request retirement, in order to be entirely sure that intel_engine_breadcrumbs_irq may no longer dereference the intel_context (ce->signals and ce->signal_link), we must wait for irq spinlock. In order to prevent the race, we use a bit in the fence.flags to signal the transfer onto the signal list inside intel_engine_breadcrumbs_irq. For simplicity, we use the DMA_FENCE_FLAG_SIGNALED_BIT as it then quickly signals to any outside observer that the fence is indeed signaled. v2: Sketch out potential dma-fence API for manual signaling v3: And the test_and_set_bit() Fixes: 52c0fdb25c7c ("drm/i915: Replace global breadcrumbs with per-context interrupt tracking") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190508112452.18942-1-chris@chris-wilson.co.uk
2019-05-08 18:24:52 +07:00
lockdep_assert_held(&rq->lock);
lockdep_assert_irqs_disabled();
drm/i915: Seal races between async GPU cancellation, retirement and signaling Currently there is an underlying assumption that i915_request_unsubmit() is synchronous wrt the GPU -- that is the request is no longer in flight as we remove it. In the near future that may change, and this may upset our signaling as we can process an interrupt for that request while it is no longer in flight. CPU0 CPU1 intel_engine_breadcrumbs_irq (queue request completion) i915_request_cancel_signaling ... ... i915_request_enable_signaling dma_fence_signal Hence in the time it took us to drop the lock to signal the request, a preemption event may have occurred and re-queued the request. In the process, that request would have seen I915_FENCE_FLAG_SIGNAL clear and so reused the rq->signal_link that was in use on CPU0, leading to bad pointer chasing in intel_engine_breadcrumbs_irq. A related issue was that if someone started listening for a signal on a completed but no longer in-flight request, we missed the opportunity to immediately signal that request. Furthermore, as intel_contexts may be immediately released during request retirement, in order to be entirely sure that intel_engine_breadcrumbs_irq may no longer dereference the intel_context (ce->signals and ce->signal_link), we must wait for irq spinlock. In order to prevent the race, we use a bit in the fence.flags to signal the transfer onto the signal list inside intel_engine_breadcrumbs_irq. For simplicity, we use the DMA_FENCE_FLAG_SIGNALED_BIT as it then quickly signals to any outside observer that the fence is indeed signaled. v2: Sketch out potential dma-fence API for manual signaling v3: And the test_and_set_bit() Fixes: 52c0fdb25c7c ("drm/i915: Replace global breadcrumbs with per-context interrupt tracking") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190508112452.18942-1-chris@chris-wilson.co.uk
2019-05-08 18:24:52 +07:00
if (test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags)) {
struct intel_breadcrumbs *b = &rq->engine->breadcrumbs;
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
struct intel_context *ce = rq->hw_context;
struct list_head *pos;
drm/i915: Seal races between async GPU cancellation, retirement and signaling Currently there is an underlying assumption that i915_request_unsubmit() is synchronous wrt the GPU -- that is the request is no longer in flight as we remove it. In the near future that may change, and this may upset our signaling as we can process an interrupt for that request while it is no longer in flight. CPU0 CPU1 intel_engine_breadcrumbs_irq (queue request completion) i915_request_cancel_signaling ... ... i915_request_enable_signaling dma_fence_signal Hence in the time it took us to drop the lock to signal the request, a preemption event may have occurred and re-queued the request. In the process, that request would have seen I915_FENCE_FLAG_SIGNAL clear and so reused the rq->signal_link that was in use on CPU0, leading to bad pointer chasing in intel_engine_breadcrumbs_irq. A related issue was that if someone started listening for a signal on a completed but no longer in-flight request, we missed the opportunity to immediately signal that request. Furthermore, as intel_contexts may be immediately released during request retirement, in order to be entirely sure that intel_engine_breadcrumbs_irq may no longer dereference the intel_context (ce->signals and ce->signal_link), we must wait for irq spinlock. In order to prevent the race, we use a bit in the fence.flags to signal the transfer onto the signal list inside intel_engine_breadcrumbs_irq. For simplicity, we use the DMA_FENCE_FLAG_SIGNALED_BIT as it then quickly signals to any outside observer that the fence is indeed signaled. v2: Sketch out potential dma-fence API for manual signaling v3: And the test_and_set_bit() Fixes: 52c0fdb25c7c ("drm/i915: Replace global breadcrumbs with per-context interrupt tracking") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190508112452.18942-1-chris@chris-wilson.co.uk
2019-05-08 18:24:52 +07:00
spin_lock(&b->irq_lock);
GEM_BUG_ON(test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags));
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
__intel_breadcrumbs_arm_irq(b);
drm/i915/breadcrumbs: Reduce signaler rbtree to a sorted list The goal here is to try and reduce the latency of signaling additional requests following the wakeup from interrupt by reducing the list of to-be-signaled requests from an rbtree to a sorted linked list. The original choice of using an rbtree was to facilitate random insertions of request into the signaler while maintaining a sorted list. However, if we assume that most new requests are added when they are submitted, we see those new requests in execution order making a insertion sort fast, and the reduction in overhead of each signaler iteration significant. Since commit 56299fb7d904 ("drm/i915: Signal first fence from irq handler if complete"), we signal most fences directly from notify_ring() in the interrupt handler greatly reducing the amount of work that actually needs to be done by the signaler kthread. All the thread is then required to do is operate as the bottom-half, cleaning up after the interrupt handler and preparing the next waiter. This includes signaling all later completed fences in a saturated system, but on a mostly idle system we only have to rebuild the wait rbtree in time for the next interrupt. With this de-emphasis of the signaler's role, we want to rejig it's datastructures to reduce the amount of work we require to both setup the signal tree and maintain it on every interrupt. References: 56299fb7d904 ("drm/i915: Signal first fence from irq handler if complete") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20180222092545.17216-1-chris@chris-wilson.co.uk
2018-02-22 16:25:44 +07:00
/*
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
* We keep the seqno in retirement order, so we can break
* inside intel_engine_breadcrumbs_irq as soon as we've passed
* the last completed request (or seen a request that hasn't
* event started). We could iterate the timeline->requests list,
* but keeping a separate signalers_list has the advantage of
* hopefully being much smaller than the full list and so
* provides faster iteration and detection when there are no
* more interrupts required for this context.
*
* We typically expect to add new signalers in order, so we
* start looking for our insertion point from the tail of
* the list.
*/
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
list_for_each_prev(pos, &ce->signals) {
struct i915_request *it =
list_entry(pos, typeof(*it), signal_link);
drm/i915/breadcrumbs: Reduce signaler rbtree to a sorted list The goal here is to try and reduce the latency of signaling additional requests following the wakeup from interrupt by reducing the list of to-be-signaled requests from an rbtree to a sorted linked list. The original choice of using an rbtree was to facilitate random insertions of request into the signaler while maintaining a sorted list. However, if we assume that most new requests are added when they are submitted, we see those new requests in execution order making a insertion sort fast, and the reduction in overhead of each signaler iteration significant. Since commit 56299fb7d904 ("drm/i915: Signal first fence from irq handler if complete"), we signal most fences directly from notify_ring() in the interrupt handler greatly reducing the amount of work that actually needs to be done by the signaler kthread. All the thread is then required to do is operate as the bottom-half, cleaning up after the interrupt handler and preparing the next waiter. This includes signaling all later completed fences in a saturated system, but on a mostly idle system we only have to rebuild the wait rbtree in time for the next interrupt. With this de-emphasis of the signaler's role, we want to rejig it's datastructures to reduce the amount of work we require to both setup the signal tree and maintain it on every interrupt. References: 56299fb7d904 ("drm/i915: Signal first fence from irq handler if complete") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20180222092545.17216-1-chris@chris-wilson.co.uk
2018-02-22 16:25:44 +07:00
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
if (i915_seqno_passed(rq->fence.seqno, it->fence.seqno))
drm/i915/breadcrumbs: Reduce signaler rbtree to a sorted list The goal here is to try and reduce the latency of signaling additional requests following the wakeup from interrupt by reducing the list of to-be-signaled requests from an rbtree to a sorted linked list. The original choice of using an rbtree was to facilitate random insertions of request into the signaler while maintaining a sorted list. However, if we assume that most new requests are added when they are submitted, we see those new requests in execution order making a insertion sort fast, and the reduction in overhead of each signaler iteration significant. Since commit 56299fb7d904 ("drm/i915: Signal first fence from irq handler if complete"), we signal most fences directly from notify_ring() in the interrupt handler greatly reducing the amount of work that actually needs to be done by the signaler kthread. All the thread is then required to do is operate as the bottom-half, cleaning up after the interrupt handler and preparing the next waiter. This includes signaling all later completed fences in a saturated system, but on a mostly idle system we only have to rebuild the wait rbtree in time for the next interrupt. With this de-emphasis of the signaler's role, we want to rejig it's datastructures to reduce the amount of work we require to both setup the signal tree and maintain it on every interrupt. References: 56299fb7d904 ("drm/i915: Signal first fence from irq handler if complete") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20180222092545.17216-1-chris@chris-wilson.co.uk
2018-02-22 16:25:44 +07:00
break;
}
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
list_add(&rq->signal_link, pos);
if (pos == &ce->signals) /* catch transitions from empty list */
list_move_tail(&ce->signal_link, &b->signalers);
GEM_BUG_ON(!check_signal_order(ce, rq));
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
set_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
drm/i915: Seal races between async GPU cancellation, retirement and signaling Currently there is an underlying assumption that i915_request_unsubmit() is synchronous wrt the GPU -- that is the request is no longer in flight as we remove it. In the near future that may change, and this may upset our signaling as we can process an interrupt for that request while it is no longer in flight. CPU0 CPU1 intel_engine_breadcrumbs_irq (queue request completion) i915_request_cancel_signaling ... ... i915_request_enable_signaling dma_fence_signal Hence in the time it took us to drop the lock to signal the request, a preemption event may have occurred and re-queued the request. In the process, that request would have seen I915_FENCE_FLAG_SIGNAL clear and so reused the rq->signal_link that was in use on CPU0, leading to bad pointer chasing in intel_engine_breadcrumbs_irq. A related issue was that if someone started listening for a signal on a completed but no longer in-flight request, we missed the opportunity to immediately signal that request. Furthermore, as intel_contexts may be immediately released during request retirement, in order to be entirely sure that intel_engine_breadcrumbs_irq may no longer dereference the intel_context (ce->signals and ce->signal_link), we must wait for irq spinlock. In order to prevent the race, we use a bit in the fence.flags to signal the transfer onto the signal list inside intel_engine_breadcrumbs_irq. For simplicity, we use the DMA_FENCE_FLAG_SIGNALED_BIT as it then quickly signals to any outside observer that the fence is indeed signaled. v2: Sketch out potential dma-fence API for manual signaling v3: And the test_and_set_bit() Fixes: 52c0fdb25c7c ("drm/i915: Replace global breadcrumbs with per-context interrupt tracking") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190508112452.18942-1-chris@chris-wilson.co.uk
2019-05-08 18:24:52 +07:00
spin_unlock(&b->irq_lock);
}
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
return !__request_completed(rq);
}
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
void i915_request_cancel_breadcrumb(struct i915_request *rq)
{
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
struct intel_breadcrumbs *b = &rq->engine->breadcrumbs;
drm/i915/breadcrumbs: Reduce signaler rbtree to a sorted list The goal here is to try and reduce the latency of signaling additional requests following the wakeup from interrupt by reducing the list of to-be-signaled requests from an rbtree to a sorted linked list. The original choice of using an rbtree was to facilitate random insertions of request into the signaler while maintaining a sorted list. However, if we assume that most new requests are added when they are submitted, we see those new requests in execution order making a insertion sort fast, and the reduction in overhead of each signaler iteration significant. Since commit 56299fb7d904 ("drm/i915: Signal first fence from irq handler if complete"), we signal most fences directly from notify_ring() in the interrupt handler greatly reducing the amount of work that actually needs to be done by the signaler kthread. All the thread is then required to do is operate as the bottom-half, cleaning up after the interrupt handler and preparing the next waiter. This includes signaling all later completed fences in a saturated system, but on a mostly idle system we only have to rebuild the wait rbtree in time for the next interrupt. With this de-emphasis of the signaler's role, we want to rejig it's datastructures to reduce the amount of work we require to both setup the signal tree and maintain it on every interrupt. References: 56299fb7d904 ("drm/i915: Signal first fence from irq handler if complete") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20180222092545.17216-1-chris@chris-wilson.co.uk
2018-02-22 16:25:44 +07:00
drm/i915: Seal races between async GPU cancellation, retirement and signaling Currently there is an underlying assumption that i915_request_unsubmit() is synchronous wrt the GPU -- that is the request is no longer in flight as we remove it. In the near future that may change, and this may upset our signaling as we can process an interrupt for that request while it is no longer in flight. CPU0 CPU1 intel_engine_breadcrumbs_irq (queue request completion) i915_request_cancel_signaling ... ... i915_request_enable_signaling dma_fence_signal Hence in the time it took us to drop the lock to signal the request, a preemption event may have occurred and re-queued the request. In the process, that request would have seen I915_FENCE_FLAG_SIGNAL clear and so reused the rq->signal_link that was in use on CPU0, leading to bad pointer chasing in intel_engine_breadcrumbs_irq. A related issue was that if someone started listening for a signal on a completed but no longer in-flight request, we missed the opportunity to immediately signal that request. Furthermore, as intel_contexts may be immediately released during request retirement, in order to be entirely sure that intel_engine_breadcrumbs_irq may no longer dereference the intel_context (ce->signals and ce->signal_link), we must wait for irq spinlock. In order to prevent the race, we use a bit in the fence.flags to signal the transfer onto the signal list inside intel_engine_breadcrumbs_irq. For simplicity, we use the DMA_FENCE_FLAG_SIGNALED_BIT as it then quickly signals to any outside observer that the fence is indeed signaled. v2: Sketch out potential dma-fence API for manual signaling v3: And the test_and_set_bit() Fixes: 52c0fdb25c7c ("drm/i915: Replace global breadcrumbs with per-context interrupt tracking") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190508112452.18942-1-chris@chris-wilson.co.uk
2019-05-08 18:24:52 +07:00
lockdep_assert_held(&rq->lock);
lockdep_assert_irqs_disabled();
drm/i915: Seal races between async GPU cancellation, retirement and signaling Currently there is an underlying assumption that i915_request_unsubmit() is synchronous wrt the GPU -- that is the request is no longer in flight as we remove it. In the near future that may change, and this may upset our signaling as we can process an interrupt for that request while it is no longer in flight. CPU0 CPU1 intel_engine_breadcrumbs_irq (queue request completion) i915_request_cancel_signaling ... ... i915_request_enable_signaling dma_fence_signal Hence in the time it took us to drop the lock to signal the request, a preemption event may have occurred and re-queued the request. In the process, that request would have seen I915_FENCE_FLAG_SIGNAL clear and so reused the rq->signal_link that was in use on CPU0, leading to bad pointer chasing in intel_engine_breadcrumbs_irq. A related issue was that if someone started listening for a signal on a completed but no longer in-flight request, we missed the opportunity to immediately signal that request. Furthermore, as intel_contexts may be immediately released during request retirement, in order to be entirely sure that intel_engine_breadcrumbs_irq may no longer dereference the intel_context (ce->signals and ce->signal_link), we must wait for irq spinlock. In order to prevent the race, we use a bit in the fence.flags to signal the transfer onto the signal list inside intel_engine_breadcrumbs_irq. For simplicity, we use the DMA_FENCE_FLAG_SIGNALED_BIT as it then quickly signals to any outside observer that the fence is indeed signaled. v2: Sketch out potential dma-fence API for manual signaling v3: And the test_and_set_bit() Fixes: 52c0fdb25c7c ("drm/i915: Replace global breadcrumbs with per-context interrupt tracking") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190508112452.18942-1-chris@chris-wilson.co.uk
2019-05-08 18:24:52 +07:00
/*
* We must wait for b->irq_lock so that we know the interrupt handler
* has released its reference to the intel_context and has completed
* the DMA_FENCE_FLAG_SIGNALED_BIT/I915_FENCE_FLAG_SIGNAL dance (if
* required).
*/
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
spin_lock(&b->irq_lock);
if (test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags)) {
struct intel_context *ce = rq->hw_context;
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
list_del(&rq->signal_link);
if (list_empty(&ce->signals))
list_del_init(&ce->signal_link);
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
clear_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
}
spin_unlock(&b->irq_lock);
}
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
void intel_engine_print_breadcrumbs(struct intel_engine_cs *engine,
struct drm_printer *p)
{
struct intel_breadcrumbs *b = &engine->breadcrumbs;
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
struct intel_context *ce;
struct i915_request *rq;
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
if (list_empty(&b->signalers))
return;
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
drm_printf(p, "Signals:\n");
drm/i915: Replace global breadcrumbs with per-context interrupt tracking A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd"), the issue of handling multiple clients waiting in parallel was brought to our attention. The requirement was that every client should be woken immediately upon its request being signaled, without incurring any cpu overhead. To handle certain fragility of our hw meant that we could not do a simple check inside the irq handler (some generations required almost unbounded delays before we could be sure of seqno coherency) and so request completion checking required delegation. Before commit 688e6c725816, the solution was simple. Every client waiting on a request would be woken on every interrupt and each would do a heavyweight check to see if their request was complete. Commit 688e6c725816 introduced an rbtree so that only the earliest waiter on the global timeline would woken, and would wake the next and so on. (Along with various complications to handle requests being reordered along the global timeline, and also a requirement for kthread to provide a delegate for fence signaling that had no process context.) The global rbtree depends on knowing the execution timeline (and global seqno). Without knowing that order, we must instead check all contexts queued to the HW to see which may have advanced. We trim that list by only checking queued contexts that are being waited on, but still we keep a list of all active contexts and their active signalers that we inspect from inside the irq handler. By moving the waiters onto the fence signal list, we can combine the client wakeup with the dma_fence signaling (a dramatic reduction in complexity, but does require the HW being coherent, the seqno must be visible from the cpu before the interrupt is raised - we keep a timer backup just in case). Having previously fixed all the issues with irq-seqno serialisation (by inserting delays onto the GPU after each request instead of random delays on the CPU after each interrupt), we can rely on the seqno state to perfom direct wakeups from the interrupt handler. This allows us to preserve our single context switch behaviour of the current routine, with the only downside that we lose the RT priority sorting of wakeups. In general, direct wakeup latency of multiple clients is about the same (about 10% better in most cases) with a reduction in total CPU time spent in the waiter (about 20-50% depending on gen). Average herd behaviour is improved, but at the cost of not delegating wakeups on task_prio. v2: Capture fence signaling state for error state and add comments to warm even the most cold of hearts. v3: Check if the request is still active before busywaiting v4: Reduce the amount of pointer misdirection with list_for_each_safe and using a local i915_request variable inside the loops v5: Add a missing pluralisation to a purely informative selftest message. References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk
2019-01-30 03:52:29 +07:00
spin_lock_irq(&b->irq_lock);
list_for_each_entry(ce, &b->signalers, signal_link) {
list_for_each_entry(rq, &ce->signals, signal_link) {
drm_printf(p, "\t[%llx:%llx%s] @ %dms\n",
rq->fence.context, rq->fence.seqno,
i915_request_completed(rq) ? "!" :
i915_request_started(rq) ? "*" :
"",
jiffies_to_msecs(jiffies - rq->emitted_jiffies));
}
}
spin_unlock_irq(&b->irq_lock);
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
}