linux_dsm_epyc7002/drivers/gpu/drm/i915/i915_active.h

235 lines
8.3 KiB
C
Raw Normal View History

/*
* SPDX-License-Identifier: MIT
*
* Copyright © 2019 Intel Corporation
*/
#ifndef _I915_ACTIVE_H_
#define _I915_ACTIVE_H_
#include <linux/lockdep.h>
#include "i915_active_types.h"
#include "i915_request.h"
struct i915_request;
struct intel_engine_cs;
struct intel_timeline;
/*
* We treat requests as fences. This is not be to confused with our
* "fence registers" but pipeline synchronisation objects ala GL_ARB_sync.
* We use the fences to synchronize access from the CPU with activity on the
* GPU, for example, we should not rewrite an object's PTE whilst the GPU
* is reading them. We also track fences at a higher level to provide
* implicit synchronisation around GEM objects, e.g. set-domain will wait
* for outstanding GPU rendering before marking the object ready for CPU
* access, or a pageflip will wait until the GPU is complete before showing
* the frame on the scanout.
*
* In order to use a fence, the object must track the fence it needs to
* serialise with. For example, GEM objects want to track both read and
* write access so that we can perform concurrent read operations between
* the CPU and GPU engines, as well as waiting for all rendering to
* complete, or waiting for the last GPU user of a "fence register". The
* object then embeds a #i915_active_fence to track the most recent (in
* retirement order) request relevant for the desired mode of access.
* The #i915_active_fence is updated with i915_active_fence_set() to
* track the most recent fence request, typically this is done as part of
* i915_vma_move_to_active().
*
* When the #i915_active_fence completes (is retired), it will
* signal its completion to the owner through a callback as well as mark
* itself as idle (i915_active_fence.request == NULL). The owner
* can then perform any action, such as delayed freeing of an active
* resource including itself.
*/
void i915_active_noop(struct dma_fence *fence, struct dma_fence_cb *cb);
/**
* __i915_active_fence_init - prepares the activity tracker for use
* @active - the active tracker
* @fence - initial fence to track, can be NULL
* @func - a callback when then the tracker is retired (becomes idle),
* can be NULL
*
* i915_active_fence_init() prepares the embedded @active struct for use as
* an activity tracker, that is for tracking the last known active fence
* associated with it. When the last fence becomes idle, when it is retired
* after completion, the optional callback @func is invoked.
*/
static inline void
__i915_active_fence_init(struct i915_active_fence *active,
void *fence,
dma_fence_func_t fn)
{
RCU_INIT_POINTER(active->fence, fence);
active->cb.func = fn ?: i915_active_noop;
}
drm/i915: Serialise i915_active_fence_set() with itself The expected downside to commit 58b4c1a07ada ("drm/i915: Reduce nested prepare_remote_context() to a trylock") was that it would need to return -EAGAIN to userspace in order to resolve potential mutex inversion. Such an unsightly round trip is unnecessary if we could atomically insert a barrier into the i915_active_fence, so make it happen. Currently, we use the timeline->mutex (or some other named outer lock) to order insertion into the i915_active_fence (and so individual nodes of i915_active). Inside __i915_active_fence_set, we only need then serialise with the interrupt handler in order to claim the timeline for ourselves. However, if we remove the outer lock, we need to ensure the order is intact between not only multiple threads trying to insert themselves into the timeline, but also with the interrupt handler completing the previous occupant. We use xchg() on insert so that we have an ordered sequence of insertions (and each caller knows the previous fence on which to wait, preserving the chain of all fences in the timeline), but we then have to cmpxchg() in the interrupt handler to avoid overwriting the new occupant. The only nasty side-effect is having to temporarily strip off the RCU-annotations to apply the atomic operations, otherwise the rules are much more conventional! Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=112402 Fixes: 58b4c1a07ada ("drm/i915: Reduce nested prepare_remote_context() to a trylock") 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/20191127134527.3438410-1-chris@chris-wilson.co.uk
2019-11-27 20:45:27 +07:00
#define INIT_ACTIVE_FENCE(A) \
__i915_active_fence_init((A), NULL, NULL)
struct dma_fence *
__i915_active_fence_set(struct i915_active_fence *active,
struct dma_fence *fence);
/**
* i915_active_fence_set - updates the tracker to watch the current fence
* @active - the active tracker
* @rq - the request to watch
*
* i915_active_fence_set() watches the given @rq for completion. While
* that @rq is busy, the @active reports busy. When that @rq is signaled
* (or else retired) the @active tracker is updated to report idle.
*/
int __must_check
i915_active_fence_set(struct i915_active_fence *active,
struct i915_request *rq);
/**
* i915_active_fence_get - return a reference to the active fence
* @active - the active tracker
*
* i915_active_fence_get() returns a reference to the active fence,
* or NULL if the active tracker is idle. The reference is obtained under RCU,
* so no locking is required by the caller.
*
* The reference should be freed with dma_fence_put().
*/
static inline struct dma_fence *
i915_active_fence_get(struct i915_active_fence *active)
{
struct dma_fence *fence;
rcu_read_lock();
fence = dma_fence_get_rcu_safe(&active->fence);
rcu_read_unlock();
return fence;
}
/**
* i915_active_fence_isset - report whether the active tracker is assigned
* @active - the active tracker
*
* i915_active_fence_isset() returns true if the active tracker is currently
* assigned to a fence. Due to the lazy retiring, that fence may be idle
* and this may report stale information.
*/
static inline bool
i915_active_fence_isset(const struct i915_active_fence *active)
{
return rcu_access_pointer(active->fence);
}
/*
* GPU activity tracking
*
* Each set of commands submitted to the GPU compromises a single request that
* signals a fence upon completion. struct i915_request combines the
* command submission, scheduling and fence signaling roles. If we want to see
* if a particular task is complete, we need to grab the fence (struct
* i915_request) for that task and check or wait for it to be signaled. More
* often though we want to track the status of a bunch of tasks, for example
* to wait for the GPU to finish accessing some memory across a variety of
* different command pipelines from different clients. We could choose to
* track every single request associated with the task, but knowing that
* each request belongs to an ordered timeline (later requests within a
* timeline must wait for earlier requests), we need only track the
* latest request in each timeline to determine the overall status of the
* task.
*
* struct i915_active provides this tracking across timelines. It builds a
* composite shared-fence, and is updated as new work is submitted to the task,
* forming a snapshot of the current status. It should be embedded into the
* different resources that need to track their associated GPU activity to
* provide a callback when that GPU activity has ceased, or otherwise to
* provide a serialisation point either for request submission or for CPU
* synchronisation.
*/
void __i915_active_init(struct i915_active *ref,
int (*active)(struct i915_active *ref),
void (*retire)(struct i915_active *ref),
struct lock_class_key *mkey,
struct lock_class_key *wkey);
/* Specialise each class of i915_active to avoid impossible lockdep cycles. */
#define i915_active_init(ref, active, retire) do { \
static struct lock_class_key __mkey; \
static struct lock_class_key __wkey; \
\
__i915_active_init(ref, active, retire, &__mkey, &__wkey); \
} while (0)
int i915_active_ref(struct i915_active *ref,
struct intel_timeline *tl,
struct dma_fence *fence);
drm/i915: Mark i915_request.timeline as a volatile, rcu pointer The request->timeline is only valid until the request is retired (i.e. before it is completed). Upon retiring the request, the context may be unpinned and freed, and along with it the timeline may be freed. We therefore need to be very careful when chasing rq->timeline that the pointer does not disappear beneath us. The vast majority of users are in a protected context, either during request construction or retirement, where the timeline->mutex is held and the timeline cannot disappear. It is those few off the beaten path (where we access a second timeline) that need extra scrutiny -- to be added in the next patch after first adding the warnings about dangerous access. One complication, where we cannot use the timeline->mutex itself, is during request submission onto hardware (under spinlocks). Here, we want to check on the timeline to finalize the breadcrumb, and so we need to impose a second rule to ensure that the request->timeline is indeed valid. As we are submitting the request, it's context and timeline must be pinned, as it will be used by the hardware. Since it is pinned, we know the request->timeline must still be valid, and we cannot submit the idle barrier until after we release the engine->active.lock, ergo while submitting and holding that spinlock, a second thread cannot release the timeline. v2: Don't be lazy inside selftests; hold the timeline->mutex for as long as we need it, and tidy up acquiring the timeline with a bit of refactoring (i915_active_add_request) 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/20190919111912.21631-1-chris@chris-wilson.co.uk
2019-09-19 18:19:10 +07:00
static inline int
i915_active_add_request(struct i915_active *ref, struct i915_request *rq)
{
return i915_active_ref(ref, i915_request_timeline(rq), &rq->fence);
drm/i915: Mark i915_request.timeline as a volatile, rcu pointer The request->timeline is only valid until the request is retired (i.e. before it is completed). Upon retiring the request, the context may be unpinned and freed, and along with it the timeline may be freed. We therefore need to be very careful when chasing rq->timeline that the pointer does not disappear beneath us. The vast majority of users are in a protected context, either during request construction or retirement, where the timeline->mutex is held and the timeline cannot disappear. It is those few off the beaten path (where we access a second timeline) that need extra scrutiny -- to be added in the next patch after first adding the warnings about dangerous access. One complication, where we cannot use the timeline->mutex itself, is during request submission onto hardware (under spinlocks). Here, we want to check on the timeline to finalize the breadcrumb, and so we need to impose a second rule to ensure that the request->timeline is indeed valid. As we are submitting the request, it's context and timeline must be pinned, as it will be used by the hardware. Since it is pinned, we know the request->timeline must still be valid, and we cannot submit the idle barrier until after we release the engine->active.lock, ergo while submitting and holding that spinlock, a second thread cannot release the timeline. v2: Don't be lazy inside selftests; hold the timeline->mutex for as long as we need it, and tidy up acquiring the timeline with a bit of refactoring (i915_active_add_request) 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/20190919111912.21631-1-chris@chris-wilson.co.uk
2019-09-19 18:19:10 +07:00
}
struct dma_fence *
i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f);
drm/i915: Pull i915_vma_pin under the vm->mutex Replace the struct_mutex requirement for pinning the i915_vma with the local vm->mutex instead. Note that the vm->mutex is tainted by the shrinker (we require unbinding from inside fs-reclaim) and so we cannot allocate while holding that mutex. Instead we have to preallocate workers to do allocate and apply the PTE updates after we have we reserved their slot in the drm_mm (using fences to order the PTE writes with the GPU work and with later unbind). In adding the asynchronous vma binding, one subtle requirement is to avoid coupling the binding fence into the backing object->resv. That is the asynchronous binding only applies to the vma timeline itself and not to the pages as that is a more global timeline (the binding of one vma does not need to be ordered with another vma, nor does the implicit GEM fencing depend on a vma, only on writes to the backing store). Keeping the vma binding distinct from the backing store timelines is verified by a number of async gem_exec_fence and gem_exec_schedule tests. The way we do this is quite simple, we keep the fence for the vma binding separate and only wait on it as required, and never add it to the obj->resv itself. Another consequence in reducing the locking around the vma is the destruction of the vma is no longer globally serialised by struct_mutex. A natural solution would be to add a kref to i915_vma, but that requires decoupling the reference cycles, possibly by introducing a new i915_mm_pages object that is own by both obj->mm and vma->pages. However, we have not taken that route due to the overshadowing lmem/ttm discussions, and instead play a series of complicated games with trylocks to (hopefully) ensure that only one destruction path is called! v2: Add some commentary, and some helpers to reduce patch churn. 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/20191004134015.13204-4-chris@chris-wilson.co.uk
2019-10-04 20:39:58 +07:00
static inline bool i915_active_has_exclusive(struct i915_active *ref)
{
return rcu_access_pointer(ref->excl.fence);
drm/i915: Pull i915_vma_pin under the vm->mutex Replace the struct_mutex requirement for pinning the i915_vma with the local vm->mutex instead. Note that the vm->mutex is tainted by the shrinker (we require unbinding from inside fs-reclaim) and so we cannot allocate while holding that mutex. Instead we have to preallocate workers to do allocate and apply the PTE updates after we have we reserved their slot in the drm_mm (using fences to order the PTE writes with the GPU work and with later unbind). In adding the asynchronous vma binding, one subtle requirement is to avoid coupling the binding fence into the backing object->resv. That is the asynchronous binding only applies to the vma timeline itself and not to the pages as that is a more global timeline (the binding of one vma does not need to be ordered with another vma, nor does the implicit GEM fencing depend on a vma, only on writes to the backing store). Keeping the vma binding distinct from the backing store timelines is verified by a number of async gem_exec_fence and gem_exec_schedule tests. The way we do this is quite simple, we keep the fence for the vma binding separate and only wait on it as required, and never add it to the obj->resv itself. Another consequence in reducing the locking around the vma is the destruction of the vma is no longer globally serialised by struct_mutex. A natural solution would be to add a kref to i915_vma, but that requires decoupling the reference cycles, possibly by introducing a new i915_mm_pages object that is own by both obj->mm and vma->pages. However, we have not taken that route due to the overshadowing lmem/ttm discussions, and instead play a series of complicated games with trylocks to (hopefully) ensure that only one destruction path is called! v2: Add some commentary, and some helpers to reduce patch churn. 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/20191004134015.13204-4-chris@chris-wilson.co.uk
2019-10-04 20:39:58 +07:00
}
int __i915_active_wait(struct i915_active *ref, int state);
static inline int i915_active_wait(struct i915_active *ref)
{
return __i915_active_wait(ref, TASK_INTERRUPTIBLE);
}
int i915_sw_fence_await_active(struct i915_sw_fence *fence,
struct i915_active *ref,
unsigned int flags);
int i915_request_await_active(struct i915_request *rq,
struct i915_active *ref,
unsigned int flags);
#define I915_ACTIVE_AWAIT_EXCL BIT(0)
#define I915_ACTIVE_AWAIT_ACTIVE BIT(1)
#define I915_ACTIVE_AWAIT_BARRIER BIT(2)
int i915_active_acquire(struct i915_active *ref);
bool i915_active_acquire_if_busy(struct i915_active *ref);
void i915_active_release(struct i915_active *ref);
drm/i915/gt: Acquire ce->active before ce->pin_count/ce->pin_mutex Similar to commit ac0e331a628b ("drm/i915: Tighten atomicity of i915_active_acquire vs i915_active_release") we have the same race of trying to pin the context underneath a mutex while allowing the decrement to be atomic outside of that mutex. This leads to the problem where two threads may simultaneously try to pin the context and the second not notice that they needed to repin the context. <2> [198.669621] kernel BUG at drivers/gpu/drm/i915/gt/intel_timeline.c:387! <4> [198.669703] invalid opcode: 0000 [#1] PREEMPT SMP PTI <4> [198.669712] CPU: 0 PID: 1246 Comm: gem_exec_create Tainted: G U W 5.5.0-rc6-CI-CI_DRM_7755+ #1 <4> [198.669723] Hardware name: /NUC7i5BNB, BIOS BNKBL357.86A.0054.2017.1025.1822 10/25/2017 <4> [198.669776] RIP: 0010:timeline_advance+0x7b/0xe0 [i915] <4> [198.669785] Code: 00 48 c7 c2 10 f1 46 a0 48 c7 c7 70 1b 32 a0 e8 bb dd e7 e0 bf 01 00 00 00 e8 d1 af e7 e0 31 f6 bf 09 00 00 00 e8 35 ef d8 e0 <0f> 0b 48 c7 c1 48 fa 49 a0 ba 84 01 00 00 48 c7 c6 10 f1 46 a0 48 <4> [198.669803] RSP: 0018:ffffc900004c3a38 EFLAGS: 00010296 <4> [198.669810] RAX: ffff888270b35140 RBX: ffff88826f32ee00 RCX: 0000000000000006 <4> [198.669818] RDX: 00000000000017c5 RSI: 0000000000000000 RDI: 0000000000000009 <4> [198.669826] RBP: ffffc900004c3a64 R08: 0000000000000000 R09: 0000000000000000 <4> [198.669834] R10: 0000000000000000 R11: 0000000000000000 R12: ffff88826f9b5980 <4> [198.669841] R13: 0000000000000cc0 R14: ffffc900004c3dc0 R15: ffff888253610068 <4> [198.669849] FS: 00007f63e663fe40(0000) GS:ffff888276c00000(0000) knlGS:0000000000000000 <4> [198.669857] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 <4> [198.669864] CR2: 00007f171f8e39a8 CR3: 000000026b1f6005 CR4: 00000000003606f0 <4> [198.669872] Call Trace: <4> [198.669924] intel_timeline_get_seqno+0x12/0x40 [i915] <4> [198.669977] __i915_request_create+0x76/0x5a0 [i915] <4> [198.670024] i915_request_create+0x86/0x1c0 [i915] <4> [198.670068] i915_gem_do_execbuffer+0xbf2/0x2500 [i915] <4> [198.670082] ? __lock_acquire+0x460/0x15d0 <4> [198.670128] i915_gem_execbuffer2_ioctl+0x11f/0x470 [i915] <4> [198.670171] ? i915_gem_execbuffer_ioctl+0x300/0x300 [i915] <4> [198.670181] drm_ioctl_kernel+0xa7/0xf0 <4> [198.670188] drm_ioctl+0x2e1/0x390 <4> [198.670233] ? i915_gem_execbuffer_ioctl+0x300/0x300 [i915] Fixes: 841350223816 ("drm/i915/gt: Drop mutex serialisation between context pin/unpin") References: ac0e331a628b ("drm/i915: Tighten atomicity of i915_active_acquire vs i915_active_release") 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/20200127152829.2842149-1-chris@chris-wilson.co.uk
2020-01-27 22:28:29 +07:00
static inline void __i915_active_acquire(struct i915_active *ref)
{
GEM_BUG_ON(!atomic_read(&ref->count));
atomic_inc(&ref->count);
}
static inline bool
i915_active_is_idle(const struct i915_active *ref)
{
return !atomic_read(&ref->count);
}
#if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
void i915_active_fini(struct i915_active *ref);
#else
static inline void i915_active_fini(struct i915_active *ref) { }
#endif
drm/i915: Keep contexts pinned until after the next kernel context switch We need to keep the context image pinned in memory until after the GPU has finished writing into it. Since it continues to write as we signal the final breadcrumb, we need to keep it pinned until the request after it is complete. Currently we know the order in which requests execute on each engine, and so to remove that presumption we need to identify a request/context-switch we know must occur after our completion. Any request queued after the signal must imply a context switch, for simplicity we use a fresh request from the kernel context. The sequence of operations for keeping the context pinned until saved is: - On context activation, we preallocate a node for each physical engine the context may operate on. This is to avoid allocations during unpinning, which may be from inside FS_RECLAIM context (aka the shrinker) - On context deactivation on retirement of the last active request (which is before we know the context has been saved), we add the preallocated node onto a barrier list on each engine - On engine idling, we emit a switch to kernel context. When this switch completes, we know that all previous contexts must have been saved, and so on retiring this request we can finally unpin all the contexts that were marked as deactivated prior to the switch. We can enhance this in future by flushing all the idle contexts on a regular heartbeat pulse of a switch to kernel context, which will also be used to check for hung engines. v2: intel_context_active_acquire/_release Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com> Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190614164606.15633-1-chris@chris-wilson.co.uk
2019-06-14 23:46:04 +07:00
int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
struct intel_engine_cs *engine);
void i915_active_acquire_barrier(struct i915_active *ref);
void i915_request_add_active_barriers(struct i915_request *rq);
drm/i915: Keep contexts pinned until after the next kernel context switch We need to keep the context image pinned in memory until after the GPU has finished writing into it. Since it continues to write as we signal the final breadcrumb, we need to keep it pinned until the request after it is complete. Currently we know the order in which requests execute on each engine, and so to remove that presumption we need to identify a request/context-switch we know must occur after our completion. Any request queued after the signal must imply a context switch, for simplicity we use a fresh request from the kernel context. The sequence of operations for keeping the context pinned until saved is: - On context activation, we preallocate a node for each physical engine the context may operate on. This is to avoid allocations during unpinning, which may be from inside FS_RECLAIM context (aka the shrinker) - On context deactivation on retirement of the last active request (which is before we know the context has been saved), we add the preallocated node onto a barrier list on each engine - On engine idling, we emit a switch to kernel context. When this switch completes, we know that all previous contexts must have been saved, and so on retiring this request we can finally unpin all the contexts that were marked as deactivated prior to the switch. We can enhance this in future by flushing all the idle contexts on a regular heartbeat pulse of a switch to kernel context, which will also be used to check for hung engines. v2: intel_context_active_acquire/_release Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com> Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190614164606.15633-1-chris@chris-wilson.co.uk
2019-06-14 23:46:04 +07:00
void i915_active_print(struct i915_active *ref, struct drm_printer *m);
void i915_active_unlock_wait(struct i915_active *ref);
struct i915_active *i915_active_create(void);
struct i915_active *i915_active_get(struct i915_active *ref);
void i915_active_put(struct i915_active *ref);
#endif /* _I915_ACTIVE_H_ */