linux_dsm_epyc7002/drivers/gpu/drm/i915/i915_vma.c

1252 lines
31 KiB
C
Raw Normal View History

/*
* Copyright © 2016 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/sched/mm.h>
#include <drm/drm_gem.h>
#include "display/intel_frontbuffer.h"
#include "gt/intel_engine.h"
#include "gt/intel_gt.h"
#include "i915_drv.h"
#include "i915_globals.h"
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
#include "i915_sw_fence_work.h"
#include "i915_trace.h"
#include "i915_vma.h"
static struct i915_global_vma {
struct i915_global base;
struct kmem_cache *slab_vmas;
} global;
struct i915_vma *i915_vma_alloc(void)
{
return kmem_cache_zalloc(global.slab_vmas, GFP_KERNEL);
}
void i915_vma_free(struct i915_vma *vma)
{
return kmem_cache_free(global.slab_vmas, vma);
}
#if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM)
#include <linux/stackdepot.h>
static void vma_print_allocator(struct i915_vma *vma, const char *reason)
{
drm: Simplify stacktrace handling Replace the indirection through struct stack_trace by using the storage array based interfaces. The original code in all printing functions is really wrong. It allocates a storage array on stack which is unused because depot_fetch_stack() does not store anything in it. It overwrites the entries pointer in the stack_trace struct so it points to the depot storage. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Josh Poimboeuf <jpoimboe@redhat.com> Acked-by: Daniel Vetter <daniel@ffwll.ch> Cc: Andy Lutomirski <luto@kernel.org> Cc: intel-gfx@lists.freedesktop.org Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Cc: dri-devel@lists.freedesktop.org Cc: David Airlie <airlied@linux.ie> Cc: Jani Nikula <jani.nikula@linux.intel.com> Cc: Rodrigo Vivi <rodrigo.vivi@intel.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Alexander Potapenko <glider@google.com> Cc: Alexey Dobriyan <adobriyan@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: linux-mm@kvack.org Cc: David Rientjes <rientjes@google.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: kasan-dev@googlegroups.com Cc: Mike Rapoport <rppt@linux.vnet.ibm.com> Cc: Akinobu Mita <akinobu.mita@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: iommu@lists.linux-foundation.org Cc: Robin Murphy <robin.murphy@arm.com> Cc: Marek Szyprowski <m.szyprowski@samsung.com> Cc: Johannes Thumshirn <jthumshirn@suse.de> Cc: David Sterba <dsterba@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: linux-btrfs@vger.kernel.org Cc: dm-devel@redhat.com Cc: Mike Snitzer <snitzer@redhat.com> Cc: Alasdair Kergon <agk@redhat.com> Cc: Tom Zanussi <tom.zanussi@linux.intel.com> Cc: Miroslav Benes <mbenes@suse.cz> Cc: linux-arch@vger.kernel.org Link: https://lkml.kernel.org/r/20190425094802.622094226@linutronix.de
2019-04-25 16:45:09 +07:00
unsigned long *entries;
unsigned int nr_entries;
char buf[512];
if (!vma->node.stack) {
DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",
vma->node.start, vma->node.size, reason);
return;
}
drm: Simplify stacktrace handling Replace the indirection through struct stack_trace by using the storage array based interfaces. The original code in all printing functions is really wrong. It allocates a storage array on stack which is unused because depot_fetch_stack() does not store anything in it. It overwrites the entries pointer in the stack_trace struct so it points to the depot storage. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Josh Poimboeuf <jpoimboe@redhat.com> Acked-by: Daniel Vetter <daniel@ffwll.ch> Cc: Andy Lutomirski <luto@kernel.org> Cc: intel-gfx@lists.freedesktop.org Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Cc: dri-devel@lists.freedesktop.org Cc: David Airlie <airlied@linux.ie> Cc: Jani Nikula <jani.nikula@linux.intel.com> Cc: Rodrigo Vivi <rodrigo.vivi@intel.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Alexander Potapenko <glider@google.com> Cc: Alexey Dobriyan <adobriyan@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: linux-mm@kvack.org Cc: David Rientjes <rientjes@google.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: kasan-dev@googlegroups.com Cc: Mike Rapoport <rppt@linux.vnet.ibm.com> Cc: Akinobu Mita <akinobu.mita@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: iommu@lists.linux-foundation.org Cc: Robin Murphy <robin.murphy@arm.com> Cc: Marek Szyprowski <m.szyprowski@samsung.com> Cc: Johannes Thumshirn <jthumshirn@suse.de> Cc: David Sterba <dsterba@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: linux-btrfs@vger.kernel.org Cc: dm-devel@redhat.com Cc: Mike Snitzer <snitzer@redhat.com> Cc: Alasdair Kergon <agk@redhat.com> Cc: Tom Zanussi <tom.zanussi@linux.intel.com> Cc: Miroslav Benes <mbenes@suse.cz> Cc: linux-arch@vger.kernel.org Link: https://lkml.kernel.org/r/20190425094802.622094226@linutronix.de
2019-04-25 16:45:09 +07:00
nr_entries = stack_depot_fetch(vma->node.stack, &entries);
stack_trace_snprint(buf, sizeof(buf), entries, nr_entries, 0);
DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",
vma->node.start, vma->node.size, reason, buf);
}
#else
static void vma_print_allocator(struct i915_vma *vma, const char *reason)
{
}
#endif
static inline struct i915_vma *active_to_vma(struct i915_active *ref)
{
return container_of(ref, typeof(struct i915_vma), active);
}
static int __i915_vma_active(struct i915_active *ref)
{
return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT;
}
__i915_active_call
static void __i915_vma_retire(struct i915_active *ref)
{
i915_vma_put(active_to_vma(ref));
}
static struct i915_vma *
vma_create(struct drm_i915_gem_object *obj,
struct i915_address_space *vm,
const struct i915_ggtt_view *view)
{
struct i915_vma *vma;
struct rb_node *rb, **p;
/* The aliasing_ppgtt should never be used directly! */
GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm);
vma = i915_vma_alloc();
if (vma == NULL)
return ERR_PTR(-ENOMEM);
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
mutex_init(&vma->pages_mutex);
vma->vm = i915_vm_get(vm);
vma->ops = &vm->vma_ops;
vma->obj = obj;
vma->resv = obj->base.resv;
vma->size = obj->base.size;
vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire);
/* Declare ourselves safe for use inside shrinkers */
if (IS_ENABLED(CONFIG_LOCKDEP)) {
fs_reclaim_acquire(GFP_KERNEL);
might_lock(&vma->active.mutex);
fs_reclaim_release(GFP_KERNEL);
}
INIT_LIST_HEAD(&vma->closed_link);
if (view && view->type != I915_GGTT_VIEW_NORMAL) {
vma->ggtt_view = *view;
if (view->type == I915_GGTT_VIEW_PARTIAL) {
GEM_BUG_ON(range_overflows_t(u64,
view->partial.offset,
view->partial.size,
obj->base.size >> PAGE_SHIFT));
vma->size = view->partial.size;
vma->size <<= PAGE_SHIFT;
GEM_BUG_ON(vma->size > obj->base.size);
} else if (view->type == I915_GGTT_VIEW_ROTATED) {
vma->size = intel_rotation_info_size(&view->rotated);
vma->size <<= PAGE_SHIFT;
} else if (view->type == I915_GGTT_VIEW_REMAPPED) {
vma->size = intel_remapped_info_size(&view->remapped);
vma->size <<= PAGE_SHIFT;
}
}
if (unlikely(vma->size > vm->total))
goto err_vma;
GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
if (i915_is_ggtt(vm)) {
if (unlikely(overflows_type(vma->size, u32)))
goto err_vma;
vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
i915_gem_object_get_tiling(obj),
i915_gem_object_get_stride(obj));
if (unlikely(vma->fence_size < vma->size || /* overflow */
vma->fence_size > vm->total))
goto err_vma;
GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
i915_gem_object_get_tiling(obj),
i915_gem_object_get_stride(obj));
GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
__set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma));
}
spin_lock(&obj->vma.lock);
rb = NULL;
p = &obj->vma.tree.rb_node;
while (*p) {
struct i915_vma *pos;
long cmp;
rb = *p;
pos = rb_entry(rb, struct i915_vma, obj_node);
/*
* If the view already exists in the tree, another thread
* already created a matching vma, so return the older instance
* and dispose of ours.
*/
cmp = i915_vma_compare(pos, vm, view);
if (cmp == 0) {
spin_unlock(&obj->vma.lock);
i915_vma_free(vma);
return pos;
}
if (cmp < 0)
p = &rb->rb_right;
else
p = &rb->rb_left;
}
rb_link_node(&vma->obj_node, rb, p);
rb_insert_color(&vma->obj_node, &obj->vma.tree);
if (i915_vma_is_ggtt(vma))
/*
* We put the GGTT vma at the start of the vma-list, followed
* by the ppGGTT vma. This allows us to break early when
* iterating over only the GGTT vma for an object, see
* for_each_ggtt_vma()
*/
list_add(&vma->obj_link, &obj->vma.list);
else
list_add_tail(&vma->obj_link, &obj->vma.list);
spin_unlock(&obj->vma.lock);
return vma;
err_vma:
i915_vma_free(vma);
return ERR_PTR(-E2BIG);
}
static struct i915_vma *
vma_lookup(struct drm_i915_gem_object *obj,
struct i915_address_space *vm,
const struct i915_ggtt_view *view)
{
struct rb_node *rb;
rb = obj->vma.tree.rb_node;
while (rb) {
struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
long cmp;
cmp = i915_vma_compare(vma, vm, view);
if (cmp == 0)
return vma;
if (cmp < 0)
rb = rb->rb_right;
else
rb = rb->rb_left;
}
return NULL;
}
/**
* i915_vma_instance - return the singleton instance of the VMA
* @obj: parent &struct drm_i915_gem_object to be mapped
* @vm: address space in which the mapping is located
* @view: additional mapping requirements
*
* i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
* the same @view characteristics. If a match is not found, one is created.
* Once created, the VMA is kept until either the object is freed, or the
* address space is closed.
*
* Returns the vma, or an error pointer.
*/
struct i915_vma *
i915_vma_instance(struct drm_i915_gem_object *obj,
struct i915_address_space *vm,
const struct i915_ggtt_view *view)
{
struct i915_vma *vma;
GEM_BUG_ON(view && !i915_is_ggtt(vm));
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
GEM_BUG_ON(!atomic_read(&vm->open));
spin_lock(&obj->vma.lock);
vma = vma_lookup(obj, vm, view);
spin_unlock(&obj->vma.lock);
/* vma_create() will resolve the race if another creates the vma */
if (unlikely(!vma))
vma = vma_create(obj, vm, view);
GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
return vma;
}
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
struct i915_vma_work {
struct dma_fence_work base;
struct i915_vma *vma;
enum i915_cache_level cache_level;
unsigned int flags;
};
static int __vma_bind(struct dma_fence_work *work)
{
struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
struct i915_vma *vma = vw->vma;
int err;
err = vma->ops->bind_vma(vma, vw->cache_level, vw->flags);
if (err)
atomic_or(I915_VMA_ERROR, &vma->flags);
if (vma->obj)
__i915_gem_object_unpin_pages(vma->obj);
return err;
}
static const struct dma_fence_work_ops bind_ops = {
.name = "bind",
.work = __vma_bind,
};
struct i915_vma_work *i915_vma_work(void)
{
struct i915_vma_work *vw;
vw = kzalloc(sizeof(*vw), GFP_KERNEL);
if (!vw)
return NULL;
dma_fence_work_init(&vw->base, &bind_ops);
vw->base.dma.error = -EAGAIN; /* disable the worker by default */
return vw;
}
/**
* i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
* @vma: VMA to map
* @cache_level: mapping cache level
* @flags: flags like global or local mapping
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
* @work: preallocated worker for allocating and binding the PTE
*
* DMA addresses are taken from the scatter-gather table of this object (or of
* this VMA in case of non-default GGTT views) and PTE entries set up.
* Note that DMA addresses are also the only part of the SG table we care about.
*/
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_vma_bind(struct i915_vma *vma,
enum i915_cache_level cache_level,
u32 flags,
struct i915_vma_work *work)
{
u32 bind_flags;
u32 vma_flags;
int ret;
GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
GEM_BUG_ON(vma->size > vma->node.size);
drm/i915: GEM_WARN_ON considered harmful GEM_WARN_ON currently has dangerous semantics where it is completely compiled out on !GEM_DEBUG builds. This can leave users who expect it to be more like a WARN_ON, just without a warning in non-debug builds, in complete ignorance. Another gotcha with it is that it cannot be used as a statement. Which is again different from a standard kernel WARN_ON. This patch fixes both problems by making it behave as one would expect. It can now be used both as an expression and as statement, and also the condition evaluates properly in all builds - code under the conditional will therefore not unexpectedly disappear. To satisfy call sites which really want the code under the conditional to completely disappear, we add GEM_DEBUG_WARN_ON and convert some of the callers to it. This one can also be used as both expression and statement. >From the above it follows GEM_DEBUG_WARN_ON should be used in situations where we are certain the condition will be hit during development, but at a place in code where error can be handled to the benefit of not crashing the machine. GEM_WARN_ON on the other hand should be used where condition may happen in production and we just want to distinguish the level of debugging output emitted between the production and debug build. v2: * Dropped BUG_ON hunk. Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Cc: Matthew Auld <matthew.auld@intel.com> Cc: Mika Kuoppala <mika.kuoppala@intel.com> Cc: Tomasz Lis <tomasz.lis@intel.com> Reviewed-by: Tomasz Lis <tomasz.lis@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20181012063142.16080-1-tvrtko.ursulin@linux.intel.com
2018-10-12 13:31:42 +07:00
if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
vma->node.size,
vma->vm->total)))
return -ENODEV;
drm/i915: GEM_WARN_ON considered harmful GEM_WARN_ON currently has dangerous semantics where it is completely compiled out on !GEM_DEBUG builds. This can leave users who expect it to be more like a WARN_ON, just without a warning in non-debug builds, in complete ignorance. Another gotcha with it is that it cannot be used as a statement. Which is again different from a standard kernel WARN_ON. This patch fixes both problems by making it behave as one would expect. It can now be used both as an expression and as statement, and also the condition evaluates properly in all builds - code under the conditional will therefore not unexpectedly disappear. To satisfy call sites which really want the code under the conditional to completely disappear, we add GEM_DEBUG_WARN_ON and convert some of the callers to it. This one can also be used as both expression and statement. >From the above it follows GEM_DEBUG_WARN_ON should be used in situations where we are certain the condition will be hit during development, but at a place in code where error can be handled to the benefit of not crashing the machine. GEM_WARN_ON on the other hand should be used where condition may happen in production and we just want to distinguish the level of debugging output emitted between the production and debug build. v2: * Dropped BUG_ON hunk. Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Cc: Matthew Auld <matthew.auld@intel.com> Cc: Mika Kuoppala <mika.kuoppala@intel.com> Cc: Tomasz Lis <tomasz.lis@intel.com> Reviewed-by: Tomasz Lis <tomasz.lis@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20181012063142.16080-1-tvrtko.ursulin@linux.intel.com
2018-10-12 13:31:42 +07:00
if (GEM_DEBUG_WARN_ON(!flags))
return -EINVAL;
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
bind_flags = flags;
bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
vma_flags = atomic_read(&vma->flags);
vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
if (flags & PIN_UPDATE)
bind_flags |= vma_flags;
else
bind_flags &= ~vma_flags;
if (bind_flags == 0)
return 0;
GEM_BUG_ON(!vma->pages);
trace_i915_vma_bind(vma, bind_flags);
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
if (work && (bind_flags & ~vma_flags) & vma->vm->bind_async_flags) {
work->vma = vma;
work->cache_level = cache_level;
work->flags = bind_flags | I915_VMA_ALLOC;
/*
* Note we only want to chain up to the migration fence on
* the pages (not the object itself). As we don't track that,
* yet, we have to use the exclusive fence instead.
*
* Also note that we do not want to track the async vma as
* part of the obj->resv->excl_fence as it only affects
* execution and not content or object's backing store lifetime.
*/
GEM_BUG_ON(i915_active_has_exclusive(&vma->active));
i915_active_set_exclusive(&vma->active, &work->base.dma);
work->base.dma.error = 0; /* enable the queue_work() */
if (vma->obj)
__i915_gem_object_pin_pages(vma->obj);
} else {
GEM_BUG_ON((bind_flags & ~vma_flags) & vma->vm->bind_async_flags);
ret = vma->ops->bind_vma(vma, cache_level, bind_flags);
if (ret)
return ret;
}
atomic_or(bind_flags, &vma->flags);
return 0;
}
void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
{
void __iomem *ptr;
int err;
/* Access through the GTT requires the device to be awake. */
assert_rpm_wakelock_held(vma->vm->gt->uncore->rpm);
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
if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
err = -ENODEV;
goto err;
}
GEM_BUG_ON(!i915_vma_is_ggtt(vma));
GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
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
ptr = READ_ONCE(vma->iomap);
if (ptr == NULL) {
ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
vma->node.start,
vma->node.size);
if (ptr == NULL) {
err = -ENOMEM;
goto err;
}
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
if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
io_mapping_unmap(ptr);
ptr = vma->iomap;
}
}
__i915_vma_pin(vma);
err = i915_vma_pin_fence(vma);
if (err)
goto err_unpin;
i915_vma_set_ggtt_write(vma);
return ptr;
err_unpin:
__i915_vma_unpin(vma);
err:
return IO_ERR_PTR(err);
}
void i915_vma_flush_writes(struct i915_vma *vma)
{
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
if (i915_vma_unset_ggtt_write(vma))
intel_gt_flush_ggtt_writes(vma->vm->gt);
}
void i915_vma_unpin_iomap(struct i915_vma *vma)
{
GEM_BUG_ON(vma->iomap == NULL);
i915_vma_flush_writes(vma);
i915_vma_unpin_fence(vma);
i915_vma_unpin(vma);
}
void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
{
struct i915_vma *vma;
struct drm_i915_gem_object *obj;
vma = fetch_and_zero(p_vma);
if (!vma)
return;
obj = vma->obj;
GEM_BUG_ON(!obj);
i915_vma_unpin(vma);
i915_vma_close(vma);
if (flags & I915_VMA_RELEASE_MAP)
i915_gem_object_unpin_map(obj);
i915_gem_object_put(obj);
}
bool i915_vma_misplaced(const struct i915_vma *vma,
u64 size, u64 alignment, u64 flags)
{
if (!drm_mm_node_allocated(&vma->node))
return false;
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
if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
return true;
if (vma->node.size < size)
return true;
GEM_BUG_ON(alignment && !is_power_of_2(alignment));
if (alignment && !IS_ALIGNED(vma->node.start, alignment))
return true;
if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
return true;
if (flags & PIN_OFFSET_BIAS &&
vma->node.start < (flags & PIN_OFFSET_MASK))
return true;
if (flags & PIN_OFFSET_FIXED &&
vma->node.start != (flags & PIN_OFFSET_MASK))
return true;
return false;
}
void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
{
bool mappable, fenceable;
GEM_BUG_ON(!i915_vma_is_ggtt(vma));
GEM_BUG_ON(!vma->fence_size);
fenceable = (vma->node.size >= vma->fence_size &&
IS_ALIGNED(vma->node.start, vma->fence_alignment));
mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
if (mappable && fenceable)
set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
else
clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
}
bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
{
struct drm_mm_node *node = &vma->node;
struct drm_mm_node *other;
/*
* On some machines we have to be careful when putting differing types
* of snoopable memory together to avoid the prefetcher crossing memory
* domains and dying. During vm initialisation, we decide whether or not
* these constraints apply and set the drm_mm.color_adjust
* appropriately.
*/
if (!i915_vm_has_cache_coloring(vma->vm))
return true;
/* Only valid to be called on an already inserted vma */
GEM_BUG_ON(!drm_mm_node_allocated(node));
GEM_BUG_ON(list_empty(&node->node_list));
other = list_prev_entry(node, node_list);
if (i915_node_color_differs(other, color) &&
!drm_mm_hole_follows(other))
return false;
other = list_next_entry(node, node_list);
if (i915_node_color_differs(other, color) &&
!drm_mm_hole_follows(node))
return false;
return true;
}
static void assert_bind_count(const struct drm_i915_gem_object *obj)
{
/*
* Combine the assertion that the object is bound and that we have
* pinned its pages. But we should never have bound the object
* more than we have pinned its pages. (For complete accuracy, we
* assume that no else is pinning the pages, but as a rough assertion
* that we will not run into problems later, this will do!)
*/
GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < atomic_read(&obj->bind_count));
}
/**
* i915_vma_insert - finds a slot for the vma in its address space
* @vma: the vma
* @size: requested size in bytes (can be larger than the VMA)
* @alignment: required alignment
* @flags: mask of PIN_* flags to use
*
* First we try to allocate some free space that meets the requirements for
* the VMA. Failiing that, if the flags permit, it will evict an old VMA,
* preferrably the oldest idle entry to make room for the new VMA.
*
* Returns:
* 0 on success, negative error code otherwise.
*/
static int
i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
{
unsigned long color;
u64 start, end;
int ret;
GEM_BUG_ON(i915_vma_is_closed(vma));
GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
size = max(size, vma->size);
alignment = max(alignment, vma->display_alignment);
if (flags & PIN_MAPPABLE) {
size = max_t(typeof(size), size, vma->fence_size);
alignment = max_t(typeof(alignment),
alignment, vma->fence_alignment);
}
GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
GEM_BUG_ON(!is_power_of_2(alignment));
start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
end = vma->vm->total;
if (flags & PIN_MAPPABLE)
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
end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
if (flags & PIN_ZONE_4G)
end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
/* If binding the object/GGTT view requires more space than the entire
* aperture has, reject it early before evicting everything in a vain
* attempt to find space.
*/
if (size > end) {
DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
size, flags & PIN_MAPPABLE ? "mappable" : "total",
end);
drm/i915: Eliminate lots of iterations over the execobjects array The major scaling bottleneck in execbuffer is the processing of the execobjects. Creating an auxiliary list is inefficient when compared to using the execobject array we already have allocated. Reservation is then split into phases. As we lookup up the VMA, we try and bind it back into active location. Only if that fails, do we add it to the unbound list for phase 2. In phase 2, we try and add all those objects that could not fit into their previous location, with fallback to retrying all objects and evicting the VM in case of severe fragmentation. (This is the same as before, except that phase 1 is now done inline with looking up the VMA to avoid an iteration over the execobject array. In the ideal case, we eliminate the separate reservation phase). During the reservation phase, we only evict from the VM between passes (rather than currently as we try to fit every new VMA). In testing with Unreal Engine's Atlantis demo which stresses the eviction logic on gen7 class hardware, this speed up the framerate by a factor of 2. The second loop amalgamation is between move_to_gpu and move_to_active. As we always submit the request, even if incomplete, we can use the current request to track active VMA as we perform the flushes and synchronisation required. The next big advancement is to avoid copying back to the user any execobjects and relocations that are not changed. v2: Add a Theory of Operation spiel. v3: Fall back to slow relocations in preparation for flushing userptrs. v4: Document struct members, factor out eb_validate_vma(), add a few more comments to explain some magic and hide other magic behind macros. Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2017-06-16 21:05:19 +07:00
return -ENOSPC;
}
color = 0;
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
if (vma->obj && i915_vm_has_cache_coloring(vma->vm))
color = vma->obj->cache_level;
if (flags & PIN_OFFSET_FIXED) {
u64 offset = flags & PIN_OFFSET_MASK;
if (!IS_ALIGNED(offset, alignment) ||
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
range_overflows(offset, size, end))
return -EINVAL;
ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
size, offset, color,
flags);
if (ret)
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
return ret;
} else {
/*
* We only support huge gtt pages through the 48b PPGTT,
* however we also don't want to force any alignment for
* objects which need to be tightly packed into the low 32bits.
*
* Note that we assume that GGTT are limited to 4GiB for the
* forseeable future. See also i915_ggtt_offset().
*/
if (upper_32_bits(end - 1) &&
vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
/*
* We can't mix 64K and 4K PTEs in the same page-table
* (2M block), and so to avoid the ugliness and
* complexity of coloring we opt for just aligning 64K
* objects to 2M.
*/
u64 page_alignment =
rounddown_pow_of_two(vma->page_sizes.sg |
I915_GTT_PAGE_SIZE_2M);
/*
* Check we don't expand for the limited Global GTT
* (mappable aperture is even more precious!). This
* also checks that we exclude the aliasing-ppgtt.
*/
GEM_BUG_ON(i915_vma_is_ggtt(vma));
alignment = max(alignment, page_alignment);
if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
size = round_up(size, I915_GTT_PAGE_SIZE_2M);
}
ret = i915_gem_gtt_insert(vma->vm, &vma->node,
size, alignment, color,
start, end, flags);
if (ret)
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
return ret;
GEM_BUG_ON(vma->node.start < start);
GEM_BUG_ON(vma->node.start + vma->node.size > end);
}
GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
if (vma->obj) {
struct drm_i915_gem_object *obj = vma->obj;
atomic_inc(&obj->bind_count);
assert_bind_count(obj);
}
list_add_tail(&vma->vm_link, &vma->vm->bound_list);
return 0;
}
static void
i915_vma_detach(struct i915_vma *vma)
{
GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
/*
* And finally now the object is completely decoupled from this
* vma, we can drop its hold on the backing storage and allow
* it to be reaped by the shrinker.
*/
list_del(&vma->vm_link);
if (vma->obj) {
struct drm_i915_gem_object *obj = vma->obj;
drm/i915: Report all objects with allocated pages to the shrinker Currently, we try to report to the shrinker the precise number of objects (pages) that are available to be reaped at this moment. This requires searching all objects with allocated pages to see if they fulfill the search criteria, and this count is performed quite frequently. (The shrinker tries to free ~128 pages on each invocation, before which we count all the objects; counting takes longer than unbinding the objects!) If we take the pragmatic view that with sufficient desire, all objects are eventually reapable (they become inactive, or no longer used as framebuffer etc), we can simply return the count of pinned pages maintained during get_pages/put_pages rather than walk the lists every time. The downside is that we may (slightly) over-report the number of objects/pages we could shrink and so penalize ourselves by shrinking more than required. This is mitigated by keeping the order in which we shrink objects such that we avoid penalizing active and frequently used objects, and if memory is so tight that we need to free them we would need to anyway. v2: Only expose shrinkable objects to the shrinker; a small reduction in not considering stolen and foreign objects. v3: Restore the tracking from a "backup" copy from before the gem/ split Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Cc: Matthew Auld <matthew.auld@intel.com> Reviewed-by: Matthew Auld <matthew.auld@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190530203500.26272-2-chris@chris-wilson.co.uk
2019-05-31 03:35:00 +07:00
assert_bind_count(obj);
atomic_dec(&obj->bind_count);
}
}
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 bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
{
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
unsigned int bound;
bool pinned = true;
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
bound = atomic_read(&vma->flags);
do {
if (unlikely(flags & ~bound))
return false;
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
if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
return false;
if (!(bound & I915_VMA_PIN_MASK))
goto unpinned;
GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
} while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
return true;
unpinned:
/*
* If pin_count==0, but we are bound, check under the lock to avoid
* racing with a concurrent i915_vma_unbind().
*/
mutex_lock(&vma->vm->mutex);
do {
if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) {
pinned = false;
break;
}
if (unlikely(flags & ~bound)) {
pinned = false;
break;
}
} while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
mutex_unlock(&vma->vm->mutex);
return pinned;
}
static int vma_get_pages(struct i915_vma *vma)
{
int err = 0;
if (atomic_add_unless(&vma->pages_count, 1, 0))
return 0;
/* Allocations ahoy! */
if (mutex_lock_interruptible(&vma->pages_mutex))
return -EINTR;
if (!atomic_read(&vma->pages_count)) {
if (vma->obj) {
err = i915_gem_object_pin_pages(vma->obj);
if (err)
goto unlock;
}
err = vma->ops->set_pages(vma);
if (err) {
if (vma->obj)
i915_gem_object_unpin_pages(vma->obj);
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
goto unlock;
}
}
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
atomic_inc(&vma->pages_count);
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
unlock:
mutex_unlock(&vma->pages_mutex);
return err;
}
static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
{
/* We allocate under vma_get_pages, so beware the shrinker */
mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING);
GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
if (atomic_sub_return(count, &vma->pages_count) == 0) {
vma->ops->clear_pages(vma);
GEM_BUG_ON(vma->pages);
if (vma->obj)
i915_gem_object_unpin_pages(vma->obj);
}
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
mutex_unlock(&vma->pages_mutex);
}
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 void vma_put_pages(struct i915_vma *vma)
{
if (atomic_add_unless(&vma->pages_count, -1, 1))
return;
__vma_put_pages(vma, 1);
}
static void vma_unbind_pages(struct i915_vma *vma)
{
unsigned int count;
lockdep_assert_held(&vma->vm->mutex);
/* The upper portion of pages_count is the number of bindings */
count = atomic_read(&vma->pages_count);
count >>= I915_VMA_PAGES_BIAS;
GEM_BUG_ON(!count);
__vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
}
int i915_vma_pin(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
{
struct i915_vma_work *work = NULL;
unsigned int bound;
int err;
BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
GEM_BUG_ON(flags & PIN_UPDATE);
GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
/* First try and grab the pin without rebinding the vma */
if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
return 0;
err = vma_get_pages(vma);
if (err)
return err;
if (flags & vma->vm->bind_async_flags) {
work = i915_vma_work();
if (!work) {
err = -ENOMEM;
goto err_pages;
}
}
/* No more allocations allowed once we hold vm->mutex */
err = mutex_lock_interruptible(&vma->vm->mutex);
if (err)
goto err_fence;
bound = atomic_read(&vma->flags);
if (unlikely(bound & I915_VMA_ERROR)) {
err = -ENOMEM;
goto err_unlock;
}
if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
err = -EAGAIN; /* pins are meant to be fairly temporary */
goto err_unlock;
}
if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
__i915_vma_pin(vma);
goto err_unlock;
}
err = i915_active_acquire(&vma->active);
if (err)
goto err_unlock;
if (!(bound & I915_VMA_BIND_MASK)) {
err = i915_vma_insert(vma, size, alignment, flags);
if (err)
goto err_active;
if (i915_is_ggtt(vma->vm))
__i915_vma_set_map_and_fenceable(vma);
}
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
GEM_BUG_ON(!vma->pages);
err = i915_vma_bind(vma,
vma->obj ? vma->obj->cache_level : 0,
flags, work);
if (err)
goto err_remove;
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
/* There should only be at most 2 active bindings (user, global) */
GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
list_move_tail(&vma->vm_link, &vma->vm->bound_list);
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
__i915_vma_pin(vma);
GEM_BUG_ON(!i915_vma_is_pinned(vma));
GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
err_remove:
if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
i915_vma_detach(vma);
drm_mm_remove_node(&vma->node);
}
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
err_active:
i915_active_release(&vma->active);
err_unlock:
mutex_unlock(&vma->vm->mutex);
err_fence:
if (work)
dma_fence_work_commit(&work->base);
err_pages:
vma_put_pages(vma);
return err;
}
drm/i915: Lazily unbind vma on close When userspace is passing around swapbuffers using DRI, we frequently have to open and close the same object in the foreign address space. This shows itself as the same object being rebound at roughly 30fps (with a second object also being rebound at 30fps), which involves us having to rewrite the page tables and maintain the drm_mm range manager every time. However, since the object still exists and it is only the local handle that disappears, if we are lazy and do not unbind the VMA immediately when the local user closes the object but defer it until the GPU is idle, then we can reuse the same VMA binding. We still have to be careful to mark the handle and lookup tables as closed to maintain the uABI, just allowing the underlying VMA to be resurrected if the user is able to access the same object from the same context again. If the object itself is destroyed (neither userspace keeping a handle to it), the VMA will be reaped immediately as usual. In the future, this will be even more useful as instantiating a new VMA for use on the GPU will become heavier. A nuisance indeed, so nip it in the bud. v2: s/__i915_vma_final_close/i915_vma_destroy/ etc. v3: Leave a hint as to why we deferred the unbind on close. 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/20180503195115.22309-1-chris@chris-wilson.co.uk
2018-05-04 02:51:14 +07:00
void i915_vma_close(struct i915_vma *vma)
{
struct intel_gt *gt = vma->vm->gt;
unsigned long flags;
drm/i915: Lazily unbind vma on close When userspace is passing around swapbuffers using DRI, we frequently have to open and close the same object in the foreign address space. This shows itself as the same object being rebound at roughly 30fps (with a second object also being rebound at 30fps), which involves us having to rewrite the page tables and maintain the drm_mm range manager every time. However, since the object still exists and it is only the local handle that disappears, if we are lazy and do not unbind the VMA immediately when the local user closes the object but defer it until the GPU is idle, then we can reuse the same VMA binding. We still have to be careful to mark the handle and lookup tables as closed to maintain the uABI, just allowing the underlying VMA to be resurrected if the user is able to access the same object from the same context again. If the object itself is destroyed (neither userspace keeping a handle to it), the VMA will be reaped immediately as usual. In the future, this will be even more useful as instantiating a new VMA for use on the GPU will become heavier. A nuisance indeed, so nip it in the bud. v2: s/__i915_vma_final_close/i915_vma_destroy/ etc. v3: Leave a hint as to why we deferred the unbind on close. 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/20180503195115.22309-1-chris@chris-wilson.co.uk
2018-05-04 02:51:14 +07:00
GEM_BUG_ON(i915_vma_is_closed(vma));
/*
* We defer actually closing, unbinding and destroying the VMA until
* the next idle point, or if the object is freed in the meantime. By
* postponing the unbind, we allow for it to be resurrected by the
* client, avoiding the work required to rebind the VMA. This is
* advantageous for DRI, where the client/server pass objects
* between themselves, temporarily opening a local VMA to the
* object, and then closing it again. The same object is then reused
* on the next frame (or two, depending on the depth of the swap queue)
* causing us to rebind the VMA once more. This ends up being a lot
* of wasted work for the steady state.
*/
spin_lock_irqsave(&gt->closed_lock, flags);
list_add(&vma->closed_link, &gt->closed_vma);
spin_unlock_irqrestore(&gt->closed_lock, flags);
drm/i915: Lazily unbind vma on close When userspace is passing around swapbuffers using DRI, we frequently have to open and close the same object in the foreign address space. This shows itself as the same object being rebound at roughly 30fps (with a second object also being rebound at 30fps), which involves us having to rewrite the page tables and maintain the drm_mm range manager every time. However, since the object still exists and it is only the local handle that disappears, if we are lazy and do not unbind the VMA immediately when the local user closes the object but defer it until the GPU is idle, then we can reuse the same VMA binding. We still have to be careful to mark the handle and lookup tables as closed to maintain the uABI, just allowing the underlying VMA to be resurrected if the user is able to access the same object from the same context again. If the object itself is destroyed (neither userspace keeping a handle to it), the VMA will be reaped immediately as usual. In the future, this will be even more useful as instantiating a new VMA for use on the GPU will become heavier. A nuisance indeed, so nip it in the bud. v2: s/__i915_vma_final_close/i915_vma_destroy/ etc. v3: Leave a hint as to why we deferred the unbind on close. 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/20180503195115.22309-1-chris@chris-wilson.co.uk
2018-05-04 02:51:14 +07:00
}
static void __i915_vma_remove_closed(struct i915_vma *vma)
drm/i915: Lazily unbind vma on close When userspace is passing around swapbuffers using DRI, we frequently have to open and close the same object in the foreign address space. This shows itself as the same object being rebound at roughly 30fps (with a second object also being rebound at 30fps), which involves us having to rewrite the page tables and maintain the drm_mm range manager every time. However, since the object still exists and it is only the local handle that disappears, if we are lazy and do not unbind the VMA immediately when the local user closes the object but defer it until the GPU is idle, then we can reuse the same VMA binding. We still have to be careful to mark the handle and lookup tables as closed to maintain the uABI, just allowing the underlying VMA to be resurrected if the user is able to access the same object from the same context again. If the object itself is destroyed (neither userspace keeping a handle to it), the VMA will be reaped immediately as usual. In the future, this will be even more useful as instantiating a new VMA for use on the GPU will become heavier. A nuisance indeed, so nip it in the bud. v2: s/__i915_vma_final_close/i915_vma_destroy/ etc. v3: Leave a hint as to why we deferred the unbind on close. 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/20180503195115.22309-1-chris@chris-wilson.co.uk
2018-05-04 02:51:14 +07:00
{
struct intel_gt *gt = vma->vm->gt;
drm/i915: Lazily unbind vma on close When userspace is passing around swapbuffers using DRI, we frequently have to open and close the same object in the foreign address space. This shows itself as the same object being rebound at roughly 30fps (with a second object also being rebound at 30fps), which involves us having to rewrite the page tables and maintain the drm_mm range manager every time. However, since the object still exists and it is only the local handle that disappears, if we are lazy and do not unbind the VMA immediately when the local user closes the object but defer it until the GPU is idle, then we can reuse the same VMA binding. We still have to be careful to mark the handle and lookup tables as closed to maintain the uABI, just allowing the underlying VMA to be resurrected if the user is able to access the same object from the same context again. If the object itself is destroyed (neither userspace keeping a handle to it), the VMA will be reaped immediately as usual. In the future, this will be even more useful as instantiating a new VMA for use on the GPU will become heavier. A nuisance indeed, so nip it in the bud. v2: s/__i915_vma_final_close/i915_vma_destroy/ etc. v3: Leave a hint as to why we deferred the unbind on close. 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/20180503195115.22309-1-chris@chris-wilson.co.uk
2018-05-04 02:51:14 +07:00
spin_lock_irq(&gt->closed_lock);
list_del_init(&vma->closed_link);
spin_unlock_irq(&gt->closed_lock);
}
void i915_vma_reopen(struct i915_vma *vma)
{
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
if (i915_vma_is_closed(vma))
__i915_vma_remove_closed(vma);
drm/i915: Lazily unbind vma on close When userspace is passing around swapbuffers using DRI, we frequently have to open and close the same object in the foreign address space. This shows itself as the same object being rebound at roughly 30fps (with a second object also being rebound at 30fps), which involves us having to rewrite the page tables and maintain the drm_mm range manager every time. However, since the object still exists and it is only the local handle that disappears, if we are lazy and do not unbind the VMA immediately when the local user closes the object but defer it until the GPU is idle, then we can reuse the same VMA binding. We still have to be careful to mark the handle and lookup tables as closed to maintain the uABI, just allowing the underlying VMA to be resurrected if the user is able to access the same object from the same context again. If the object itself is destroyed (neither userspace keeping a handle to it), the VMA will be reaped immediately as usual. In the future, this will be even more useful as instantiating a new VMA for use on the GPU will become heavier. A nuisance indeed, so nip it in the bud. v2: s/__i915_vma_final_close/i915_vma_destroy/ etc. v3: Leave a hint as to why we deferred the unbind on close. 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/20180503195115.22309-1-chris@chris-wilson.co.uk
2018-05-04 02:51:14 +07:00
}
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
void i915_vma_destroy(struct i915_vma *vma)
{
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
if (drm_mm_node_allocated(&vma->node)) {
mutex_lock(&vma->vm->mutex);
atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
WARN_ON(__i915_vma_unbind(vma));
mutex_unlock(&vma->vm->mutex);
GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
}
GEM_BUG_ON(i915_vma_is_active(vma));
if (vma->obj) {
struct drm_i915_gem_object *obj = vma->obj;
spin_lock(&obj->vma.lock);
list_del(&vma->obj_link);
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
rb_erase(&vma->obj_node, &obj->vma.tree);
spin_unlock(&obj->vma.lock);
}
__i915_vma_remove_closed(vma);
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
i915_vm_put(vma->vm);
drm/i915: Lazily unbind vma on close When userspace is passing around swapbuffers using DRI, we frequently have to open and close the same object in the foreign address space. This shows itself as the same object being rebound at roughly 30fps (with a second object also being rebound at 30fps), which involves us having to rewrite the page tables and maintain the drm_mm range manager every time. However, since the object still exists and it is only the local handle that disappears, if we are lazy and do not unbind the VMA immediately when the local user closes the object but defer it until the GPU is idle, then we can reuse the same VMA binding. We still have to be careful to mark the handle and lookup tables as closed to maintain the uABI, just allowing the underlying VMA to be resurrected if the user is able to access the same object from the same context again. If the object itself is destroyed (neither userspace keeping a handle to it), the VMA will be reaped immediately as usual. In the future, this will be even more useful as instantiating a new VMA for use on the GPU will become heavier. A nuisance indeed, so nip it in the bud. v2: s/__i915_vma_final_close/i915_vma_destroy/ etc. v3: Leave a hint as to why we deferred the unbind on close. 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/20180503195115.22309-1-chris@chris-wilson.co.uk
2018-05-04 02:51:14 +07:00
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
i915_active_fini(&vma->active);
i915_vma_free(vma);
drm/i915: Lazily unbind vma on close When userspace is passing around swapbuffers using DRI, we frequently have to open and close the same object in the foreign address space. This shows itself as the same object being rebound at roughly 30fps (with a second object also being rebound at 30fps), which involves us having to rewrite the page tables and maintain the drm_mm range manager every time. However, since the object still exists and it is only the local handle that disappears, if we are lazy and do not unbind the VMA immediately when the local user closes the object but defer it until the GPU is idle, then we can reuse the same VMA binding. We still have to be careful to mark the handle and lookup tables as closed to maintain the uABI, just allowing the underlying VMA to be resurrected if the user is able to access the same object from the same context again. If the object itself is destroyed (neither userspace keeping a handle to it), the VMA will be reaped immediately as usual. In the future, this will be even more useful as instantiating a new VMA for use on the GPU will become heavier. A nuisance indeed, so nip it in the bud. v2: s/__i915_vma_final_close/i915_vma_destroy/ etc. v3: Leave a hint as to why we deferred the unbind on close. 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/20180503195115.22309-1-chris@chris-wilson.co.uk
2018-05-04 02:51:14 +07:00
}
void i915_vma_parked(struct intel_gt *gt)
drm/i915: Lazily unbind vma on close When userspace is passing around swapbuffers using DRI, we frequently have to open and close the same object in the foreign address space. This shows itself as the same object being rebound at roughly 30fps (with a second object also being rebound at 30fps), which involves us having to rewrite the page tables and maintain the drm_mm range manager every time. However, since the object still exists and it is only the local handle that disappears, if we are lazy and do not unbind the VMA immediately when the local user closes the object but defer it until the GPU is idle, then we can reuse the same VMA binding. We still have to be careful to mark the handle and lookup tables as closed to maintain the uABI, just allowing the underlying VMA to be resurrected if the user is able to access the same object from the same context again. If the object itself is destroyed (neither userspace keeping a handle to it), the VMA will be reaped immediately as usual. In the future, this will be even more useful as instantiating a new VMA for use on the GPU will become heavier. A nuisance indeed, so nip it in the bud. v2: s/__i915_vma_final_close/i915_vma_destroy/ etc. v3: Leave a hint as to why we deferred the unbind on close. 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/20180503195115.22309-1-chris@chris-wilson.co.uk
2018-05-04 02:51:14 +07:00
{
struct i915_vma *vma, *next;
spin_lock_irq(&gt->closed_lock);
list_for_each_entry_safe(vma, next, &gt->closed_vma, closed_link) {
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
struct drm_i915_gem_object *obj = vma->obj;
struct i915_address_space *vm = vma->vm;
/* XXX All to avoid keeping a reference on i915_vma itself */
if (!kref_get_unless_zero(&obj->base.refcount))
continue;
if (!i915_vm_tryopen(vm)) {
i915_gem_object_put(obj);
obj = NULL;
}
spin_unlock_irq(&gt->closed_lock);
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
if (obj) {
i915_vma_destroy(vma);
i915_gem_object_put(obj);
}
drm/i915: Lazily unbind vma on close When userspace is passing around swapbuffers using DRI, we frequently have to open and close the same object in the foreign address space. This shows itself as the same object being rebound at roughly 30fps (with a second object also being rebound at 30fps), which involves us having to rewrite the page tables and maintain the drm_mm range manager every time. However, since the object still exists and it is only the local handle that disappears, if we are lazy and do not unbind the VMA immediately when the local user closes the object but defer it until the GPU is idle, then we can reuse the same VMA binding. We still have to be careful to mark the handle and lookup tables as closed to maintain the uABI, just allowing the underlying VMA to be resurrected if the user is able to access the same object from the same context again. If the object itself is destroyed (neither userspace keeping a handle to it), the VMA will be reaped immediately as usual. In the future, this will be even more useful as instantiating a new VMA for use on the GPU will become heavier. A nuisance indeed, so nip it in the bud. v2: s/__i915_vma_final_close/i915_vma_destroy/ etc. v3: Leave a hint as to why we deferred the unbind on close. 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/20180503195115.22309-1-chris@chris-wilson.co.uk
2018-05-04 02:51:14 +07:00
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
i915_vm_close(vm);
/* Restart after dropping lock */
spin_lock_irq(&gt->closed_lock);
next = list_first_entry(&gt->closed_vma,
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
typeof(*next), closed_link);
}
spin_unlock_irq(&gt->closed_lock);
}
static void __i915_vma_iounmap(struct i915_vma *vma)
{
GEM_BUG_ON(i915_vma_is_pinned(vma));
if (vma->iomap == NULL)
return;
io_mapping_unmap(vma->iomap);
vma->iomap = NULL;
}
void i915_vma_revoke_mmap(struct i915_vma *vma)
{
struct drm_vma_offset_node *node = &vma->obj->base.vma_node;
u64 vma_offset;
lockdep_assert_held(&vma->vm->mutex);
if (!i915_vma_has_userfault(vma))
return;
GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
GEM_BUG_ON(!vma->obj->userfault_count);
vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
drm_vma_node_offset_addr(node) + vma_offset,
vma->size,
1);
i915_vma_unset_userfault(vma);
if (!--vma->obj->userfault_count)
list_del(&vma->obj->userfault_link);
}
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_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
{
int err;
GEM_BUG_ON(!i915_vma_is_pinned(vma));
/* Wait for the vma to be bound before we start! */
err = i915_request_await_active(rq, &vma->active);
if (err)
return err;
return i915_active_add_request(&vma->active, rq);
}
int i915_vma_move_to_active(struct i915_vma *vma,
struct i915_request *rq,
unsigned int flags)
{
struct drm_i915_gem_object *obj = vma->obj;
int err;
assert_object_held(obj);
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
err = __i915_vma_move_to_active(vma, rq);
if (unlikely(err))
return err;
if (flags & EXEC_OBJECT_WRITE) {
if (intel_frontbuffer_invalidate(obj->frontbuffer, ORIGIN_CS))
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
i915_active_add_request(&obj->frontbuffer->write, rq);
dma_resv_add_excl_fence(vma->resv, &rq->fence);
obj->write_domain = I915_GEM_DOMAIN_RENDER;
obj->read_domains = 0;
} else {
err = dma_resv_reserve_shared(vma->resv, 1);
if (unlikely(err))
return err;
dma_resv_add_shared_fence(vma->resv, &rq->fence);
obj->write_domain = 0;
}
obj->read_domains |= I915_GEM_GPU_DOMAINS;
obj->mm.dirty = true;
GEM_BUG_ON(!i915_vma_is_active(vma));
return 0;
}
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_vma_unbind(struct i915_vma *vma)
{
int ret;
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
lockdep_assert_held(&vma->vm->mutex);
/*
* First wait upon any activity as retiring the request may
* have side-effects such as unpinning or even unbinding this vma.
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
*
* XXX Actually waiting under the vm->mutex is a hinderance and
* should be pipelined wherever possible. In cases where that is
* unavoidable, we should lift the wait to before the mutex.
*/
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
ret = i915_vma_sync(vma);
if (ret)
return ret;
GEM_BUG_ON(i915_vma_is_active(vma));
if (i915_vma_is_pinned(vma)) {
vma_print_allocator(vma, "is pinned");
return -EBUSY;
}
GEM_BUG_ON(i915_vma_is_active(vma));
if (!drm_mm_node_allocated(&vma->node))
drm/i915: Lazily unbind vma on close When userspace is passing around swapbuffers using DRI, we frequently have to open and close the same object in the foreign address space. This shows itself as the same object being rebound at roughly 30fps (with a second object also being rebound at 30fps), which involves us having to rewrite the page tables and maintain the drm_mm range manager every time. However, since the object still exists and it is only the local handle that disappears, if we are lazy and do not unbind the VMA immediately when the local user closes the object but defer it until the GPU is idle, then we can reuse the same VMA binding. We still have to be careful to mark the handle and lookup tables as closed to maintain the uABI, just allowing the underlying VMA to be resurrected if the user is able to access the same object from the same context again. If the object itself is destroyed (neither userspace keeping a handle to it), the VMA will be reaped immediately as usual. In the future, this will be even more useful as instantiating a new VMA for use on the GPU will become heavier. A nuisance indeed, so nip it in the bud. v2: s/__i915_vma_final_close/i915_vma_destroy/ etc. v3: Leave a hint as to why we deferred the unbind on close. 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/20180503195115.22309-1-chris@chris-wilson.co.uk
2018-05-04 02:51:14 +07:00
return 0;
if (i915_vma_is_map_and_fenceable(vma)) {
/*
* Check that we have flushed all writes through the GGTT
* before the unbind, other due to non-strict nature of those
* indirect writes they may end up referencing the GGTT PTE
* after the unbind.
*/
i915_vma_flush_writes(vma);
GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
/* release the fence reg _after_ flushing */
ret = i915_vma_revoke_fence(vma);
if (ret)
return ret;
/* Force a pagefault for domain tracking on next user access */
i915_vma_revoke_mmap(vma);
__i915_vma_iounmap(vma);
clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
}
GEM_BUG_ON(vma->fence);
GEM_BUG_ON(i915_vma_has_userfault(vma));
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
if (likely(atomic_read(&vma->vm->open))) {
trace_i915_vma_unbind(vma);
vma->ops->unbind_vma(vma);
}
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
atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR), &vma->flags);
i915_vma_detach(vma);
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
vma_unbind_pages(vma);
drm_mm_remove_node(&vma->node); /* pairs with i915_vma_destroy() */
return 0;
}
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_vma_unbind(struct i915_vma *vma)
{
struct i915_address_space *vm = vma->vm;
int err;
err = mutex_lock_interruptible(&vm->mutex);
if (err)
return err;
err = __i915_vma_unbind(vma);
mutex_unlock(&vm->mutex);
return err;
}
struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
{
i915_gem_object_make_unshrinkable(vma->obj);
return vma;
}
void i915_vma_make_shrinkable(struct i915_vma *vma)
{
i915_gem_object_make_shrinkable(vma->obj);
}
void i915_vma_make_purgeable(struct i915_vma *vma)
{
i915_gem_object_make_purgeable(vma->obj);
}
#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
#include "selftests/i915_vma.c"
#endif
static void i915_global_vma_shrink(void)
{
kmem_cache_shrink(global.slab_vmas);
}
static void i915_global_vma_exit(void)
{
kmem_cache_destroy(global.slab_vmas);
}
static struct i915_global_vma global = { {
.shrink = i915_global_vma_shrink,
.exit = i915_global_vma_exit,
} };
int __init i915_global_vma_init(void)
{
global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
if (!global.slab_vmas)
return -ENOMEM;
i915_global_register(&global.base);
return 0;
}