2 * Copyright © 2008-2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eric Anholt <eric@anholt.net>
29 #include <drm/drm_vma_manager.h>
30 #include <drm/i915_drm.h>
32 #include "i915_gem_clflush.h"
33 #include "i915_vgpu.h"
34 #include "i915_trace.h"
35 #include "intel_drv.h"
36 #include "intel_frontbuffer.h"
37 #include "intel_mocs.h"
38 #include "intel_workarounds.h"
39 #include "i915_gemfs.h"
40 #include <linux/dma-fence-array.h>
41 #include <linux/kthread.h>
42 #include <linux/reservation.h>
43 #include <linux/shmem_fs.h>
44 #include <linux/slab.h>
45 #include <linux/stop_machine.h>
46 #include <linux/swap.h>
47 #include <linux/pci.h>
48 #include <linux/dma-buf.h>
50 static void i915_gem_flush_free_objects(struct drm_i915_private *i915);
52 static bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj)
57 if (!(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE))
60 return obj->pin_global; /* currently in use by HW, keep flushed */
64 insert_mappable_node(struct i915_ggtt *ggtt,
65 struct drm_mm_node *node, u32 size)
67 memset(node, 0, sizeof(*node));
68 return drm_mm_insert_node_in_range(&ggtt->vm.mm, node,
69 size, 0, I915_COLOR_UNEVICTABLE,
70 0, ggtt->mappable_end,
75 remove_mappable_node(struct drm_mm_node *node)
77 drm_mm_remove_node(node);
80 /* some bookkeeping */
81 static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
84 spin_lock(&dev_priv->mm.object_stat_lock);
85 dev_priv->mm.object_count++;
86 dev_priv->mm.object_memory += size;
87 spin_unlock(&dev_priv->mm.object_stat_lock);
90 static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
93 spin_lock(&dev_priv->mm.object_stat_lock);
94 dev_priv->mm.object_count--;
95 dev_priv->mm.object_memory -= size;
96 spin_unlock(&dev_priv->mm.object_stat_lock);
100 i915_gem_wait_for_error(struct i915_gpu_error *error)
107 * Only wait 10 seconds for the gpu reset to complete to avoid hanging
108 * userspace. If it takes that long something really bad is going on and
109 * we should simply try to bail out and fail as gracefully as possible.
111 ret = wait_event_interruptible_timeout(error->reset_queue,
112 !i915_reset_backoff(error),
115 DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
117 } else if (ret < 0) {
124 int i915_mutex_lock_interruptible(struct drm_device *dev)
126 struct drm_i915_private *dev_priv = to_i915(dev);
129 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
133 ret = mutex_lock_interruptible(&dev->struct_mutex);
140 static u32 __i915_gem_park(struct drm_i915_private *i915)
144 lockdep_assert_held(&i915->drm.struct_mutex);
145 GEM_BUG_ON(i915->gt.active_requests);
146 GEM_BUG_ON(!list_empty(&i915->gt.active_rings));
149 return I915_EPOCH_INVALID;
151 GEM_BUG_ON(i915->gt.epoch == I915_EPOCH_INVALID);
154 * Be paranoid and flush a concurrent interrupt to make sure
155 * we don't reactivate any irq tasklets after parking.
157 * FIXME: Note that even though we have waited for execlists to be idle,
158 * there may still be an in-flight interrupt even though the CSB
159 * is now empty. synchronize_irq() makes sure that a residual interrupt
160 * is completed before we continue, but it doesn't prevent the HW from
161 * raising a spurious interrupt later. To complete the shield we should
162 * coordinate disabling the CS irq with flushing the interrupts.
164 synchronize_irq(i915->drm.irq);
166 intel_engines_park(i915);
167 i915_timelines_park(i915);
169 i915_pmu_gt_parked(i915);
170 i915_vma_parked(i915);
172 i915->gt.awake = false;
174 if (INTEL_GEN(i915) >= 6)
177 intel_display_power_put(i915, POWER_DOMAIN_GT_IRQ);
179 intel_runtime_pm_put(i915);
181 return i915->gt.epoch;
184 void i915_gem_park(struct drm_i915_private *i915)
188 lockdep_assert_held(&i915->drm.struct_mutex);
189 GEM_BUG_ON(i915->gt.active_requests);
194 /* Defer the actual call to __i915_gem_park() to prevent ping-pongs */
195 mod_delayed_work(i915->wq, &i915->gt.idle_work, msecs_to_jiffies(100));
198 void i915_gem_unpark(struct drm_i915_private *i915)
202 lockdep_assert_held(&i915->drm.struct_mutex);
203 GEM_BUG_ON(!i915->gt.active_requests);
208 intel_runtime_pm_get_noresume(i915);
211 * It seems that the DMC likes to transition between the DC states a lot
212 * when there are no connected displays (no active power domains) during
213 * command submission.
215 * This activity has negative impact on the performance of the chip with
216 * huge latencies observed in the interrupt handler and elsewhere.
218 * Work around it by grabbing a GT IRQ power domain whilst there is any
219 * GT activity, preventing any DC state transitions.
221 intel_display_power_get(i915, POWER_DOMAIN_GT_IRQ);
223 i915->gt.awake = true;
224 if (unlikely(++i915->gt.epoch == 0)) /* keep 0 as invalid */
227 intel_enable_gt_powersave(i915);
228 i915_update_gfx_val(i915);
229 if (INTEL_GEN(i915) >= 6)
231 i915_pmu_gt_unparked(i915);
233 intel_engines_unpark(i915);
235 i915_queue_hangcheck(i915);
237 queue_delayed_work(i915->wq,
238 &i915->gt.retire_work,
239 round_jiffies_up_relative(HZ));
243 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
244 struct drm_file *file)
246 struct drm_i915_private *dev_priv = to_i915(dev);
247 struct i915_ggtt *ggtt = &dev_priv->ggtt;
248 struct drm_i915_gem_get_aperture *args = data;
249 struct i915_vma *vma;
252 pinned = ggtt->vm.reserved;
253 mutex_lock(&dev->struct_mutex);
254 list_for_each_entry(vma, &ggtt->vm.active_list, vm_link)
255 if (i915_vma_is_pinned(vma))
256 pinned += vma->node.size;
257 list_for_each_entry(vma, &ggtt->vm.inactive_list, vm_link)
258 if (i915_vma_is_pinned(vma))
259 pinned += vma->node.size;
260 mutex_unlock(&dev->struct_mutex);
262 args->aper_size = ggtt->vm.total;
263 args->aper_available_size = args->aper_size - pinned;
268 static int i915_gem_object_get_pages_phys(struct drm_i915_gem_object *obj)
270 struct address_space *mapping = obj->base.filp->f_mapping;
271 drm_dma_handle_t *phys;
273 struct scatterlist *sg;
278 if (WARN_ON(i915_gem_object_needs_bit17_swizzle(obj)))
281 /* Always aligning to the object size, allows a single allocation
282 * to handle all possible callers, and given typical object sizes,
283 * the alignment of the buddy allocation will naturally match.
285 phys = drm_pci_alloc(obj->base.dev,
286 roundup_pow_of_two(obj->base.size),
287 roundup_pow_of_two(obj->base.size));
292 for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
296 page = shmem_read_mapping_page(mapping, i);
302 src = kmap_atomic(page);
303 memcpy(vaddr, src, PAGE_SIZE);
304 drm_clflush_virt_range(vaddr, PAGE_SIZE);
311 i915_gem_chipset_flush(to_i915(obj->base.dev));
313 st = kmalloc(sizeof(*st), GFP_KERNEL);
319 if (sg_alloc_table(st, 1, GFP_KERNEL)) {
327 sg->length = obj->base.size;
329 sg_dma_address(sg) = phys->busaddr;
330 sg_dma_len(sg) = obj->base.size;
332 obj->phys_handle = phys;
334 __i915_gem_object_set_pages(obj, st, sg->length);
339 drm_pci_free(obj->base.dev, phys);
344 static void __start_cpu_write(struct drm_i915_gem_object *obj)
346 obj->read_domains = I915_GEM_DOMAIN_CPU;
347 obj->write_domain = I915_GEM_DOMAIN_CPU;
348 if (cpu_write_needs_clflush(obj))
349 obj->cache_dirty = true;
353 __i915_gem_object_release_shmem(struct drm_i915_gem_object *obj,
354 struct sg_table *pages,
357 GEM_BUG_ON(obj->mm.madv == __I915_MADV_PURGED);
359 if (obj->mm.madv == I915_MADV_DONTNEED)
360 obj->mm.dirty = false;
363 (obj->read_domains & I915_GEM_DOMAIN_CPU) == 0 &&
364 !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ))
365 drm_clflush_sg(pages);
367 __start_cpu_write(obj);
371 i915_gem_object_put_pages_phys(struct drm_i915_gem_object *obj,
372 struct sg_table *pages)
374 __i915_gem_object_release_shmem(obj, pages, false);
377 struct address_space *mapping = obj->base.filp->f_mapping;
378 char *vaddr = obj->phys_handle->vaddr;
381 for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
385 page = shmem_read_mapping_page(mapping, i);
389 dst = kmap_atomic(page);
390 drm_clflush_virt_range(vaddr, PAGE_SIZE);
391 memcpy(dst, vaddr, PAGE_SIZE);
394 set_page_dirty(page);
395 if (obj->mm.madv == I915_MADV_WILLNEED)
396 mark_page_accessed(page);
400 obj->mm.dirty = false;
403 sg_free_table(pages);
406 drm_pci_free(obj->base.dev, obj->phys_handle);
410 i915_gem_object_release_phys(struct drm_i915_gem_object *obj)
412 i915_gem_object_unpin_pages(obj);
415 static const struct drm_i915_gem_object_ops i915_gem_phys_ops = {
416 .get_pages = i915_gem_object_get_pages_phys,
417 .put_pages = i915_gem_object_put_pages_phys,
418 .release = i915_gem_object_release_phys,
421 static const struct drm_i915_gem_object_ops i915_gem_object_ops;
423 int i915_gem_object_unbind(struct drm_i915_gem_object *obj)
425 struct i915_vma *vma;
426 LIST_HEAD(still_in_list);
429 lockdep_assert_held(&obj->base.dev->struct_mutex);
431 /* Closed vma are removed from the obj->vma_list - but they may
432 * still have an active binding on the object. To remove those we
433 * must wait for all rendering to complete to the object (as unbinding
434 * must anyway), and retire the requests.
436 ret = i915_gem_object_set_to_cpu_domain(obj, false);
440 while ((vma = list_first_entry_or_null(&obj->vma_list,
443 list_move_tail(&vma->obj_link, &still_in_list);
444 ret = i915_vma_unbind(vma);
448 list_splice(&still_in_list, &obj->vma_list);
454 i915_gem_object_wait_fence(struct dma_fence *fence,
457 struct intel_rps_client *rps_client)
459 struct i915_request *rq;
461 BUILD_BUG_ON(I915_WAIT_INTERRUPTIBLE != 0x1);
463 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
466 if (!dma_fence_is_i915(fence))
467 return dma_fence_wait_timeout(fence,
468 flags & I915_WAIT_INTERRUPTIBLE,
471 rq = to_request(fence);
472 if (i915_request_completed(rq))
476 * This client is about to stall waiting for the GPU. In many cases
477 * this is undesirable and limits the throughput of the system, as
478 * many clients cannot continue processing user input/output whilst
479 * blocked. RPS autotuning may take tens of milliseconds to respond
480 * to the GPU load and thus incurs additional latency for the client.
481 * We can circumvent that by promoting the GPU frequency to maximum
482 * before we wait. This makes the GPU throttle up much more quickly
483 * (good for benchmarks and user experience, e.g. window animations),
484 * but at a cost of spending more power processing the workload
485 * (bad for battery). Not all clients even want their results
486 * immediately and for them we should just let the GPU select its own
487 * frequency to maximise efficiency. To prevent a single client from
488 * forcing the clocks too high for the whole system, we only allow
489 * each client to waitboost once in a busy period.
491 if (rps_client && !i915_request_started(rq)) {
492 if (INTEL_GEN(rq->i915) >= 6)
493 gen6_rps_boost(rq, rps_client);
496 timeout = i915_request_wait(rq, flags, timeout);
499 if (flags & I915_WAIT_LOCKED && i915_request_completed(rq))
500 i915_request_retire_upto(rq);
506 i915_gem_object_wait_reservation(struct reservation_object *resv,
509 struct intel_rps_client *rps_client)
511 unsigned int seq = __read_seqcount_begin(&resv->seq);
512 struct dma_fence *excl;
513 bool prune_fences = false;
515 if (flags & I915_WAIT_ALL) {
516 struct dma_fence **shared;
517 unsigned int count, i;
520 ret = reservation_object_get_fences_rcu(resv,
521 &excl, &count, &shared);
525 for (i = 0; i < count; i++) {
526 timeout = i915_gem_object_wait_fence(shared[i],
532 dma_fence_put(shared[i]);
535 for (; i < count; i++)
536 dma_fence_put(shared[i]);
540 * If both shared fences and an exclusive fence exist,
541 * then by construction the shared fences must be later
542 * than the exclusive fence. If we successfully wait for
543 * all the shared fences, we know that the exclusive fence
544 * must all be signaled. If all the shared fences are
545 * signaled, we can prune the array and recover the
546 * floating references on the fences/requests.
548 prune_fences = count && timeout >= 0;
550 excl = reservation_object_get_excl_rcu(resv);
553 if (excl && timeout >= 0)
554 timeout = i915_gem_object_wait_fence(excl, flags, timeout,
560 * Opportunistically prune the fences iff we know they have *all* been
561 * signaled and that the reservation object has not been changed (i.e.
562 * no new fences have been added).
564 if (prune_fences && !__read_seqcount_retry(&resv->seq, seq)) {
565 if (reservation_object_trylock(resv)) {
566 if (!__read_seqcount_retry(&resv->seq, seq))
567 reservation_object_add_excl_fence(resv, NULL);
568 reservation_object_unlock(resv);
575 static void __fence_set_priority(struct dma_fence *fence,
576 const struct i915_sched_attr *attr)
578 struct i915_request *rq;
579 struct intel_engine_cs *engine;
581 if (dma_fence_is_signaled(fence) || !dma_fence_is_i915(fence))
584 rq = to_request(fence);
588 rcu_read_lock(); /* RCU serialisation for set-wedged protection */
589 if (engine->schedule)
590 engine->schedule(rq, attr);
592 local_bh_enable(); /* kick the tasklets if queues were reprioritised */
595 static void fence_set_priority(struct dma_fence *fence,
596 const struct i915_sched_attr *attr)
598 /* Recurse once into a fence-array */
599 if (dma_fence_is_array(fence)) {
600 struct dma_fence_array *array = to_dma_fence_array(fence);
603 for (i = 0; i < array->num_fences; i++)
604 __fence_set_priority(array->fences[i], attr);
606 __fence_set_priority(fence, attr);
611 i915_gem_object_wait_priority(struct drm_i915_gem_object *obj,
613 const struct i915_sched_attr *attr)
615 struct dma_fence *excl;
617 if (flags & I915_WAIT_ALL) {
618 struct dma_fence **shared;
619 unsigned int count, i;
622 ret = reservation_object_get_fences_rcu(obj->resv,
623 &excl, &count, &shared);
627 for (i = 0; i < count; i++) {
628 fence_set_priority(shared[i], attr);
629 dma_fence_put(shared[i]);
634 excl = reservation_object_get_excl_rcu(obj->resv);
638 fence_set_priority(excl, attr);
645 * Waits for rendering to the object to be completed
646 * @obj: i915 gem object
647 * @flags: how to wait (under a lock, for all rendering or just for writes etc)
648 * @timeout: how long to wait
649 * @rps_client: client (user process) to charge for any waitboosting
652 i915_gem_object_wait(struct drm_i915_gem_object *obj,
655 struct intel_rps_client *rps_client)
658 #if IS_ENABLED(CONFIG_LOCKDEP)
659 GEM_BUG_ON(debug_locks &&
660 !!lockdep_is_held(&obj->base.dev->struct_mutex) !=
661 !!(flags & I915_WAIT_LOCKED));
663 GEM_BUG_ON(timeout < 0);
665 timeout = i915_gem_object_wait_reservation(obj->resv,
668 return timeout < 0 ? timeout : 0;
671 static struct intel_rps_client *to_rps_client(struct drm_file *file)
673 struct drm_i915_file_private *fpriv = file->driver_priv;
675 return &fpriv->rps_client;
679 i915_gem_phys_pwrite(struct drm_i915_gem_object *obj,
680 struct drm_i915_gem_pwrite *args,
681 struct drm_file *file)
683 void *vaddr = obj->phys_handle->vaddr + args->offset;
684 char __user *user_data = u64_to_user_ptr(args->data_ptr);
686 /* We manually control the domain here and pretend that it
687 * remains coherent i.e. in the GTT domain, like shmem_pwrite.
689 intel_fb_obj_invalidate(obj, ORIGIN_CPU);
690 if (copy_from_user(vaddr, user_data, args->size))
693 drm_clflush_virt_range(vaddr, args->size);
694 i915_gem_chipset_flush(to_i915(obj->base.dev));
696 intel_fb_obj_flush(obj, ORIGIN_CPU);
700 void *i915_gem_object_alloc(struct drm_i915_private *dev_priv)
702 return kmem_cache_zalloc(dev_priv->objects, GFP_KERNEL);
705 void i915_gem_object_free(struct drm_i915_gem_object *obj)
707 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
708 kmem_cache_free(dev_priv->objects, obj);
712 i915_gem_create(struct drm_file *file,
713 struct drm_i915_private *dev_priv,
717 struct drm_i915_gem_object *obj;
721 size = roundup(size, PAGE_SIZE);
725 /* Allocate the new object */
726 obj = i915_gem_object_create(dev_priv, size);
730 ret = drm_gem_handle_create(file, &obj->base, &handle);
731 /* drop reference from allocate - handle holds it now */
732 i915_gem_object_put(obj);
741 i915_gem_dumb_create(struct drm_file *file,
742 struct drm_device *dev,
743 struct drm_mode_create_dumb *args)
745 /* have to work out size/pitch and return them */
746 args->pitch = ALIGN(args->width * DIV_ROUND_UP(args->bpp, 8), 64);
747 args->size = args->pitch * args->height;
748 return i915_gem_create(file, to_i915(dev),
749 args->size, &args->handle);
752 static bool gpu_write_needs_clflush(struct drm_i915_gem_object *obj)
754 return !(obj->cache_level == I915_CACHE_NONE ||
755 obj->cache_level == I915_CACHE_WT);
759 * Creates a new mm object and returns a handle to it.
760 * @dev: drm device pointer
761 * @data: ioctl data blob
762 * @file: drm file pointer
765 i915_gem_create_ioctl(struct drm_device *dev, void *data,
766 struct drm_file *file)
768 struct drm_i915_private *dev_priv = to_i915(dev);
769 struct drm_i915_gem_create *args = data;
771 i915_gem_flush_free_objects(dev_priv);
773 return i915_gem_create(file, dev_priv,
774 args->size, &args->handle);
777 static inline enum fb_op_origin
778 fb_write_origin(struct drm_i915_gem_object *obj, unsigned int domain)
780 return (domain == I915_GEM_DOMAIN_GTT ?
781 obj->frontbuffer_ggtt_origin : ORIGIN_CPU);
784 void i915_gem_flush_ggtt_writes(struct drm_i915_private *dev_priv)
787 * No actual flushing is required for the GTT write domain for reads
788 * from the GTT domain. Writes to it "immediately" go to main memory
789 * as far as we know, so there's no chipset flush. It also doesn't
790 * land in the GPU render cache.
792 * However, we do have to enforce the order so that all writes through
793 * the GTT land before any writes to the device, such as updates to
796 * We also have to wait a bit for the writes to land from the GTT.
797 * An uncached read (i.e. mmio) seems to be ideal for the round-trip
798 * timing. This issue has only been observed when switching quickly
799 * between GTT writes and CPU reads from inside the kernel on recent hw,
800 * and it appears to only affect discrete GTT blocks (i.e. on LLC
801 * system agents we cannot reproduce this behaviour, until Cannonlake
807 if (INTEL_INFO(dev_priv)->has_coherent_ggtt)
810 i915_gem_chipset_flush(dev_priv);
812 intel_runtime_pm_get(dev_priv);
813 spin_lock_irq(&dev_priv->uncore.lock);
815 POSTING_READ_FW(RING_HEAD(RENDER_RING_BASE));
817 spin_unlock_irq(&dev_priv->uncore.lock);
818 intel_runtime_pm_put(dev_priv);
822 flush_write_domain(struct drm_i915_gem_object *obj, unsigned int flush_domains)
824 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
825 struct i915_vma *vma;
827 if (!(obj->write_domain & flush_domains))
830 switch (obj->write_domain) {
831 case I915_GEM_DOMAIN_GTT:
832 i915_gem_flush_ggtt_writes(dev_priv);
834 intel_fb_obj_flush(obj,
835 fb_write_origin(obj, I915_GEM_DOMAIN_GTT));
837 for_each_ggtt_vma(vma, obj) {
841 i915_vma_unset_ggtt_write(vma);
845 case I915_GEM_DOMAIN_WC:
849 case I915_GEM_DOMAIN_CPU:
850 i915_gem_clflush_object(obj, I915_CLFLUSH_SYNC);
853 case I915_GEM_DOMAIN_RENDER:
854 if (gpu_write_needs_clflush(obj))
855 obj->cache_dirty = true;
859 obj->write_domain = 0;
863 __copy_to_user_swizzled(char __user *cpu_vaddr,
864 const char *gpu_vaddr, int gpu_offset,
867 int ret, cpu_offset = 0;
870 int cacheline_end = ALIGN(gpu_offset + 1, 64);
871 int this_length = min(cacheline_end - gpu_offset, length);
872 int swizzled_gpu_offset = gpu_offset ^ 64;
874 ret = __copy_to_user(cpu_vaddr + cpu_offset,
875 gpu_vaddr + swizzled_gpu_offset,
880 cpu_offset += this_length;
881 gpu_offset += this_length;
882 length -= this_length;
889 __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset,
890 const char __user *cpu_vaddr,
893 int ret, cpu_offset = 0;
896 int cacheline_end = ALIGN(gpu_offset + 1, 64);
897 int this_length = min(cacheline_end - gpu_offset, length);
898 int swizzled_gpu_offset = gpu_offset ^ 64;
900 ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset,
901 cpu_vaddr + cpu_offset,
906 cpu_offset += this_length;
907 gpu_offset += this_length;
908 length -= this_length;
915 * Pins the specified object's pages and synchronizes the object with
916 * GPU accesses. Sets needs_clflush to non-zero if the caller should
917 * flush the object from the CPU cache.
919 int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj,
920 unsigned int *needs_clflush)
924 lockdep_assert_held(&obj->base.dev->struct_mutex);
927 if (!i915_gem_object_has_struct_page(obj))
930 ret = i915_gem_object_wait(obj,
931 I915_WAIT_INTERRUPTIBLE |
933 MAX_SCHEDULE_TIMEOUT,
938 ret = i915_gem_object_pin_pages(obj);
942 if (obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ ||
943 !static_cpu_has(X86_FEATURE_CLFLUSH)) {
944 ret = i915_gem_object_set_to_cpu_domain(obj, false);
951 flush_write_domain(obj, ~I915_GEM_DOMAIN_CPU);
953 /* If we're not in the cpu read domain, set ourself into the gtt
954 * read domain and manually flush cachelines (if required). This
955 * optimizes for the case when the gpu will dirty the data
956 * anyway again before the next pread happens.
958 if (!obj->cache_dirty &&
959 !(obj->read_domains & I915_GEM_DOMAIN_CPU))
960 *needs_clflush = CLFLUSH_BEFORE;
963 /* return with the pages pinned */
967 i915_gem_object_unpin_pages(obj);
971 int i915_gem_obj_prepare_shmem_write(struct drm_i915_gem_object *obj,
972 unsigned int *needs_clflush)
976 lockdep_assert_held(&obj->base.dev->struct_mutex);
979 if (!i915_gem_object_has_struct_page(obj))
982 ret = i915_gem_object_wait(obj,
983 I915_WAIT_INTERRUPTIBLE |
986 MAX_SCHEDULE_TIMEOUT,
991 ret = i915_gem_object_pin_pages(obj);
995 if (obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE ||
996 !static_cpu_has(X86_FEATURE_CLFLUSH)) {
997 ret = i915_gem_object_set_to_cpu_domain(obj, true);
1004 flush_write_domain(obj, ~I915_GEM_DOMAIN_CPU);
1006 /* If we're not in the cpu write domain, set ourself into the
1007 * gtt write domain and manually flush cachelines (as required).
1008 * This optimizes for the case when the gpu will use the data
1009 * right away and we therefore have to clflush anyway.
1011 if (!obj->cache_dirty) {
1012 *needs_clflush |= CLFLUSH_AFTER;
1015 * Same trick applies to invalidate partially written
1016 * cachelines read before writing.
1018 if (!(obj->read_domains & I915_GEM_DOMAIN_CPU))
1019 *needs_clflush |= CLFLUSH_BEFORE;
1023 intel_fb_obj_invalidate(obj, ORIGIN_CPU);
1024 obj->mm.dirty = true;
1025 /* return with the pages pinned */
1029 i915_gem_object_unpin_pages(obj);
1034 shmem_clflush_swizzled_range(char *addr, unsigned long length,
1037 if (unlikely(swizzled)) {
1038 unsigned long start = (unsigned long) addr;
1039 unsigned long end = (unsigned long) addr + length;
1041 /* For swizzling simply ensure that we always flush both
1042 * channels. Lame, but simple and it works. Swizzled
1043 * pwrite/pread is far from a hotpath - current userspace
1044 * doesn't use it at all. */
1045 start = round_down(start, 128);
1046 end = round_up(end, 128);
1048 drm_clflush_virt_range((void *)start, end - start);
1050 drm_clflush_virt_range(addr, length);
1055 /* Only difference to the fast-path function is that this can handle bit17
1056 * and uses non-atomic copy and kmap functions. */
1058 shmem_pread_slow(struct page *page, int offset, int length,
1059 char __user *user_data,
1060 bool page_do_bit17_swizzling, bool needs_clflush)
1067 shmem_clflush_swizzled_range(vaddr + offset, length,
1068 page_do_bit17_swizzling);
1070 if (page_do_bit17_swizzling)
1071 ret = __copy_to_user_swizzled(user_data, vaddr, offset, length);
1073 ret = __copy_to_user(user_data, vaddr + offset, length);
1076 return ret ? - EFAULT : 0;
1080 shmem_pread(struct page *page, int offset, int length, char __user *user_data,
1081 bool page_do_bit17_swizzling, bool needs_clflush)
1086 if (!page_do_bit17_swizzling) {
1087 char *vaddr = kmap_atomic(page);
1090 drm_clflush_virt_range(vaddr + offset, length);
1091 ret = __copy_to_user_inatomic(user_data, vaddr + offset, length);
1092 kunmap_atomic(vaddr);
1097 return shmem_pread_slow(page, offset, length, user_data,
1098 page_do_bit17_swizzling, needs_clflush);
1102 i915_gem_shmem_pread(struct drm_i915_gem_object *obj,
1103 struct drm_i915_gem_pread *args)
1105 char __user *user_data;
1107 unsigned int obj_do_bit17_swizzling;
1108 unsigned int needs_clflush;
1109 unsigned int idx, offset;
1112 obj_do_bit17_swizzling = 0;
1113 if (i915_gem_object_needs_bit17_swizzle(obj))
1114 obj_do_bit17_swizzling = BIT(17);
1116 ret = mutex_lock_interruptible(&obj->base.dev->struct_mutex);
1120 ret = i915_gem_obj_prepare_shmem_read(obj, &needs_clflush);
1121 mutex_unlock(&obj->base.dev->struct_mutex);
1125 remain = args->size;
1126 user_data = u64_to_user_ptr(args->data_ptr);
1127 offset = offset_in_page(args->offset);
1128 for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
1129 struct page *page = i915_gem_object_get_page(obj, idx);
1130 unsigned int length = min_t(u64, remain, PAGE_SIZE - offset);
1132 ret = shmem_pread(page, offset, length, user_data,
1133 page_to_phys(page) & obj_do_bit17_swizzling,
1139 user_data += length;
1143 i915_gem_obj_finish_shmem_access(obj);
1148 gtt_user_read(struct io_mapping *mapping,
1149 loff_t base, int offset,
1150 char __user *user_data, int length)
1152 void __iomem *vaddr;
1153 unsigned long unwritten;
1155 /* We can use the cpu mem copy function because this is X86. */
1156 vaddr = io_mapping_map_atomic_wc(mapping, base);
1157 unwritten = __copy_to_user_inatomic(user_data,
1158 (void __force *)vaddr + offset,
1160 io_mapping_unmap_atomic(vaddr);
1162 vaddr = io_mapping_map_wc(mapping, base, PAGE_SIZE);
1163 unwritten = copy_to_user(user_data,
1164 (void __force *)vaddr + offset,
1166 io_mapping_unmap(vaddr);
1172 i915_gem_gtt_pread(struct drm_i915_gem_object *obj,
1173 const struct drm_i915_gem_pread *args)
1175 struct drm_i915_private *i915 = to_i915(obj->base.dev);
1176 struct i915_ggtt *ggtt = &i915->ggtt;
1177 struct drm_mm_node node;
1178 struct i915_vma *vma;
1179 void __user *user_data;
1183 ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
1187 intel_runtime_pm_get(i915);
1188 vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
1193 node.start = i915_ggtt_offset(vma);
1194 node.allocated = false;
1195 ret = i915_vma_put_fence(vma);
1197 i915_vma_unpin(vma);
1202 ret = insert_mappable_node(ggtt, &node, PAGE_SIZE);
1205 GEM_BUG_ON(!node.allocated);
1208 ret = i915_gem_object_set_to_gtt_domain(obj, false);
1212 mutex_unlock(&i915->drm.struct_mutex);
1214 user_data = u64_to_user_ptr(args->data_ptr);
1215 remain = args->size;
1216 offset = args->offset;
1218 while (remain > 0) {
1219 /* Operation in this page
1221 * page_base = page offset within aperture
1222 * page_offset = offset within page
1223 * page_length = bytes to copy for this page
1225 u32 page_base = node.start;
1226 unsigned page_offset = offset_in_page(offset);
1227 unsigned page_length = PAGE_SIZE - page_offset;
1228 page_length = remain < page_length ? remain : page_length;
1229 if (node.allocated) {
1231 ggtt->vm.insert_page(&ggtt->vm,
1232 i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT),
1233 node.start, I915_CACHE_NONE, 0);
1236 page_base += offset & PAGE_MASK;
1239 if (gtt_user_read(&ggtt->iomap, page_base, page_offset,
1240 user_data, page_length)) {
1245 remain -= page_length;
1246 user_data += page_length;
1247 offset += page_length;
1250 mutex_lock(&i915->drm.struct_mutex);
1252 if (node.allocated) {
1254 ggtt->vm.clear_range(&ggtt->vm, node.start, node.size);
1255 remove_mappable_node(&node);
1257 i915_vma_unpin(vma);
1260 intel_runtime_pm_put(i915);
1261 mutex_unlock(&i915->drm.struct_mutex);
1267 * Reads data from the object referenced by handle.
1268 * @dev: drm device pointer
1269 * @data: ioctl data blob
1270 * @file: drm file pointer
1272 * On error, the contents of *data are undefined.
1275 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
1276 struct drm_file *file)
1278 struct drm_i915_gem_pread *args = data;
1279 struct drm_i915_gem_object *obj;
1282 if (args->size == 0)
1285 if (!access_ok(VERIFY_WRITE,
1286 u64_to_user_ptr(args->data_ptr),
1290 obj = i915_gem_object_lookup(file, args->handle);
1294 /* Bounds check source. */
1295 if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) {
1300 trace_i915_gem_object_pread(obj, args->offset, args->size);
1302 ret = i915_gem_object_wait(obj,
1303 I915_WAIT_INTERRUPTIBLE,
1304 MAX_SCHEDULE_TIMEOUT,
1305 to_rps_client(file));
1309 ret = i915_gem_object_pin_pages(obj);
1313 ret = i915_gem_shmem_pread(obj, args);
1314 if (ret == -EFAULT || ret == -ENODEV)
1315 ret = i915_gem_gtt_pread(obj, args);
1317 i915_gem_object_unpin_pages(obj);
1319 i915_gem_object_put(obj);
1323 /* This is the fast write path which cannot handle
1324 * page faults in the source data
1328 ggtt_write(struct io_mapping *mapping,
1329 loff_t base, int offset,
1330 char __user *user_data, int length)
1332 void __iomem *vaddr;
1333 unsigned long unwritten;
1335 /* We can use the cpu mem copy function because this is X86. */
1336 vaddr = io_mapping_map_atomic_wc(mapping, base);
1337 unwritten = __copy_from_user_inatomic_nocache((void __force *)vaddr + offset,
1339 io_mapping_unmap_atomic(vaddr);
1341 vaddr = io_mapping_map_wc(mapping, base, PAGE_SIZE);
1342 unwritten = copy_from_user((void __force *)vaddr + offset,
1344 io_mapping_unmap(vaddr);
1351 * This is the fast pwrite path, where we copy the data directly from the
1352 * user into the GTT, uncached.
1353 * @obj: i915 GEM object
1354 * @args: pwrite arguments structure
1357 i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
1358 const struct drm_i915_gem_pwrite *args)
1360 struct drm_i915_private *i915 = to_i915(obj->base.dev);
1361 struct i915_ggtt *ggtt = &i915->ggtt;
1362 struct drm_mm_node node;
1363 struct i915_vma *vma;
1365 void __user *user_data;
1368 ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
1372 if (i915_gem_object_has_struct_page(obj)) {
1374 * Avoid waking the device up if we can fallback, as
1375 * waking/resuming is very slow (worst-case 10-100 ms
1376 * depending on PCI sleeps and our own resume time).
1377 * This easily dwarfs any performance advantage from
1378 * using the cache bypass of indirect GGTT access.
1380 if (!intel_runtime_pm_get_if_in_use(i915)) {
1385 /* No backing pages, no fallback, we must force GGTT access */
1386 intel_runtime_pm_get(i915);
1389 vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
1394 node.start = i915_ggtt_offset(vma);
1395 node.allocated = false;
1396 ret = i915_vma_put_fence(vma);
1398 i915_vma_unpin(vma);
1403 ret = insert_mappable_node(ggtt, &node, PAGE_SIZE);
1406 GEM_BUG_ON(!node.allocated);
1409 ret = i915_gem_object_set_to_gtt_domain(obj, true);
1413 mutex_unlock(&i915->drm.struct_mutex);
1415 intel_fb_obj_invalidate(obj, ORIGIN_CPU);
1417 user_data = u64_to_user_ptr(args->data_ptr);
1418 offset = args->offset;
1419 remain = args->size;
1421 /* Operation in this page
1423 * page_base = page offset within aperture
1424 * page_offset = offset within page
1425 * page_length = bytes to copy for this page
1427 u32 page_base = node.start;
1428 unsigned int page_offset = offset_in_page(offset);
1429 unsigned int page_length = PAGE_SIZE - page_offset;
1430 page_length = remain < page_length ? remain : page_length;
1431 if (node.allocated) {
1432 wmb(); /* flush the write before we modify the GGTT */
1433 ggtt->vm.insert_page(&ggtt->vm,
1434 i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT),
1435 node.start, I915_CACHE_NONE, 0);
1436 wmb(); /* flush modifications to the GGTT (insert_page) */
1438 page_base += offset & PAGE_MASK;
1440 /* If we get a fault while copying data, then (presumably) our
1441 * source page isn't available. Return the error and we'll
1442 * retry in the slow path.
1443 * If the object is non-shmem backed, we retry again with the
1444 * path that handles page fault.
1446 if (ggtt_write(&ggtt->iomap, page_base, page_offset,
1447 user_data, page_length)) {
1452 remain -= page_length;
1453 user_data += page_length;
1454 offset += page_length;
1456 intel_fb_obj_flush(obj, ORIGIN_CPU);
1458 mutex_lock(&i915->drm.struct_mutex);
1460 if (node.allocated) {
1462 ggtt->vm.clear_range(&ggtt->vm, node.start, node.size);
1463 remove_mappable_node(&node);
1465 i915_vma_unpin(vma);
1468 intel_runtime_pm_put(i915);
1470 mutex_unlock(&i915->drm.struct_mutex);
1475 shmem_pwrite_slow(struct page *page, int offset, int length,
1476 char __user *user_data,
1477 bool page_do_bit17_swizzling,
1478 bool needs_clflush_before,
1479 bool needs_clflush_after)
1485 if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
1486 shmem_clflush_swizzled_range(vaddr + offset, length,
1487 page_do_bit17_swizzling);
1488 if (page_do_bit17_swizzling)
1489 ret = __copy_from_user_swizzled(vaddr, offset, user_data,
1492 ret = __copy_from_user(vaddr + offset, user_data, length);
1493 if (needs_clflush_after)
1494 shmem_clflush_swizzled_range(vaddr + offset, length,
1495 page_do_bit17_swizzling);
1498 return ret ? -EFAULT : 0;
1501 /* Per-page copy function for the shmem pwrite fastpath.
1502 * Flushes invalid cachelines before writing to the target if
1503 * needs_clflush_before is set and flushes out any written cachelines after
1504 * writing if needs_clflush is set.
1507 shmem_pwrite(struct page *page, int offset, int len, char __user *user_data,
1508 bool page_do_bit17_swizzling,
1509 bool needs_clflush_before,
1510 bool needs_clflush_after)
1515 if (!page_do_bit17_swizzling) {
1516 char *vaddr = kmap_atomic(page);
1518 if (needs_clflush_before)
1519 drm_clflush_virt_range(vaddr + offset, len);
1520 ret = __copy_from_user_inatomic(vaddr + offset, user_data, len);
1521 if (needs_clflush_after)
1522 drm_clflush_virt_range(vaddr + offset, len);
1524 kunmap_atomic(vaddr);
1529 return shmem_pwrite_slow(page, offset, len, user_data,
1530 page_do_bit17_swizzling,
1531 needs_clflush_before,
1532 needs_clflush_after);
1536 i915_gem_shmem_pwrite(struct drm_i915_gem_object *obj,
1537 const struct drm_i915_gem_pwrite *args)
1539 struct drm_i915_private *i915 = to_i915(obj->base.dev);
1540 void __user *user_data;
1542 unsigned int obj_do_bit17_swizzling;
1543 unsigned int partial_cacheline_write;
1544 unsigned int needs_clflush;
1545 unsigned int offset, idx;
1548 ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
1552 ret = i915_gem_obj_prepare_shmem_write(obj, &needs_clflush);
1553 mutex_unlock(&i915->drm.struct_mutex);
1557 obj_do_bit17_swizzling = 0;
1558 if (i915_gem_object_needs_bit17_swizzle(obj))
1559 obj_do_bit17_swizzling = BIT(17);
1561 /* If we don't overwrite a cacheline completely we need to be
1562 * careful to have up-to-date data by first clflushing. Don't
1563 * overcomplicate things and flush the entire patch.
1565 partial_cacheline_write = 0;
1566 if (needs_clflush & CLFLUSH_BEFORE)
1567 partial_cacheline_write = boot_cpu_data.x86_clflush_size - 1;
1569 user_data = u64_to_user_ptr(args->data_ptr);
1570 remain = args->size;
1571 offset = offset_in_page(args->offset);
1572 for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
1573 struct page *page = i915_gem_object_get_page(obj, idx);
1574 unsigned int length = min_t(u64, remain, PAGE_SIZE - offset);
1576 ret = shmem_pwrite(page, offset, length, user_data,
1577 page_to_phys(page) & obj_do_bit17_swizzling,
1578 (offset | length) & partial_cacheline_write,
1579 needs_clflush & CLFLUSH_AFTER);
1584 user_data += length;
1588 intel_fb_obj_flush(obj, ORIGIN_CPU);
1589 i915_gem_obj_finish_shmem_access(obj);
1594 * Writes data to the object referenced by handle.
1596 * @data: ioctl data blob
1599 * On error, the contents of the buffer that were to be modified are undefined.
1602 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
1603 struct drm_file *file)
1605 struct drm_i915_gem_pwrite *args = data;
1606 struct drm_i915_gem_object *obj;
1609 if (args->size == 0)
1612 if (!access_ok(VERIFY_READ,
1613 u64_to_user_ptr(args->data_ptr),
1617 obj = i915_gem_object_lookup(file, args->handle);
1621 /* Bounds check destination. */
1622 if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) {
1627 /* Writes not allowed into this read-only object */
1628 if (i915_gem_object_is_readonly(obj)) {
1633 trace_i915_gem_object_pwrite(obj, args->offset, args->size);
1636 if (obj->ops->pwrite)
1637 ret = obj->ops->pwrite(obj, args);
1641 ret = i915_gem_object_wait(obj,
1642 I915_WAIT_INTERRUPTIBLE |
1644 MAX_SCHEDULE_TIMEOUT,
1645 to_rps_client(file));
1649 ret = i915_gem_object_pin_pages(obj);
1654 /* We can only do the GTT pwrite on untiled buffers, as otherwise
1655 * it would end up going through the fenced access, and we'll get
1656 * different detiling behavior between reading and writing.
1657 * pread/pwrite currently are reading and writing from the CPU
1658 * perspective, requiring manual detiling by the client.
1660 if (!i915_gem_object_has_struct_page(obj) ||
1661 cpu_write_needs_clflush(obj))
1662 /* Note that the gtt paths might fail with non-page-backed user
1663 * pointers (e.g. gtt mappings when moving data between
1664 * textures). Fallback to the shmem path in that case.
1666 ret = i915_gem_gtt_pwrite_fast(obj, args);
1668 if (ret == -EFAULT || ret == -ENOSPC) {
1669 if (obj->phys_handle)
1670 ret = i915_gem_phys_pwrite(obj, args, file);
1672 ret = i915_gem_shmem_pwrite(obj, args);
1675 i915_gem_object_unpin_pages(obj);
1677 i915_gem_object_put(obj);
1681 static void i915_gem_object_bump_inactive_ggtt(struct drm_i915_gem_object *obj)
1683 struct drm_i915_private *i915;
1684 struct list_head *list;
1685 struct i915_vma *vma;
1687 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
1689 for_each_ggtt_vma(vma, obj) {
1690 if (i915_vma_is_active(vma))
1693 if (!drm_mm_node_allocated(&vma->node))
1696 list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
1699 i915 = to_i915(obj->base.dev);
1700 spin_lock(&i915->mm.obj_lock);
1701 list = obj->bind_count ? &i915->mm.bound_list : &i915->mm.unbound_list;
1702 list_move_tail(&obj->mm.link, list);
1703 spin_unlock(&i915->mm.obj_lock);
1707 * Called when user space prepares to use an object with the CPU, either
1708 * through the mmap ioctl's mapping or a GTT mapping.
1710 * @data: ioctl data blob
1714 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1715 struct drm_file *file)
1717 struct drm_i915_gem_set_domain *args = data;
1718 struct drm_i915_gem_object *obj;
1719 uint32_t read_domains = args->read_domains;
1720 uint32_t write_domain = args->write_domain;
1723 /* Only handle setting domains to types used by the CPU. */
1724 if ((write_domain | read_domains) & I915_GEM_GPU_DOMAINS)
1727 /* Having something in the write domain implies it's in the read
1728 * domain, and only that read domain. Enforce that in the request.
1730 if (write_domain != 0 && read_domains != write_domain)
1733 obj = i915_gem_object_lookup(file, args->handle);
1737 /* Try to flush the object off the GPU without holding the lock.
1738 * We will repeat the flush holding the lock in the normal manner
1739 * to catch cases where we are gazumped.
1741 err = i915_gem_object_wait(obj,
1742 I915_WAIT_INTERRUPTIBLE |
1743 I915_WAIT_PRIORITY |
1744 (write_domain ? I915_WAIT_ALL : 0),
1745 MAX_SCHEDULE_TIMEOUT,
1746 to_rps_client(file));
1751 * Proxy objects do not control access to the backing storage, ergo
1752 * they cannot be used as a means to manipulate the cache domain
1753 * tracking for that backing storage. The proxy object is always
1754 * considered to be outside of any cache domain.
1756 if (i915_gem_object_is_proxy(obj)) {
1762 * Flush and acquire obj->pages so that we are coherent through
1763 * direct access in memory with previous cached writes through
1764 * shmemfs and that our cache domain tracking remains valid.
1765 * For example, if the obj->filp was moved to swap without us
1766 * being notified and releasing the pages, we would mistakenly
1767 * continue to assume that the obj remained out of the CPU cached
1770 err = i915_gem_object_pin_pages(obj);
1774 err = i915_mutex_lock_interruptible(dev);
1778 if (read_domains & I915_GEM_DOMAIN_WC)
1779 err = i915_gem_object_set_to_wc_domain(obj, write_domain);
1780 else if (read_domains & I915_GEM_DOMAIN_GTT)
1781 err = i915_gem_object_set_to_gtt_domain(obj, write_domain);
1783 err = i915_gem_object_set_to_cpu_domain(obj, write_domain);
1785 /* And bump the LRU for this access */
1786 i915_gem_object_bump_inactive_ggtt(obj);
1788 mutex_unlock(&dev->struct_mutex);
1790 if (write_domain != 0)
1791 intel_fb_obj_invalidate(obj,
1792 fb_write_origin(obj, write_domain));
1795 i915_gem_object_unpin_pages(obj);
1797 i915_gem_object_put(obj);
1802 * Called when user space has done writes to this buffer
1804 * @data: ioctl data blob
1808 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1809 struct drm_file *file)
1811 struct drm_i915_gem_sw_finish *args = data;
1812 struct drm_i915_gem_object *obj;
1814 obj = i915_gem_object_lookup(file, args->handle);
1819 * Proxy objects are barred from CPU access, so there is no
1820 * need to ban sw_finish as it is a nop.
1823 /* Pinned buffers may be scanout, so flush the cache */
1824 i915_gem_object_flush_if_display(obj);
1825 i915_gem_object_put(obj);
1831 * i915_gem_mmap_ioctl - Maps the contents of an object, returning the address
1834 * @data: ioctl data blob
1837 * While the mapping holds a reference on the contents of the object, it doesn't
1838 * imply a ref on the object itself.
1842 * DRM driver writers who look a this function as an example for how to do GEM
1843 * mmap support, please don't implement mmap support like here. The modern way
1844 * to implement DRM mmap support is with an mmap offset ioctl (like
1845 * i915_gem_mmap_gtt) and then using the mmap syscall on the DRM fd directly.
1846 * That way debug tooling like valgrind will understand what's going on, hiding
1847 * the mmap call in a driver private ioctl will break that. The i915 driver only
1848 * does cpu mmaps this way because we didn't know better.
1851 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1852 struct drm_file *file)
1854 struct drm_i915_gem_mmap *args = data;
1855 struct drm_i915_gem_object *obj;
1858 if (args->flags & ~(I915_MMAP_WC))
1861 if (args->flags & I915_MMAP_WC && !boot_cpu_has(X86_FEATURE_PAT))
1864 obj = i915_gem_object_lookup(file, args->handle);
1868 /* prime objects have no backing filp to GEM mmap
1871 if (!obj->base.filp) {
1872 i915_gem_object_put(obj);
1876 addr = vm_mmap(obj->base.filp, 0, args->size,
1877 PROT_READ | PROT_WRITE, MAP_SHARED,
1879 if (args->flags & I915_MMAP_WC) {
1880 struct mm_struct *mm = current->mm;
1881 struct vm_area_struct *vma;
1883 if (down_write_killable(&mm->mmap_sem)) {
1884 i915_gem_object_put(obj);
1887 vma = find_vma(mm, addr);
1890 pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
1893 up_write(&mm->mmap_sem);
1895 /* This may race, but that's ok, it only gets set */
1896 WRITE_ONCE(obj->frontbuffer_ggtt_origin, ORIGIN_CPU);
1898 i915_gem_object_put(obj);
1899 if (IS_ERR((void *)addr))
1902 args->addr_ptr = (uint64_t) addr;
1907 static unsigned int tile_row_pages(const struct drm_i915_gem_object *obj)
1909 return i915_gem_object_get_tile_row_size(obj) >> PAGE_SHIFT;
1913 * i915_gem_mmap_gtt_version - report the current feature set for GTT mmaps
1915 * A history of the GTT mmap interface:
1917 * 0 - Everything had to fit into the GTT. Both parties of a memcpy had to
1918 * aligned and suitable for fencing, and still fit into the available
1919 * mappable space left by the pinned display objects. A classic problem
1920 * we called the page-fault-of-doom where we would ping-pong between
1921 * two objects that could not fit inside the GTT and so the memcpy
1922 * would page one object in at the expense of the other between every
1925 * 1 - Objects can be any size, and have any compatible fencing (X Y, or none
1926 * as set via i915_gem_set_tiling() [DRM_I915_GEM_SET_TILING]). If the
1927 * object is too large for the available space (or simply too large
1928 * for the mappable aperture!), a view is created instead and faulted
1929 * into userspace. (This view is aligned and sized appropriately for
1932 * 2 - Recognise WC as a separate cache domain so that we can flush the
1933 * delayed writes via GTT before performing direct access via WC.
1937 * * snoopable objects cannot be accessed via the GTT. It can cause machine
1938 * hangs on some architectures, corruption on others. An attempt to service
1939 * a GTT page fault from a snoopable object will generate a SIGBUS.
1941 * * the object must be able to fit into RAM (physical memory, though no
1942 * limited to the mappable aperture).
1947 * * a new GTT page fault will synchronize rendering from the GPU and flush
1948 * all data to system memory. Subsequent access will not be synchronized.
1950 * * all mappings are revoked on runtime device suspend.
1952 * * there are only 8, 16 or 32 fence registers to share between all users
1953 * (older machines require fence register for display and blitter access
1954 * as well). Contention of the fence registers will cause the previous users
1955 * to be unmapped and any new access will generate new page faults.
1957 * * running out of memory while servicing a fault may generate a SIGBUS,
1958 * rather than the expected SIGSEGV.
1960 int i915_gem_mmap_gtt_version(void)
1965 static inline struct i915_ggtt_view
1966 compute_partial_view(const struct drm_i915_gem_object *obj,
1967 pgoff_t page_offset,
1970 struct i915_ggtt_view view;
1972 if (i915_gem_object_is_tiled(obj))
1973 chunk = roundup(chunk, tile_row_pages(obj));
1975 view.type = I915_GGTT_VIEW_PARTIAL;
1976 view.partial.offset = rounddown(page_offset, chunk);
1978 min_t(unsigned int, chunk,
1979 (obj->base.size >> PAGE_SHIFT) - view.partial.offset);
1981 /* If the partial covers the entire object, just create a normal VMA. */
1982 if (chunk >= obj->base.size >> PAGE_SHIFT)
1983 view.type = I915_GGTT_VIEW_NORMAL;
1989 * i915_gem_fault - fault a page into the GTT
1992 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1993 * from userspace. The fault handler takes care of binding the object to
1994 * the GTT (if needed), allocating and programming a fence register (again,
1995 * only if needed based on whether the old reg is still valid or the object
1996 * is tiled) and inserting a new PTE into the faulting process.
1998 * Note that the faulting process may involve evicting existing objects
1999 * from the GTT and/or fence registers to make room. So performance may
2000 * suffer if the GTT working set is large or there are few fence registers
2003 * The current feature set supported by i915_gem_fault() and thus GTT mmaps
2004 * is exposed via I915_PARAM_MMAP_GTT_VERSION (see i915_gem_mmap_gtt_version).
2006 vm_fault_t i915_gem_fault(struct vm_fault *vmf)
2008 #define MIN_CHUNK_PAGES (SZ_1M >> PAGE_SHIFT)
2009 struct vm_area_struct *area = vmf->vma;
2010 struct drm_i915_gem_object *obj = to_intel_bo(area->vm_private_data);
2011 struct drm_device *dev = obj->base.dev;
2012 struct drm_i915_private *dev_priv = to_i915(dev);
2013 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2014 bool write = area->vm_flags & VM_WRITE;
2015 struct i915_vma *vma;
2016 pgoff_t page_offset;
2019 /* Sanity check that we allow writing into this object */
2020 if (i915_gem_object_is_readonly(obj) && write)
2021 return VM_FAULT_SIGBUS;
2023 /* We don't use vmf->pgoff since that has the fake offset */
2024 page_offset = (vmf->address - area->vm_start) >> PAGE_SHIFT;
2026 trace_i915_gem_object_fault(obj, page_offset, true, write);
2028 /* Try to flush the object off the GPU first without holding the lock.
2029 * Upon acquiring the lock, we will perform our sanity checks and then
2030 * repeat the flush holding the lock in the normal manner to catch cases
2031 * where we are gazumped.
2033 ret = i915_gem_object_wait(obj,
2034 I915_WAIT_INTERRUPTIBLE,
2035 MAX_SCHEDULE_TIMEOUT,
2040 ret = i915_gem_object_pin_pages(obj);
2044 intel_runtime_pm_get(dev_priv);
2046 ret = i915_mutex_lock_interruptible(dev);
2050 /* Access to snoopable pages through the GTT is incoherent. */
2051 if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev_priv)) {
2057 /* Now pin it into the GTT as needed */
2058 vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
2063 /* Use a partial view if it is bigger than available space */
2064 struct i915_ggtt_view view =
2065 compute_partial_view(obj, page_offset, MIN_CHUNK_PAGES);
2068 flags = PIN_MAPPABLE;
2069 if (view.type == I915_GGTT_VIEW_NORMAL)
2070 flags |= PIN_NONBLOCK; /* avoid warnings for pinned */
2073 * Userspace is now writing through an untracked VMA, abandon
2074 * all hope that the hardware is able to track future writes.
2076 obj->frontbuffer_ggtt_origin = ORIGIN_CPU;
2078 vma = i915_gem_object_ggtt_pin(obj, &view, 0, 0, flags);
2079 if (IS_ERR(vma) && !view.type) {
2080 flags = PIN_MAPPABLE;
2081 view.type = I915_GGTT_VIEW_PARTIAL;
2082 vma = i915_gem_object_ggtt_pin(obj, &view, 0, 0, flags);
2090 ret = i915_gem_object_set_to_gtt_domain(obj, write);
2094 ret = i915_vma_pin_fence(vma);
2098 /* Finally, remap it using the new GTT offset */
2099 ret = remap_io_mapping(area,
2100 area->vm_start + (vma->ggtt_view.partial.offset << PAGE_SHIFT),
2101 (ggtt->gmadr.start + vma->node.start) >> PAGE_SHIFT,
2102 min_t(u64, vma->size, area->vm_end - area->vm_start),
2107 /* Mark as being mmapped into userspace for later revocation */
2108 assert_rpm_wakelock_held(dev_priv);
2109 if (!i915_vma_set_userfault(vma) && !obj->userfault_count++)
2110 list_add(&obj->userfault_link, &dev_priv->mm.userfault_list);
2111 GEM_BUG_ON(!obj->userfault_count);
2113 i915_vma_set_ggtt_write(vma);
2116 i915_vma_unpin_fence(vma);
2118 __i915_vma_unpin(vma);
2120 mutex_unlock(&dev->struct_mutex);
2122 intel_runtime_pm_put(dev_priv);
2123 i915_gem_object_unpin_pages(obj);
2128 * We eat errors when the gpu is terminally wedged to avoid
2129 * userspace unduly crashing (gl has no provisions for mmaps to
2130 * fail). But any other -EIO isn't ours (e.g. swap in failure)
2131 * and so needs to be reported.
2133 if (!i915_terminally_wedged(&dev_priv->gpu_error))
2134 return VM_FAULT_SIGBUS;
2135 /* else: fall through */
2138 * EAGAIN means the gpu is hung and we'll wait for the error
2139 * handler to reset everything when re-faulting in
2140 * i915_mutex_lock_interruptible.
2147 * EBUSY is ok: this just means that another thread
2148 * already did the job.
2150 return VM_FAULT_NOPAGE;
2152 return VM_FAULT_OOM;
2155 return VM_FAULT_SIGBUS;
2157 WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret);
2158 return VM_FAULT_SIGBUS;
2162 static void __i915_gem_object_release_mmap(struct drm_i915_gem_object *obj)
2164 struct i915_vma *vma;
2166 GEM_BUG_ON(!obj->userfault_count);
2168 obj->userfault_count = 0;
2169 list_del(&obj->userfault_link);
2170 drm_vma_node_unmap(&obj->base.vma_node,
2171 obj->base.dev->anon_inode->i_mapping);
2173 for_each_ggtt_vma(vma, obj)
2174 i915_vma_unset_userfault(vma);
2178 * i915_gem_release_mmap - remove physical page mappings
2179 * @obj: obj in question
2181 * Preserve the reservation of the mmapping with the DRM core code, but
2182 * relinquish ownership of the pages back to the system.
2184 * It is vital that we remove the page mapping if we have mapped a tiled
2185 * object through the GTT and then lose the fence register due to
2186 * resource pressure. Similarly if the object has been moved out of the
2187 * aperture, than pages mapped into userspace must be revoked. Removing the
2188 * mapping will then trigger a page fault on the next user access, allowing
2189 * fixup by i915_gem_fault().
2192 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
2194 struct drm_i915_private *i915 = to_i915(obj->base.dev);
2196 /* Serialisation between user GTT access and our code depends upon
2197 * revoking the CPU's PTE whilst the mutex is held. The next user
2198 * pagefault then has to wait until we release the mutex.
2200 * Note that RPM complicates somewhat by adding an additional
2201 * requirement that operations to the GGTT be made holding the RPM
2204 lockdep_assert_held(&i915->drm.struct_mutex);
2205 intel_runtime_pm_get(i915);
2207 if (!obj->userfault_count)
2210 __i915_gem_object_release_mmap(obj);
2212 /* Ensure that the CPU's PTE are revoked and there are not outstanding
2213 * memory transactions from userspace before we return. The TLB
2214 * flushing implied above by changing the PTE above *should* be
2215 * sufficient, an extra barrier here just provides us with a bit
2216 * of paranoid documentation about our requirement to serialise
2217 * memory writes before touching registers / GSM.
2222 intel_runtime_pm_put(i915);
2225 void i915_gem_runtime_suspend(struct drm_i915_private *dev_priv)
2227 struct drm_i915_gem_object *obj, *on;
2231 * Only called during RPM suspend. All users of the userfault_list
2232 * must be holding an RPM wakeref to ensure that this can not
2233 * run concurrently with themselves (and use the struct_mutex for
2234 * protection between themselves).
2237 list_for_each_entry_safe(obj, on,
2238 &dev_priv->mm.userfault_list, userfault_link)
2239 __i915_gem_object_release_mmap(obj);
2241 /* The fence will be lost when the device powers down. If any were
2242 * in use by hardware (i.e. they are pinned), we should not be powering
2243 * down! All other fences will be reacquired by the user upon waking.
2245 for (i = 0; i < dev_priv->num_fence_regs; i++) {
2246 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
2248 /* Ideally we want to assert that the fence register is not
2249 * live at this point (i.e. that no piece of code will be
2250 * trying to write through fence + GTT, as that both violates
2251 * our tracking of activity and associated locking/barriers,
2252 * but also is illegal given that the hw is powered down).
2254 * Previously we used reg->pin_count as a "liveness" indicator.
2255 * That is not sufficient, and we need a more fine-grained
2256 * tool if we want to have a sanity check here.
2262 GEM_BUG_ON(i915_vma_has_userfault(reg->vma));
2267 static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj)
2269 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
2272 err = drm_gem_create_mmap_offset(&obj->base);
2276 /* Attempt to reap some mmap space from dead objects */
2278 err = i915_gem_wait_for_idle(dev_priv,
2279 I915_WAIT_INTERRUPTIBLE,
2280 MAX_SCHEDULE_TIMEOUT);
2284 i915_gem_drain_freed_objects(dev_priv);
2285 err = drm_gem_create_mmap_offset(&obj->base);
2289 } while (flush_delayed_work(&dev_priv->gt.retire_work));
2294 static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)
2296 drm_gem_free_mmap_offset(&obj->base);
2300 i915_gem_mmap_gtt(struct drm_file *file,
2301 struct drm_device *dev,
2305 struct drm_i915_gem_object *obj;
2308 obj = i915_gem_object_lookup(file, handle);
2312 ret = i915_gem_object_create_mmap_offset(obj);
2314 *offset = drm_vma_node_offset_addr(&obj->base.vma_node);
2316 i915_gem_object_put(obj);
2321 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
2323 * @data: GTT mapping ioctl data
2324 * @file: GEM object info
2326 * Simply returns the fake offset to userspace so it can mmap it.
2327 * The mmap call will end up in drm_gem_mmap(), which will set things
2328 * up so we can get faults in the handler above.
2330 * The fault handler will take care of binding the object into the GTT
2331 * (since it may have been evicted to make room for something), allocating
2332 * a fence register, and mapping the appropriate aperture address into
2336 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
2337 struct drm_file *file)
2339 struct drm_i915_gem_mmap_gtt *args = data;
2341 return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
2344 /* Immediately discard the backing storage */
2346 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
2348 i915_gem_object_free_mmap_offset(obj);
2350 if (obj->base.filp == NULL)
2353 /* Our goal here is to return as much of the memory as
2354 * is possible back to the system as we are called from OOM.
2355 * To do this we must instruct the shmfs to drop all of its
2356 * backing pages, *now*.
2358 shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1);
2359 obj->mm.madv = __I915_MADV_PURGED;
2360 obj->mm.pages = ERR_PTR(-EFAULT);
2363 /* Try to discard unwanted pages */
2364 void __i915_gem_object_invalidate(struct drm_i915_gem_object *obj)
2366 struct address_space *mapping;
2368 lockdep_assert_held(&obj->mm.lock);
2369 GEM_BUG_ON(i915_gem_object_has_pages(obj));
2371 switch (obj->mm.madv) {
2372 case I915_MADV_DONTNEED:
2373 i915_gem_object_truncate(obj);
2374 case __I915_MADV_PURGED:
2378 if (obj->base.filp == NULL)
2381 mapping = obj->base.filp->f_mapping,
2382 invalidate_mapping_pages(mapping, 0, (loff_t)-1);
2386 * Move pages to appropriate lru and release the pagevec, decrementing the
2387 * ref count of those pages.
2389 static void check_release_pagevec(struct pagevec *pvec)
2391 check_move_unevictable_pages(pvec);
2392 __pagevec_release(pvec);
2397 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj,
2398 struct sg_table *pages)
2400 struct sgt_iter sgt_iter;
2401 struct pagevec pvec;
2404 __i915_gem_object_release_shmem(obj, pages, true);
2406 i915_gem_gtt_finish_pages(obj, pages);
2408 if (i915_gem_object_needs_bit17_swizzle(obj))
2409 i915_gem_object_save_bit_17_swizzle(obj, pages);
2411 mapping_clear_unevictable(file_inode(obj->base.filp)->i_mapping);
2413 pagevec_init(&pvec);
2414 for_each_sgt_page(page, sgt_iter, pages) {
2416 set_page_dirty(page);
2418 if (obj->mm.madv == I915_MADV_WILLNEED)
2419 mark_page_accessed(page);
2421 if (!pagevec_add(&pvec, page))
2422 check_release_pagevec(&pvec);
2424 if (pagevec_count(&pvec))
2425 check_release_pagevec(&pvec);
2426 obj->mm.dirty = false;
2428 sg_free_table(pages);
2432 static void __i915_gem_object_reset_page_iter(struct drm_i915_gem_object *obj)
2434 struct radix_tree_iter iter;
2438 radix_tree_for_each_slot(slot, &obj->mm.get_page.radix, &iter, 0)
2439 radix_tree_delete(&obj->mm.get_page.radix, iter.index);
2443 static struct sg_table *
2444 __i915_gem_object_unset_pages(struct drm_i915_gem_object *obj)
2446 struct drm_i915_private *i915 = to_i915(obj->base.dev);
2447 struct sg_table *pages;
2449 pages = fetch_and_zero(&obj->mm.pages);
2453 spin_lock(&i915->mm.obj_lock);
2454 list_del(&obj->mm.link);
2455 spin_unlock(&i915->mm.obj_lock);
2457 if (obj->mm.mapping) {
2460 ptr = page_mask_bits(obj->mm.mapping);
2461 if (is_vmalloc_addr(ptr))
2464 kunmap(kmap_to_page(ptr));
2466 obj->mm.mapping = NULL;
2469 __i915_gem_object_reset_page_iter(obj);
2470 obj->mm.page_sizes.phys = obj->mm.page_sizes.sg = 0;
2475 void __i915_gem_object_put_pages(struct drm_i915_gem_object *obj,
2476 enum i915_mm_subclass subclass)
2478 struct sg_table *pages;
2480 if (i915_gem_object_has_pinned_pages(obj))
2483 GEM_BUG_ON(obj->bind_count);
2484 if (!i915_gem_object_has_pages(obj))
2487 /* May be called by shrinker from within get_pages() (on another bo) */
2488 mutex_lock_nested(&obj->mm.lock, subclass);
2489 if (unlikely(atomic_read(&obj->mm.pages_pin_count)))
2493 * ->put_pages might need to allocate memory for the bit17 swizzle
2494 * array, hence protect them from being reaped by removing them from gtt
2497 pages = __i915_gem_object_unset_pages(obj);
2499 obj->ops->put_pages(obj, pages);
2502 mutex_unlock(&obj->mm.lock);
2505 bool i915_sg_trim(struct sg_table *orig_st)
2507 struct sg_table new_st;
2508 struct scatterlist *sg, *new_sg;
2511 if (orig_st->nents == orig_st->orig_nents)
2514 if (sg_alloc_table(&new_st, orig_st->nents, GFP_KERNEL | __GFP_NOWARN))
2517 new_sg = new_st.sgl;
2518 for_each_sg(orig_st->sgl, sg, orig_st->nents, i) {
2519 sg_set_page(new_sg, sg_page(sg), sg->length, 0);
2520 sg_dma_address(new_sg) = sg_dma_address(sg);
2521 sg_dma_len(new_sg) = sg_dma_len(sg);
2523 new_sg = sg_next(new_sg);
2525 GEM_BUG_ON(new_sg); /* Should walk exactly nents and hit the end */
2527 sg_free_table(orig_st);
2533 static int i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
2535 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
2536 const unsigned long page_count = obj->base.size / PAGE_SIZE;
2538 struct address_space *mapping;
2539 struct sg_table *st;
2540 struct scatterlist *sg;
2541 struct sgt_iter sgt_iter;
2543 unsigned long last_pfn = 0; /* suppress gcc warning */
2544 unsigned int max_segment = i915_sg_segment_size();
2545 unsigned int sg_page_sizes;
2546 struct pagevec pvec;
2551 * Assert that the object is not currently in any GPU domain. As it
2552 * wasn't in the GTT, there shouldn't be any way it could have been in
2555 GEM_BUG_ON(obj->read_domains & I915_GEM_GPU_DOMAINS);
2556 GEM_BUG_ON(obj->write_domain & I915_GEM_GPU_DOMAINS);
2559 * If there's no chance of allocating enough pages for the whole
2560 * object, bail early.
2562 if (page_count > totalram_pages)
2565 st = kmalloc(sizeof(*st), GFP_KERNEL);
2570 if (sg_alloc_table(st, page_count, GFP_KERNEL)) {
2576 * Get the list of pages out of our struct file. They'll be pinned
2577 * at this point until we release them.
2579 * Fail silently without starting the shrinker
2581 mapping = obj->base.filp->f_mapping;
2582 mapping_set_unevictable(mapping);
2583 noreclaim = mapping_gfp_constraint(mapping, ~__GFP_RECLAIM);
2584 noreclaim |= __GFP_NORETRY | __GFP_NOWARN;
2589 for (i = 0; i < page_count; i++) {
2590 const unsigned int shrink[] = {
2591 I915_SHRINK_BOUND | I915_SHRINK_UNBOUND | I915_SHRINK_PURGEABLE,
2594 gfp_t gfp = noreclaim;
2598 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2599 if (likely(!IS_ERR(page)))
2603 ret = PTR_ERR(page);
2607 i915_gem_shrink(dev_priv, 2 * page_count, NULL, *s++);
2610 * We've tried hard to allocate the memory by reaping
2611 * our own buffer, now let the real VM do its job and
2612 * go down in flames if truly OOM.
2614 * However, since graphics tend to be disposable,
2615 * defer the oom here by reporting the ENOMEM back
2619 /* reclaim and warn, but no oom */
2620 gfp = mapping_gfp_mask(mapping);
2623 * Our bo are always dirty and so we require
2624 * kswapd to reclaim our pages (direct reclaim
2625 * does not effectively begin pageout of our
2626 * buffers on its own). However, direct reclaim
2627 * only waits for kswapd when under allocation
2628 * congestion. So as a result __GFP_RECLAIM is
2629 * unreliable and fails to actually reclaim our
2630 * dirty pages -- unless you try over and over
2631 * again with !__GFP_NORETRY. However, we still
2632 * want to fail this allocation rather than
2633 * trigger the out-of-memory killer and for
2634 * this we want __GFP_RETRY_MAYFAIL.
2636 gfp |= __GFP_RETRY_MAYFAIL;
2641 sg->length >= max_segment ||
2642 page_to_pfn(page) != last_pfn + 1) {
2644 sg_page_sizes |= sg->length;
2648 sg_set_page(sg, page, PAGE_SIZE, 0);
2650 sg->length += PAGE_SIZE;
2652 last_pfn = page_to_pfn(page);
2654 /* Check that the i965g/gm workaround works. */
2655 WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
2657 if (sg) { /* loop terminated early; short sg table */
2658 sg_page_sizes |= sg->length;
2662 /* Trim unused sg entries to avoid wasting memory. */
2665 ret = i915_gem_gtt_prepare_pages(obj, st);
2668 * DMA remapping failed? One possible cause is that
2669 * it could not reserve enough large entries, asking
2670 * for PAGE_SIZE chunks instead may be helpful.
2672 if (max_segment > PAGE_SIZE) {
2673 for_each_sgt_page(page, sgt_iter, st)
2677 max_segment = PAGE_SIZE;
2680 dev_warn(&dev_priv->drm.pdev->dev,
2681 "Failed to DMA remap %lu pages\n",
2687 if (i915_gem_object_needs_bit17_swizzle(obj))
2688 i915_gem_object_do_bit_17_swizzle(obj, st);
2690 __i915_gem_object_set_pages(obj, st, sg_page_sizes);
2697 mapping_clear_unevictable(mapping);
2698 pagevec_init(&pvec);
2699 for_each_sgt_page(page, sgt_iter, st) {
2700 if (!pagevec_add(&pvec, page))
2701 check_release_pagevec(&pvec);
2703 if (pagevec_count(&pvec))
2704 check_release_pagevec(&pvec);
2709 * shmemfs first checks if there is enough memory to allocate the page
2710 * and reports ENOSPC should there be insufficient, along with the usual
2711 * ENOMEM for a genuine allocation failure.
2713 * We use ENOSPC in our driver to mean that we have run out of aperture
2714 * space and so want to translate the error from shmemfs back to our
2715 * usual understanding of ENOMEM.
2723 void __i915_gem_object_set_pages(struct drm_i915_gem_object *obj,
2724 struct sg_table *pages,
2725 unsigned int sg_page_sizes)
2727 struct drm_i915_private *i915 = to_i915(obj->base.dev);
2728 unsigned long supported = INTEL_INFO(i915)->page_sizes;
2731 lockdep_assert_held(&obj->mm.lock);
2733 obj->mm.get_page.sg_pos = pages->sgl;
2734 obj->mm.get_page.sg_idx = 0;
2736 obj->mm.pages = pages;
2738 if (i915_gem_object_is_tiled(obj) &&
2739 i915->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
2740 GEM_BUG_ON(obj->mm.quirked);
2741 __i915_gem_object_pin_pages(obj);
2742 obj->mm.quirked = true;
2745 GEM_BUG_ON(!sg_page_sizes);
2746 obj->mm.page_sizes.phys = sg_page_sizes;
2749 * Calculate the supported page-sizes which fit into the given
2750 * sg_page_sizes. This will give us the page-sizes which we may be able
2751 * to use opportunistically when later inserting into the GTT. For
2752 * example if phys=2G, then in theory we should be able to use 1G, 2M,
2753 * 64K or 4K pages, although in practice this will depend on a number of
2756 obj->mm.page_sizes.sg = 0;
2757 for_each_set_bit(i, &supported, ilog2(I915_GTT_MAX_PAGE_SIZE) + 1) {
2758 if (obj->mm.page_sizes.phys & ~0u << i)
2759 obj->mm.page_sizes.sg |= BIT(i);
2761 GEM_BUG_ON(!HAS_PAGE_SIZES(i915, obj->mm.page_sizes.sg));
2763 spin_lock(&i915->mm.obj_lock);
2764 list_add(&obj->mm.link, &i915->mm.unbound_list);
2765 spin_unlock(&i915->mm.obj_lock);
2768 static int ____i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
2772 if (unlikely(obj->mm.madv != I915_MADV_WILLNEED)) {
2773 DRM_DEBUG("Attempting to obtain a purgeable object\n");
2777 err = obj->ops->get_pages(obj);
2778 GEM_BUG_ON(!err && !i915_gem_object_has_pages(obj));
2783 /* Ensure that the associated pages are gathered from the backing storage
2784 * and pinned into our object. i915_gem_object_pin_pages() may be called
2785 * multiple times before they are released by a single call to
2786 * i915_gem_object_unpin_pages() - once the pages are no longer referenced
2787 * either as a result of memory pressure (reaping pages under the shrinker)
2788 * or as the object is itself released.
2790 int __i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
2794 err = mutex_lock_interruptible(&obj->mm.lock);
2798 if (unlikely(!i915_gem_object_has_pages(obj))) {
2799 GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj));
2801 err = ____i915_gem_object_get_pages(obj);
2805 smp_mb__before_atomic();
2807 atomic_inc(&obj->mm.pages_pin_count);
2810 mutex_unlock(&obj->mm.lock);
2814 /* The 'mapping' part of i915_gem_object_pin_map() below */
2815 static void *i915_gem_object_map(const struct drm_i915_gem_object *obj,
2816 enum i915_map_type type)
2818 unsigned long n_pages = obj->base.size >> PAGE_SHIFT;
2819 struct sg_table *sgt = obj->mm.pages;
2820 struct sgt_iter sgt_iter;
2822 struct page *stack_pages[32];
2823 struct page **pages = stack_pages;
2824 unsigned long i = 0;
2828 /* A single page can always be kmapped */
2829 if (n_pages == 1 && type == I915_MAP_WB)
2830 return kmap(sg_page(sgt->sgl));
2832 if (n_pages > ARRAY_SIZE(stack_pages)) {
2833 /* Too big for stack -- allocate temporary array instead */
2834 pages = kvmalloc_array(n_pages, sizeof(*pages), GFP_KERNEL);
2839 for_each_sgt_page(page, sgt_iter, sgt)
2842 /* Check that we have the expected number of pages */
2843 GEM_BUG_ON(i != n_pages);
2848 /* fallthrough to use PAGE_KERNEL anyway */
2850 pgprot = PAGE_KERNEL;
2853 pgprot = pgprot_writecombine(PAGE_KERNEL_IO);
2856 addr = vmap(pages, n_pages, 0, pgprot);
2858 if (pages != stack_pages)
2864 /* get, pin, and map the pages of the object into kernel space */
2865 void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj,
2866 enum i915_map_type type)
2868 enum i915_map_type has_type;
2873 if (unlikely(!i915_gem_object_has_struct_page(obj)))
2874 return ERR_PTR(-ENXIO);
2876 ret = mutex_lock_interruptible(&obj->mm.lock);
2878 return ERR_PTR(ret);
2880 pinned = !(type & I915_MAP_OVERRIDE);
2881 type &= ~I915_MAP_OVERRIDE;
2883 if (!atomic_inc_not_zero(&obj->mm.pages_pin_count)) {
2884 if (unlikely(!i915_gem_object_has_pages(obj))) {
2885 GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj));
2887 ret = ____i915_gem_object_get_pages(obj);
2891 smp_mb__before_atomic();
2893 atomic_inc(&obj->mm.pages_pin_count);
2896 GEM_BUG_ON(!i915_gem_object_has_pages(obj));
2898 ptr = page_unpack_bits(obj->mm.mapping, &has_type);
2899 if (ptr && has_type != type) {
2905 if (is_vmalloc_addr(ptr))
2908 kunmap(kmap_to_page(ptr));
2910 ptr = obj->mm.mapping = NULL;
2914 ptr = i915_gem_object_map(obj, type);
2920 obj->mm.mapping = page_pack_bits(ptr, type);
2924 mutex_unlock(&obj->mm.lock);
2928 atomic_dec(&obj->mm.pages_pin_count);
2935 i915_gem_object_pwrite_gtt(struct drm_i915_gem_object *obj,
2936 const struct drm_i915_gem_pwrite *arg)
2938 struct address_space *mapping = obj->base.filp->f_mapping;
2939 char __user *user_data = u64_to_user_ptr(arg->data_ptr);
2943 /* Before we instantiate/pin the backing store for our use, we
2944 * can prepopulate the shmemfs filp efficiently using a write into
2945 * the pagecache. We avoid the penalty of instantiating all the
2946 * pages, important if the user is just writing to a few and never
2947 * uses the object on the GPU, and using a direct write into shmemfs
2948 * allows it to avoid the cost of retrieving a page (either swapin
2949 * or clearing-before-use) before it is overwritten.
2951 if (i915_gem_object_has_pages(obj))
2954 if (obj->mm.madv != I915_MADV_WILLNEED)
2957 /* Before the pages are instantiated the object is treated as being
2958 * in the CPU domain. The pages will be clflushed as required before
2959 * use, and we can freely write into the pages directly. If userspace
2960 * races pwrite with any other operation; corruption will ensue -
2961 * that is userspace's prerogative!
2965 offset = arg->offset;
2966 pg = offset_in_page(offset);
2969 unsigned int len, unwritten;
2974 len = PAGE_SIZE - pg;
2978 err = pagecache_write_begin(obj->base.filp, mapping,
2985 unwritten = copy_from_user(vaddr + pg, user_data, len);
2988 err = pagecache_write_end(obj->base.filp, mapping,
2989 offset, len, len - unwritten,
3006 static void i915_gem_client_mark_guilty(struct drm_i915_file_private *file_priv,
3007 const struct i915_gem_context *ctx)
3010 unsigned long prev_hang;
3012 if (i915_gem_context_is_banned(ctx))
3013 score = I915_CLIENT_SCORE_CONTEXT_BAN;
3017 prev_hang = xchg(&file_priv->hang_timestamp, jiffies);
3018 if (time_before(jiffies, prev_hang + I915_CLIENT_FAST_HANG_JIFFIES))
3019 score += I915_CLIENT_SCORE_HANG_FAST;
3022 atomic_add(score, &file_priv->ban_score);
3024 DRM_DEBUG_DRIVER("client %s: gained %u ban score, now %u\n",
3026 atomic_read(&file_priv->ban_score));
3030 static void i915_gem_context_mark_guilty(struct i915_gem_context *ctx)
3033 bool banned, bannable;
3035 atomic_inc(&ctx->guilty_count);
3037 bannable = i915_gem_context_is_bannable(ctx);
3038 score = atomic_add_return(CONTEXT_SCORE_GUILTY, &ctx->ban_score);
3039 banned = score >= CONTEXT_SCORE_BAN_THRESHOLD;
3041 /* Cool contexts don't accumulate client ban score */
3046 DRM_DEBUG_DRIVER("context %s: guilty %d, score %u, banned\n",
3047 ctx->name, atomic_read(&ctx->guilty_count),
3049 i915_gem_context_set_banned(ctx);
3052 if (!IS_ERR_OR_NULL(ctx->file_priv))
3053 i915_gem_client_mark_guilty(ctx->file_priv, ctx);
3056 static void i915_gem_context_mark_innocent(struct i915_gem_context *ctx)
3058 atomic_inc(&ctx->active_count);
3061 struct i915_request *
3062 i915_gem_find_active_request(struct intel_engine_cs *engine)
3064 struct i915_request *request, *active = NULL;
3065 unsigned long flags;
3068 * We are called by the error capture, reset and to dump engine
3069 * state at random points in time. In particular, note that neither is
3070 * crucially ordered with an interrupt. After a hang, the GPU is dead
3071 * and we assume that no more writes can happen (we waited long enough
3072 * for all writes that were in transaction to be flushed) - adding an
3073 * extra delay for a recent interrupt is pointless. Hence, we do
3074 * not need an engine->irq_seqno_barrier() before the seqno reads.
3075 * At all other times, we must assume the GPU is still running, but
3076 * we only care about the snapshot of this moment.
3078 spin_lock_irqsave(&engine->timeline.lock, flags);
3079 list_for_each_entry(request, &engine->timeline.requests, link) {
3080 if (__i915_request_completed(request, request->global_seqno))
3086 spin_unlock_irqrestore(&engine->timeline.lock, flags);
3092 * Ensure irq handler finishes, and not run again.
3093 * Also return the active request so that we only search for it once.
3095 struct i915_request *
3096 i915_gem_reset_prepare_engine(struct intel_engine_cs *engine)
3098 struct i915_request *request;
3101 * During the reset sequence, we must prevent the engine from
3102 * entering RC6. As the context state is undefined until we restart
3103 * the engine, if it does enter RC6 during the reset, the state
3104 * written to the powercontext is undefined and so we may lose
3105 * GPU state upon resume, i.e. fail to restart after a reset.
3107 intel_uncore_forcewake_get(engine->i915, FORCEWAKE_ALL);
3109 request = engine->reset.prepare(engine);
3110 if (request && request->fence.error == -EIO)
3111 request = ERR_PTR(-EIO); /* Previous reset failed! */
3116 int i915_gem_reset_prepare(struct drm_i915_private *dev_priv)
3118 struct intel_engine_cs *engine;
3119 struct i915_request *request;
3120 enum intel_engine_id id;
3123 for_each_engine(engine, dev_priv, id) {
3124 request = i915_gem_reset_prepare_engine(engine);
3125 if (IS_ERR(request)) {
3126 err = PTR_ERR(request);
3130 engine->hangcheck.active_request = request;
3133 i915_gem_revoke_fences(dev_priv);
3134 intel_uc_sanitize(dev_priv);
3139 static void engine_skip_context(struct i915_request *request)
3141 struct intel_engine_cs *engine = request->engine;
3142 struct i915_gem_context *hung_ctx = request->gem_context;
3143 struct i915_timeline *timeline = request->timeline;
3144 unsigned long flags;
3146 GEM_BUG_ON(timeline == &engine->timeline);
3148 spin_lock_irqsave(&engine->timeline.lock, flags);
3149 spin_lock(&timeline->lock);
3151 list_for_each_entry_continue(request, &engine->timeline.requests, link)
3152 if (request->gem_context == hung_ctx)
3153 i915_request_skip(request, -EIO);
3155 list_for_each_entry(request, &timeline->requests, link)
3156 i915_request_skip(request, -EIO);
3158 spin_unlock(&timeline->lock);
3159 spin_unlock_irqrestore(&engine->timeline.lock, flags);
3162 /* Returns the request if it was guilty of the hang */
3163 static struct i915_request *
3164 i915_gem_reset_request(struct intel_engine_cs *engine,
3165 struct i915_request *request,
3168 /* The guilty request will get skipped on a hung engine.
3170 * Users of client default contexts do not rely on logical
3171 * state preserved between batches so it is safe to execute
3172 * queued requests following the hang. Non default contexts
3173 * rely on preserved state, so skipping a batch loses the
3174 * evolution of the state and it needs to be considered corrupted.
3175 * Executing more queued batches on top of corrupted state is
3176 * risky. But we take the risk by trying to advance through
3177 * the queued requests in order to make the client behaviour
3178 * more predictable around resets, by not throwing away random
3179 * amount of batches it has prepared for execution. Sophisticated
3180 * clients can use gem_reset_stats_ioctl and dma fence status
3181 * (exported via sync_file info ioctl on explicit fences) to observe
3182 * when it loses the context state and should rebuild accordingly.
3184 * The context ban, and ultimately the client ban, mechanism are safety
3185 * valves if client submission ends up resulting in nothing more than
3189 if (i915_request_completed(request)) {
3190 GEM_TRACE("%s pardoned global=%d (fence %llx:%d), current %d\n",
3191 engine->name, request->global_seqno,
3192 request->fence.context, request->fence.seqno,
3193 intel_engine_get_seqno(engine));
3198 i915_gem_context_mark_guilty(request->gem_context);
3199 i915_request_skip(request, -EIO);
3201 /* If this context is now banned, skip all pending requests. */
3202 if (i915_gem_context_is_banned(request->gem_context))
3203 engine_skip_context(request);
3206 * Since this is not the hung engine, it may have advanced
3207 * since the hang declaration. Double check by refinding
3208 * the active request at the time of the reset.
3210 request = i915_gem_find_active_request(engine);
3212 unsigned long flags;
3214 i915_gem_context_mark_innocent(request->gem_context);
3215 dma_fence_set_error(&request->fence, -EAGAIN);
3217 /* Rewind the engine to replay the incomplete rq */
3218 spin_lock_irqsave(&engine->timeline.lock, flags);
3219 request = list_prev_entry(request, link);
3220 if (&request->link == &engine->timeline.requests)
3222 spin_unlock_irqrestore(&engine->timeline.lock, flags);
3229 void i915_gem_reset_engine(struct intel_engine_cs *engine,
3230 struct i915_request *request,
3234 * Make sure this write is visible before we re-enable the interrupt
3235 * handlers on another CPU, as tasklet_enable() resolves to just
3236 * a compiler barrier which is insufficient for our purpose here.
3238 smp_store_mb(engine->irq_posted, 0);
3241 request = i915_gem_reset_request(engine, request, stalled);
3243 /* Setup the CS to resume from the breadcrumb of the hung request */
3244 engine->reset.reset(engine, request);
3247 void i915_gem_reset(struct drm_i915_private *dev_priv,
3248 unsigned int stalled_mask)
3250 struct intel_engine_cs *engine;
3251 enum intel_engine_id id;
3253 lockdep_assert_held(&dev_priv->drm.struct_mutex);
3255 i915_retire_requests(dev_priv);
3257 for_each_engine(engine, dev_priv, id) {
3258 struct intel_context *ce;
3260 i915_gem_reset_engine(engine,
3261 engine->hangcheck.active_request,
3262 stalled_mask & ENGINE_MASK(id));
3263 ce = fetch_and_zero(&engine->last_retired_context);
3265 intel_context_unpin(ce);
3268 * Ostensibily, we always want a context loaded for powersaving,
3269 * so if the engine is idle after the reset, send a request
3270 * to load our scratch kernel_context.
3272 * More mysteriously, if we leave the engine idle after a reset,
3273 * the next userspace batch may hang, with what appears to be
3274 * an incoherent read by the CS (presumably stale TLB). An
3275 * empty request appears sufficient to paper over the glitch.
3277 if (intel_engine_is_idle(engine)) {
3278 struct i915_request *rq;
3280 rq = i915_request_alloc(engine,
3281 dev_priv->kernel_context);
3283 i915_request_add(rq);
3287 i915_gem_restore_fences(dev_priv);
3290 void i915_gem_reset_finish_engine(struct intel_engine_cs *engine)
3292 engine->reset.finish(engine);
3294 intel_uncore_forcewake_put(engine->i915, FORCEWAKE_ALL);
3297 void i915_gem_reset_finish(struct drm_i915_private *dev_priv)
3299 struct intel_engine_cs *engine;
3300 enum intel_engine_id id;
3302 lockdep_assert_held(&dev_priv->drm.struct_mutex);
3304 for_each_engine(engine, dev_priv, id) {
3305 engine->hangcheck.active_request = NULL;
3306 i915_gem_reset_finish_engine(engine);
3310 static void nop_submit_request(struct i915_request *request)
3312 GEM_TRACE("%s fence %llx:%d -> -EIO\n",
3313 request->engine->name,
3314 request->fence.context, request->fence.seqno);
3315 dma_fence_set_error(&request->fence, -EIO);
3317 i915_request_submit(request);
3320 static void nop_complete_submit_request(struct i915_request *request)
3322 unsigned long flags;
3324 GEM_TRACE("%s fence %llx:%d -> -EIO\n",
3325 request->engine->name,
3326 request->fence.context, request->fence.seqno);
3327 dma_fence_set_error(&request->fence, -EIO);
3329 spin_lock_irqsave(&request->engine->timeline.lock, flags);
3330 __i915_request_submit(request);
3331 intel_engine_init_global_seqno(request->engine, request->global_seqno);
3332 spin_unlock_irqrestore(&request->engine->timeline.lock, flags);
3335 void i915_gem_set_wedged(struct drm_i915_private *i915)
3337 struct intel_engine_cs *engine;
3338 enum intel_engine_id id;
3340 GEM_TRACE("start\n");
3342 if (GEM_SHOW_DEBUG()) {
3343 struct drm_printer p = drm_debug_printer(__func__);
3345 for_each_engine(engine, i915, id)
3346 intel_engine_dump(engine, &p, "%s\n", engine->name);
3349 if (test_and_set_bit(I915_WEDGED, &i915->gpu_error.flags))
3353 * First, stop submission to hw, but do not yet complete requests by
3354 * rolling the global seqno forward (since this would complete requests
3355 * for which we haven't set the fence error to EIO yet).
3357 for_each_engine(engine, i915, id) {
3358 i915_gem_reset_prepare_engine(engine);
3360 engine->submit_request = nop_submit_request;
3361 engine->schedule = NULL;
3363 i915->caps.scheduler = 0;
3365 /* Even if the GPU reset fails, it should still stop the engines */
3366 if (INTEL_GEN(i915) >= 5)
3367 intel_gpu_reset(i915, ALL_ENGINES);
3370 * Make sure no one is running the old callback before we proceed with
3371 * cancelling requests and resetting the completion tracking. Otherwise
3372 * we might submit a request to the hardware which never completes.
3376 for_each_engine(engine, i915, id) {
3377 /* Mark all executing requests as skipped */
3378 engine->cancel_requests(engine);
3381 * Only once we've force-cancelled all in-flight requests can we
3382 * start to complete all requests.
3384 engine->submit_request = nop_complete_submit_request;
3388 * Make sure no request can slip through without getting completed by
3389 * either this call here to intel_engine_init_global_seqno, or the one
3390 * in nop_complete_submit_request.
3394 for_each_engine(engine, i915, id) {
3395 unsigned long flags;
3398 * Mark all pending requests as complete so that any concurrent
3399 * (lockless) lookup doesn't try and wait upon the request as we
3402 spin_lock_irqsave(&engine->timeline.lock, flags);
3403 intel_engine_init_global_seqno(engine,
3404 intel_engine_last_submit(engine));
3405 spin_unlock_irqrestore(&engine->timeline.lock, flags);
3407 i915_gem_reset_finish_engine(engine);
3413 wake_up_all(&i915->gpu_error.reset_queue);
3416 bool i915_gem_unset_wedged(struct drm_i915_private *i915)
3418 struct i915_timeline *tl;
3420 lockdep_assert_held(&i915->drm.struct_mutex);
3421 if (!test_bit(I915_WEDGED, &i915->gpu_error.flags))
3424 GEM_TRACE("start\n");
3427 * Before unwedging, make sure that all pending operations
3428 * are flushed and errored out - we may have requests waiting upon
3429 * third party fences. We marked all inflight requests as EIO, and
3430 * every execbuf since returned EIO, for consistency we want all
3431 * the currently pending requests to also be marked as EIO, which
3432 * is done inside our nop_submit_request - and so we must wait.
3434 * No more can be submitted until we reset the wedged bit.
3436 list_for_each_entry(tl, &i915->gt.timelines, link) {
3437 struct i915_request *rq;
3439 rq = i915_gem_active_peek(&tl->last_request,
3440 &i915->drm.struct_mutex);
3445 * We can't use our normal waiter as we want to
3446 * avoid recursively trying to handle the current
3447 * reset. The basic dma_fence_default_wait() installs
3448 * a callback for dma_fence_signal(), which is
3449 * triggered by our nop handler (indirectly, the
3450 * callback enables the signaler thread which is
3451 * woken by the nop_submit_request() advancing the seqno
3452 * and when the seqno passes the fence, the signaler
3453 * then signals the fence waking us up).
3455 if (dma_fence_default_wait(&rq->fence, true,
3456 MAX_SCHEDULE_TIMEOUT) < 0)
3459 i915_retire_requests(i915);
3460 GEM_BUG_ON(i915->gt.active_requests);
3462 if (!intel_gpu_reset(i915, ALL_ENGINES))
3463 intel_engines_sanitize(i915);
3466 * Undo nop_submit_request. We prevent all new i915 requests from
3467 * being queued (by disallowing execbuf whilst wedged) so having
3468 * waited for all active requests above, we know the system is idle
3469 * and do not have to worry about a thread being inside
3470 * engine->submit_request() as we swap over. So unlike installing
3471 * the nop_submit_request on reset, we can do this from normal
3472 * context and do not require stop_machine().
3474 intel_engines_reset_default_submission(i915);
3475 i915_gem_contexts_lost(i915);
3479 smp_mb__before_atomic(); /* complete takeover before enabling execbuf */
3480 clear_bit(I915_WEDGED, &i915->gpu_error.flags);
3486 i915_gem_retire_work_handler(struct work_struct *work)
3488 struct drm_i915_private *dev_priv =
3489 container_of(work, typeof(*dev_priv), gt.retire_work.work);
3490 struct drm_device *dev = &dev_priv->drm;
3492 /* Come back later if the device is busy... */
3493 if (mutex_trylock(&dev->struct_mutex)) {
3494 i915_retire_requests(dev_priv);
3495 mutex_unlock(&dev->struct_mutex);
3499 * Keep the retire handler running until we are finally idle.
3500 * We do not need to do this test under locking as in the worst-case
3501 * we queue the retire worker once too often.
3503 if (READ_ONCE(dev_priv->gt.awake))
3504 queue_delayed_work(dev_priv->wq,
3505 &dev_priv->gt.retire_work,
3506 round_jiffies_up_relative(HZ));
3509 static void shrink_caches(struct drm_i915_private *i915)
3512 * kmem_cache_shrink() discards empty slabs and reorders partially
3513 * filled slabs to prioritise allocating from the mostly full slabs,
3514 * with the aim of reducing fragmentation.
3516 kmem_cache_shrink(i915->priorities);
3517 kmem_cache_shrink(i915->dependencies);
3518 kmem_cache_shrink(i915->requests);
3519 kmem_cache_shrink(i915->luts);
3520 kmem_cache_shrink(i915->vmas);
3521 kmem_cache_shrink(i915->objects);
3524 struct sleep_rcu_work {
3526 struct rcu_head rcu;
3527 struct work_struct work;
3529 struct drm_i915_private *i915;
3534 same_epoch(struct drm_i915_private *i915, unsigned int epoch)
3537 * There is a small chance that the epoch wrapped since we started
3538 * sleeping. If we assume that epoch is at least a u32, then it will
3539 * take at least 2^32 * 100ms for it to wrap, or about 326 years.
3541 return epoch == READ_ONCE(i915->gt.epoch);
3544 static void __sleep_work(struct work_struct *work)
3546 struct sleep_rcu_work *s = container_of(work, typeof(*s), work);
3547 struct drm_i915_private *i915 = s->i915;
3548 unsigned int epoch = s->epoch;
3551 if (same_epoch(i915, epoch))
3552 shrink_caches(i915);
3555 static void __sleep_rcu(struct rcu_head *rcu)
3557 struct sleep_rcu_work *s = container_of(rcu, typeof(*s), rcu);
3558 struct drm_i915_private *i915 = s->i915;
3560 destroy_rcu_head(&s->rcu);
3562 if (same_epoch(i915, s->epoch)) {
3563 INIT_WORK(&s->work, __sleep_work);
3564 queue_work(i915->wq, &s->work);
3571 new_requests_since_last_retire(const struct drm_i915_private *i915)
3573 return (READ_ONCE(i915->gt.active_requests) ||
3574 work_pending(&i915->gt.idle_work.work));
3577 static void assert_kernel_context_is_current(struct drm_i915_private *i915)
3579 struct intel_engine_cs *engine;
3580 enum intel_engine_id id;
3582 if (i915_terminally_wedged(&i915->gpu_error))
3585 GEM_BUG_ON(i915->gt.active_requests);
3586 for_each_engine(engine, i915, id) {
3587 GEM_BUG_ON(__i915_gem_active_peek(&engine->timeline.last_request));
3588 GEM_BUG_ON(engine->last_retired_context !=
3589 to_intel_context(i915->kernel_context, engine));
3594 i915_gem_idle_work_handler(struct work_struct *work)
3596 struct drm_i915_private *dev_priv =
3597 container_of(work, typeof(*dev_priv), gt.idle_work.work);
3598 unsigned int epoch = I915_EPOCH_INVALID;
3599 bool rearm_hangcheck;
3601 if (!READ_ONCE(dev_priv->gt.awake))
3604 if (READ_ONCE(dev_priv->gt.active_requests))
3608 * Flush out the last user context, leaving only the pinned
3609 * kernel context resident. When we are idling on the kernel_context,
3610 * no more new requests (with a context switch) are emitted and we
3611 * can finally rest. A consequence is that the idle work handler is
3612 * always called at least twice before idling (and if the system is
3613 * idle that implies a round trip through the retire worker).
3615 mutex_lock(&dev_priv->drm.struct_mutex);
3616 i915_gem_switch_to_kernel_context(dev_priv);
3617 mutex_unlock(&dev_priv->drm.struct_mutex);
3619 GEM_TRACE("active_requests=%d (after switch-to-kernel-context)\n",
3620 READ_ONCE(dev_priv->gt.active_requests));
3623 * Wait for last execlists context complete, but bail out in case a
3624 * new request is submitted. As we don't trust the hardware, we
3625 * continue on if the wait times out. This is necessary to allow
3626 * the machine to suspend even if the hardware dies, and we will
3627 * try to recover in resume (after depriving the hardware of power,
3628 * it may be in a better mmod).
3630 __wait_for(if (new_requests_since_last_retire(dev_priv)) return,
3631 intel_engines_are_idle(dev_priv),
3632 I915_IDLE_ENGINES_TIMEOUT * 1000,
3636 cancel_delayed_work_sync(&dev_priv->gpu_error.hangcheck_work);
3638 if (!mutex_trylock(&dev_priv->drm.struct_mutex)) {
3639 /* Currently busy, come back later */
3640 mod_delayed_work(dev_priv->wq,
3641 &dev_priv->gt.idle_work,
3642 msecs_to_jiffies(50));
3647 * New request retired after this work handler started, extend active
3648 * period until next instance of the work.
3650 if (new_requests_since_last_retire(dev_priv))
3653 epoch = __i915_gem_park(dev_priv);
3655 assert_kernel_context_is_current(dev_priv);
3657 rearm_hangcheck = false;
3659 mutex_unlock(&dev_priv->drm.struct_mutex);
3662 if (rearm_hangcheck) {
3663 GEM_BUG_ON(!dev_priv->gt.awake);
3664 i915_queue_hangcheck(dev_priv);
3668 * When we are idle, it is an opportune time to reap our caches.
3669 * However, we have many objects that utilise RCU and the ordered
3670 * i915->wq that this work is executing on. To try and flush any
3671 * pending frees now we are idle, we first wait for an RCU grace
3672 * period, and then queue a task (that will run last on the wq) to
3673 * shrink and re-optimize the caches.
3675 if (same_epoch(dev_priv, epoch)) {
3676 struct sleep_rcu_work *s = kmalloc(sizeof(*s), GFP_KERNEL);
3678 init_rcu_head(&s->rcu);
3681 call_rcu(&s->rcu, __sleep_rcu);
3686 void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file)
3688 struct drm_i915_private *i915 = to_i915(gem->dev);
3689 struct drm_i915_gem_object *obj = to_intel_bo(gem);
3690 struct drm_i915_file_private *fpriv = file->driver_priv;
3691 struct i915_lut_handle *lut, *ln;
3693 mutex_lock(&i915->drm.struct_mutex);
3695 list_for_each_entry_safe(lut, ln, &obj->lut_list, obj_link) {
3696 struct i915_gem_context *ctx = lut->ctx;
3697 struct i915_vma *vma;
3699 GEM_BUG_ON(ctx->file_priv == ERR_PTR(-EBADF));
3700 if (ctx->file_priv != fpriv)
3703 vma = radix_tree_delete(&ctx->handles_vma, lut->handle);
3704 GEM_BUG_ON(vma->obj != obj);
3706 /* We allow the process to have multiple handles to the same
3707 * vma, in the same fd namespace, by virtue of flink/open.
3709 GEM_BUG_ON(!vma->open_count);
3710 if (!--vma->open_count && !i915_vma_is_ggtt(vma))
3711 i915_vma_close(vma);
3713 list_del(&lut->obj_link);
3714 list_del(&lut->ctx_link);
3716 kmem_cache_free(i915->luts, lut);
3717 __i915_gem_object_release_unless_active(obj);
3720 mutex_unlock(&i915->drm.struct_mutex);
3723 static unsigned long to_wait_timeout(s64 timeout_ns)
3726 return MAX_SCHEDULE_TIMEOUT;
3728 if (timeout_ns == 0)
3731 return nsecs_to_jiffies_timeout(timeout_ns);
3735 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
3736 * @dev: drm device pointer
3737 * @data: ioctl data blob
3738 * @file: drm file pointer
3740 * Returns 0 if successful, else an error is returned with the remaining time in
3741 * the timeout parameter.
3742 * -ETIME: object is still busy after timeout
3743 * -ERESTARTSYS: signal interrupted the wait
3744 * -ENONENT: object doesn't exist
3745 * Also possible, but rare:
3746 * -EAGAIN: incomplete, restart syscall
3748 * -ENODEV: Internal IRQ fail
3749 * -E?: The add request failed
3751 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
3752 * non-zero timeout parameter the wait ioctl will wait for the given number of
3753 * nanoseconds on an object becoming unbusy. Since the wait itself does so
3754 * without holding struct_mutex the object may become re-busied before this
3755 * function completes. A similar but shorter * race condition exists in the busy
3759 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
3761 struct drm_i915_gem_wait *args = data;
3762 struct drm_i915_gem_object *obj;
3766 if (args->flags != 0)
3769 obj = i915_gem_object_lookup(file, args->bo_handle);
3773 start = ktime_get();
3775 ret = i915_gem_object_wait(obj,
3776 I915_WAIT_INTERRUPTIBLE |
3777 I915_WAIT_PRIORITY |
3779 to_wait_timeout(args->timeout_ns),
3780 to_rps_client(file));
3782 if (args->timeout_ns > 0) {
3783 args->timeout_ns -= ktime_to_ns(ktime_sub(ktime_get(), start));
3784 if (args->timeout_ns < 0)
3785 args->timeout_ns = 0;
3788 * Apparently ktime isn't accurate enough and occasionally has a
3789 * bit of mismatch in the jiffies<->nsecs<->ktime loop. So patch
3790 * things up to make the test happy. We allow up to 1 jiffy.
3792 * This is a regression from the timespec->ktime conversion.
3794 if (ret == -ETIME && !nsecs_to_jiffies(args->timeout_ns))
3795 args->timeout_ns = 0;
3797 /* Asked to wait beyond the jiffie/scheduler precision? */
3798 if (ret == -ETIME && args->timeout_ns)
3802 i915_gem_object_put(obj);
3806 static long wait_for_timeline(struct i915_timeline *tl,
3807 unsigned int flags, long timeout)
3809 struct i915_request *rq;
3811 rq = i915_gem_active_get_unlocked(&tl->last_request);
3818 * Switching to the kernel context is often used a synchronous
3819 * step prior to idling, e.g. in suspend for flushing all
3820 * current operations to memory before sleeping. These we
3821 * want to complete as quickly as possible to avoid prolonged
3822 * stalls, so allow the gpu to boost to maximum clocks.
3824 if (flags & I915_WAIT_FOR_IDLE_BOOST)
3825 gen6_rps_boost(rq, NULL);
3827 timeout = i915_request_wait(rq, flags, timeout);
3828 i915_request_put(rq);
3833 static int wait_for_engines(struct drm_i915_private *i915)
3835 if (wait_for(intel_engines_are_idle(i915), I915_IDLE_ENGINES_TIMEOUT)) {
3836 dev_err(i915->drm.dev,
3837 "Failed to idle engines, declaring wedged!\n");
3839 i915_gem_set_wedged(i915);
3846 int i915_gem_wait_for_idle(struct drm_i915_private *i915,
3847 unsigned int flags, long timeout)
3849 GEM_TRACE("flags=%x (%s), timeout=%ld%s\n",
3850 flags, flags & I915_WAIT_LOCKED ? "locked" : "unlocked",
3851 timeout, timeout == MAX_SCHEDULE_TIMEOUT ? " (forever)" : "");
3853 /* If the device is asleep, we have no requests outstanding */
3854 if (!READ_ONCE(i915->gt.awake))
3857 if (flags & I915_WAIT_LOCKED) {
3858 struct i915_timeline *tl;
3861 lockdep_assert_held(&i915->drm.struct_mutex);
3863 list_for_each_entry(tl, &i915->gt.timelines, link) {
3864 timeout = wait_for_timeline(tl, flags, timeout);
3868 if (GEM_SHOW_DEBUG() && !timeout) {
3869 /* Presume that timeout was non-zero to begin with! */
3870 dev_warn(&i915->drm.pdev->dev,
3871 "Missed idle-completion interrupt!\n");
3875 err = wait_for_engines(i915);
3879 i915_retire_requests(i915);
3880 GEM_BUG_ON(i915->gt.active_requests);
3882 struct intel_engine_cs *engine;
3883 enum intel_engine_id id;
3885 for_each_engine(engine, i915, id) {
3886 struct i915_timeline *tl = &engine->timeline;
3888 timeout = wait_for_timeline(tl, flags, timeout);
3897 static void __i915_gem_object_flush_for_display(struct drm_i915_gem_object *obj)
3900 * We manually flush the CPU domain so that we can override and
3901 * force the flush for the display, and perform it asyncrhonously.
3903 flush_write_domain(obj, ~I915_GEM_DOMAIN_CPU);
3904 if (obj->cache_dirty)
3905 i915_gem_clflush_object(obj, I915_CLFLUSH_FORCE);
3906 obj->write_domain = 0;
3909 void i915_gem_object_flush_if_display(struct drm_i915_gem_object *obj)
3911 if (!READ_ONCE(obj->pin_global))
3914 mutex_lock(&obj->base.dev->struct_mutex);
3915 __i915_gem_object_flush_for_display(obj);
3916 mutex_unlock(&obj->base.dev->struct_mutex);
3920 * Moves a single object to the WC read, and possibly write domain.
3921 * @obj: object to act on
3922 * @write: ask for write access or read only
3924 * This function returns when the move is complete, including waiting on
3928 i915_gem_object_set_to_wc_domain(struct drm_i915_gem_object *obj, bool write)
3932 lockdep_assert_held(&obj->base.dev->struct_mutex);
3934 ret = i915_gem_object_wait(obj,
3935 I915_WAIT_INTERRUPTIBLE |
3937 (write ? I915_WAIT_ALL : 0),
3938 MAX_SCHEDULE_TIMEOUT,
3943 if (obj->write_domain == I915_GEM_DOMAIN_WC)
3946 /* Flush and acquire obj->pages so that we are coherent through
3947 * direct access in memory with previous cached writes through
3948 * shmemfs and that our cache domain tracking remains valid.
3949 * For example, if the obj->filp was moved to swap without us
3950 * being notified and releasing the pages, we would mistakenly
3951 * continue to assume that the obj remained out of the CPU cached
3954 ret = i915_gem_object_pin_pages(obj);
3958 flush_write_domain(obj, ~I915_GEM_DOMAIN_WC);
3960 /* Serialise direct access to this object with the barriers for
3961 * coherent writes from the GPU, by effectively invalidating the
3962 * WC domain upon first access.
3964 if ((obj->read_domains & I915_GEM_DOMAIN_WC) == 0)
3967 /* It should now be out of any other write domains, and we can update
3968 * the domain values for our changes.
3970 GEM_BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_WC) != 0);
3971 obj->read_domains |= I915_GEM_DOMAIN_WC;
3973 obj->read_domains = I915_GEM_DOMAIN_WC;
3974 obj->write_domain = I915_GEM_DOMAIN_WC;
3975 obj->mm.dirty = true;
3978 i915_gem_object_unpin_pages(obj);
3983 * Moves a single object to the GTT read, and possibly write domain.
3984 * @obj: object to act on
3985 * @write: ask for write access or read only
3987 * This function returns when the move is complete, including waiting on
3991 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
3995 lockdep_assert_held(&obj->base.dev->struct_mutex);
3997 ret = i915_gem_object_wait(obj,
3998 I915_WAIT_INTERRUPTIBLE |
4000 (write ? I915_WAIT_ALL : 0),
4001 MAX_SCHEDULE_TIMEOUT,
4006 if (obj->write_domain == I915_GEM_DOMAIN_GTT)
4009 /* Flush and acquire obj->pages so that we are coherent through
4010 * direct access in memory with previous cached writes through
4011 * shmemfs and that our cache domain tracking remains valid.
4012 * For example, if the obj->filp was moved to swap without us
4013 * being notified and releasing the pages, we would mistakenly
4014 * continue to assume that the obj remained out of the CPU cached
4017 ret = i915_gem_object_pin_pages(obj);
4021 flush_write_domain(obj, ~I915_GEM_DOMAIN_GTT);
4023 /* Serialise direct access to this object with the barriers for
4024 * coherent writes from the GPU, by effectively invalidating the
4025 * GTT domain upon first access.
4027 if ((obj->read_domains & I915_GEM_DOMAIN_GTT) == 0)
4030 /* It should now be out of any other write domains, and we can update
4031 * the domain values for our changes.
4033 GEM_BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
4034 obj->read_domains |= I915_GEM_DOMAIN_GTT;
4036 obj->read_domains = I915_GEM_DOMAIN_GTT;
4037 obj->write_domain = I915_GEM_DOMAIN_GTT;
4038 obj->mm.dirty = true;
4041 i915_gem_object_unpin_pages(obj);
4046 * Changes the cache-level of an object across all VMA.
4047 * @obj: object to act on
4048 * @cache_level: new cache level to set for the object
4050 * After this function returns, the object will be in the new cache-level
4051 * across all GTT and the contents of the backing storage will be coherent,
4052 * with respect to the new cache-level. In order to keep the backing storage
4053 * coherent for all users, we only allow a single cache level to be set
4054 * globally on the object and prevent it from being changed whilst the
4055 * hardware is reading from the object. That is if the object is currently
4056 * on the scanout it will be set to uncached (or equivalent display
4057 * cache coherency) and all non-MOCS GPU access will also be uncached so
4058 * that all direct access to the scanout remains coherent.
4060 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
4061 enum i915_cache_level cache_level)
4063 struct i915_vma *vma;
4066 lockdep_assert_held(&obj->base.dev->struct_mutex);
4068 if (obj->cache_level == cache_level)
4071 /* Inspect the list of currently bound VMA and unbind any that would
4072 * be invalid given the new cache-level. This is principally to
4073 * catch the issue of the CS prefetch crossing page boundaries and
4074 * reading an invalid PTE on older architectures.
4077 list_for_each_entry(vma, &obj->vma_list, obj_link) {
4078 if (!drm_mm_node_allocated(&vma->node))
4081 if (i915_vma_is_pinned(vma)) {
4082 DRM_DEBUG("can not change the cache level of pinned objects\n");
4086 if (!i915_vma_is_closed(vma) &&
4087 i915_gem_valid_gtt_space(vma, cache_level))
4090 ret = i915_vma_unbind(vma);
4094 /* As unbinding may affect other elements in the
4095 * obj->vma_list (due to side-effects from retiring
4096 * an active vma), play safe and restart the iterator.
4101 /* We can reuse the existing drm_mm nodes but need to change the
4102 * cache-level on the PTE. We could simply unbind them all and
4103 * rebind with the correct cache-level on next use. However since
4104 * we already have a valid slot, dma mapping, pages etc, we may as
4105 * rewrite the PTE in the belief that doing so tramples upon less
4106 * state and so involves less work.
4108 if (obj->bind_count) {
4109 /* Before we change the PTE, the GPU must not be accessing it.
4110 * If we wait upon the object, we know that all the bound
4111 * VMA are no longer active.
4113 ret = i915_gem_object_wait(obj,
4114 I915_WAIT_INTERRUPTIBLE |
4117 MAX_SCHEDULE_TIMEOUT,
4122 if (!HAS_LLC(to_i915(obj->base.dev)) &&
4123 cache_level != I915_CACHE_NONE) {
4124 /* Access to snoopable pages through the GTT is
4125 * incoherent and on some machines causes a hard
4126 * lockup. Relinquish the CPU mmaping to force
4127 * userspace to refault in the pages and we can
4128 * then double check if the GTT mapping is still
4129 * valid for that pointer access.
4131 i915_gem_release_mmap(obj);
4133 /* As we no longer need a fence for GTT access,
4134 * we can relinquish it now (and so prevent having
4135 * to steal a fence from someone else on the next
4136 * fence request). Note GPU activity would have
4137 * dropped the fence as all snoopable access is
4138 * supposed to be linear.
4140 for_each_ggtt_vma(vma, obj) {
4141 ret = i915_vma_put_fence(vma);
4146 /* We either have incoherent backing store and
4147 * so no GTT access or the architecture is fully
4148 * coherent. In such cases, existing GTT mmaps
4149 * ignore the cache bit in the PTE and we can
4150 * rewrite it without confusing the GPU or having
4151 * to force userspace to fault back in its mmaps.
4155 list_for_each_entry(vma, &obj->vma_list, obj_link) {
4156 if (!drm_mm_node_allocated(&vma->node))
4159 ret = i915_vma_bind(vma, cache_level, PIN_UPDATE);
4165 list_for_each_entry(vma, &obj->vma_list, obj_link)
4166 vma->node.color = cache_level;
4167 i915_gem_object_set_cache_coherency(obj, cache_level);
4168 obj->cache_dirty = true; /* Always invalidate stale cachelines */
4173 int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data,
4174 struct drm_file *file)
4176 struct drm_i915_gem_caching *args = data;
4177 struct drm_i915_gem_object *obj;
4181 obj = i915_gem_object_lookup_rcu(file, args->handle);
4187 switch (obj->cache_level) {
4188 case I915_CACHE_LLC:
4189 case I915_CACHE_L3_LLC:
4190 args->caching = I915_CACHING_CACHED;
4194 args->caching = I915_CACHING_DISPLAY;
4198 args->caching = I915_CACHING_NONE;
4206 int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
4207 struct drm_file *file)
4209 struct drm_i915_private *i915 = to_i915(dev);
4210 struct drm_i915_gem_caching *args = data;
4211 struct drm_i915_gem_object *obj;
4212 enum i915_cache_level level;
4215 switch (args->caching) {
4216 case I915_CACHING_NONE:
4217 level = I915_CACHE_NONE;
4219 case I915_CACHING_CACHED:
4221 * Due to a HW issue on BXT A stepping, GPU stores via a
4222 * snooped mapping may leave stale data in a corresponding CPU
4223 * cacheline, whereas normally such cachelines would get
4226 if (!HAS_LLC(i915) && !HAS_SNOOP(i915))
4229 level = I915_CACHE_LLC;
4231 case I915_CACHING_DISPLAY:
4232 level = HAS_WT(i915) ? I915_CACHE_WT : I915_CACHE_NONE;
4238 obj = i915_gem_object_lookup(file, args->handle);
4243 * The caching mode of proxy object is handled by its generator, and
4244 * not allowed to be changed by userspace.
4246 if (i915_gem_object_is_proxy(obj)) {
4251 if (obj->cache_level == level)
4254 ret = i915_gem_object_wait(obj,
4255 I915_WAIT_INTERRUPTIBLE,
4256 MAX_SCHEDULE_TIMEOUT,
4257 to_rps_client(file));
4261 ret = i915_mutex_lock_interruptible(dev);
4265 ret = i915_gem_object_set_cache_level(obj, level);
4266 mutex_unlock(&dev->struct_mutex);
4269 i915_gem_object_put(obj);
4274 * Prepare buffer for display plane (scanout, cursors, etc). Can be called from
4275 * an uninterruptible phase (modesetting) and allows any flushes to be pipelined
4276 * (for pageflips). We only flush the caches while preparing the buffer for
4277 * display, the callers are responsible for frontbuffer flush.
4280 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
4282 const struct i915_ggtt_view *view,
4285 struct i915_vma *vma;
4288 lockdep_assert_held(&obj->base.dev->struct_mutex);
4290 /* Mark the global pin early so that we account for the
4291 * display coherency whilst setting up the cache domains.
4295 /* The display engine is not coherent with the LLC cache on gen6. As
4296 * a result, we make sure that the pinning that is about to occur is
4297 * done with uncached PTEs. This is lowest common denominator for all
4300 * However for gen6+, we could do better by using the GFDT bit instead
4301 * of uncaching, which would allow us to flush all the LLC-cached data
4302 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
4304 ret = i915_gem_object_set_cache_level(obj,
4305 HAS_WT(to_i915(obj->base.dev)) ?
4306 I915_CACHE_WT : I915_CACHE_NONE);
4309 goto err_unpin_global;
4312 /* As the user may map the buffer once pinned in the display plane
4313 * (e.g. libkms for the bootup splash), we have to ensure that we
4314 * always use map_and_fenceable for all scanout buffers. However,
4315 * it may simply be too big to fit into mappable, in which case
4316 * put it anyway and hope that userspace can cope (but always first
4317 * try to preserve the existing ABI).
4319 vma = ERR_PTR(-ENOSPC);
4320 if ((flags & PIN_MAPPABLE) == 0 &&
4321 (!view || view->type == I915_GGTT_VIEW_NORMAL))
4322 vma = i915_gem_object_ggtt_pin(obj, view, 0, alignment,
4327 vma = i915_gem_object_ggtt_pin(obj, view, 0, alignment, flags);
4329 goto err_unpin_global;
4331 vma->display_alignment = max_t(u64, vma->display_alignment, alignment);
4333 __i915_gem_object_flush_for_display(obj);
4335 /* It should now be out of any other write domains, and we can update
4336 * the domain values for our changes.
4338 obj->read_domains |= I915_GEM_DOMAIN_GTT;
4348 i915_gem_object_unpin_from_display_plane(struct i915_vma *vma)
4350 lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
4352 if (WARN_ON(vma->obj->pin_global == 0))
4355 if (--vma->obj->pin_global == 0)
4356 vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
4358 /* Bump the LRU to try and avoid premature eviction whilst flipping */
4359 i915_gem_object_bump_inactive_ggtt(vma->obj);
4361 i915_vma_unpin(vma);
4365 * Moves a single object to the CPU read, and possibly write domain.
4366 * @obj: object to act on
4367 * @write: requesting write or read-only access
4369 * This function returns when the move is complete, including waiting on
4373 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
4377 lockdep_assert_held(&obj->base.dev->struct_mutex);
4379 ret = i915_gem_object_wait(obj,
4380 I915_WAIT_INTERRUPTIBLE |
4382 (write ? I915_WAIT_ALL : 0),
4383 MAX_SCHEDULE_TIMEOUT,
4388 flush_write_domain(obj, ~I915_GEM_DOMAIN_CPU);
4390 /* Flush the CPU cache if it's still invalid. */
4391 if ((obj->read_domains & I915_GEM_DOMAIN_CPU) == 0) {
4392 i915_gem_clflush_object(obj, I915_CLFLUSH_SYNC);
4393 obj->read_domains |= I915_GEM_DOMAIN_CPU;
4396 /* It should now be out of any other write domains, and we can update
4397 * the domain values for our changes.
4399 GEM_BUG_ON(obj->write_domain & ~I915_GEM_DOMAIN_CPU);
4401 /* If we're writing through the CPU, then the GPU read domains will
4402 * need to be invalidated at next use.
4405 __start_cpu_write(obj);
4410 /* Throttle our rendering by waiting until the ring has completed our requests
4411 * emitted over 20 msec ago.
4413 * Note that if we were to use the current jiffies each time around the loop,
4414 * we wouldn't escape the function with any frames outstanding if the time to
4415 * render a frame was over 20ms.
4417 * This should get us reasonable parallelism between CPU and GPU but also
4418 * relatively low latency when blocking on a particular request to finish.
4421 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
4423 struct drm_i915_private *dev_priv = to_i915(dev);
4424 struct drm_i915_file_private *file_priv = file->driver_priv;
4425 unsigned long recent_enough = jiffies - DRM_I915_THROTTLE_JIFFIES;
4426 struct i915_request *request, *target = NULL;
4429 /* ABI: return -EIO if already wedged */
4430 if (i915_terminally_wedged(&dev_priv->gpu_error))
4433 spin_lock(&file_priv->mm.lock);
4434 list_for_each_entry(request, &file_priv->mm.request_list, client_link) {
4435 if (time_after_eq(request->emitted_jiffies, recent_enough))
4439 list_del(&target->client_link);
4440 target->file_priv = NULL;
4446 i915_request_get(target);
4447 spin_unlock(&file_priv->mm.lock);
4452 ret = i915_request_wait(target,
4453 I915_WAIT_INTERRUPTIBLE,
4454 MAX_SCHEDULE_TIMEOUT);
4455 i915_request_put(target);
4457 return ret < 0 ? ret : 0;
4461 i915_gem_object_ggtt_pin(struct drm_i915_gem_object *obj,
4462 const struct i915_ggtt_view *view,
4467 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
4468 struct i915_address_space *vm = &dev_priv->ggtt.vm;
4469 struct i915_vma *vma;
4472 lockdep_assert_held(&obj->base.dev->struct_mutex);
4474 if (flags & PIN_MAPPABLE &&
4475 (!view || view->type == I915_GGTT_VIEW_NORMAL)) {
4476 /* If the required space is larger than the available
4477 * aperture, we will not able to find a slot for the
4478 * object and unbinding the object now will be in
4479 * vain. Worse, doing so may cause us to ping-pong
4480 * the object in and out of the Global GTT and
4481 * waste a lot of cycles under the mutex.
4483 if (obj->base.size > dev_priv->ggtt.mappable_end)
4484 return ERR_PTR(-E2BIG);
4486 /* If NONBLOCK is set the caller is optimistically
4487 * trying to cache the full object within the mappable
4488 * aperture, and *must* have a fallback in place for
4489 * situations where we cannot bind the object. We
4490 * can be a little more lax here and use the fallback
4491 * more often to avoid costly migrations of ourselves
4492 * and other objects within the aperture.
4494 * Half-the-aperture is used as a simple heuristic.
4495 * More interesting would to do search for a free
4496 * block prior to making the commitment to unbind.
4497 * That caters for the self-harm case, and with a
4498 * little more heuristics (e.g. NOFAULT, NOEVICT)
4499 * we could try to minimise harm to others.
4501 if (flags & PIN_NONBLOCK &&
4502 obj->base.size > dev_priv->ggtt.mappable_end / 2)
4503 return ERR_PTR(-ENOSPC);
4506 vma = i915_vma_instance(obj, vm, view);
4507 if (unlikely(IS_ERR(vma)))
4510 if (i915_vma_misplaced(vma, size, alignment, flags)) {
4511 if (flags & PIN_NONBLOCK) {
4512 if (i915_vma_is_pinned(vma) || i915_vma_is_active(vma))
4513 return ERR_PTR(-ENOSPC);
4515 if (flags & PIN_MAPPABLE &&
4516 vma->fence_size > dev_priv->ggtt.mappable_end / 2)
4517 return ERR_PTR(-ENOSPC);
4520 WARN(i915_vma_is_pinned(vma),
4521 "bo is already pinned in ggtt with incorrect alignment:"
4522 " offset=%08x, req.alignment=%llx,"
4523 " req.map_and_fenceable=%d, vma->map_and_fenceable=%d\n",
4524 i915_ggtt_offset(vma), alignment,
4525 !!(flags & PIN_MAPPABLE),
4526 i915_vma_is_map_and_fenceable(vma));
4527 ret = i915_vma_unbind(vma);
4529 return ERR_PTR(ret);
4532 ret = i915_vma_pin(vma, size, alignment, flags | PIN_GLOBAL);
4534 return ERR_PTR(ret);
4539 static __always_inline unsigned int __busy_read_flag(unsigned int id)
4541 /* Note that we could alias engines in the execbuf API, but
4542 * that would be very unwise as it prevents userspace from
4543 * fine control over engine selection. Ahem.
4545 * This should be something like EXEC_MAX_ENGINE instead of
4548 BUILD_BUG_ON(I915_NUM_ENGINES > 16);
4549 return 0x10000 << id;
4552 static __always_inline unsigned int __busy_write_id(unsigned int id)
4554 /* The uABI guarantees an active writer is also amongst the read
4555 * engines. This would be true if we accessed the activity tracking
4556 * under the lock, but as we perform the lookup of the object and
4557 * its activity locklessly we can not guarantee that the last_write
4558 * being active implies that we have set the same engine flag from
4559 * last_read - hence we always set both read and write busy for
4562 return id | __busy_read_flag(id);
4565 static __always_inline unsigned int
4566 __busy_set_if_active(const struct dma_fence *fence,
4567 unsigned int (*flag)(unsigned int id))
4569 struct i915_request *rq;
4571 /* We have to check the current hw status of the fence as the uABI
4572 * guarantees forward progress. We could rely on the idle worker
4573 * to eventually flush us, but to minimise latency just ask the
4576 * Note we only report on the status of native fences.
4578 if (!dma_fence_is_i915(fence))
4581 /* opencode to_request() in order to avoid const warnings */
4582 rq = container_of(fence, struct i915_request, fence);
4583 if (i915_request_completed(rq))
4586 return flag(rq->engine->uabi_id);
4589 static __always_inline unsigned int
4590 busy_check_reader(const struct dma_fence *fence)
4592 return __busy_set_if_active(fence, __busy_read_flag);
4595 static __always_inline unsigned int
4596 busy_check_writer(const struct dma_fence *fence)
4601 return __busy_set_if_active(fence, __busy_write_id);
4605 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
4606 struct drm_file *file)
4608 struct drm_i915_gem_busy *args = data;
4609 struct drm_i915_gem_object *obj;
4610 struct reservation_object_list *list;
4616 obj = i915_gem_object_lookup_rcu(file, args->handle);
4620 /* A discrepancy here is that we do not report the status of
4621 * non-i915 fences, i.e. even though we may report the object as idle,
4622 * a call to set-domain may still stall waiting for foreign rendering.
4623 * This also means that wait-ioctl may report an object as busy,
4624 * where busy-ioctl considers it idle.
4626 * We trade the ability to warn of foreign fences to report on which
4627 * i915 engines are active for the object.
4629 * Alternatively, we can trade that extra information on read/write
4632 * !reservation_object_test_signaled_rcu(obj->resv, true);
4633 * to report the overall busyness. This is what the wait-ioctl does.
4637 seq = raw_read_seqcount(&obj->resv->seq);
4639 /* Translate the exclusive fence to the READ *and* WRITE engine */
4640 args->busy = busy_check_writer(rcu_dereference(obj->resv->fence_excl));
4642 /* Translate shared fences to READ set of engines */
4643 list = rcu_dereference(obj->resv->fence);
4645 unsigned int shared_count = list->shared_count, i;
4647 for (i = 0; i < shared_count; ++i) {
4648 struct dma_fence *fence =
4649 rcu_dereference(list->shared[i]);
4651 args->busy |= busy_check_reader(fence);
4655 if (args->busy && read_seqcount_retry(&obj->resv->seq, seq))
4665 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
4666 struct drm_file *file_priv)
4668 return i915_gem_ring_throttle(dev, file_priv);
4672 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
4673 struct drm_file *file_priv)
4675 struct drm_i915_private *dev_priv = to_i915(dev);
4676 struct drm_i915_gem_madvise *args = data;
4677 struct drm_i915_gem_object *obj;
4680 switch (args->madv) {
4681 case I915_MADV_DONTNEED:
4682 case I915_MADV_WILLNEED:
4688 obj = i915_gem_object_lookup(file_priv, args->handle);
4692 err = mutex_lock_interruptible(&obj->mm.lock);
4696 if (i915_gem_object_has_pages(obj) &&
4697 i915_gem_object_is_tiled(obj) &&
4698 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
4699 if (obj->mm.madv == I915_MADV_WILLNEED) {
4700 GEM_BUG_ON(!obj->mm.quirked);
4701 __i915_gem_object_unpin_pages(obj);
4702 obj->mm.quirked = false;
4704 if (args->madv == I915_MADV_WILLNEED) {
4705 GEM_BUG_ON(obj->mm.quirked);
4706 __i915_gem_object_pin_pages(obj);
4707 obj->mm.quirked = true;
4711 if (obj->mm.madv != __I915_MADV_PURGED)
4712 obj->mm.madv = args->madv;
4714 /* if the object is no longer attached, discard its backing storage */
4715 if (obj->mm.madv == I915_MADV_DONTNEED &&
4716 !i915_gem_object_has_pages(obj))
4717 i915_gem_object_truncate(obj);
4719 args->retained = obj->mm.madv != __I915_MADV_PURGED;
4720 mutex_unlock(&obj->mm.lock);
4723 i915_gem_object_put(obj);
4728 frontbuffer_retire(struct i915_gem_active *active, struct i915_request *request)
4730 struct drm_i915_gem_object *obj =
4731 container_of(active, typeof(*obj), frontbuffer_write);
4733 intel_fb_obj_flush(obj, ORIGIN_CS);
4736 void i915_gem_object_init(struct drm_i915_gem_object *obj,
4737 const struct drm_i915_gem_object_ops *ops)
4739 mutex_init(&obj->mm.lock);
4741 INIT_LIST_HEAD(&obj->vma_list);
4742 INIT_LIST_HEAD(&obj->lut_list);
4743 INIT_LIST_HEAD(&obj->batch_pool_link);
4745 init_rcu_head(&obj->rcu);
4749 reservation_object_init(&obj->__builtin_resv);
4750 obj->resv = &obj->__builtin_resv;
4752 obj->frontbuffer_ggtt_origin = ORIGIN_GTT;
4753 init_request_active(&obj->frontbuffer_write, frontbuffer_retire);
4755 obj->mm.madv = I915_MADV_WILLNEED;
4756 INIT_RADIX_TREE(&obj->mm.get_page.radix, GFP_KERNEL | __GFP_NOWARN);
4757 mutex_init(&obj->mm.get_page.lock);
4759 i915_gem_info_add_obj(to_i915(obj->base.dev), obj->base.size);
4762 static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
4763 .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE |
4764 I915_GEM_OBJECT_IS_SHRINKABLE,
4766 .get_pages = i915_gem_object_get_pages_gtt,
4767 .put_pages = i915_gem_object_put_pages_gtt,
4769 .pwrite = i915_gem_object_pwrite_gtt,
4772 static int i915_gem_object_create_shmem(struct drm_device *dev,
4773 struct drm_gem_object *obj,
4776 struct drm_i915_private *i915 = to_i915(dev);
4777 unsigned long flags = VM_NORESERVE;
4780 drm_gem_private_object_init(dev, obj, size);
4783 filp = shmem_file_setup_with_mnt(i915->mm.gemfs, "i915", size,
4786 filp = shmem_file_setup("i915", size, flags);
4789 return PTR_ERR(filp);
4796 struct drm_i915_gem_object *
4797 i915_gem_object_create(struct drm_i915_private *dev_priv, u64 size)
4799 struct drm_i915_gem_object *obj;
4800 struct address_space *mapping;
4801 unsigned int cache_level;
4805 /* There is a prevalence of the assumption that we fit the object's
4806 * page count inside a 32bit _signed_ variable. Let's document this and
4807 * catch if we ever need to fix it. In the meantime, if you do spot
4808 * such a local variable, please consider fixing!
4810 if (size >> PAGE_SHIFT > INT_MAX)
4811 return ERR_PTR(-E2BIG);
4813 if (overflows_type(size, obj->base.size))
4814 return ERR_PTR(-E2BIG);
4816 obj = i915_gem_object_alloc(dev_priv);
4818 return ERR_PTR(-ENOMEM);
4820 ret = i915_gem_object_create_shmem(&dev_priv->drm, &obj->base, size);
4824 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
4825 if (IS_I965GM(dev_priv) || IS_I965G(dev_priv)) {
4826 /* 965gm cannot relocate objects above 4GiB. */
4827 mask &= ~__GFP_HIGHMEM;
4828 mask |= __GFP_DMA32;
4831 mapping = obj->base.filp->f_mapping;
4832 mapping_set_gfp_mask(mapping, mask);
4833 GEM_BUG_ON(!(mapping_gfp_mask(mapping) & __GFP_RECLAIM));
4835 i915_gem_object_init(obj, &i915_gem_object_ops);
4837 obj->write_domain = I915_GEM_DOMAIN_CPU;
4838 obj->read_domains = I915_GEM_DOMAIN_CPU;
4840 if (HAS_LLC(dev_priv))
4841 /* On some devices, we can have the GPU use the LLC (the CPU
4842 * cache) for about a 10% performance improvement
4843 * compared to uncached. Graphics requests other than
4844 * display scanout are coherent with the CPU in
4845 * accessing this cache. This means in this mode we
4846 * don't need to clflush on the CPU side, and on the
4847 * GPU side we only need to flush internal caches to
4848 * get data visible to the CPU.
4850 * However, we maintain the display planes as UC, and so
4851 * need to rebind when first used as such.
4853 cache_level = I915_CACHE_LLC;
4855 cache_level = I915_CACHE_NONE;
4857 i915_gem_object_set_cache_coherency(obj, cache_level);
4859 trace_i915_gem_object_create(obj);
4864 i915_gem_object_free(obj);
4865 return ERR_PTR(ret);
4868 static bool discard_backing_storage(struct drm_i915_gem_object *obj)
4870 /* If we are the last user of the backing storage (be it shmemfs
4871 * pages or stolen etc), we know that the pages are going to be
4872 * immediately released. In this case, we can then skip copying
4873 * back the contents from the GPU.
4876 if (obj->mm.madv != I915_MADV_WILLNEED)
4879 if (obj->base.filp == NULL)
4882 /* At first glance, this looks racy, but then again so would be
4883 * userspace racing mmap against close. However, the first external
4884 * reference to the filp can only be obtained through the
4885 * i915_gem_mmap_ioctl() which safeguards us against the user
4886 * acquiring such a reference whilst we are in the middle of
4887 * freeing the object.
4889 return atomic_long_read(&obj->base.filp->f_count) == 1;
4892 static void __i915_gem_free_objects(struct drm_i915_private *i915,
4893 struct llist_node *freed)
4895 struct drm_i915_gem_object *obj, *on;
4897 intel_runtime_pm_get(i915);
4898 llist_for_each_entry_safe(obj, on, freed, freed) {
4899 struct i915_vma *vma, *vn;
4901 trace_i915_gem_object_destroy(obj);
4903 mutex_lock(&i915->drm.struct_mutex);
4905 GEM_BUG_ON(i915_gem_object_is_active(obj));
4906 list_for_each_entry_safe(vma, vn,
4907 &obj->vma_list, obj_link) {
4908 GEM_BUG_ON(i915_vma_is_active(vma));
4909 vma->flags &= ~I915_VMA_PIN_MASK;
4910 i915_vma_destroy(vma);
4912 GEM_BUG_ON(!list_empty(&obj->vma_list));
4913 GEM_BUG_ON(!RB_EMPTY_ROOT(&obj->vma_tree));
4915 /* This serializes freeing with the shrinker. Since the free
4916 * is delayed, first by RCU then by the workqueue, we want the
4917 * shrinker to be able to free pages of unreferenced objects,
4918 * or else we may oom whilst there are plenty of deferred
4921 if (i915_gem_object_has_pages(obj)) {
4922 spin_lock(&i915->mm.obj_lock);
4923 list_del_init(&obj->mm.link);
4924 spin_unlock(&i915->mm.obj_lock);
4927 mutex_unlock(&i915->drm.struct_mutex);
4929 GEM_BUG_ON(obj->bind_count);
4930 GEM_BUG_ON(obj->userfault_count);
4931 GEM_BUG_ON(atomic_read(&obj->frontbuffer_bits));
4932 GEM_BUG_ON(!list_empty(&obj->lut_list));
4934 if (obj->ops->release)
4935 obj->ops->release(obj);
4937 if (WARN_ON(i915_gem_object_has_pinned_pages(obj)))
4938 atomic_set(&obj->mm.pages_pin_count, 0);
4939 __i915_gem_object_put_pages(obj, I915_MM_NORMAL);
4940 GEM_BUG_ON(i915_gem_object_has_pages(obj));
4942 if (obj->base.import_attach)
4943 drm_prime_gem_destroy(&obj->base, NULL);
4945 reservation_object_fini(&obj->__builtin_resv);
4946 drm_gem_object_release(&obj->base);
4947 i915_gem_info_remove_obj(i915, obj->base.size);
4950 i915_gem_object_free(obj);
4952 GEM_BUG_ON(!atomic_read(&i915->mm.free_count));
4953 atomic_dec(&i915->mm.free_count);
4958 intel_runtime_pm_put(i915);
4961 static void i915_gem_flush_free_objects(struct drm_i915_private *i915)
4963 struct llist_node *freed;
4965 /* Free the oldest, most stale object to keep the free_list short */
4967 if (!llist_empty(&i915->mm.free_list)) { /* quick test for hotpath */
4968 /* Only one consumer of llist_del_first() allowed */
4969 spin_lock(&i915->mm.free_lock);
4970 freed = llist_del_first(&i915->mm.free_list);
4971 spin_unlock(&i915->mm.free_lock);
4973 if (unlikely(freed)) {
4975 __i915_gem_free_objects(i915, freed);
4979 static void __i915_gem_free_work(struct work_struct *work)
4981 struct drm_i915_private *i915 =
4982 container_of(work, struct drm_i915_private, mm.free_work);
4983 struct llist_node *freed;
4986 * All file-owned VMA should have been released by this point through
4987 * i915_gem_close_object(), or earlier by i915_gem_context_close().
4988 * However, the object may also be bound into the global GTT (e.g.
4989 * older GPUs without per-process support, or for direct access through
4990 * the GTT either for the user or for scanout). Those VMA still need to
4994 spin_lock(&i915->mm.free_lock);
4995 while ((freed = llist_del_all(&i915->mm.free_list))) {
4996 spin_unlock(&i915->mm.free_lock);
4998 __i915_gem_free_objects(i915, freed);
5002 spin_lock(&i915->mm.free_lock);
5004 spin_unlock(&i915->mm.free_lock);
5007 static void __i915_gem_free_object_rcu(struct rcu_head *head)
5009 struct drm_i915_gem_object *obj =
5010 container_of(head, typeof(*obj), rcu);
5011 struct drm_i915_private *i915 = to_i915(obj->base.dev);
5014 * We reuse obj->rcu for the freed list, so we had better not treat
5015 * it like a rcu_head from this point forwards. And we expect all
5016 * objects to be freed via this path.
5018 destroy_rcu_head(&obj->rcu);
5021 * Since we require blocking on struct_mutex to unbind the freed
5022 * object from the GPU before releasing resources back to the
5023 * system, we can not do that directly from the RCU callback (which may
5024 * be a softirq context), but must instead then defer that work onto a
5025 * kthread. We use the RCU callback rather than move the freed object
5026 * directly onto the work queue so that we can mix between using the
5027 * worker and performing frees directly from subsequent allocations for
5028 * crude but effective memory throttling.
5030 if (llist_add(&obj->freed, &i915->mm.free_list))
5031 queue_work(i915->wq, &i915->mm.free_work);
5034 void i915_gem_free_object(struct drm_gem_object *gem_obj)
5036 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
5038 if (obj->mm.quirked)
5039 __i915_gem_object_unpin_pages(obj);
5041 if (discard_backing_storage(obj))
5042 obj->mm.madv = I915_MADV_DONTNEED;
5045 * Before we free the object, make sure any pure RCU-only
5046 * read-side critical sections are complete, e.g.
5047 * i915_gem_busy_ioctl(). For the corresponding synchronized
5048 * lookup see i915_gem_object_lookup_rcu().
5050 atomic_inc(&to_i915(obj->base.dev)->mm.free_count);
5051 call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
5054 void __i915_gem_object_release_unless_active(struct drm_i915_gem_object *obj)
5056 lockdep_assert_held(&obj->base.dev->struct_mutex);
5058 if (!i915_gem_object_has_active_reference(obj) &&
5059 i915_gem_object_is_active(obj))
5060 i915_gem_object_set_active_reference(obj);
5062 i915_gem_object_put(obj);
5065 void i915_gem_sanitize(struct drm_i915_private *i915)
5071 mutex_lock(&i915->drm.struct_mutex);
5073 intel_runtime_pm_get(i915);
5074 intel_uncore_forcewake_get(i915, FORCEWAKE_ALL);
5077 * As we have just resumed the machine and woken the device up from
5078 * deep PCI sleep (presumably D3_cold), assume the HW has been reset
5079 * back to defaults, recovering from whatever wedged state we left it
5080 * in and so worth trying to use the device once more.
5082 if (i915_terminally_wedged(&i915->gpu_error))
5083 i915_gem_unset_wedged(i915);
5086 * If we inherit context state from the BIOS or earlier occupants
5087 * of the GPU, the GPU may be in an inconsistent state when we
5088 * try to take over. The only way to remove the earlier state
5089 * is by resetting. However, resetting on earlier gen is tricky as
5090 * it may impact the display and we are uncertain about the stability
5091 * of the reset, so this could be applied to even earlier gen.
5094 if (INTEL_GEN(i915) >= 5 && intel_has_gpu_reset(i915))
5095 err = WARN_ON(intel_gpu_reset(i915, ALL_ENGINES));
5097 intel_engines_sanitize(i915);
5099 intel_uncore_forcewake_put(i915, FORCEWAKE_ALL);
5100 intel_runtime_pm_put(i915);
5102 i915_gem_contexts_lost(i915);
5103 mutex_unlock(&i915->drm.struct_mutex);
5106 int i915_gem_suspend(struct drm_i915_private *i915)
5112 intel_runtime_pm_get(i915);
5113 intel_suspend_gt_powersave(i915);
5115 mutex_lock(&i915->drm.struct_mutex);
5118 * We have to flush all the executing contexts to main memory so
5119 * that they can saved in the hibernation image. To ensure the last
5120 * context image is coherent, we have to switch away from it. That
5121 * leaves the i915->kernel_context still active when
5122 * we actually suspend, and its image in memory may not match the GPU
5123 * state. Fortunately, the kernel_context is disposable and we do
5124 * not rely on its state.
5126 if (!i915_terminally_wedged(&i915->gpu_error)) {
5127 ret = i915_gem_switch_to_kernel_context(i915);
5131 ret = i915_gem_wait_for_idle(i915,
5132 I915_WAIT_INTERRUPTIBLE |
5134 I915_WAIT_FOR_IDLE_BOOST,
5135 MAX_SCHEDULE_TIMEOUT);
5136 if (ret && ret != -EIO)
5139 assert_kernel_context_is_current(i915);
5141 i915_retire_requests(i915); /* ensure we flush after wedging */
5143 mutex_unlock(&i915->drm.struct_mutex);
5145 intel_uc_suspend(i915);
5147 cancel_delayed_work_sync(&i915->gpu_error.hangcheck_work);
5148 cancel_delayed_work_sync(&i915->gt.retire_work);
5151 * As the idle_work is rearming if it detects a race, play safe and
5152 * repeat the flush until it is definitely idle.
5154 drain_delayed_work(&i915->gt.idle_work);
5157 * Assert that we successfully flushed all the work and
5158 * reset the GPU back to its idle, low power state.
5160 WARN_ON(i915->gt.awake);
5161 if (WARN_ON(!intel_engines_are_idle(i915)))
5162 i915_gem_set_wedged(i915); /* no hope, discard everything */
5164 intel_runtime_pm_put(i915);
5168 mutex_unlock(&i915->drm.struct_mutex);
5169 intel_runtime_pm_put(i915);
5173 void i915_gem_suspend_late(struct drm_i915_private *i915)
5175 struct drm_i915_gem_object *obj;
5176 struct list_head *phases[] = {
5177 &i915->mm.unbound_list,
5178 &i915->mm.bound_list,
5183 * Neither the BIOS, ourselves or any other kernel
5184 * expects the system to be in execlists mode on startup,
5185 * so we need to reset the GPU back to legacy mode. And the only
5186 * known way to disable logical contexts is through a GPU reset.
5188 * So in order to leave the system in a known default configuration,
5189 * always reset the GPU upon unload and suspend. Afterwards we then
5190 * clean up the GEM state tracking, flushing off the requests and
5191 * leaving the system in a known idle state.
5193 * Note that is of the upmost importance that the GPU is idle and
5194 * all stray writes are flushed *before* we dismantle the backing
5195 * storage for the pinned objects.
5197 * However, since we are uncertain that resetting the GPU on older
5198 * machines is a good idea, we don't - just in case it leaves the
5199 * machine in an unusable condition.
5202 mutex_lock(&i915->drm.struct_mutex);
5203 for (phase = phases; *phase; phase++) {
5204 list_for_each_entry(obj, *phase, mm.link)
5205 WARN_ON(i915_gem_object_set_to_gtt_domain(obj, false));
5207 mutex_unlock(&i915->drm.struct_mutex);
5209 intel_uc_sanitize(i915);
5210 i915_gem_sanitize(i915);
5213 void i915_gem_resume(struct drm_i915_private *i915)
5217 WARN_ON(i915->gt.awake);
5219 mutex_lock(&i915->drm.struct_mutex);
5220 intel_uncore_forcewake_get(i915, FORCEWAKE_ALL);
5222 i915_gem_restore_gtt_mappings(i915);
5223 i915_gem_restore_fences(i915);
5226 * As we didn't flush the kernel context before suspend, we cannot
5227 * guarantee that the context image is complete. So let's just reset
5228 * it and start again.
5230 i915->gt.resume(i915);
5232 if (i915_gem_init_hw(i915))
5235 intel_uc_resume(i915);
5237 /* Always reload a context for powersaving. */
5238 if (i915_gem_switch_to_kernel_context(i915))
5242 intel_uncore_forcewake_put(i915, FORCEWAKE_ALL);
5243 mutex_unlock(&i915->drm.struct_mutex);
5247 if (!i915_terminally_wedged(&i915->gpu_error)) {
5248 DRM_ERROR("failed to re-initialize GPU, declaring wedged!\n");
5249 i915_gem_set_wedged(i915);
5254 void i915_gem_init_swizzling(struct drm_i915_private *dev_priv)
5256 if (INTEL_GEN(dev_priv) < 5 ||
5257 dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
5260 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
5261 DISP_TILE_SURFACE_SWIZZLING);
5263 if (IS_GEN5(dev_priv))
5266 I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
5267 if (IS_GEN6(dev_priv))
5268 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
5269 else if (IS_GEN7(dev_priv))
5270 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
5271 else if (IS_GEN8(dev_priv))
5272 I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW));
5277 static void init_unused_ring(struct drm_i915_private *dev_priv, u32 base)
5279 I915_WRITE(RING_CTL(base), 0);
5280 I915_WRITE(RING_HEAD(base), 0);
5281 I915_WRITE(RING_TAIL(base), 0);
5282 I915_WRITE(RING_START(base), 0);
5285 static void init_unused_rings(struct drm_i915_private *dev_priv)
5287 if (IS_I830(dev_priv)) {
5288 init_unused_ring(dev_priv, PRB1_BASE);
5289 init_unused_ring(dev_priv, SRB0_BASE);
5290 init_unused_ring(dev_priv, SRB1_BASE);
5291 init_unused_ring(dev_priv, SRB2_BASE);
5292 init_unused_ring(dev_priv, SRB3_BASE);
5293 } else if (IS_GEN2(dev_priv)) {
5294 init_unused_ring(dev_priv, SRB0_BASE);
5295 init_unused_ring(dev_priv, SRB1_BASE);
5296 } else if (IS_GEN3(dev_priv)) {
5297 init_unused_ring(dev_priv, PRB1_BASE);
5298 init_unused_ring(dev_priv, PRB2_BASE);
5302 static int __i915_gem_restart_engines(void *data)
5304 struct drm_i915_private *i915 = data;
5305 struct intel_engine_cs *engine;
5306 enum intel_engine_id id;
5309 for_each_engine(engine, i915, id) {
5310 err = engine->init_hw(engine);
5312 DRM_ERROR("Failed to restart %s (%d)\n",
5321 int i915_gem_init_hw(struct drm_i915_private *dev_priv)
5325 dev_priv->gt.last_init_time = ktime_get();
5327 /* Double layer security blanket, see i915_gem_init() */
5328 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5330 if (HAS_EDRAM(dev_priv) && INTEL_GEN(dev_priv) < 9)
5331 I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf));
5333 if (IS_HASWELL(dev_priv))
5334 I915_WRITE(MI_PREDICATE_RESULT_2, IS_HSW_GT3(dev_priv) ?
5335 LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED);
5337 intel_gt_workarounds_apply(dev_priv);
5339 i915_gem_init_swizzling(dev_priv);
5342 * At least 830 can leave some of the unused rings
5343 * "active" (ie. head != tail) after resume which
5344 * will prevent c3 entry. Makes sure all unused rings
5347 init_unused_rings(dev_priv);
5349 BUG_ON(!dev_priv->kernel_context);
5350 if (i915_terminally_wedged(&dev_priv->gpu_error)) {
5355 ret = i915_ppgtt_init_hw(dev_priv);
5357 DRM_ERROR("Enabling PPGTT failed (%d)\n", ret);
5361 ret = intel_wopcm_init_hw(&dev_priv->wopcm);
5363 DRM_ERROR("Enabling WOPCM failed (%d)\n", ret);
5367 /* We can't enable contexts until all firmware is loaded */
5368 ret = intel_uc_init_hw(dev_priv);
5370 DRM_ERROR("Enabling uc failed (%d)\n", ret);
5374 intel_mocs_init_l3cc_table(dev_priv);
5376 /* Only when the HW is re-initialised, can we replay the requests */
5377 ret = __i915_gem_restart_engines(dev_priv);
5381 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5386 intel_uc_fini_hw(dev_priv);
5388 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5393 static int __intel_engines_record_defaults(struct drm_i915_private *i915)
5395 struct i915_gem_context *ctx;
5396 struct intel_engine_cs *engine;
5397 enum intel_engine_id id;
5401 * As we reset the gpu during very early sanitisation, the current
5402 * register state on the GPU should reflect its defaults values.
5403 * We load a context onto the hw (with restore-inhibit), then switch
5404 * over to a second context to save that default register state. We
5405 * can then prime every new context with that state so they all start
5406 * from the same default HW values.
5409 ctx = i915_gem_context_create_kernel(i915, 0);
5411 return PTR_ERR(ctx);
5413 for_each_engine(engine, i915, id) {
5414 struct i915_request *rq;
5416 rq = i915_request_alloc(engine, ctx);
5423 if (engine->init_context)
5424 err = engine->init_context(rq);
5426 i915_request_add(rq);
5431 err = i915_gem_switch_to_kernel_context(i915);
5435 if (i915_gem_wait_for_idle(i915, I915_WAIT_LOCKED, HZ / 5)) {
5436 i915_gem_set_wedged(i915);
5437 err = -EIO; /* Caller will declare us wedged */
5441 assert_kernel_context_is_current(i915);
5444 * Immediately park the GPU so that we enable powersaving and
5445 * treat it as idle. The next time we issue a request, we will
5446 * unpark and start using the engine->pinned_default_state, otherwise
5447 * it is in limbo and an early reset may fail.
5449 __i915_gem_park(i915);
5451 for_each_engine(engine, i915, id) {
5452 struct i915_vma *state;
5455 GEM_BUG_ON(to_intel_context(ctx, engine)->pin_count);
5457 state = to_intel_context(ctx, engine)->state;
5462 * As we will hold a reference to the logical state, it will
5463 * not be torn down with the context, and importantly the
5464 * object will hold onto its vma (making it possible for a
5465 * stray GTT write to corrupt our defaults). Unmap the vma
5466 * from the GTT to prevent such accidents and reclaim the
5469 err = i915_vma_unbind(state);
5473 err = i915_gem_object_set_to_cpu_domain(state->obj, false);
5477 engine->default_state = i915_gem_object_get(state->obj);
5479 /* Check we can acquire the image of the context state */
5480 vaddr = i915_gem_object_pin_map(engine->default_state,
5482 if (IS_ERR(vaddr)) {
5483 err = PTR_ERR(vaddr);
5487 i915_gem_object_unpin_map(engine->default_state);
5490 if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)) {
5491 unsigned int found = intel_engines_has_context_isolation(i915);
5494 * Make sure that classes with multiple engine instances all
5495 * share the same basic configuration.
5497 for_each_engine(engine, i915, id) {
5498 unsigned int bit = BIT(engine->uabi_class);
5499 unsigned int expected = engine->default_state ? bit : 0;
5501 if ((found & bit) != expected) {
5502 DRM_ERROR("mismatching default context state for class %d on engine %s\n",
5503 engine->uabi_class, engine->name);
5509 i915_gem_context_set_closed(ctx);
5510 i915_gem_context_put(ctx);
5515 * If we have to abandon now, we expect the engines to be idle
5516 * and ready to be torn-down. First try to flush any remaining
5517 * request, ensure we are pointing at the kernel context and
5520 if (WARN_ON(i915_gem_switch_to_kernel_context(i915)))
5523 if (WARN_ON(i915_gem_wait_for_idle(i915,
5525 MAX_SCHEDULE_TIMEOUT)))
5528 i915_gem_contexts_lost(i915);
5532 int i915_gem_init(struct drm_i915_private *dev_priv)
5536 /* We need to fallback to 4K pages if host doesn't support huge gtt. */
5537 if (intel_vgpu_active(dev_priv) && !intel_vgpu_has_huge_gtt(dev_priv))
5538 mkwrite_device_info(dev_priv)->page_sizes =
5539 I915_GTT_PAGE_SIZE_4K;
5541 dev_priv->mm.unordered_timeline = dma_fence_context_alloc(1);
5543 if (HAS_LOGICAL_RING_CONTEXTS(dev_priv)) {
5544 dev_priv->gt.resume = intel_lr_context_resume;
5545 dev_priv->gt.cleanup_engine = intel_logical_ring_cleanup;
5547 dev_priv->gt.resume = intel_legacy_submission_resume;
5548 dev_priv->gt.cleanup_engine = intel_engine_cleanup;
5551 ret = i915_gem_init_userptr(dev_priv);
5555 ret = intel_uc_init_misc(dev_priv);
5559 ret = intel_wopcm_init(&dev_priv->wopcm);
5563 /* This is just a security blanket to placate dragons.
5564 * On some systems, we very sporadically observe that the first TLBs
5565 * used by the CS may be stale, despite us poking the TLB reset. If
5566 * we hold the forcewake during initialisation these problems
5567 * just magically go away.
5569 mutex_lock(&dev_priv->drm.struct_mutex);
5570 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5572 ret = i915_gem_init_ggtt(dev_priv);
5574 GEM_BUG_ON(ret == -EIO);
5578 ret = i915_gem_contexts_init(dev_priv);
5580 GEM_BUG_ON(ret == -EIO);
5584 ret = intel_engines_init(dev_priv);
5586 GEM_BUG_ON(ret == -EIO);
5590 intel_init_gt_powersave(dev_priv);
5592 ret = intel_uc_init(dev_priv);
5596 ret = i915_gem_init_hw(dev_priv);
5601 * Despite its name intel_init_clock_gating applies both display
5602 * clock gating workarounds; GT mmio workarounds and the occasional
5603 * GT power context workaround. Worse, sometimes it includes a context
5604 * register workaround which we need to apply before we record the
5605 * default HW state for all contexts.
5607 * FIXME: break up the workarounds and apply them at the right time!
5609 intel_init_clock_gating(dev_priv);
5611 ret = __intel_engines_record_defaults(dev_priv);
5615 if (i915_inject_load_failure()) {
5620 if (i915_inject_load_failure()) {
5625 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5626 mutex_unlock(&dev_priv->drm.struct_mutex);
5631 * Unwinding is complicated by that we want to handle -EIO to mean
5632 * disable GPU submission but keep KMS alive. We want to mark the
5633 * HW as irrevisibly wedged, but keep enough state around that the
5634 * driver doesn't explode during runtime.
5637 mutex_unlock(&dev_priv->drm.struct_mutex);
5639 WARN_ON(i915_gem_suspend(dev_priv));
5640 i915_gem_suspend_late(dev_priv);
5642 i915_gem_drain_workqueue(dev_priv);
5644 mutex_lock(&dev_priv->drm.struct_mutex);
5645 intel_uc_fini_hw(dev_priv);
5647 intel_uc_fini(dev_priv);
5650 intel_cleanup_gt_powersave(dev_priv);
5651 i915_gem_cleanup_engines(dev_priv);
5655 i915_gem_contexts_fini(dev_priv);
5658 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5659 mutex_unlock(&dev_priv->drm.struct_mutex);
5662 intel_uc_fini_misc(dev_priv);
5665 i915_gem_cleanup_userptr(dev_priv);
5668 mutex_lock(&dev_priv->drm.struct_mutex);
5671 * Allow engine initialisation to fail by marking the GPU as
5672 * wedged. But we only want to do this where the GPU is angry,
5673 * for all other failure, such as an allocation failure, bail.
5675 if (!i915_terminally_wedged(&dev_priv->gpu_error)) {
5676 i915_load_error(dev_priv,
5677 "Failed to initialize GPU, declaring it wedged!\n");
5678 i915_gem_set_wedged(dev_priv);
5681 /* Minimal basic recovery for KMS */
5682 ret = i915_ggtt_enable_hw(dev_priv);
5683 i915_gem_restore_gtt_mappings(dev_priv);
5684 i915_gem_restore_fences(dev_priv);
5685 intel_init_clock_gating(dev_priv);
5687 mutex_unlock(&dev_priv->drm.struct_mutex);
5690 i915_gem_drain_freed_objects(dev_priv);
5694 void i915_gem_fini(struct drm_i915_private *dev_priv)
5696 i915_gem_suspend_late(dev_priv);
5697 intel_disable_gt_powersave(dev_priv);
5699 /* Flush any outstanding unpin_work. */
5700 i915_gem_drain_workqueue(dev_priv);
5702 mutex_lock(&dev_priv->drm.struct_mutex);
5703 intel_uc_fini_hw(dev_priv);
5704 intel_uc_fini(dev_priv);
5705 i915_gem_cleanup_engines(dev_priv);
5706 i915_gem_contexts_fini(dev_priv);
5707 mutex_unlock(&dev_priv->drm.struct_mutex);
5709 intel_cleanup_gt_powersave(dev_priv);
5711 intel_uc_fini_misc(dev_priv);
5712 i915_gem_cleanup_userptr(dev_priv);
5714 i915_gem_drain_freed_objects(dev_priv);
5716 WARN_ON(!list_empty(&dev_priv->contexts.list));
5719 void i915_gem_init_mmio(struct drm_i915_private *i915)
5721 i915_gem_sanitize(i915);
5725 i915_gem_cleanup_engines(struct drm_i915_private *dev_priv)
5727 struct intel_engine_cs *engine;
5728 enum intel_engine_id id;
5730 for_each_engine(engine, dev_priv, id)
5731 dev_priv->gt.cleanup_engine(engine);
5735 i915_gem_load_init_fences(struct drm_i915_private *dev_priv)
5739 if (INTEL_GEN(dev_priv) >= 7 && !IS_VALLEYVIEW(dev_priv) &&
5740 !IS_CHERRYVIEW(dev_priv))
5741 dev_priv->num_fence_regs = 32;
5742 else if (INTEL_GEN(dev_priv) >= 4 ||
5743 IS_I945G(dev_priv) || IS_I945GM(dev_priv) ||
5744 IS_G33(dev_priv) || IS_PINEVIEW(dev_priv))
5745 dev_priv->num_fence_regs = 16;
5747 dev_priv->num_fence_regs = 8;
5749 if (intel_vgpu_active(dev_priv))
5750 dev_priv->num_fence_regs =
5751 I915_READ(vgtif_reg(avail_rs.fence_num));
5753 /* Initialize fence registers to zero */
5754 for (i = 0; i < dev_priv->num_fence_regs; i++) {
5755 struct drm_i915_fence_reg *fence = &dev_priv->fence_regs[i];
5757 fence->i915 = dev_priv;
5759 list_add_tail(&fence->link, &dev_priv->mm.fence_list);
5761 i915_gem_restore_fences(dev_priv);
5763 i915_gem_detect_bit_6_swizzle(dev_priv);
5766 static void i915_gem_init__mm(struct drm_i915_private *i915)
5768 spin_lock_init(&i915->mm.object_stat_lock);
5769 spin_lock_init(&i915->mm.obj_lock);
5770 spin_lock_init(&i915->mm.free_lock);
5772 init_llist_head(&i915->mm.free_list);
5774 INIT_LIST_HEAD(&i915->mm.unbound_list);
5775 INIT_LIST_HEAD(&i915->mm.bound_list);
5776 INIT_LIST_HEAD(&i915->mm.fence_list);
5777 INIT_LIST_HEAD(&i915->mm.userfault_list);
5779 INIT_WORK(&i915->mm.free_work, __i915_gem_free_work);
5782 int i915_gem_init_early(struct drm_i915_private *dev_priv)
5786 dev_priv->objects = KMEM_CACHE(drm_i915_gem_object, SLAB_HWCACHE_ALIGN);
5787 if (!dev_priv->objects)
5790 dev_priv->vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
5791 if (!dev_priv->vmas)
5794 dev_priv->luts = KMEM_CACHE(i915_lut_handle, 0);
5795 if (!dev_priv->luts)
5798 dev_priv->requests = KMEM_CACHE(i915_request,
5799 SLAB_HWCACHE_ALIGN |
5800 SLAB_RECLAIM_ACCOUNT |
5801 SLAB_TYPESAFE_BY_RCU);
5802 if (!dev_priv->requests)
5805 dev_priv->dependencies = KMEM_CACHE(i915_dependency,
5806 SLAB_HWCACHE_ALIGN |
5807 SLAB_RECLAIM_ACCOUNT);
5808 if (!dev_priv->dependencies)
5811 dev_priv->priorities = KMEM_CACHE(i915_priolist, SLAB_HWCACHE_ALIGN);
5812 if (!dev_priv->priorities)
5813 goto err_dependencies;
5815 INIT_LIST_HEAD(&dev_priv->gt.timelines);
5816 INIT_LIST_HEAD(&dev_priv->gt.active_rings);
5817 INIT_LIST_HEAD(&dev_priv->gt.closed_vma);
5819 i915_gem_init__mm(dev_priv);
5821 INIT_DELAYED_WORK(&dev_priv->gt.retire_work,
5822 i915_gem_retire_work_handler);
5823 INIT_DELAYED_WORK(&dev_priv->gt.idle_work,
5824 i915_gem_idle_work_handler);
5825 init_waitqueue_head(&dev_priv->gpu_error.wait_queue);
5826 init_waitqueue_head(&dev_priv->gpu_error.reset_queue);
5828 atomic_set(&dev_priv->mm.bsd_engine_dispatch_index, 0);
5830 spin_lock_init(&dev_priv->fb_tracking.lock);
5832 err = i915_gemfs_init(dev_priv);
5834 DRM_NOTE("Unable to create a private tmpfs mount, hugepage support will be disabled(%d).\n", err);
5839 kmem_cache_destroy(dev_priv->dependencies);
5841 kmem_cache_destroy(dev_priv->requests);
5843 kmem_cache_destroy(dev_priv->luts);
5845 kmem_cache_destroy(dev_priv->vmas);
5847 kmem_cache_destroy(dev_priv->objects);
5852 void i915_gem_cleanup_early(struct drm_i915_private *dev_priv)
5854 i915_gem_drain_freed_objects(dev_priv);
5855 GEM_BUG_ON(!llist_empty(&dev_priv->mm.free_list));
5856 GEM_BUG_ON(atomic_read(&dev_priv->mm.free_count));
5857 WARN_ON(dev_priv->mm.object_count);
5858 WARN_ON(!list_empty(&dev_priv->gt.timelines));
5860 kmem_cache_destroy(dev_priv->priorities);
5861 kmem_cache_destroy(dev_priv->dependencies);
5862 kmem_cache_destroy(dev_priv->requests);
5863 kmem_cache_destroy(dev_priv->luts);
5864 kmem_cache_destroy(dev_priv->vmas);
5865 kmem_cache_destroy(dev_priv->objects);
5867 /* And ensure that our DESTROY_BY_RCU slabs are truly destroyed */
5870 i915_gemfs_fini(dev_priv);
5873 int i915_gem_freeze(struct drm_i915_private *dev_priv)
5875 /* Discard all purgeable objects, let userspace recover those as
5876 * required after resuming.
5878 i915_gem_shrink_all(dev_priv);
5883 int i915_gem_freeze_late(struct drm_i915_private *i915)
5885 struct drm_i915_gem_object *obj;
5886 struct list_head *phases[] = {
5887 &i915->mm.unbound_list,
5888 &i915->mm.bound_list,
5893 * Called just before we write the hibernation image.
5895 * We need to update the domain tracking to reflect that the CPU
5896 * will be accessing all the pages to create and restore from the
5897 * hibernation, and so upon restoration those pages will be in the
5900 * To make sure the hibernation image contains the latest state,
5901 * we update that state just before writing out the image.
5903 * To try and reduce the hibernation image, we manually shrink
5904 * the objects as well, see i915_gem_freeze()
5907 i915_gem_shrink(i915, -1UL, NULL, I915_SHRINK_UNBOUND);
5908 i915_gem_drain_freed_objects(i915);
5910 mutex_lock(&i915->drm.struct_mutex);
5911 for (phase = phases; *phase; phase++) {
5912 list_for_each_entry(obj, *phase, mm.link)
5913 WARN_ON(i915_gem_object_set_to_cpu_domain(obj, true));
5915 mutex_unlock(&i915->drm.struct_mutex);
5920 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
5922 struct drm_i915_file_private *file_priv = file->driver_priv;
5923 struct i915_request *request;
5925 /* Clean up our request list when the client is going away, so that
5926 * later retire_requests won't dereference our soon-to-be-gone
5929 spin_lock(&file_priv->mm.lock);
5930 list_for_each_entry(request, &file_priv->mm.request_list, client_link)
5931 request->file_priv = NULL;
5932 spin_unlock(&file_priv->mm.lock);
5935 int i915_gem_open(struct drm_i915_private *i915, struct drm_file *file)
5937 struct drm_i915_file_private *file_priv;
5942 file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL);
5946 file->driver_priv = file_priv;
5947 file_priv->dev_priv = i915;
5948 file_priv->file = file;
5950 spin_lock_init(&file_priv->mm.lock);
5951 INIT_LIST_HEAD(&file_priv->mm.request_list);
5953 file_priv->bsd_engine = -1;
5954 file_priv->hang_timestamp = jiffies;
5956 ret = i915_gem_context_open(i915, file);
5964 * i915_gem_track_fb - update frontbuffer tracking
5965 * @old: current GEM buffer for the frontbuffer slots
5966 * @new: new GEM buffer for the frontbuffer slots
5967 * @frontbuffer_bits: bitmask of frontbuffer slots
5969 * This updates the frontbuffer tracking bits @frontbuffer_bits by clearing them
5970 * from @old and setting them in @new. Both @old and @new can be NULL.
5972 void i915_gem_track_fb(struct drm_i915_gem_object *old,
5973 struct drm_i915_gem_object *new,
5974 unsigned frontbuffer_bits)
5976 /* Control of individual bits within the mask are guarded by
5977 * the owning plane->mutex, i.e. we can never see concurrent
5978 * manipulation of individual bits. But since the bitfield as a whole
5979 * is updated using RMW, we need to use atomics in order to update
5982 BUILD_BUG_ON(INTEL_FRONTBUFFER_BITS_PER_PIPE * I915_MAX_PIPES >
5983 BITS_PER_TYPE(atomic_t));
5986 WARN_ON(!(atomic_read(&old->frontbuffer_bits) & frontbuffer_bits));
5987 atomic_andnot(frontbuffer_bits, &old->frontbuffer_bits);
5991 WARN_ON(atomic_read(&new->frontbuffer_bits) & frontbuffer_bits);
5992 atomic_or(frontbuffer_bits, &new->frontbuffer_bits);
5996 /* Allocate a new GEM object and fill it with the supplied data */
5997 struct drm_i915_gem_object *
5998 i915_gem_object_create_from_data(struct drm_i915_private *dev_priv,
5999 const void *data, size_t size)
6001 struct drm_i915_gem_object *obj;
6006 obj = i915_gem_object_create(dev_priv, round_up(size, PAGE_SIZE));
6010 GEM_BUG_ON(obj->write_domain != I915_GEM_DOMAIN_CPU);
6012 file = obj->base.filp;
6015 unsigned int len = min_t(typeof(size), size, PAGE_SIZE);
6017 void *pgdata, *vaddr;
6019 err = pagecache_write_begin(file, file->f_mapping,
6026 memcpy(vaddr, data, len);
6029 err = pagecache_write_end(file, file->f_mapping,
6043 i915_gem_object_put(obj);
6044 return ERR_PTR(err);
6047 struct scatterlist *
6048 i915_gem_object_get_sg(struct drm_i915_gem_object *obj,
6050 unsigned int *offset)
6052 struct i915_gem_object_page_iter *iter = &obj->mm.get_page;
6053 struct scatterlist *sg;
6054 unsigned int idx, count;
6057 GEM_BUG_ON(n >= obj->base.size >> PAGE_SHIFT);
6058 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
6060 /* As we iterate forward through the sg, we record each entry in a
6061 * radixtree for quick repeated (backwards) lookups. If we have seen
6062 * this index previously, we will have an entry for it.
6064 * Initial lookup is O(N), but this is amortized to O(1) for
6065 * sequential page access (where each new request is consecutive
6066 * to the previous one). Repeated lookups are O(lg(obj->base.size)),
6067 * i.e. O(1) with a large constant!
6069 if (n < READ_ONCE(iter->sg_idx))
6072 mutex_lock(&iter->lock);
6074 /* We prefer to reuse the last sg so that repeated lookup of this
6075 * (or the subsequent) sg are fast - comparing against the last
6076 * sg is faster than going through the radixtree.
6081 count = __sg_page_count(sg);
6083 while (idx + count <= n) {
6088 /* If we cannot allocate and insert this entry, or the
6089 * individual pages from this range, cancel updating the
6090 * sg_idx so that on this lookup we are forced to linearly
6091 * scan onwards, but on future lookups we will try the
6092 * insertion again (in which case we need to be careful of
6093 * the error return reporting that we have already inserted
6096 ret = radix_tree_insert(&iter->radix, idx, sg);
6097 if (ret && ret != -EEXIST)
6100 entry = xa_mk_value(idx);
6101 for (i = 1; i < count; i++) {
6102 ret = radix_tree_insert(&iter->radix, idx + i, entry);
6103 if (ret && ret != -EEXIST)
6108 sg = ____sg_next(sg);
6109 count = __sg_page_count(sg);
6116 mutex_unlock(&iter->lock);
6118 if (unlikely(n < idx)) /* insertion completed by another thread */
6121 /* In case we failed to insert the entry into the radixtree, we need
6122 * to look beyond the current sg.
6124 while (idx + count <= n) {
6126 sg = ____sg_next(sg);
6127 count = __sg_page_count(sg);
6136 sg = radix_tree_lookup(&iter->radix, n);
6139 /* If this index is in the middle of multi-page sg entry,
6140 * the radix tree will contain a value entry that points
6141 * to the start of that range. We will return the pointer to
6142 * the base page and the offset of this page within the
6146 if (unlikely(xa_is_value(sg))) {
6147 unsigned long base = xa_to_value(sg);
6149 sg = radix_tree_lookup(&iter->radix, base);
6161 i915_gem_object_get_page(struct drm_i915_gem_object *obj, unsigned int n)
6163 struct scatterlist *sg;
6164 unsigned int offset;
6166 GEM_BUG_ON(!i915_gem_object_has_struct_page(obj));
6168 sg = i915_gem_object_get_sg(obj, n, &offset);
6169 return nth_page(sg_page(sg), offset);
6172 /* Like i915_gem_object_get_page(), but mark the returned page dirty */
6174 i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj,
6179 page = i915_gem_object_get_page(obj, n);
6181 set_page_dirty(page);
6187 i915_gem_object_get_dma_address(struct drm_i915_gem_object *obj,
6190 struct scatterlist *sg;
6191 unsigned int offset;
6193 sg = i915_gem_object_get_sg(obj, n, &offset);
6194 return sg_dma_address(sg) + (offset << PAGE_SHIFT);
6197 int i915_gem_object_attach_phys(struct drm_i915_gem_object *obj, int align)
6199 struct sg_table *pages;
6202 if (align > obj->base.size)
6205 if (obj->ops == &i915_gem_phys_ops)
6208 if (obj->ops != &i915_gem_object_ops)
6211 err = i915_gem_object_unbind(obj);
6215 mutex_lock(&obj->mm.lock);
6217 if (obj->mm.madv != I915_MADV_WILLNEED) {
6222 if (obj->mm.quirked) {
6227 if (obj->mm.mapping) {
6232 pages = __i915_gem_object_unset_pages(obj);
6234 obj->ops = &i915_gem_phys_ops;
6236 err = ____i915_gem_object_get_pages(obj);
6240 /* Perma-pin (until release) the physical set of pages */
6241 __i915_gem_object_pin_pages(obj);
6243 if (!IS_ERR_OR_NULL(pages))
6244 i915_gem_object_ops.put_pages(obj, pages);
6245 mutex_unlock(&obj->mm.lock);
6249 obj->ops = &i915_gem_object_ops;
6250 if (!IS_ERR_OR_NULL(pages)) {
6251 unsigned int sg_page_sizes = i915_sg_page_sizes(pages->sgl);
6253 __i915_gem_object_set_pages(obj, pages, sg_page_sizes);
6256 mutex_unlock(&obj->mm.lock);
6260 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
6261 #include "selftests/scatterlist.c"
6262 #include "selftests/mock_gem_device.c"
6263 #include "selftests/huge_gem_object.c"
6264 #include "selftests/huge_pages.c"
6265 #include "selftests/i915_gem_object.c"
6266 #include "selftests/i915_gem_coherency.c"
6267 #include "selftests/i915_gem.c"