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_vgpu.h"
33 #include "i915_trace.h"
34 #include "intel_drv.h"
35 #include "intel_frontbuffer.h"
36 #include "intel_mocs.h"
37 #include <linux/dma-fence-array.h>
38 #include <linux/reservation.h>
39 #include <linux/shmem_fs.h>
40 #include <linux/slab.h>
41 #include <linux/swap.h>
42 #include <linux/pci.h>
43 #include <linux/dma-buf.h>
45 static void i915_gem_flush_free_objects(struct drm_i915_private *i915);
46 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);
47 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj);
49 static bool cpu_cache_is_coherent(struct drm_device *dev,
50 enum i915_cache_level level)
52 return HAS_LLC(to_i915(dev)) || level != I915_CACHE_NONE;
55 static bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj)
57 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
60 if (!cpu_cache_is_coherent(obj->base.dev, obj->cache_level))
63 return obj->pin_display;
67 insert_mappable_node(struct i915_ggtt *ggtt,
68 struct drm_mm_node *node, u32 size)
70 memset(node, 0, sizeof(*node));
71 return drm_mm_insert_node_in_range_generic(&ggtt->base.mm, node,
73 0, ggtt->mappable_end,
74 DRM_MM_SEARCH_DEFAULT,
75 DRM_MM_CREATE_DEFAULT);
79 remove_mappable_node(struct drm_mm_node *node)
81 drm_mm_remove_node(node);
84 /* some bookkeeping */
85 static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
88 spin_lock(&dev_priv->mm.object_stat_lock);
89 dev_priv->mm.object_count++;
90 dev_priv->mm.object_memory += size;
91 spin_unlock(&dev_priv->mm.object_stat_lock);
94 static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
97 spin_lock(&dev_priv->mm.object_stat_lock);
98 dev_priv->mm.object_count--;
99 dev_priv->mm.object_memory -= size;
100 spin_unlock(&dev_priv->mm.object_stat_lock);
104 i915_gem_wait_for_error(struct i915_gpu_error *error)
110 if (!i915_reset_in_progress(error))
114 * Only wait 10 seconds for the gpu reset to complete to avoid hanging
115 * userspace. If it takes that long something really bad is going on and
116 * we should simply try to bail out and fail as gracefully as possible.
118 ret = wait_event_interruptible_timeout(error->reset_queue,
119 !i915_reset_in_progress(error),
122 DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
124 } else if (ret < 0) {
131 int i915_mutex_lock_interruptible(struct drm_device *dev)
133 struct drm_i915_private *dev_priv = to_i915(dev);
136 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
140 ret = mutex_lock_interruptible(&dev->struct_mutex);
148 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
149 struct drm_file *file)
151 struct drm_i915_private *dev_priv = to_i915(dev);
152 struct i915_ggtt *ggtt = &dev_priv->ggtt;
153 struct drm_i915_gem_get_aperture *args = data;
154 struct i915_vma *vma;
158 mutex_lock(&dev->struct_mutex);
159 list_for_each_entry(vma, &ggtt->base.active_list, vm_link)
160 if (i915_vma_is_pinned(vma))
161 pinned += vma->node.size;
162 list_for_each_entry(vma, &ggtt->base.inactive_list, vm_link)
163 if (i915_vma_is_pinned(vma))
164 pinned += vma->node.size;
165 mutex_unlock(&dev->struct_mutex);
167 args->aper_size = ggtt->base.total;
168 args->aper_available_size = args->aper_size - pinned;
173 static struct sg_table *
174 i915_gem_object_get_pages_phys(struct drm_i915_gem_object *obj)
176 struct address_space *mapping = obj->base.filp->f_mapping;
177 char *vaddr = obj->phys_handle->vaddr;
179 struct scatterlist *sg;
182 if (WARN_ON(i915_gem_object_needs_bit17_swizzle(obj)))
183 return ERR_PTR(-EINVAL);
185 for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
189 page = shmem_read_mapping_page(mapping, i);
191 return ERR_CAST(page);
193 src = kmap_atomic(page);
194 memcpy(vaddr, src, PAGE_SIZE);
195 drm_clflush_virt_range(vaddr, PAGE_SIZE);
202 i915_gem_chipset_flush(to_i915(obj->base.dev));
204 st = kmalloc(sizeof(*st), GFP_KERNEL);
206 return ERR_PTR(-ENOMEM);
208 if (sg_alloc_table(st, 1, GFP_KERNEL)) {
210 return ERR_PTR(-ENOMEM);
215 sg->length = obj->base.size;
217 sg_dma_address(sg) = obj->phys_handle->busaddr;
218 sg_dma_len(sg) = obj->base.size;
224 __i915_gem_object_release_shmem(struct drm_i915_gem_object *obj,
225 struct sg_table *pages)
227 GEM_BUG_ON(obj->mm.madv == __I915_MADV_PURGED);
229 if (obj->mm.madv == I915_MADV_DONTNEED)
230 obj->mm.dirty = false;
232 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0 &&
233 !cpu_cache_is_coherent(obj->base.dev, obj->cache_level))
234 drm_clflush_sg(pages);
236 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
237 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
241 i915_gem_object_put_pages_phys(struct drm_i915_gem_object *obj,
242 struct sg_table *pages)
244 __i915_gem_object_release_shmem(obj, pages);
247 struct address_space *mapping = obj->base.filp->f_mapping;
248 char *vaddr = obj->phys_handle->vaddr;
251 for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
255 page = shmem_read_mapping_page(mapping, i);
259 dst = kmap_atomic(page);
260 drm_clflush_virt_range(vaddr, PAGE_SIZE);
261 memcpy(dst, vaddr, PAGE_SIZE);
264 set_page_dirty(page);
265 if (obj->mm.madv == I915_MADV_WILLNEED)
266 mark_page_accessed(page);
270 obj->mm.dirty = false;
273 sg_free_table(pages);
278 i915_gem_object_release_phys(struct drm_i915_gem_object *obj)
280 drm_pci_free(obj->base.dev, obj->phys_handle);
281 i915_gem_object_unpin_pages(obj);
284 static const struct drm_i915_gem_object_ops i915_gem_phys_ops = {
285 .get_pages = i915_gem_object_get_pages_phys,
286 .put_pages = i915_gem_object_put_pages_phys,
287 .release = i915_gem_object_release_phys,
290 int i915_gem_object_unbind(struct drm_i915_gem_object *obj)
292 struct i915_vma *vma;
293 LIST_HEAD(still_in_list);
296 lockdep_assert_held(&obj->base.dev->struct_mutex);
298 /* Closed vma are removed from the obj->vma_list - but they may
299 * still have an active binding on the object. To remove those we
300 * must wait for all rendering to complete to the object (as unbinding
301 * must anyway), and retire the requests.
303 ret = i915_gem_object_wait(obj,
304 I915_WAIT_INTERRUPTIBLE |
307 MAX_SCHEDULE_TIMEOUT,
312 i915_gem_retire_requests(to_i915(obj->base.dev));
314 while ((vma = list_first_entry_or_null(&obj->vma_list,
317 list_move_tail(&vma->obj_link, &still_in_list);
318 ret = i915_vma_unbind(vma);
322 list_splice(&still_in_list, &obj->vma_list);
328 i915_gem_object_wait_fence(struct dma_fence *fence,
331 struct intel_rps_client *rps)
333 struct drm_i915_gem_request *rq;
335 BUILD_BUG_ON(I915_WAIT_INTERRUPTIBLE != 0x1);
337 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
340 if (!dma_fence_is_i915(fence))
341 return dma_fence_wait_timeout(fence,
342 flags & I915_WAIT_INTERRUPTIBLE,
345 rq = to_request(fence);
346 if (i915_gem_request_completed(rq))
349 /* This client is about to stall waiting for the GPU. In many cases
350 * this is undesirable and limits the throughput of the system, as
351 * many clients cannot continue processing user input/output whilst
352 * blocked. RPS autotuning may take tens of milliseconds to respond
353 * to the GPU load and thus incurs additional latency for the client.
354 * We can circumvent that by promoting the GPU frequency to maximum
355 * before we wait. This makes the GPU throttle up much more quickly
356 * (good for benchmarks and user experience, e.g. window animations),
357 * but at a cost of spending more power processing the workload
358 * (bad for battery). Not all clients even want their results
359 * immediately and for them we should just let the GPU select its own
360 * frequency to maximise efficiency. To prevent a single client from
361 * forcing the clocks too high for the whole system, we only allow
362 * each client to waitboost once in a busy period.
365 if (INTEL_GEN(rq->i915) >= 6)
366 gen6_rps_boost(rq->i915, rps, rq->emitted_jiffies);
371 timeout = i915_wait_request(rq, flags, timeout);
374 if (flags & I915_WAIT_LOCKED && i915_gem_request_completed(rq))
375 i915_gem_request_retire_upto(rq);
377 if (rps && rq->global_seqno == intel_engine_last_submit(rq->engine)) {
378 /* The GPU is now idle and this client has stalled.
379 * Since no other client has submitted a request in the
380 * meantime, assume that this client is the only one
381 * supplying work to the GPU but is unable to keep that
382 * work supplied because it is waiting. Since the GPU is
383 * then never kept fully busy, RPS autoclocking will
384 * keep the clocks relatively low, causing further delays.
385 * Compensate by giving the synchronous client credit for
386 * a waitboost next time.
388 spin_lock(&rq->i915->rps.client_lock);
389 list_del_init(&rps->link);
390 spin_unlock(&rq->i915->rps.client_lock);
397 i915_gem_object_wait_reservation(struct reservation_object *resv,
400 struct intel_rps_client *rps)
402 struct dma_fence *excl;
404 if (flags & I915_WAIT_ALL) {
405 struct dma_fence **shared;
406 unsigned int count, i;
409 ret = reservation_object_get_fences_rcu(resv,
410 &excl, &count, &shared);
414 for (i = 0; i < count; i++) {
415 timeout = i915_gem_object_wait_fence(shared[i],
421 dma_fence_put(shared[i]);
424 for (; i < count; i++)
425 dma_fence_put(shared[i]);
428 excl = reservation_object_get_excl_rcu(resv);
431 if (excl && timeout > 0)
432 timeout = i915_gem_object_wait_fence(excl, flags, timeout, rps);
439 static void __fence_set_priority(struct dma_fence *fence, int prio)
441 struct drm_i915_gem_request *rq;
442 struct intel_engine_cs *engine;
444 if (!dma_fence_is_i915(fence))
447 rq = to_request(fence);
449 if (!engine->schedule)
452 engine->schedule(rq, prio);
455 static void fence_set_priority(struct dma_fence *fence, int prio)
457 /* Recurse once into a fence-array */
458 if (dma_fence_is_array(fence)) {
459 struct dma_fence_array *array = to_dma_fence_array(fence);
462 for (i = 0; i < array->num_fences; i++)
463 __fence_set_priority(array->fences[i], prio);
465 __fence_set_priority(fence, prio);
470 i915_gem_object_wait_priority(struct drm_i915_gem_object *obj,
474 struct dma_fence *excl;
476 if (flags & I915_WAIT_ALL) {
477 struct dma_fence **shared;
478 unsigned int count, i;
481 ret = reservation_object_get_fences_rcu(obj->resv,
482 &excl, &count, &shared);
486 for (i = 0; i < count; i++) {
487 fence_set_priority(shared[i], prio);
488 dma_fence_put(shared[i]);
493 excl = reservation_object_get_excl_rcu(obj->resv);
497 fence_set_priority(excl, prio);
504 * Waits for rendering to the object to be completed
505 * @obj: i915 gem object
506 * @flags: how to wait (under a lock, for all rendering or just for writes etc)
507 * @timeout: how long to wait
508 * @rps: client (user process) to charge for any waitboosting
511 i915_gem_object_wait(struct drm_i915_gem_object *obj,
514 struct intel_rps_client *rps)
517 #if IS_ENABLED(CONFIG_LOCKDEP)
518 GEM_BUG_ON(debug_locks &&
519 !!lockdep_is_held(&obj->base.dev->struct_mutex) !=
520 !!(flags & I915_WAIT_LOCKED));
522 GEM_BUG_ON(timeout < 0);
524 timeout = i915_gem_object_wait_reservation(obj->resv,
527 return timeout < 0 ? timeout : 0;
530 static struct intel_rps_client *to_rps_client(struct drm_file *file)
532 struct drm_i915_file_private *fpriv = file->driver_priv;
538 i915_gem_object_attach_phys(struct drm_i915_gem_object *obj,
541 drm_dma_handle_t *phys;
544 if (obj->phys_handle) {
545 if ((unsigned long)obj->phys_handle->vaddr & (align -1))
551 if (obj->mm.madv != I915_MADV_WILLNEED)
554 if (obj->base.filp == NULL)
557 ret = i915_gem_object_unbind(obj);
561 __i915_gem_object_put_pages(obj, I915_MM_NORMAL);
565 /* create a new object */
566 phys = drm_pci_alloc(obj->base.dev, obj->base.size, align);
570 obj->phys_handle = phys;
571 obj->ops = &i915_gem_phys_ops;
573 return i915_gem_object_pin_pages(obj);
577 i915_gem_phys_pwrite(struct drm_i915_gem_object *obj,
578 struct drm_i915_gem_pwrite *args,
579 struct drm_file *file)
581 struct drm_device *dev = obj->base.dev;
582 void *vaddr = obj->phys_handle->vaddr + args->offset;
583 char __user *user_data = u64_to_user_ptr(args->data_ptr);
586 /* We manually control the domain here and pretend that it
587 * remains coherent i.e. in the GTT domain, like shmem_pwrite.
589 lockdep_assert_held(&obj->base.dev->struct_mutex);
590 ret = i915_gem_object_wait(obj,
591 I915_WAIT_INTERRUPTIBLE |
594 MAX_SCHEDULE_TIMEOUT,
595 to_rps_client(file));
599 intel_fb_obj_invalidate(obj, ORIGIN_CPU);
600 if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
601 unsigned long unwritten;
603 /* The physical object once assigned is fixed for the lifetime
604 * of the obj, so we can safely drop the lock and continue
607 mutex_unlock(&dev->struct_mutex);
608 unwritten = copy_from_user(vaddr, user_data, args->size);
609 mutex_lock(&dev->struct_mutex);
616 drm_clflush_virt_range(vaddr, args->size);
617 i915_gem_chipset_flush(to_i915(dev));
620 intel_fb_obj_flush(obj, false, ORIGIN_CPU);
624 void *i915_gem_object_alloc(struct drm_device *dev)
626 struct drm_i915_private *dev_priv = to_i915(dev);
627 return kmem_cache_zalloc(dev_priv->objects, GFP_KERNEL);
630 void i915_gem_object_free(struct drm_i915_gem_object *obj)
632 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
633 kmem_cache_free(dev_priv->objects, obj);
637 i915_gem_create(struct drm_file *file,
638 struct drm_device *dev,
642 struct drm_i915_gem_object *obj;
646 size = roundup(size, PAGE_SIZE);
650 /* Allocate the new object */
651 obj = i915_gem_object_create(dev, size);
655 ret = drm_gem_handle_create(file, &obj->base, &handle);
656 /* drop reference from allocate - handle holds it now */
657 i915_gem_object_put(obj);
666 i915_gem_dumb_create(struct drm_file *file,
667 struct drm_device *dev,
668 struct drm_mode_create_dumb *args)
670 /* have to work out size/pitch and return them */
671 args->pitch = ALIGN(args->width * DIV_ROUND_UP(args->bpp, 8), 64);
672 args->size = args->pitch * args->height;
673 return i915_gem_create(file, dev,
674 args->size, &args->handle);
678 * Creates a new mm object and returns a handle to it.
679 * @dev: drm device pointer
680 * @data: ioctl data blob
681 * @file: drm file pointer
684 i915_gem_create_ioctl(struct drm_device *dev, void *data,
685 struct drm_file *file)
687 struct drm_i915_gem_create *args = data;
689 i915_gem_flush_free_objects(to_i915(dev));
691 return i915_gem_create(file, dev,
692 args->size, &args->handle);
696 __copy_to_user_swizzled(char __user *cpu_vaddr,
697 const char *gpu_vaddr, int gpu_offset,
700 int ret, cpu_offset = 0;
703 int cacheline_end = ALIGN(gpu_offset + 1, 64);
704 int this_length = min(cacheline_end - gpu_offset, length);
705 int swizzled_gpu_offset = gpu_offset ^ 64;
707 ret = __copy_to_user(cpu_vaddr + cpu_offset,
708 gpu_vaddr + swizzled_gpu_offset,
713 cpu_offset += this_length;
714 gpu_offset += this_length;
715 length -= this_length;
722 __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset,
723 const char __user *cpu_vaddr,
726 int ret, cpu_offset = 0;
729 int cacheline_end = ALIGN(gpu_offset + 1, 64);
730 int this_length = min(cacheline_end - gpu_offset, length);
731 int swizzled_gpu_offset = gpu_offset ^ 64;
733 ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset,
734 cpu_vaddr + cpu_offset,
739 cpu_offset += this_length;
740 gpu_offset += this_length;
741 length -= this_length;
748 * Pins the specified object's pages and synchronizes the object with
749 * GPU accesses. Sets needs_clflush to non-zero if the caller should
750 * flush the object from the CPU cache.
752 int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj,
753 unsigned int *needs_clflush)
757 lockdep_assert_held(&obj->base.dev->struct_mutex);
760 if (!i915_gem_object_has_struct_page(obj))
763 ret = i915_gem_object_wait(obj,
764 I915_WAIT_INTERRUPTIBLE |
766 MAX_SCHEDULE_TIMEOUT,
771 ret = i915_gem_object_pin_pages(obj);
775 i915_gem_object_flush_gtt_write_domain(obj);
777 /* If we're not in the cpu read domain, set ourself into the gtt
778 * read domain and manually flush cachelines (if required). This
779 * optimizes for the case when the gpu will dirty the data
780 * anyway again before the next pread happens.
782 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU))
783 *needs_clflush = !cpu_cache_is_coherent(obj->base.dev,
786 if (*needs_clflush && !static_cpu_has(X86_FEATURE_CLFLUSH)) {
787 ret = i915_gem_object_set_to_cpu_domain(obj, false);
794 /* return with the pages pinned */
798 i915_gem_object_unpin_pages(obj);
802 int i915_gem_obj_prepare_shmem_write(struct drm_i915_gem_object *obj,
803 unsigned int *needs_clflush)
807 lockdep_assert_held(&obj->base.dev->struct_mutex);
810 if (!i915_gem_object_has_struct_page(obj))
813 ret = i915_gem_object_wait(obj,
814 I915_WAIT_INTERRUPTIBLE |
817 MAX_SCHEDULE_TIMEOUT,
822 ret = i915_gem_object_pin_pages(obj);
826 i915_gem_object_flush_gtt_write_domain(obj);
828 /* If we're not in the cpu write domain, set ourself into the
829 * gtt write domain and manually flush cachelines (as required).
830 * This optimizes for the case when the gpu will use the data
831 * right away and we therefore have to clflush anyway.
833 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
834 *needs_clflush |= cpu_write_needs_clflush(obj) << 1;
836 /* Same trick applies to invalidate partially written cachelines read
839 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU))
840 *needs_clflush |= !cpu_cache_is_coherent(obj->base.dev,
843 if (*needs_clflush && !static_cpu_has(X86_FEATURE_CLFLUSH)) {
844 ret = i915_gem_object_set_to_cpu_domain(obj, true);
851 if ((*needs_clflush & CLFLUSH_AFTER) == 0)
852 obj->cache_dirty = true;
854 intel_fb_obj_invalidate(obj, ORIGIN_CPU);
855 obj->mm.dirty = true;
856 /* return with the pages pinned */
860 i915_gem_object_unpin_pages(obj);
865 shmem_clflush_swizzled_range(char *addr, unsigned long length,
868 if (unlikely(swizzled)) {
869 unsigned long start = (unsigned long) addr;
870 unsigned long end = (unsigned long) addr + length;
872 /* For swizzling simply ensure that we always flush both
873 * channels. Lame, but simple and it works. Swizzled
874 * pwrite/pread is far from a hotpath - current userspace
875 * doesn't use it at all. */
876 start = round_down(start, 128);
877 end = round_up(end, 128);
879 drm_clflush_virt_range((void *)start, end - start);
881 drm_clflush_virt_range(addr, length);
886 /* Only difference to the fast-path function is that this can handle bit17
887 * and uses non-atomic copy and kmap functions. */
889 shmem_pread_slow(struct page *page, int offset, int length,
890 char __user *user_data,
891 bool page_do_bit17_swizzling, bool needs_clflush)
898 shmem_clflush_swizzled_range(vaddr + offset, length,
899 page_do_bit17_swizzling);
901 if (page_do_bit17_swizzling)
902 ret = __copy_to_user_swizzled(user_data, vaddr, offset, length);
904 ret = __copy_to_user(user_data, vaddr + offset, length);
907 return ret ? - EFAULT : 0;
911 shmem_pread(struct page *page, int offset, int length, char __user *user_data,
912 bool page_do_bit17_swizzling, bool needs_clflush)
917 if (!page_do_bit17_swizzling) {
918 char *vaddr = kmap_atomic(page);
921 drm_clflush_virt_range(vaddr + offset, length);
922 ret = __copy_to_user_inatomic(user_data, vaddr + offset, length);
923 kunmap_atomic(vaddr);
928 return shmem_pread_slow(page, offset, length, user_data,
929 page_do_bit17_swizzling, needs_clflush);
933 i915_gem_shmem_pread(struct drm_i915_gem_object *obj,
934 struct drm_i915_gem_pread *args)
936 char __user *user_data;
938 unsigned int obj_do_bit17_swizzling;
939 unsigned int needs_clflush;
940 unsigned int idx, offset;
943 obj_do_bit17_swizzling = 0;
944 if (i915_gem_object_needs_bit17_swizzle(obj))
945 obj_do_bit17_swizzling = BIT(17);
947 ret = mutex_lock_interruptible(&obj->base.dev->struct_mutex);
951 ret = i915_gem_obj_prepare_shmem_read(obj, &needs_clflush);
952 mutex_unlock(&obj->base.dev->struct_mutex);
957 user_data = u64_to_user_ptr(args->data_ptr);
958 offset = offset_in_page(args->offset);
959 for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
960 struct page *page = i915_gem_object_get_page(obj, idx);
964 if (offset + length > PAGE_SIZE)
965 length = PAGE_SIZE - offset;
967 ret = shmem_pread(page, offset, length, user_data,
968 page_to_phys(page) & obj_do_bit17_swizzling,
978 i915_gem_obj_finish_shmem_access(obj);
983 gtt_user_read(struct io_mapping *mapping,
984 loff_t base, int offset,
985 char __user *user_data, int length)
988 unsigned long unwritten;
990 /* We can use the cpu mem copy function because this is X86. */
991 vaddr = (void __force *)io_mapping_map_atomic_wc(mapping, base);
992 unwritten = __copy_to_user_inatomic(user_data, vaddr + offset, length);
993 io_mapping_unmap_atomic(vaddr);
995 vaddr = (void __force *)
996 io_mapping_map_wc(mapping, base, PAGE_SIZE);
997 unwritten = copy_to_user(user_data, vaddr + offset, length);
998 io_mapping_unmap(vaddr);
1004 i915_gem_gtt_pread(struct drm_i915_gem_object *obj,
1005 const struct drm_i915_gem_pread *args)
1007 struct drm_i915_private *i915 = to_i915(obj->base.dev);
1008 struct i915_ggtt *ggtt = &i915->ggtt;
1009 struct drm_mm_node node;
1010 struct i915_vma *vma;
1011 void __user *user_data;
1015 ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
1019 intel_runtime_pm_get(i915);
1020 vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
1021 PIN_MAPPABLE | PIN_NONBLOCK);
1023 node.start = i915_ggtt_offset(vma);
1024 node.allocated = false;
1025 ret = i915_vma_put_fence(vma);
1027 i915_vma_unpin(vma);
1032 ret = insert_mappable_node(ggtt, &node, PAGE_SIZE);
1035 GEM_BUG_ON(!node.allocated);
1038 ret = i915_gem_object_set_to_gtt_domain(obj, false);
1042 mutex_unlock(&i915->drm.struct_mutex);
1044 user_data = u64_to_user_ptr(args->data_ptr);
1045 remain = args->size;
1046 offset = args->offset;
1048 while (remain > 0) {
1049 /* Operation in this page
1051 * page_base = page offset within aperture
1052 * page_offset = offset within page
1053 * page_length = bytes to copy for this page
1055 u32 page_base = node.start;
1056 unsigned page_offset = offset_in_page(offset);
1057 unsigned page_length = PAGE_SIZE - page_offset;
1058 page_length = remain < page_length ? remain : page_length;
1059 if (node.allocated) {
1061 ggtt->base.insert_page(&ggtt->base,
1062 i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT),
1063 node.start, I915_CACHE_NONE, 0);
1066 page_base += offset & PAGE_MASK;
1069 if (gtt_user_read(&ggtt->mappable, page_base, page_offset,
1070 user_data, page_length)) {
1075 remain -= page_length;
1076 user_data += page_length;
1077 offset += page_length;
1080 mutex_lock(&i915->drm.struct_mutex);
1082 if (node.allocated) {
1084 ggtt->base.clear_range(&ggtt->base,
1085 node.start, node.size);
1086 remove_mappable_node(&node);
1088 i915_vma_unpin(vma);
1091 intel_runtime_pm_put(i915);
1092 mutex_unlock(&i915->drm.struct_mutex);
1098 * Reads data from the object referenced by handle.
1099 * @dev: drm device pointer
1100 * @data: ioctl data blob
1101 * @file: drm file pointer
1103 * On error, the contents of *data are undefined.
1106 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
1107 struct drm_file *file)
1109 struct drm_i915_gem_pread *args = data;
1110 struct drm_i915_gem_object *obj;
1113 if (args->size == 0)
1116 if (!access_ok(VERIFY_WRITE,
1117 u64_to_user_ptr(args->data_ptr),
1121 obj = i915_gem_object_lookup(file, args->handle);
1125 /* Bounds check source. */
1126 if (args->offset > obj->base.size ||
1127 args->size > obj->base.size - args->offset) {
1132 trace_i915_gem_object_pread(obj, args->offset, args->size);
1134 ret = i915_gem_object_wait(obj,
1135 I915_WAIT_INTERRUPTIBLE,
1136 MAX_SCHEDULE_TIMEOUT,
1137 to_rps_client(file));
1141 ret = i915_gem_object_pin_pages(obj);
1145 ret = i915_gem_shmem_pread(obj, args);
1146 if (ret == -EFAULT || ret == -ENODEV)
1147 ret = i915_gem_gtt_pread(obj, args);
1149 i915_gem_object_unpin_pages(obj);
1151 i915_gem_object_put(obj);
1155 /* This is the fast write path which cannot handle
1156 * page faults in the source data
1160 ggtt_write(struct io_mapping *mapping,
1161 loff_t base, int offset,
1162 char __user *user_data, int length)
1165 unsigned long unwritten;
1167 /* We can use the cpu mem copy function because this is X86. */
1168 vaddr = (void __force *)io_mapping_map_atomic_wc(mapping, base);
1169 unwritten = __copy_from_user_inatomic_nocache(vaddr + offset,
1171 io_mapping_unmap_atomic(vaddr);
1173 vaddr = (void __force *)
1174 io_mapping_map_wc(mapping, base, PAGE_SIZE);
1175 unwritten = copy_from_user(vaddr + offset, user_data, length);
1176 io_mapping_unmap(vaddr);
1183 * This is the fast pwrite path, where we copy the data directly from the
1184 * user into the GTT, uncached.
1185 * @obj: i915 GEM object
1186 * @args: pwrite arguments structure
1189 i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
1190 const struct drm_i915_gem_pwrite *args)
1192 struct drm_i915_private *i915 = to_i915(obj->base.dev);
1193 struct i915_ggtt *ggtt = &i915->ggtt;
1194 struct drm_mm_node node;
1195 struct i915_vma *vma;
1197 void __user *user_data;
1200 ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
1204 intel_runtime_pm_get(i915);
1205 vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
1206 PIN_MAPPABLE | PIN_NONBLOCK);
1208 node.start = i915_ggtt_offset(vma);
1209 node.allocated = false;
1210 ret = i915_vma_put_fence(vma);
1212 i915_vma_unpin(vma);
1217 ret = insert_mappable_node(ggtt, &node, PAGE_SIZE);
1220 GEM_BUG_ON(!node.allocated);
1223 ret = i915_gem_object_set_to_gtt_domain(obj, true);
1227 mutex_unlock(&i915->drm.struct_mutex);
1229 intel_fb_obj_invalidate(obj, ORIGIN_CPU);
1231 user_data = u64_to_user_ptr(args->data_ptr);
1232 offset = args->offset;
1233 remain = args->size;
1235 /* Operation in this page
1237 * page_base = page offset within aperture
1238 * page_offset = offset within page
1239 * page_length = bytes to copy for this page
1241 u32 page_base = node.start;
1242 unsigned int page_offset = offset_in_page(offset);
1243 unsigned int page_length = PAGE_SIZE - page_offset;
1244 page_length = remain < page_length ? remain : page_length;
1245 if (node.allocated) {
1246 wmb(); /* flush the write before we modify the GGTT */
1247 ggtt->base.insert_page(&ggtt->base,
1248 i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT),
1249 node.start, I915_CACHE_NONE, 0);
1250 wmb(); /* flush modifications to the GGTT (insert_page) */
1252 page_base += offset & PAGE_MASK;
1254 /* If we get a fault while copying data, then (presumably) our
1255 * source page isn't available. Return the error and we'll
1256 * retry in the slow path.
1257 * If the object is non-shmem backed, we retry again with the
1258 * path that handles page fault.
1260 if (ggtt_write(&ggtt->mappable, page_base, page_offset,
1261 user_data, page_length)) {
1266 remain -= page_length;
1267 user_data += page_length;
1268 offset += page_length;
1270 intel_fb_obj_flush(obj, false, ORIGIN_CPU);
1272 mutex_lock(&i915->drm.struct_mutex);
1274 if (node.allocated) {
1276 ggtt->base.clear_range(&ggtt->base,
1277 node.start, node.size);
1278 remove_mappable_node(&node);
1280 i915_vma_unpin(vma);
1283 intel_runtime_pm_put(i915);
1284 mutex_unlock(&i915->drm.struct_mutex);
1289 shmem_pwrite_slow(struct page *page, int offset, int length,
1290 char __user *user_data,
1291 bool page_do_bit17_swizzling,
1292 bool needs_clflush_before,
1293 bool needs_clflush_after)
1299 if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
1300 shmem_clflush_swizzled_range(vaddr + offset, length,
1301 page_do_bit17_swizzling);
1302 if (page_do_bit17_swizzling)
1303 ret = __copy_from_user_swizzled(vaddr, offset, user_data,
1306 ret = __copy_from_user(vaddr + offset, user_data, length);
1307 if (needs_clflush_after)
1308 shmem_clflush_swizzled_range(vaddr + offset, length,
1309 page_do_bit17_swizzling);
1312 return ret ? -EFAULT : 0;
1315 /* Per-page copy function for the shmem pwrite fastpath.
1316 * Flushes invalid cachelines before writing to the target if
1317 * needs_clflush_before is set and flushes out any written cachelines after
1318 * writing if needs_clflush is set.
1321 shmem_pwrite(struct page *page, int offset, int len, char __user *user_data,
1322 bool page_do_bit17_swizzling,
1323 bool needs_clflush_before,
1324 bool needs_clflush_after)
1329 if (!page_do_bit17_swizzling) {
1330 char *vaddr = kmap_atomic(page);
1332 if (needs_clflush_before)
1333 drm_clflush_virt_range(vaddr + offset, len);
1334 ret = __copy_from_user_inatomic(vaddr + offset, user_data, len);
1335 if (needs_clflush_after)
1336 drm_clflush_virt_range(vaddr + offset, len);
1338 kunmap_atomic(vaddr);
1343 return shmem_pwrite_slow(page, offset, len, user_data,
1344 page_do_bit17_swizzling,
1345 needs_clflush_before,
1346 needs_clflush_after);
1350 i915_gem_shmem_pwrite(struct drm_i915_gem_object *obj,
1351 const struct drm_i915_gem_pwrite *args)
1353 struct drm_i915_private *i915 = to_i915(obj->base.dev);
1354 void __user *user_data;
1356 unsigned int obj_do_bit17_swizzling;
1357 unsigned int partial_cacheline_write;
1358 unsigned int needs_clflush;
1359 unsigned int offset, idx;
1362 ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
1366 ret = i915_gem_obj_prepare_shmem_write(obj, &needs_clflush);
1367 mutex_unlock(&i915->drm.struct_mutex);
1371 obj_do_bit17_swizzling = 0;
1372 if (i915_gem_object_needs_bit17_swizzle(obj))
1373 obj_do_bit17_swizzling = BIT(17);
1375 /* If we don't overwrite a cacheline completely we need to be
1376 * careful to have up-to-date data by first clflushing. Don't
1377 * overcomplicate things and flush the entire patch.
1379 partial_cacheline_write = 0;
1380 if (needs_clflush & CLFLUSH_BEFORE)
1381 partial_cacheline_write = boot_cpu_data.x86_clflush_size - 1;
1383 user_data = u64_to_user_ptr(args->data_ptr);
1384 remain = args->size;
1385 offset = offset_in_page(args->offset);
1386 for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
1387 struct page *page = i915_gem_object_get_page(obj, idx);
1391 if (offset + length > PAGE_SIZE)
1392 length = PAGE_SIZE - offset;
1394 ret = shmem_pwrite(page, offset, length, user_data,
1395 page_to_phys(page) & obj_do_bit17_swizzling,
1396 (offset | length) & partial_cacheline_write,
1397 needs_clflush & CLFLUSH_AFTER);
1402 user_data += length;
1406 intel_fb_obj_flush(obj, false, ORIGIN_CPU);
1407 i915_gem_obj_finish_shmem_access(obj);
1412 * Writes data to the object referenced by handle.
1414 * @data: ioctl data blob
1417 * On error, the contents of the buffer that were to be modified are undefined.
1420 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
1421 struct drm_file *file)
1423 struct drm_i915_gem_pwrite *args = data;
1424 struct drm_i915_gem_object *obj;
1427 if (args->size == 0)
1430 if (!access_ok(VERIFY_READ,
1431 u64_to_user_ptr(args->data_ptr),
1435 obj = i915_gem_object_lookup(file, args->handle);
1439 /* Bounds check destination. */
1440 if (args->offset > obj->base.size ||
1441 args->size > obj->base.size - args->offset) {
1446 trace_i915_gem_object_pwrite(obj, args->offset, args->size);
1448 ret = i915_gem_object_wait(obj,
1449 I915_WAIT_INTERRUPTIBLE |
1451 MAX_SCHEDULE_TIMEOUT,
1452 to_rps_client(file));
1456 ret = i915_gem_object_pin_pages(obj);
1461 /* We can only do the GTT pwrite on untiled buffers, as otherwise
1462 * it would end up going through the fenced access, and we'll get
1463 * different detiling behavior between reading and writing.
1464 * pread/pwrite currently are reading and writing from the CPU
1465 * perspective, requiring manual detiling by the client.
1467 if (!i915_gem_object_has_struct_page(obj) ||
1468 cpu_write_needs_clflush(obj))
1469 /* Note that the gtt paths might fail with non-page-backed user
1470 * pointers (e.g. gtt mappings when moving data between
1471 * textures). Fallback to the shmem path in that case.
1473 ret = i915_gem_gtt_pwrite_fast(obj, args);
1475 if (ret == -EFAULT || ret == -ENOSPC) {
1476 if (obj->phys_handle)
1477 ret = i915_gem_phys_pwrite(obj, args, file);
1479 ret = i915_gem_shmem_pwrite(obj, args);
1482 i915_gem_object_unpin_pages(obj);
1484 i915_gem_object_put(obj);
1488 static inline enum fb_op_origin
1489 write_origin(struct drm_i915_gem_object *obj, unsigned domain)
1491 return (domain == I915_GEM_DOMAIN_GTT ?
1492 obj->frontbuffer_ggtt_origin : ORIGIN_CPU);
1495 static void i915_gem_object_bump_inactive_ggtt(struct drm_i915_gem_object *obj)
1497 struct drm_i915_private *i915;
1498 struct list_head *list;
1499 struct i915_vma *vma;
1501 list_for_each_entry(vma, &obj->vma_list, obj_link) {
1502 if (!i915_vma_is_ggtt(vma))
1505 if (i915_vma_is_active(vma))
1508 if (!drm_mm_node_allocated(&vma->node))
1511 list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
1514 i915 = to_i915(obj->base.dev);
1515 list = obj->bind_count ? &i915->mm.bound_list : &i915->mm.unbound_list;
1516 list_move_tail(&obj->global_link, list);
1520 * Called when user space prepares to use an object with the CPU, either
1521 * through the mmap ioctl's mapping or a GTT mapping.
1523 * @data: ioctl data blob
1527 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1528 struct drm_file *file)
1530 struct drm_i915_gem_set_domain *args = data;
1531 struct drm_i915_gem_object *obj;
1532 uint32_t read_domains = args->read_domains;
1533 uint32_t write_domain = args->write_domain;
1536 /* Only handle setting domains to types used by the CPU. */
1537 if ((write_domain | read_domains) & I915_GEM_GPU_DOMAINS)
1540 /* Having something in the write domain implies it's in the read
1541 * domain, and only that read domain. Enforce that in the request.
1543 if (write_domain != 0 && read_domains != write_domain)
1546 obj = i915_gem_object_lookup(file, args->handle);
1550 /* Try to flush the object off the GPU without holding the lock.
1551 * We will repeat the flush holding the lock in the normal manner
1552 * to catch cases where we are gazumped.
1554 err = i915_gem_object_wait(obj,
1555 I915_WAIT_INTERRUPTIBLE |
1556 (write_domain ? I915_WAIT_ALL : 0),
1557 MAX_SCHEDULE_TIMEOUT,
1558 to_rps_client(file));
1562 /* Flush and acquire obj->pages so that we are coherent through
1563 * direct access in memory with previous cached writes through
1564 * shmemfs and that our cache domain tracking remains valid.
1565 * For example, if the obj->filp was moved to swap without us
1566 * being notified and releasing the pages, we would mistakenly
1567 * continue to assume that the obj remained out of the CPU cached
1570 err = i915_gem_object_pin_pages(obj);
1574 err = i915_mutex_lock_interruptible(dev);
1578 if (read_domains & I915_GEM_DOMAIN_GTT)
1579 err = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1581 err = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1583 /* And bump the LRU for this access */
1584 i915_gem_object_bump_inactive_ggtt(obj);
1586 mutex_unlock(&dev->struct_mutex);
1588 if (write_domain != 0)
1589 intel_fb_obj_invalidate(obj, write_origin(obj, write_domain));
1592 i915_gem_object_unpin_pages(obj);
1594 i915_gem_object_put(obj);
1599 * Called when user space has done writes to this buffer
1601 * @data: ioctl data blob
1605 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1606 struct drm_file *file)
1608 struct drm_i915_gem_sw_finish *args = data;
1609 struct drm_i915_gem_object *obj;
1612 obj = i915_gem_object_lookup(file, args->handle);
1616 /* Pinned buffers may be scanout, so flush the cache */
1617 if (READ_ONCE(obj->pin_display)) {
1618 err = i915_mutex_lock_interruptible(dev);
1620 i915_gem_object_flush_cpu_write_domain(obj);
1621 mutex_unlock(&dev->struct_mutex);
1625 i915_gem_object_put(obj);
1630 * i915_gem_mmap_ioctl - Maps the contents of an object, returning the address
1633 * @data: ioctl data blob
1636 * While the mapping holds a reference on the contents of the object, it doesn't
1637 * imply a ref on the object itself.
1641 * DRM driver writers who look a this function as an example for how to do GEM
1642 * mmap support, please don't implement mmap support like here. The modern way
1643 * to implement DRM mmap support is with an mmap offset ioctl (like
1644 * i915_gem_mmap_gtt) and then using the mmap syscall on the DRM fd directly.
1645 * That way debug tooling like valgrind will understand what's going on, hiding
1646 * the mmap call in a driver private ioctl will break that. The i915 driver only
1647 * does cpu mmaps this way because we didn't know better.
1650 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1651 struct drm_file *file)
1653 struct drm_i915_gem_mmap *args = data;
1654 struct drm_i915_gem_object *obj;
1657 if (args->flags & ~(I915_MMAP_WC))
1660 if (args->flags & I915_MMAP_WC && !boot_cpu_has(X86_FEATURE_PAT))
1663 obj = i915_gem_object_lookup(file, args->handle);
1667 /* prime objects have no backing filp to GEM mmap
1670 if (!obj->base.filp) {
1671 i915_gem_object_put(obj);
1675 addr = vm_mmap(obj->base.filp, 0, args->size,
1676 PROT_READ | PROT_WRITE, MAP_SHARED,
1678 if (args->flags & I915_MMAP_WC) {
1679 struct mm_struct *mm = current->mm;
1680 struct vm_area_struct *vma;
1682 if (down_write_killable(&mm->mmap_sem)) {
1683 i915_gem_object_put(obj);
1686 vma = find_vma(mm, addr);
1689 pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
1692 up_write(&mm->mmap_sem);
1694 /* This may race, but that's ok, it only gets set */
1695 WRITE_ONCE(obj->frontbuffer_ggtt_origin, ORIGIN_CPU);
1697 i915_gem_object_put(obj);
1698 if (IS_ERR((void *)addr))
1701 args->addr_ptr = (uint64_t) addr;
1706 static unsigned int tile_row_pages(struct drm_i915_gem_object *obj)
1710 size = i915_gem_object_get_stride(obj);
1711 size *= i915_gem_object_get_tiling(obj) == I915_TILING_Y ? 32 : 8;
1713 return size >> PAGE_SHIFT;
1717 * i915_gem_mmap_gtt_version - report the current feature set for GTT mmaps
1719 * A history of the GTT mmap interface:
1721 * 0 - Everything had to fit into the GTT. Both parties of a memcpy had to
1722 * aligned and suitable for fencing, and still fit into the available
1723 * mappable space left by the pinned display objects. A classic problem
1724 * we called the page-fault-of-doom where we would ping-pong between
1725 * two objects that could not fit inside the GTT and so the memcpy
1726 * would page one object in at the expense of the other between every
1729 * 1 - Objects can be any size, and have any compatible fencing (X Y, or none
1730 * as set via i915_gem_set_tiling() [DRM_I915_GEM_SET_TILING]). If the
1731 * object is too large for the available space (or simply too large
1732 * for the mappable aperture!), a view is created instead and faulted
1733 * into userspace. (This view is aligned and sized appropriately for
1738 * * snoopable objects cannot be accessed via the GTT. It can cause machine
1739 * hangs on some architectures, corruption on others. An attempt to service
1740 * a GTT page fault from a snoopable object will generate a SIGBUS.
1742 * * the object must be able to fit into RAM (physical memory, though no
1743 * limited to the mappable aperture).
1748 * * a new GTT page fault will synchronize rendering from the GPU and flush
1749 * all data to system memory. Subsequent access will not be synchronized.
1751 * * all mappings are revoked on runtime device suspend.
1753 * * there are only 8, 16 or 32 fence registers to share between all users
1754 * (older machines require fence register for display and blitter access
1755 * as well). Contention of the fence registers will cause the previous users
1756 * to be unmapped and any new access will generate new page faults.
1758 * * running out of memory while servicing a fault may generate a SIGBUS,
1759 * rather than the expected SIGSEGV.
1761 int i915_gem_mmap_gtt_version(void)
1767 * i915_gem_fault - fault a page into the GTT
1768 * @area: CPU VMA in question
1771 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1772 * from userspace. The fault handler takes care of binding the object to
1773 * the GTT (if needed), allocating and programming a fence register (again,
1774 * only if needed based on whether the old reg is still valid or the object
1775 * is tiled) and inserting a new PTE into the faulting process.
1777 * Note that the faulting process may involve evicting existing objects
1778 * from the GTT and/or fence registers to make room. So performance may
1779 * suffer if the GTT working set is large or there are few fence registers
1782 * The current feature set supported by i915_gem_fault() and thus GTT mmaps
1783 * is exposed via I915_PARAM_MMAP_GTT_VERSION (see i915_gem_mmap_gtt_version).
1785 int i915_gem_fault(struct vm_area_struct *area, struct vm_fault *vmf)
1787 #define MIN_CHUNK_PAGES ((1 << 20) >> PAGE_SHIFT) /* 1 MiB */
1788 struct drm_i915_gem_object *obj = to_intel_bo(area->vm_private_data);
1789 struct drm_device *dev = obj->base.dev;
1790 struct drm_i915_private *dev_priv = to_i915(dev);
1791 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1792 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1793 struct i915_vma *vma;
1794 pgoff_t page_offset;
1798 /* We don't use vmf->pgoff since that has the fake offset */
1799 page_offset = ((unsigned long)vmf->virtual_address - area->vm_start) >>
1802 trace_i915_gem_object_fault(obj, page_offset, true, write);
1804 /* Try to flush the object off the GPU first without holding the lock.
1805 * Upon acquiring the lock, we will perform our sanity checks and then
1806 * repeat the flush holding the lock in the normal manner to catch cases
1807 * where we are gazumped.
1809 ret = i915_gem_object_wait(obj,
1810 I915_WAIT_INTERRUPTIBLE,
1811 MAX_SCHEDULE_TIMEOUT,
1816 ret = i915_gem_object_pin_pages(obj);
1820 intel_runtime_pm_get(dev_priv);
1822 ret = i915_mutex_lock_interruptible(dev);
1826 /* Access to snoopable pages through the GTT is incoherent. */
1827 if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev_priv)) {
1832 /* If the object is smaller than a couple of partial vma, it is
1833 * not worth only creating a single partial vma - we may as well
1834 * clear enough space for the full object.
1836 flags = PIN_MAPPABLE;
1837 if (obj->base.size > 2 * MIN_CHUNK_PAGES << PAGE_SHIFT)
1838 flags |= PIN_NONBLOCK | PIN_NONFAULT;
1840 /* Now pin it into the GTT as needed */
1841 vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, flags);
1843 struct i915_ggtt_view view;
1844 unsigned int chunk_size;
1846 /* Use a partial view if it is bigger than available space */
1847 chunk_size = MIN_CHUNK_PAGES;
1848 if (i915_gem_object_is_tiled(obj))
1849 chunk_size = roundup(chunk_size, tile_row_pages(obj));
1851 memset(&view, 0, sizeof(view));
1852 view.type = I915_GGTT_VIEW_PARTIAL;
1853 view.params.partial.offset = rounddown(page_offset, chunk_size);
1854 view.params.partial.size =
1855 min_t(unsigned int, chunk_size,
1856 vma_pages(area) - view.params.partial.offset);
1858 /* If the partial covers the entire object, just create a
1861 if (chunk_size >= obj->base.size >> PAGE_SHIFT)
1862 view.type = I915_GGTT_VIEW_NORMAL;
1864 /* Userspace is now writing through an untracked VMA, abandon
1865 * all hope that the hardware is able to track future writes.
1867 obj->frontbuffer_ggtt_origin = ORIGIN_CPU;
1869 vma = i915_gem_object_ggtt_pin(obj, &view, 0, 0, PIN_MAPPABLE);
1876 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1880 ret = i915_vma_get_fence(vma);
1884 /* Mark as being mmapped into userspace for later revocation */
1885 assert_rpm_wakelock_held(dev_priv);
1886 if (list_empty(&obj->userfault_link))
1887 list_add(&obj->userfault_link, &dev_priv->mm.userfault_list);
1889 /* Finally, remap it using the new GTT offset */
1890 ret = remap_io_mapping(area,
1891 area->vm_start + (vma->ggtt_view.params.partial.offset << PAGE_SHIFT),
1892 (ggtt->mappable_base + vma->node.start) >> PAGE_SHIFT,
1893 min_t(u64, vma->size, area->vm_end - area->vm_start),
1897 __i915_vma_unpin(vma);
1899 mutex_unlock(&dev->struct_mutex);
1901 intel_runtime_pm_put(dev_priv);
1902 i915_gem_object_unpin_pages(obj);
1907 * We eat errors when the gpu is terminally wedged to avoid
1908 * userspace unduly crashing (gl has no provisions for mmaps to
1909 * fail). But any other -EIO isn't ours (e.g. swap in failure)
1910 * and so needs to be reported.
1912 if (!i915_terminally_wedged(&dev_priv->gpu_error)) {
1913 ret = VM_FAULT_SIGBUS;
1918 * EAGAIN means the gpu is hung and we'll wait for the error
1919 * handler to reset everything when re-faulting in
1920 * i915_mutex_lock_interruptible.
1927 * EBUSY is ok: this just means that another thread
1928 * already did the job.
1930 ret = VM_FAULT_NOPAGE;
1937 ret = VM_FAULT_SIGBUS;
1940 WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret);
1941 ret = VM_FAULT_SIGBUS;
1948 * i915_gem_release_mmap - remove physical page mappings
1949 * @obj: obj in question
1951 * Preserve the reservation of the mmapping with the DRM core code, but
1952 * relinquish ownership of the pages back to the system.
1954 * It is vital that we remove the page mapping if we have mapped a tiled
1955 * object through the GTT and then lose the fence register due to
1956 * resource pressure. Similarly if the object has been moved out of the
1957 * aperture, than pages mapped into userspace must be revoked. Removing the
1958 * mapping will then trigger a page fault on the next user access, allowing
1959 * fixup by i915_gem_fault().
1962 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1964 struct drm_i915_private *i915 = to_i915(obj->base.dev);
1966 /* Serialisation between user GTT access and our code depends upon
1967 * revoking the CPU's PTE whilst the mutex is held. The next user
1968 * pagefault then has to wait until we release the mutex.
1970 * Note that RPM complicates somewhat by adding an additional
1971 * requirement that operations to the GGTT be made holding the RPM
1974 lockdep_assert_held(&i915->drm.struct_mutex);
1975 intel_runtime_pm_get(i915);
1977 if (list_empty(&obj->userfault_link))
1980 list_del_init(&obj->userfault_link);
1981 drm_vma_node_unmap(&obj->base.vma_node,
1982 obj->base.dev->anon_inode->i_mapping);
1984 /* Ensure that the CPU's PTE are revoked and there are not outstanding
1985 * memory transactions from userspace before we return. The TLB
1986 * flushing implied above by changing the PTE above *should* be
1987 * sufficient, an extra barrier here just provides us with a bit
1988 * of paranoid documentation about our requirement to serialise
1989 * memory writes before touching registers / GSM.
1994 intel_runtime_pm_put(i915);
1997 void i915_gem_runtime_suspend(struct drm_i915_private *dev_priv)
1999 struct drm_i915_gem_object *obj, *on;
2003 * Only called during RPM suspend. All users of the userfault_list
2004 * must be holding an RPM wakeref to ensure that this can not
2005 * run concurrently with themselves (and use the struct_mutex for
2006 * protection between themselves).
2009 list_for_each_entry_safe(obj, on,
2010 &dev_priv->mm.userfault_list, userfault_link) {
2011 list_del_init(&obj->userfault_link);
2012 drm_vma_node_unmap(&obj->base.vma_node,
2013 obj->base.dev->anon_inode->i_mapping);
2016 /* The fence will be lost when the device powers down. If any were
2017 * in use by hardware (i.e. they are pinned), we should not be powering
2018 * down! All other fences will be reacquired by the user upon waking.
2020 for (i = 0; i < dev_priv->num_fence_regs; i++) {
2021 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
2023 if (WARN_ON(reg->pin_count))
2029 GEM_BUG_ON(!list_empty(®->vma->obj->userfault_link));
2035 * i915_gem_get_ggtt_size - return required global GTT size for an object
2036 * @dev_priv: i915 device
2037 * @size: object size
2038 * @tiling_mode: tiling mode
2040 * Return the required global GTT size for an object, taking into account
2041 * potential fence register mapping.
2043 u64 i915_gem_get_ggtt_size(struct drm_i915_private *dev_priv,
2044 u64 size, int tiling_mode)
2048 GEM_BUG_ON(size == 0);
2050 if (INTEL_GEN(dev_priv) >= 4 ||
2051 tiling_mode == I915_TILING_NONE)
2054 /* Previous chips need a power-of-two fence region when tiling */
2055 if (IS_GEN3(dev_priv))
2056 ggtt_size = 1024*1024;
2058 ggtt_size = 512*1024;
2060 while (ggtt_size < size)
2067 * i915_gem_get_ggtt_alignment - return required global GTT alignment
2068 * @dev_priv: i915 device
2069 * @size: object size
2070 * @tiling_mode: tiling mode
2071 * @fenced: is fenced alignment required or not
2073 * Return the required global GTT alignment for an object, taking into account
2074 * potential fence register mapping.
2076 u64 i915_gem_get_ggtt_alignment(struct drm_i915_private *dev_priv, u64 size,
2077 int tiling_mode, bool fenced)
2079 GEM_BUG_ON(size == 0);
2082 * Minimum alignment is 4k (GTT page size), but might be greater
2083 * if a fence register is needed for the object.
2085 if (INTEL_GEN(dev_priv) >= 4 || (!fenced && IS_G33(dev_priv)) ||
2086 tiling_mode == I915_TILING_NONE)
2090 * Previous chips need to be aligned to the size of the smallest
2091 * fence register that can contain the object.
2093 return i915_gem_get_ggtt_size(dev_priv, size, tiling_mode);
2096 static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj)
2098 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
2101 err = drm_gem_create_mmap_offset(&obj->base);
2105 /* We can idle the GPU locklessly to flush stale objects, but in order
2106 * to claim that space for ourselves, we need to take the big
2107 * struct_mutex to free the requests+objects and allocate our slot.
2109 err = i915_gem_wait_for_idle(dev_priv, I915_WAIT_INTERRUPTIBLE);
2113 err = i915_mutex_lock_interruptible(&dev_priv->drm);
2115 i915_gem_retire_requests(dev_priv);
2116 err = drm_gem_create_mmap_offset(&obj->base);
2117 mutex_unlock(&dev_priv->drm.struct_mutex);
2123 static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)
2125 drm_gem_free_mmap_offset(&obj->base);
2129 i915_gem_mmap_gtt(struct drm_file *file,
2130 struct drm_device *dev,
2134 struct drm_i915_gem_object *obj;
2137 obj = i915_gem_object_lookup(file, handle);
2141 ret = i915_gem_object_create_mmap_offset(obj);
2143 *offset = drm_vma_node_offset_addr(&obj->base.vma_node);
2145 i915_gem_object_put(obj);
2150 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
2152 * @data: GTT mapping ioctl data
2153 * @file: GEM object info
2155 * Simply returns the fake offset to userspace so it can mmap it.
2156 * The mmap call will end up in drm_gem_mmap(), which will set things
2157 * up so we can get faults in the handler above.
2159 * The fault handler will take care of binding the object into the GTT
2160 * (since it may have been evicted to make room for something), allocating
2161 * a fence register, and mapping the appropriate aperture address into
2165 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
2166 struct drm_file *file)
2168 struct drm_i915_gem_mmap_gtt *args = data;
2170 return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
2173 /* Immediately discard the backing storage */
2175 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
2177 i915_gem_object_free_mmap_offset(obj);
2179 if (obj->base.filp == NULL)
2182 /* Our goal here is to return as much of the memory as
2183 * is possible back to the system as we are called from OOM.
2184 * To do this we must instruct the shmfs to drop all of its
2185 * backing pages, *now*.
2187 shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1);
2188 obj->mm.madv = __I915_MADV_PURGED;
2191 /* Try to discard unwanted pages */
2192 void __i915_gem_object_invalidate(struct drm_i915_gem_object *obj)
2194 struct address_space *mapping;
2196 lockdep_assert_held(&obj->mm.lock);
2197 GEM_BUG_ON(obj->mm.pages);
2199 switch (obj->mm.madv) {
2200 case I915_MADV_DONTNEED:
2201 i915_gem_object_truncate(obj);
2202 case __I915_MADV_PURGED:
2206 if (obj->base.filp == NULL)
2209 mapping = obj->base.filp->f_mapping,
2210 invalidate_mapping_pages(mapping, 0, (loff_t)-1);
2214 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj,
2215 struct sg_table *pages)
2217 struct sgt_iter sgt_iter;
2220 __i915_gem_object_release_shmem(obj, pages);
2222 i915_gem_gtt_finish_pages(obj, pages);
2224 if (i915_gem_object_needs_bit17_swizzle(obj))
2225 i915_gem_object_save_bit_17_swizzle(obj, pages);
2227 for_each_sgt_page(page, sgt_iter, pages) {
2229 set_page_dirty(page);
2231 if (obj->mm.madv == I915_MADV_WILLNEED)
2232 mark_page_accessed(page);
2236 obj->mm.dirty = false;
2238 sg_free_table(pages);
2242 static void __i915_gem_object_reset_page_iter(struct drm_i915_gem_object *obj)
2244 struct radix_tree_iter iter;
2247 radix_tree_for_each_slot(slot, &obj->mm.get_page.radix, &iter, 0)
2248 radix_tree_delete(&obj->mm.get_page.radix, iter.index);
2251 void __i915_gem_object_put_pages(struct drm_i915_gem_object *obj,
2252 enum i915_mm_subclass subclass)
2254 struct sg_table *pages;
2256 if (i915_gem_object_has_pinned_pages(obj))
2259 GEM_BUG_ON(obj->bind_count);
2260 if (!READ_ONCE(obj->mm.pages))
2263 /* May be called by shrinker from within get_pages() (on another bo) */
2264 mutex_lock_nested(&obj->mm.lock, subclass);
2265 if (unlikely(atomic_read(&obj->mm.pages_pin_count)))
2268 /* ->put_pages might need to allocate memory for the bit17 swizzle
2269 * array, hence protect them from being reaped by removing them from gtt
2271 pages = fetch_and_zero(&obj->mm.pages);
2274 if (obj->mm.mapping) {
2277 ptr = ptr_mask_bits(obj->mm.mapping);
2278 if (is_vmalloc_addr(ptr))
2281 kunmap(kmap_to_page(ptr));
2283 obj->mm.mapping = NULL;
2286 __i915_gem_object_reset_page_iter(obj);
2288 obj->ops->put_pages(obj, pages);
2290 mutex_unlock(&obj->mm.lock);
2293 static unsigned int swiotlb_max_size(void)
2295 #if IS_ENABLED(CONFIG_SWIOTLB)
2296 return rounddown(swiotlb_nr_tbl() << IO_TLB_SHIFT, PAGE_SIZE);
2302 static void i915_sg_trim(struct sg_table *orig_st)
2304 struct sg_table new_st;
2305 struct scatterlist *sg, *new_sg;
2308 if (orig_st->nents == orig_st->orig_nents)
2311 if (sg_alloc_table(&new_st, orig_st->nents, GFP_KERNEL))
2314 new_sg = new_st.sgl;
2315 for_each_sg(orig_st->sgl, sg, orig_st->nents, i) {
2316 sg_set_page(new_sg, sg_page(sg), sg->length, 0);
2317 /* called before being DMA mapped, no need to copy sg->dma_* */
2318 new_sg = sg_next(new_sg);
2321 sg_free_table(orig_st);
2326 static struct sg_table *
2327 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
2329 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
2331 struct address_space *mapping;
2332 struct sg_table *st;
2333 struct scatterlist *sg;
2334 struct sgt_iter sgt_iter;
2336 unsigned long last_pfn = 0; /* suppress gcc warning */
2337 unsigned int max_segment;
2341 /* Assert that the object is not currently in any GPU domain. As it
2342 * wasn't in the GTT, there shouldn't be any way it could have been in
2345 GEM_BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
2346 GEM_BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
2348 max_segment = swiotlb_max_size();
2350 max_segment = rounddown(UINT_MAX, PAGE_SIZE);
2352 st = kmalloc(sizeof(*st), GFP_KERNEL);
2354 return ERR_PTR(-ENOMEM);
2356 page_count = obj->base.size / PAGE_SIZE;
2357 if (sg_alloc_table(st, page_count, GFP_KERNEL)) {
2359 return ERR_PTR(-ENOMEM);
2362 /* Get the list of pages out of our struct file. They'll be pinned
2363 * at this point until we release them.
2365 * Fail silently without starting the shrinker
2367 mapping = obj->base.filp->f_mapping;
2368 gfp = mapping_gfp_constraint(mapping, ~(__GFP_IO | __GFP_RECLAIM));
2369 gfp |= __GFP_NORETRY | __GFP_NOWARN;
2372 for (i = 0; i < page_count; i++) {
2373 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2375 i915_gem_shrink(dev_priv,
2378 I915_SHRINK_UNBOUND |
2379 I915_SHRINK_PURGEABLE);
2380 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2383 /* We've tried hard to allocate the memory by reaping
2384 * our own buffer, now let the real VM do its job and
2385 * go down in flames if truly OOM.
2387 page = shmem_read_mapping_page(mapping, i);
2389 ret = PTR_ERR(page);
2394 sg->length >= max_segment ||
2395 page_to_pfn(page) != last_pfn + 1) {
2399 sg_set_page(sg, page, PAGE_SIZE, 0);
2401 sg->length += PAGE_SIZE;
2403 last_pfn = page_to_pfn(page);
2405 /* Check that the i965g/gm workaround works. */
2406 WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
2408 if (sg) /* loop terminated early; short sg table */
2411 /* Trim unused sg entries to avoid wasting memory. */
2414 ret = i915_gem_gtt_prepare_pages(obj, st);
2418 if (i915_gem_object_needs_bit17_swizzle(obj))
2419 i915_gem_object_do_bit_17_swizzle(obj, st);
2426 for_each_sgt_page(page, sgt_iter, st)
2431 /* shmemfs first checks if there is enough memory to allocate the page
2432 * and reports ENOSPC should there be insufficient, along with the usual
2433 * ENOMEM for a genuine allocation failure.
2435 * We use ENOSPC in our driver to mean that we have run out of aperture
2436 * space and so want to translate the error from shmemfs back to our
2437 * usual understanding of ENOMEM.
2442 return ERR_PTR(ret);
2445 void __i915_gem_object_set_pages(struct drm_i915_gem_object *obj,
2446 struct sg_table *pages)
2448 lockdep_assert_held(&obj->mm.lock);
2450 obj->mm.get_page.sg_pos = pages->sgl;
2451 obj->mm.get_page.sg_idx = 0;
2453 obj->mm.pages = pages;
2455 if (i915_gem_object_is_tiled(obj) &&
2456 to_i915(obj->base.dev)->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
2457 GEM_BUG_ON(obj->mm.quirked);
2458 __i915_gem_object_pin_pages(obj);
2459 obj->mm.quirked = true;
2463 static int ____i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
2465 struct sg_table *pages;
2467 GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj));
2469 if (unlikely(obj->mm.madv != I915_MADV_WILLNEED)) {
2470 DRM_DEBUG("Attempting to obtain a purgeable object\n");
2474 pages = obj->ops->get_pages(obj);
2475 if (unlikely(IS_ERR(pages)))
2476 return PTR_ERR(pages);
2478 __i915_gem_object_set_pages(obj, pages);
2482 /* Ensure that the associated pages are gathered from the backing storage
2483 * and pinned into our object. i915_gem_object_pin_pages() may be called
2484 * multiple times before they are released by a single call to
2485 * i915_gem_object_unpin_pages() - once the pages are no longer referenced
2486 * either as a result of memory pressure (reaping pages under the shrinker)
2487 * or as the object is itself released.
2489 int __i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
2493 err = mutex_lock_interruptible(&obj->mm.lock);
2497 if (unlikely(!obj->mm.pages)) {
2498 err = ____i915_gem_object_get_pages(obj);
2502 smp_mb__before_atomic();
2504 atomic_inc(&obj->mm.pages_pin_count);
2507 mutex_unlock(&obj->mm.lock);
2511 /* The 'mapping' part of i915_gem_object_pin_map() below */
2512 static void *i915_gem_object_map(const struct drm_i915_gem_object *obj,
2513 enum i915_map_type type)
2515 unsigned long n_pages = obj->base.size >> PAGE_SHIFT;
2516 struct sg_table *sgt = obj->mm.pages;
2517 struct sgt_iter sgt_iter;
2519 struct page *stack_pages[32];
2520 struct page **pages = stack_pages;
2521 unsigned long i = 0;
2525 /* A single page can always be kmapped */
2526 if (n_pages == 1 && type == I915_MAP_WB)
2527 return kmap(sg_page(sgt->sgl));
2529 if (n_pages > ARRAY_SIZE(stack_pages)) {
2530 /* Too big for stack -- allocate temporary array instead */
2531 pages = drm_malloc_gfp(n_pages, sizeof(*pages), GFP_TEMPORARY);
2536 for_each_sgt_page(page, sgt_iter, sgt)
2539 /* Check that we have the expected number of pages */
2540 GEM_BUG_ON(i != n_pages);
2544 pgprot = PAGE_KERNEL;
2547 pgprot = pgprot_writecombine(PAGE_KERNEL_IO);
2550 addr = vmap(pages, n_pages, 0, pgprot);
2552 if (pages != stack_pages)
2553 drm_free_large(pages);
2558 /* get, pin, and map the pages of the object into kernel space */
2559 void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj,
2560 enum i915_map_type type)
2562 enum i915_map_type has_type;
2567 GEM_BUG_ON(!i915_gem_object_has_struct_page(obj));
2569 ret = mutex_lock_interruptible(&obj->mm.lock);
2571 return ERR_PTR(ret);
2574 if (!atomic_inc_not_zero(&obj->mm.pages_pin_count)) {
2575 if (unlikely(!obj->mm.pages)) {
2576 ret = ____i915_gem_object_get_pages(obj);
2580 smp_mb__before_atomic();
2582 atomic_inc(&obj->mm.pages_pin_count);
2585 GEM_BUG_ON(!obj->mm.pages);
2587 ptr = ptr_unpack_bits(obj->mm.mapping, has_type);
2588 if (ptr && has_type != type) {
2594 if (is_vmalloc_addr(ptr))
2597 kunmap(kmap_to_page(ptr));
2599 ptr = obj->mm.mapping = NULL;
2603 ptr = i915_gem_object_map(obj, type);
2609 obj->mm.mapping = ptr_pack_bits(ptr, type);
2613 mutex_unlock(&obj->mm.lock);
2617 atomic_dec(&obj->mm.pages_pin_count);
2623 static bool i915_context_is_banned(const struct i915_gem_context *ctx)
2625 unsigned long elapsed;
2627 if (ctx->hang_stats.banned)
2630 elapsed = get_seconds() - ctx->hang_stats.guilty_ts;
2631 if (ctx->hang_stats.ban_period_seconds &&
2632 elapsed <= ctx->hang_stats.ban_period_seconds) {
2633 DRM_DEBUG("context hanging too fast, banning!\n");
2640 static void i915_set_reset_status(struct i915_gem_context *ctx,
2643 struct i915_ctx_hang_stats *hs = &ctx->hang_stats;
2646 hs->banned = i915_context_is_banned(ctx);
2648 hs->guilty_ts = get_seconds();
2650 hs->batch_pending++;
2654 struct drm_i915_gem_request *
2655 i915_gem_find_active_request(struct intel_engine_cs *engine)
2657 struct drm_i915_gem_request *request;
2659 /* We are called by the error capture and reset at a random
2660 * point in time. In particular, note that neither is crucially
2661 * ordered with an interrupt. After a hang, the GPU is dead and we
2662 * assume that no more writes can happen (we waited long enough for
2663 * all writes that were in transaction to be flushed) - adding an
2664 * extra delay for a recent interrupt is pointless. Hence, we do
2665 * not need an engine->irq_seqno_barrier() before the seqno reads.
2667 list_for_each_entry(request, &engine->timeline->requests, link) {
2668 if (__i915_gem_request_completed(request))
2677 static void reset_request(struct drm_i915_gem_request *request)
2679 void *vaddr = request->ring->vaddr;
2682 /* As this request likely depends on state from the lost
2683 * context, clear out all the user operations leaving the
2684 * breadcrumb at the end (so we get the fence notifications).
2686 head = request->head;
2687 if (request->postfix < head) {
2688 memset(vaddr + head, 0, request->ring->size - head);
2691 memset(vaddr + head, 0, request->postfix - head);
2694 static void i915_gem_reset_engine(struct intel_engine_cs *engine)
2696 struct drm_i915_gem_request *request;
2697 struct i915_gem_context *incomplete_ctx;
2698 struct intel_timeline *timeline;
2701 if (engine->irq_seqno_barrier)
2702 engine->irq_seqno_barrier(engine);
2704 request = i915_gem_find_active_request(engine);
2708 ring_hung = engine->hangcheck.score >= HANGCHECK_SCORE_RING_HUNG;
2709 if (engine->hangcheck.seqno != intel_engine_get_seqno(engine))
2712 i915_set_reset_status(request->ctx, ring_hung);
2716 DRM_DEBUG_DRIVER("resetting %s to restart from tail of request 0x%x\n",
2717 engine->name, request->global_seqno);
2719 /* Setup the CS to resume from the breadcrumb of the hung request */
2720 engine->reset_hw(engine, request);
2722 /* Users of the default context do not rely on logical state
2723 * preserved between batches. They have to emit full state on
2724 * every batch and so it is safe to execute queued requests following
2727 * Other contexts preserve state, now corrupt. We want to skip all
2728 * queued requests that reference the corrupt context.
2730 incomplete_ctx = request->ctx;
2731 if (i915_gem_context_is_default(incomplete_ctx))
2734 list_for_each_entry_continue(request, &engine->timeline->requests, link)
2735 if (request->ctx == incomplete_ctx)
2736 reset_request(request);
2738 timeline = i915_gem_context_lookup_timeline(incomplete_ctx, engine);
2739 list_for_each_entry(request, &timeline->requests, link)
2740 reset_request(request);
2743 void i915_gem_reset(struct drm_i915_private *dev_priv)
2745 struct intel_engine_cs *engine;
2746 enum intel_engine_id id;
2748 lockdep_assert_held(&dev_priv->drm.struct_mutex);
2750 i915_gem_retire_requests(dev_priv);
2752 for_each_engine(engine, dev_priv, id)
2753 i915_gem_reset_engine(engine);
2755 i915_gem_restore_fences(dev_priv);
2757 if (dev_priv->gt.awake) {
2758 intel_sanitize_gt_powersave(dev_priv);
2759 intel_enable_gt_powersave(dev_priv);
2760 if (INTEL_GEN(dev_priv) >= 6)
2761 gen6_rps_busy(dev_priv);
2765 static void nop_submit_request(struct drm_i915_gem_request *request)
2767 i915_gem_request_submit(request);
2768 intel_engine_init_global_seqno(request->engine, request->global_seqno);
2771 static void i915_gem_cleanup_engine(struct intel_engine_cs *engine)
2773 engine->submit_request = nop_submit_request;
2775 /* Mark all pending requests as complete so that any concurrent
2776 * (lockless) lookup doesn't try and wait upon the request as we
2779 intel_engine_init_global_seqno(engine,
2780 intel_engine_last_submit(engine));
2783 * Clear the execlists queue up before freeing the requests, as those
2784 * are the ones that keep the context and ringbuffer backing objects
2788 if (i915.enable_execlists) {
2789 unsigned long flags;
2791 spin_lock_irqsave(&engine->timeline->lock, flags);
2793 i915_gem_request_put(engine->execlist_port[0].request);
2794 i915_gem_request_put(engine->execlist_port[1].request);
2795 memset(engine->execlist_port, 0, sizeof(engine->execlist_port));
2796 engine->execlist_queue = RB_ROOT;
2797 engine->execlist_first = NULL;
2799 spin_unlock_irqrestore(&engine->timeline->lock, flags);
2803 void i915_gem_set_wedged(struct drm_i915_private *dev_priv)
2805 struct intel_engine_cs *engine;
2806 enum intel_engine_id id;
2808 lockdep_assert_held(&dev_priv->drm.struct_mutex);
2809 set_bit(I915_WEDGED, &dev_priv->gpu_error.flags);
2811 i915_gem_context_lost(dev_priv);
2812 for_each_engine(engine, dev_priv, id)
2813 i915_gem_cleanup_engine(engine);
2814 mod_delayed_work(dev_priv->wq, &dev_priv->gt.idle_work, 0);
2816 i915_gem_retire_requests(dev_priv);
2820 i915_gem_retire_work_handler(struct work_struct *work)
2822 struct drm_i915_private *dev_priv =
2823 container_of(work, typeof(*dev_priv), gt.retire_work.work);
2824 struct drm_device *dev = &dev_priv->drm;
2826 /* Come back later if the device is busy... */
2827 if (mutex_trylock(&dev->struct_mutex)) {
2828 i915_gem_retire_requests(dev_priv);
2829 mutex_unlock(&dev->struct_mutex);
2832 /* Keep the retire handler running until we are finally idle.
2833 * We do not need to do this test under locking as in the worst-case
2834 * we queue the retire worker once too often.
2836 if (READ_ONCE(dev_priv->gt.awake)) {
2837 i915_queue_hangcheck(dev_priv);
2838 queue_delayed_work(dev_priv->wq,
2839 &dev_priv->gt.retire_work,
2840 round_jiffies_up_relative(HZ));
2845 i915_gem_idle_work_handler(struct work_struct *work)
2847 struct drm_i915_private *dev_priv =
2848 container_of(work, typeof(*dev_priv), gt.idle_work.work);
2849 struct drm_device *dev = &dev_priv->drm;
2850 struct intel_engine_cs *engine;
2851 enum intel_engine_id id;
2852 bool rearm_hangcheck;
2854 if (!READ_ONCE(dev_priv->gt.awake))
2858 * Wait for last execlists context complete, but bail out in case a
2859 * new request is submitted.
2861 wait_for(READ_ONCE(dev_priv->gt.active_requests) ||
2862 intel_execlists_idle(dev_priv), 10);
2864 if (READ_ONCE(dev_priv->gt.active_requests))
2868 cancel_delayed_work_sync(&dev_priv->gpu_error.hangcheck_work);
2870 if (!mutex_trylock(&dev->struct_mutex)) {
2871 /* Currently busy, come back later */
2872 mod_delayed_work(dev_priv->wq,
2873 &dev_priv->gt.idle_work,
2874 msecs_to_jiffies(50));
2879 * New request retired after this work handler started, extend active
2880 * period until next instance of the work.
2882 if (work_pending(work))
2885 if (dev_priv->gt.active_requests)
2888 if (wait_for(intel_execlists_idle(dev_priv), 10))
2889 DRM_ERROR("Timeout waiting for engines to idle\n");
2891 for_each_engine(engine, dev_priv, id)
2892 i915_gem_batch_pool_fini(&engine->batch_pool);
2894 GEM_BUG_ON(!dev_priv->gt.awake);
2895 dev_priv->gt.awake = false;
2896 rearm_hangcheck = false;
2898 if (INTEL_GEN(dev_priv) >= 6)
2899 gen6_rps_idle(dev_priv);
2900 intel_runtime_pm_put(dev_priv);
2902 mutex_unlock(&dev->struct_mutex);
2905 if (rearm_hangcheck) {
2906 GEM_BUG_ON(!dev_priv->gt.awake);
2907 i915_queue_hangcheck(dev_priv);
2911 void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file)
2913 struct drm_i915_gem_object *obj = to_intel_bo(gem);
2914 struct drm_i915_file_private *fpriv = file->driver_priv;
2915 struct i915_vma *vma, *vn;
2917 mutex_lock(&obj->base.dev->struct_mutex);
2918 list_for_each_entry_safe(vma, vn, &obj->vma_list, obj_link)
2919 if (vma->vm->file == fpriv)
2920 i915_vma_close(vma);
2922 if (i915_gem_object_is_active(obj) &&
2923 !i915_gem_object_has_active_reference(obj)) {
2924 i915_gem_object_set_active_reference(obj);
2925 i915_gem_object_get(obj);
2927 mutex_unlock(&obj->base.dev->struct_mutex);
2930 static unsigned long to_wait_timeout(s64 timeout_ns)
2933 return MAX_SCHEDULE_TIMEOUT;
2935 if (timeout_ns == 0)
2938 return nsecs_to_jiffies_timeout(timeout_ns);
2942 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
2943 * @dev: drm device pointer
2944 * @data: ioctl data blob
2945 * @file: drm file pointer
2947 * Returns 0 if successful, else an error is returned with the remaining time in
2948 * the timeout parameter.
2949 * -ETIME: object is still busy after timeout
2950 * -ERESTARTSYS: signal interrupted the wait
2951 * -ENONENT: object doesn't exist
2952 * Also possible, but rare:
2953 * -EAGAIN: GPU wedged
2955 * -ENODEV: Internal IRQ fail
2956 * -E?: The add request failed
2958 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
2959 * non-zero timeout parameter the wait ioctl will wait for the given number of
2960 * nanoseconds on an object becoming unbusy. Since the wait itself does so
2961 * without holding struct_mutex the object may become re-busied before this
2962 * function completes. A similar but shorter * race condition exists in the busy
2966 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
2968 struct drm_i915_gem_wait *args = data;
2969 struct drm_i915_gem_object *obj;
2973 if (args->flags != 0)
2976 obj = i915_gem_object_lookup(file, args->bo_handle);
2980 start = ktime_get();
2982 ret = i915_gem_object_wait(obj,
2983 I915_WAIT_INTERRUPTIBLE | I915_WAIT_ALL,
2984 to_wait_timeout(args->timeout_ns),
2985 to_rps_client(file));
2987 if (args->timeout_ns > 0) {
2988 args->timeout_ns -= ktime_to_ns(ktime_sub(ktime_get(), start));
2989 if (args->timeout_ns < 0)
2990 args->timeout_ns = 0;
2993 i915_gem_object_put(obj);
2997 static int wait_for_timeline(struct i915_gem_timeline *tl, unsigned int flags)
3001 for (i = 0; i < ARRAY_SIZE(tl->engine); i++) {
3002 ret = i915_gem_active_wait(&tl->engine[i].last_request, flags);
3010 int i915_gem_wait_for_idle(struct drm_i915_private *i915, unsigned int flags)
3014 if (flags & I915_WAIT_LOCKED) {
3015 struct i915_gem_timeline *tl;
3017 lockdep_assert_held(&i915->drm.struct_mutex);
3019 list_for_each_entry(tl, &i915->gt.timelines, link) {
3020 ret = wait_for_timeline(tl, flags);
3025 ret = wait_for_timeline(&i915->gt.global_timeline, flags);
3033 void i915_gem_clflush_object(struct drm_i915_gem_object *obj,
3036 /* If we don't have a page list set up, then we're not pinned
3037 * to GPU, and we can ignore the cache flush because it'll happen
3038 * again at bind time.
3044 * Stolen memory is always coherent with the GPU as it is explicitly
3045 * marked as wc by the system, or the system is cache-coherent.
3047 if (obj->stolen || obj->phys_handle)
3050 /* If the GPU is snooping the contents of the CPU cache,
3051 * we do not need to manually clear the CPU cache lines. However,
3052 * the caches are only snooped when the render cache is
3053 * flushed/invalidated. As we always have to emit invalidations
3054 * and flushes when moving into and out of the RENDER domain, correct
3055 * snooping behaviour occurs naturally as the result of our domain
3058 if (!force && cpu_cache_is_coherent(obj->base.dev, obj->cache_level)) {
3059 obj->cache_dirty = true;
3063 trace_i915_gem_object_clflush(obj);
3064 drm_clflush_sg(obj->mm.pages);
3065 obj->cache_dirty = false;
3068 /** Flushes the GTT write domain for the object if it's dirty. */
3070 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
3072 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
3074 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
3077 /* No actual flushing is required for the GTT write domain. Writes
3078 * to it "immediately" go to main memory as far as we know, so there's
3079 * no chipset flush. It also doesn't land in render cache.
3081 * However, we do have to enforce the order so that all writes through
3082 * the GTT land before any writes to the device, such as updates to
3085 * We also have to wait a bit for the writes to land from the GTT.
3086 * An uncached read (i.e. mmio) seems to be ideal for the round-trip
3087 * timing. This issue has only been observed when switching quickly
3088 * between GTT writes and CPU reads from inside the kernel on recent hw,
3089 * and it appears to only affect discrete GTT blocks (i.e. on LLC
3090 * system agents we cannot reproduce this behaviour).
3093 if (INTEL_GEN(dev_priv) >= 6 && !HAS_LLC(dev_priv))
3094 POSTING_READ(RING_ACTHD(dev_priv->engine[RCS]->mmio_base));
3096 intel_fb_obj_flush(obj, false, write_origin(obj, I915_GEM_DOMAIN_GTT));
3098 obj->base.write_domain = 0;
3099 trace_i915_gem_object_change_domain(obj,
3100 obj->base.read_domains,
3101 I915_GEM_DOMAIN_GTT);
3104 /** Flushes the CPU write domain for the object if it's dirty. */
3106 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj)
3108 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
3111 i915_gem_clflush_object(obj, obj->pin_display);
3112 intel_fb_obj_flush(obj, false, ORIGIN_CPU);
3114 obj->base.write_domain = 0;
3115 trace_i915_gem_object_change_domain(obj,
3116 obj->base.read_domains,
3117 I915_GEM_DOMAIN_CPU);
3121 * Moves a single object to the GTT read, and possibly write domain.
3122 * @obj: object to act on
3123 * @write: ask for write access or read only
3125 * This function returns when the move is complete, including waiting on
3129 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
3131 uint32_t old_write_domain, old_read_domains;
3134 lockdep_assert_held(&obj->base.dev->struct_mutex);
3136 ret = i915_gem_object_wait(obj,
3137 I915_WAIT_INTERRUPTIBLE |
3139 (write ? I915_WAIT_ALL : 0),
3140 MAX_SCHEDULE_TIMEOUT,
3145 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
3148 /* Flush and acquire obj->pages so that we are coherent through
3149 * direct access in memory with previous cached writes through
3150 * shmemfs and that our cache domain tracking remains valid.
3151 * For example, if the obj->filp was moved to swap without us
3152 * being notified and releasing the pages, we would mistakenly
3153 * continue to assume that the obj remained out of the CPU cached
3156 ret = i915_gem_object_pin_pages(obj);
3160 i915_gem_object_flush_cpu_write_domain(obj);
3162 /* Serialise direct access to this object with the barriers for
3163 * coherent writes from the GPU, by effectively invalidating the
3164 * GTT domain upon first access.
3166 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
3169 old_write_domain = obj->base.write_domain;
3170 old_read_domains = obj->base.read_domains;
3172 /* It should now be out of any other write domains, and we can update
3173 * the domain values for our changes.
3175 GEM_BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3176 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3178 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
3179 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
3180 obj->mm.dirty = true;
3183 trace_i915_gem_object_change_domain(obj,
3187 i915_gem_object_unpin_pages(obj);
3192 * Changes the cache-level of an object across all VMA.
3193 * @obj: object to act on
3194 * @cache_level: new cache level to set for the object
3196 * After this function returns, the object will be in the new cache-level
3197 * across all GTT and the contents of the backing storage will be coherent,
3198 * with respect to the new cache-level. In order to keep the backing storage
3199 * coherent for all users, we only allow a single cache level to be set
3200 * globally on the object and prevent it from being changed whilst the
3201 * hardware is reading from the object. That is if the object is currently
3202 * on the scanout it will be set to uncached (or equivalent display
3203 * cache coherency) and all non-MOCS GPU access will also be uncached so
3204 * that all direct access to the scanout remains coherent.
3206 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
3207 enum i915_cache_level cache_level)
3209 struct i915_vma *vma;
3212 lockdep_assert_held(&obj->base.dev->struct_mutex);
3214 if (obj->cache_level == cache_level)
3217 /* Inspect the list of currently bound VMA and unbind any that would
3218 * be invalid given the new cache-level. This is principally to
3219 * catch the issue of the CS prefetch crossing page boundaries and
3220 * reading an invalid PTE on older architectures.
3223 list_for_each_entry(vma, &obj->vma_list, obj_link) {
3224 if (!drm_mm_node_allocated(&vma->node))
3227 if (i915_vma_is_pinned(vma)) {
3228 DRM_DEBUG("can not change the cache level of pinned objects\n");
3232 if (i915_gem_valid_gtt_space(vma, cache_level))
3235 ret = i915_vma_unbind(vma);
3239 /* As unbinding may affect other elements in the
3240 * obj->vma_list (due to side-effects from retiring
3241 * an active vma), play safe and restart the iterator.
3246 /* We can reuse the existing drm_mm nodes but need to change the
3247 * cache-level on the PTE. We could simply unbind them all and
3248 * rebind with the correct cache-level on next use. However since
3249 * we already have a valid slot, dma mapping, pages etc, we may as
3250 * rewrite the PTE in the belief that doing so tramples upon less
3251 * state and so involves less work.
3253 if (obj->bind_count) {
3254 /* Before we change the PTE, the GPU must not be accessing it.
3255 * If we wait upon the object, we know that all the bound
3256 * VMA are no longer active.
3258 ret = i915_gem_object_wait(obj,
3259 I915_WAIT_INTERRUPTIBLE |
3262 MAX_SCHEDULE_TIMEOUT,
3267 if (!HAS_LLC(to_i915(obj->base.dev)) &&
3268 cache_level != I915_CACHE_NONE) {
3269 /* Access to snoopable pages through the GTT is
3270 * incoherent and on some machines causes a hard
3271 * lockup. Relinquish the CPU mmaping to force
3272 * userspace to refault in the pages and we can
3273 * then double check if the GTT mapping is still
3274 * valid for that pointer access.
3276 i915_gem_release_mmap(obj);
3278 /* As we no longer need a fence for GTT access,
3279 * we can relinquish it now (and so prevent having
3280 * to steal a fence from someone else on the next
3281 * fence request). Note GPU activity would have
3282 * dropped the fence as all snoopable access is
3283 * supposed to be linear.
3285 list_for_each_entry(vma, &obj->vma_list, obj_link) {
3286 ret = i915_vma_put_fence(vma);
3291 /* We either have incoherent backing store and
3292 * so no GTT access or the architecture is fully
3293 * coherent. In such cases, existing GTT mmaps
3294 * ignore the cache bit in the PTE and we can
3295 * rewrite it without confusing the GPU or having
3296 * to force userspace to fault back in its mmaps.
3300 list_for_each_entry(vma, &obj->vma_list, obj_link) {
3301 if (!drm_mm_node_allocated(&vma->node))
3304 ret = i915_vma_bind(vma, cache_level, PIN_UPDATE);
3310 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU &&
3311 cpu_cache_is_coherent(obj->base.dev, obj->cache_level))
3312 obj->cache_dirty = true;
3314 list_for_each_entry(vma, &obj->vma_list, obj_link)
3315 vma->node.color = cache_level;
3316 obj->cache_level = cache_level;
3321 int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data,
3322 struct drm_file *file)
3324 struct drm_i915_gem_caching *args = data;
3325 struct drm_i915_gem_object *obj;
3329 obj = i915_gem_object_lookup_rcu(file, args->handle);
3335 switch (obj->cache_level) {
3336 case I915_CACHE_LLC:
3337 case I915_CACHE_L3_LLC:
3338 args->caching = I915_CACHING_CACHED;
3342 args->caching = I915_CACHING_DISPLAY;
3346 args->caching = I915_CACHING_NONE;
3354 int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
3355 struct drm_file *file)
3357 struct drm_i915_private *i915 = to_i915(dev);
3358 struct drm_i915_gem_caching *args = data;
3359 struct drm_i915_gem_object *obj;
3360 enum i915_cache_level level;
3363 switch (args->caching) {
3364 case I915_CACHING_NONE:
3365 level = I915_CACHE_NONE;
3367 case I915_CACHING_CACHED:
3369 * Due to a HW issue on BXT A stepping, GPU stores via a
3370 * snooped mapping may leave stale data in a corresponding CPU
3371 * cacheline, whereas normally such cachelines would get
3374 if (!HAS_LLC(i915) && !HAS_SNOOP(i915))
3377 level = I915_CACHE_LLC;
3379 case I915_CACHING_DISPLAY:
3380 level = HAS_WT(i915) ? I915_CACHE_WT : I915_CACHE_NONE;
3386 ret = i915_mutex_lock_interruptible(dev);
3390 obj = i915_gem_object_lookup(file, args->handle);
3396 ret = i915_gem_object_set_cache_level(obj, level);
3397 i915_gem_object_put(obj);
3399 mutex_unlock(&dev->struct_mutex);
3404 * Prepare buffer for display plane (scanout, cursors, etc).
3405 * Can be called from an uninterruptible phase (modesetting) and allows
3406 * any flushes to be pipelined (for pageflips).
3409 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
3411 const struct i915_ggtt_view *view)
3413 struct i915_vma *vma;
3414 u32 old_read_domains, old_write_domain;
3417 lockdep_assert_held(&obj->base.dev->struct_mutex);
3419 /* Mark the pin_display early so that we account for the
3420 * display coherency whilst setting up the cache domains.
3424 /* The display engine is not coherent with the LLC cache on gen6. As
3425 * a result, we make sure that the pinning that is about to occur is
3426 * done with uncached PTEs. This is lowest common denominator for all
3429 * However for gen6+, we could do better by using the GFDT bit instead
3430 * of uncaching, which would allow us to flush all the LLC-cached data
3431 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3433 ret = i915_gem_object_set_cache_level(obj,
3434 HAS_WT(to_i915(obj->base.dev)) ?
3435 I915_CACHE_WT : I915_CACHE_NONE);
3438 goto err_unpin_display;
3441 /* As the user may map the buffer once pinned in the display plane
3442 * (e.g. libkms for the bootup splash), we have to ensure that we
3443 * always use map_and_fenceable for all scanout buffers. However,
3444 * it may simply be too big to fit into mappable, in which case
3445 * put it anyway and hope that userspace can cope (but always first
3446 * try to preserve the existing ABI).
3448 vma = ERR_PTR(-ENOSPC);
3449 if (view->type == I915_GGTT_VIEW_NORMAL)
3450 vma = i915_gem_object_ggtt_pin(obj, view, 0, alignment,
3451 PIN_MAPPABLE | PIN_NONBLOCK);
3453 struct drm_i915_private *i915 = to_i915(obj->base.dev);
3456 /* Valleyview is definitely limited to scanning out the first
3457 * 512MiB. Lets presume this behaviour was inherited from the
3458 * g4x display engine and that all earlier gen are similarly
3459 * limited. Testing suggests that it is a little more
3460 * complicated than this. For example, Cherryview appears quite
3461 * happy to scanout from anywhere within its global aperture.
3464 if (HAS_GMCH_DISPLAY(i915))
3465 flags = PIN_MAPPABLE;
3466 vma = i915_gem_object_ggtt_pin(obj, view, 0, alignment, flags);
3469 goto err_unpin_display;
3471 vma->display_alignment = max_t(u64, vma->display_alignment, alignment);
3473 /* Treat this as an end-of-frame, like intel_user_framebuffer_dirty() */
3474 if (obj->cache_dirty) {
3475 i915_gem_clflush_object(obj, true);
3476 intel_fb_obj_flush(obj, false, ORIGIN_DIRTYFB);
3479 old_write_domain = obj->base.write_domain;
3480 old_read_domains = obj->base.read_domains;
3482 /* It should now be out of any other write domains, and we can update
3483 * the domain values for our changes.
3485 obj->base.write_domain = 0;
3486 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3488 trace_i915_gem_object_change_domain(obj,
3500 i915_gem_object_unpin_from_display_plane(struct i915_vma *vma)
3502 lockdep_assert_held(&vma->vm->dev->struct_mutex);
3504 if (WARN_ON(vma->obj->pin_display == 0))
3507 if (--vma->obj->pin_display == 0)
3508 vma->display_alignment = 0;
3510 /* Bump the LRU to try and avoid premature eviction whilst flipping */
3511 if (!i915_vma_is_active(vma))
3512 list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
3514 i915_vma_unpin(vma);
3518 * Moves a single object to the CPU read, and possibly write domain.
3519 * @obj: object to act on
3520 * @write: requesting write or read-only access
3522 * This function returns when the move is complete, including waiting on
3526 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
3528 uint32_t old_write_domain, old_read_domains;
3531 lockdep_assert_held(&obj->base.dev->struct_mutex);
3533 ret = i915_gem_object_wait(obj,
3534 I915_WAIT_INTERRUPTIBLE |
3536 (write ? I915_WAIT_ALL : 0),
3537 MAX_SCHEDULE_TIMEOUT,
3542 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
3545 i915_gem_object_flush_gtt_write_domain(obj);
3547 old_write_domain = obj->base.write_domain;
3548 old_read_domains = obj->base.read_domains;
3550 /* Flush the CPU cache if it's still invalid. */
3551 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
3552 i915_gem_clflush_object(obj, false);
3554 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3557 /* It should now be out of any other write domains, and we can update
3558 * the domain values for our changes.
3560 GEM_BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3562 /* If we're writing through the CPU, then the GPU read domains will
3563 * need to be invalidated at next use.
3566 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3567 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3570 trace_i915_gem_object_change_domain(obj,
3577 /* Throttle our rendering by waiting until the ring has completed our requests
3578 * emitted over 20 msec ago.
3580 * Note that if we were to use the current jiffies each time around the loop,
3581 * we wouldn't escape the function with any frames outstanding if the time to
3582 * render a frame was over 20ms.
3584 * This should get us reasonable parallelism between CPU and GPU but also
3585 * relatively low latency when blocking on a particular request to finish.
3588 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
3590 struct drm_i915_private *dev_priv = to_i915(dev);
3591 struct drm_i915_file_private *file_priv = file->driver_priv;
3592 unsigned long recent_enough = jiffies - DRM_I915_THROTTLE_JIFFIES;
3593 struct drm_i915_gem_request *request, *target = NULL;
3596 /* ABI: return -EIO if already wedged */
3597 if (i915_terminally_wedged(&dev_priv->gpu_error))
3600 spin_lock(&file_priv->mm.lock);
3601 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
3602 if (time_after_eq(request->emitted_jiffies, recent_enough))
3606 * Note that the request might not have been submitted yet.
3607 * In which case emitted_jiffies will be zero.
3609 if (!request->emitted_jiffies)
3615 i915_gem_request_get(target);
3616 spin_unlock(&file_priv->mm.lock);
3621 ret = i915_wait_request(target,
3622 I915_WAIT_INTERRUPTIBLE,
3623 MAX_SCHEDULE_TIMEOUT);
3624 i915_gem_request_put(target);
3626 return ret < 0 ? ret : 0;
3630 i915_gem_object_ggtt_pin(struct drm_i915_gem_object *obj,
3631 const struct i915_ggtt_view *view,
3636 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
3637 struct i915_address_space *vm = &dev_priv->ggtt.base;
3638 struct i915_vma *vma;
3641 lockdep_assert_held(&obj->base.dev->struct_mutex);
3643 vma = i915_gem_obj_lookup_or_create_vma(obj, vm, view);
3647 if (i915_vma_misplaced(vma, size, alignment, flags)) {
3648 if (flags & PIN_NONBLOCK &&
3649 (i915_vma_is_pinned(vma) || i915_vma_is_active(vma)))
3650 return ERR_PTR(-ENOSPC);
3652 if (flags & PIN_MAPPABLE) {
3655 fence_size = i915_gem_get_ggtt_size(dev_priv, vma->size,
3656 i915_gem_object_get_tiling(obj));
3657 /* If the required space is larger than the available
3658 * aperture, we will not able to find a slot for the
3659 * object and unbinding the object now will be in
3660 * vain. Worse, doing so may cause us to ping-pong
3661 * the object in and out of the Global GTT and
3662 * waste a lot of cycles under the mutex.
3664 if (fence_size > dev_priv->ggtt.mappable_end)
3665 return ERR_PTR(-E2BIG);
3667 /* If NONBLOCK is set the caller is optimistically
3668 * trying to cache the full object within the mappable
3669 * aperture, and *must* have a fallback in place for
3670 * situations where we cannot bind the object. We
3671 * can be a little more lax here and use the fallback
3672 * more often to avoid costly migrations of ourselves
3673 * and other objects within the aperture.
3675 * Half-the-aperture is used as a simple heuristic.
3676 * More interesting would to do search for a free
3677 * block prior to making the commitment to unbind.
3678 * That caters for the self-harm case, and with a
3679 * little more heuristics (e.g. NOFAULT, NOEVICT)
3680 * we could try to minimise harm to others.
3682 if (flags & PIN_NONBLOCK &&
3683 fence_size > dev_priv->ggtt.mappable_end / 2)
3684 return ERR_PTR(-ENOSPC);
3687 WARN(i915_vma_is_pinned(vma),
3688 "bo is already pinned in ggtt with incorrect alignment:"
3689 " offset=%08x, req.alignment=%llx,"
3690 " req.map_and_fenceable=%d, vma->map_and_fenceable=%d\n",
3691 i915_ggtt_offset(vma), alignment,
3692 !!(flags & PIN_MAPPABLE),
3693 i915_vma_is_map_and_fenceable(vma));
3694 ret = i915_vma_unbind(vma);
3696 return ERR_PTR(ret);
3699 ret = i915_vma_pin(vma, size, alignment, flags | PIN_GLOBAL);
3701 return ERR_PTR(ret);
3706 static __always_inline unsigned int __busy_read_flag(unsigned int id)
3708 /* Note that we could alias engines in the execbuf API, but
3709 * that would be very unwise as it prevents userspace from
3710 * fine control over engine selection. Ahem.
3712 * This should be something like EXEC_MAX_ENGINE instead of
3715 BUILD_BUG_ON(I915_NUM_ENGINES > 16);
3716 return 0x10000 << id;
3719 static __always_inline unsigned int __busy_write_id(unsigned int id)
3721 /* The uABI guarantees an active writer is also amongst the read
3722 * engines. This would be true if we accessed the activity tracking
3723 * under the lock, but as we perform the lookup of the object and
3724 * its activity locklessly we can not guarantee that the last_write
3725 * being active implies that we have set the same engine flag from
3726 * last_read - hence we always set both read and write busy for
3729 return id | __busy_read_flag(id);
3732 static __always_inline unsigned int
3733 __busy_set_if_active(const struct dma_fence *fence,
3734 unsigned int (*flag)(unsigned int id))
3736 struct drm_i915_gem_request *rq;
3738 /* We have to check the current hw status of the fence as the uABI
3739 * guarantees forward progress. We could rely on the idle worker
3740 * to eventually flush us, but to minimise latency just ask the
3743 * Note we only report on the status of native fences.
3745 if (!dma_fence_is_i915(fence))
3748 /* opencode to_request() in order to avoid const warnings */
3749 rq = container_of(fence, struct drm_i915_gem_request, fence);
3750 if (i915_gem_request_completed(rq))
3753 return flag(rq->engine->exec_id);
3756 static __always_inline unsigned int
3757 busy_check_reader(const struct dma_fence *fence)
3759 return __busy_set_if_active(fence, __busy_read_flag);
3762 static __always_inline unsigned int
3763 busy_check_writer(const struct dma_fence *fence)
3768 return __busy_set_if_active(fence, __busy_write_id);
3772 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
3773 struct drm_file *file)
3775 struct drm_i915_gem_busy *args = data;
3776 struct drm_i915_gem_object *obj;
3777 struct reservation_object_list *list;
3783 obj = i915_gem_object_lookup_rcu(file, args->handle);
3787 /* A discrepancy here is that we do not report the status of
3788 * non-i915 fences, i.e. even though we may report the object as idle,
3789 * a call to set-domain may still stall waiting for foreign rendering.
3790 * This also means that wait-ioctl may report an object as busy,
3791 * where busy-ioctl considers it idle.
3793 * We trade the ability to warn of foreign fences to report on which
3794 * i915 engines are active for the object.
3796 * Alternatively, we can trade that extra information on read/write
3799 * !reservation_object_test_signaled_rcu(obj->resv, true);
3800 * to report the overall busyness. This is what the wait-ioctl does.
3804 seq = raw_read_seqcount(&obj->resv->seq);
3806 /* Translate the exclusive fence to the READ *and* WRITE engine */
3807 args->busy = busy_check_writer(rcu_dereference(obj->resv->fence_excl));
3809 /* Translate shared fences to READ set of engines */
3810 list = rcu_dereference(obj->resv->fence);
3812 unsigned int shared_count = list->shared_count, i;
3814 for (i = 0; i < shared_count; ++i) {
3815 struct dma_fence *fence =
3816 rcu_dereference(list->shared[i]);
3818 args->busy |= busy_check_reader(fence);
3822 if (args->busy && read_seqcount_retry(&obj->resv->seq, seq))
3832 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
3833 struct drm_file *file_priv)
3835 return i915_gem_ring_throttle(dev, file_priv);
3839 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
3840 struct drm_file *file_priv)
3842 struct drm_i915_private *dev_priv = to_i915(dev);
3843 struct drm_i915_gem_madvise *args = data;
3844 struct drm_i915_gem_object *obj;
3847 switch (args->madv) {
3848 case I915_MADV_DONTNEED:
3849 case I915_MADV_WILLNEED:
3855 obj = i915_gem_object_lookup(file_priv, args->handle);
3859 err = mutex_lock_interruptible(&obj->mm.lock);
3863 if (obj->mm.pages &&
3864 i915_gem_object_is_tiled(obj) &&
3865 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
3866 if (obj->mm.madv == I915_MADV_WILLNEED) {
3867 GEM_BUG_ON(!obj->mm.quirked);
3868 __i915_gem_object_unpin_pages(obj);
3869 obj->mm.quirked = false;
3871 if (args->madv == I915_MADV_WILLNEED) {
3872 GEM_BUG_ON(obj->mm.quirked);
3873 __i915_gem_object_pin_pages(obj);
3874 obj->mm.quirked = true;
3878 if (obj->mm.madv != __I915_MADV_PURGED)
3879 obj->mm.madv = args->madv;
3881 /* if the object is no longer attached, discard its backing storage */
3882 if (obj->mm.madv == I915_MADV_DONTNEED && !obj->mm.pages)
3883 i915_gem_object_truncate(obj);
3885 args->retained = obj->mm.madv != __I915_MADV_PURGED;
3886 mutex_unlock(&obj->mm.lock);
3889 i915_gem_object_put(obj);
3894 frontbuffer_retire(struct i915_gem_active *active,
3895 struct drm_i915_gem_request *request)
3897 struct drm_i915_gem_object *obj =
3898 container_of(active, typeof(*obj), frontbuffer_write);
3900 intel_fb_obj_flush(obj, true, ORIGIN_CS);
3903 void i915_gem_object_init(struct drm_i915_gem_object *obj,
3904 const struct drm_i915_gem_object_ops *ops)
3906 mutex_init(&obj->mm.lock);
3908 INIT_LIST_HEAD(&obj->global_link);
3909 INIT_LIST_HEAD(&obj->userfault_link);
3910 INIT_LIST_HEAD(&obj->obj_exec_link);
3911 INIT_LIST_HEAD(&obj->vma_list);
3912 INIT_LIST_HEAD(&obj->batch_pool_link);
3916 reservation_object_init(&obj->__builtin_resv);
3917 obj->resv = &obj->__builtin_resv;
3919 obj->frontbuffer_ggtt_origin = ORIGIN_GTT;
3920 init_request_active(&obj->frontbuffer_write, frontbuffer_retire);
3922 obj->mm.madv = I915_MADV_WILLNEED;
3923 INIT_RADIX_TREE(&obj->mm.get_page.radix, GFP_KERNEL | __GFP_NOWARN);
3924 mutex_init(&obj->mm.get_page.lock);
3926 i915_gem_info_add_obj(to_i915(obj->base.dev), obj->base.size);
3929 static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
3930 .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE |
3931 I915_GEM_OBJECT_IS_SHRINKABLE,
3932 .get_pages = i915_gem_object_get_pages_gtt,
3933 .put_pages = i915_gem_object_put_pages_gtt,
3936 /* Note we don't consider signbits :| */
3937 #define overflows_type(x, T) \
3938 (sizeof(x) > sizeof(T) && (x) >> (sizeof(T) * BITS_PER_BYTE))
3940 struct drm_i915_gem_object *
3941 i915_gem_object_create(struct drm_device *dev, u64 size)
3943 struct drm_i915_private *dev_priv = to_i915(dev);
3944 struct drm_i915_gem_object *obj;
3945 struct address_space *mapping;
3949 /* There is a prevalence of the assumption that we fit the object's
3950 * page count inside a 32bit _signed_ variable. Let's document this and
3951 * catch if we ever need to fix it. In the meantime, if you do spot
3952 * such a local variable, please consider fixing!
3954 if (WARN_ON(size >> PAGE_SHIFT > INT_MAX))
3955 return ERR_PTR(-E2BIG);
3957 if (overflows_type(size, obj->base.size))
3958 return ERR_PTR(-E2BIG);
3960 obj = i915_gem_object_alloc(dev);
3962 return ERR_PTR(-ENOMEM);
3964 ret = drm_gem_object_init(dev, &obj->base, size);
3968 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
3969 if (IS_CRESTLINE(dev_priv) || IS_BROADWATER(dev_priv)) {
3970 /* 965gm cannot relocate objects above 4GiB. */
3971 mask &= ~__GFP_HIGHMEM;
3972 mask |= __GFP_DMA32;
3975 mapping = obj->base.filp->f_mapping;
3976 mapping_set_gfp_mask(mapping, mask);
3978 i915_gem_object_init(obj, &i915_gem_object_ops);
3980 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3981 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3983 if (HAS_LLC(dev_priv)) {
3984 /* On some devices, we can have the GPU use the LLC (the CPU
3985 * cache) for about a 10% performance improvement
3986 * compared to uncached. Graphics requests other than
3987 * display scanout are coherent with the CPU in
3988 * accessing this cache. This means in this mode we
3989 * don't need to clflush on the CPU side, and on the
3990 * GPU side we only need to flush internal caches to
3991 * get data visible to the CPU.
3993 * However, we maintain the display planes as UC, and so
3994 * need to rebind when first used as such.
3996 obj->cache_level = I915_CACHE_LLC;
3998 obj->cache_level = I915_CACHE_NONE;
4000 trace_i915_gem_object_create(obj);
4005 i915_gem_object_free(obj);
4006 return ERR_PTR(ret);
4009 static bool discard_backing_storage(struct drm_i915_gem_object *obj)
4011 /* If we are the last user of the backing storage (be it shmemfs
4012 * pages or stolen etc), we know that the pages are going to be
4013 * immediately released. In this case, we can then skip copying
4014 * back the contents from the GPU.
4017 if (obj->mm.madv != I915_MADV_WILLNEED)
4020 if (obj->base.filp == NULL)
4023 /* At first glance, this looks racy, but then again so would be
4024 * userspace racing mmap against close. However, the first external
4025 * reference to the filp can only be obtained through the
4026 * i915_gem_mmap_ioctl() which safeguards us against the user
4027 * acquiring such a reference whilst we are in the middle of
4028 * freeing the object.
4030 return atomic_long_read(&obj->base.filp->f_count) == 1;
4033 static void __i915_gem_free_objects(struct drm_i915_private *i915,
4034 struct llist_node *freed)
4036 struct drm_i915_gem_object *obj, *on;
4038 mutex_lock(&i915->drm.struct_mutex);
4039 intel_runtime_pm_get(i915);
4040 llist_for_each_entry(obj, freed, freed) {
4041 struct i915_vma *vma, *vn;
4043 trace_i915_gem_object_destroy(obj);
4045 GEM_BUG_ON(i915_gem_object_is_active(obj));
4046 list_for_each_entry_safe(vma, vn,
4047 &obj->vma_list, obj_link) {
4048 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
4049 GEM_BUG_ON(i915_vma_is_active(vma));
4050 vma->flags &= ~I915_VMA_PIN_MASK;
4051 i915_vma_close(vma);
4053 GEM_BUG_ON(!list_empty(&obj->vma_list));
4054 GEM_BUG_ON(!RB_EMPTY_ROOT(&obj->vma_tree));
4056 list_del(&obj->global_link);
4058 intel_runtime_pm_put(i915);
4059 mutex_unlock(&i915->drm.struct_mutex);
4061 llist_for_each_entry_safe(obj, on, freed, freed) {
4062 GEM_BUG_ON(obj->bind_count);
4063 GEM_BUG_ON(atomic_read(&obj->frontbuffer_bits));
4065 if (obj->ops->release)
4066 obj->ops->release(obj);
4068 if (WARN_ON(i915_gem_object_has_pinned_pages(obj)))
4069 atomic_set(&obj->mm.pages_pin_count, 0);
4070 __i915_gem_object_put_pages(obj, I915_MM_NORMAL);
4071 GEM_BUG_ON(obj->mm.pages);
4073 if (obj->base.import_attach)
4074 drm_prime_gem_destroy(&obj->base, NULL);
4076 reservation_object_fini(&obj->__builtin_resv);
4077 drm_gem_object_release(&obj->base);
4078 i915_gem_info_remove_obj(i915, obj->base.size);
4081 i915_gem_object_free(obj);
4085 static void i915_gem_flush_free_objects(struct drm_i915_private *i915)
4087 struct llist_node *freed;
4089 freed = llist_del_all(&i915->mm.free_list);
4090 if (unlikely(freed))
4091 __i915_gem_free_objects(i915, freed);
4094 static void __i915_gem_free_work(struct work_struct *work)
4096 struct drm_i915_private *i915 =
4097 container_of(work, struct drm_i915_private, mm.free_work);
4098 struct llist_node *freed;
4100 /* All file-owned VMA should have been released by this point through
4101 * i915_gem_close_object(), or earlier by i915_gem_context_close().
4102 * However, the object may also be bound into the global GTT (e.g.
4103 * older GPUs without per-process support, or for direct access through
4104 * the GTT either for the user or for scanout). Those VMA still need to
4108 while ((freed = llist_del_all(&i915->mm.free_list)))
4109 __i915_gem_free_objects(i915, freed);
4112 static void __i915_gem_free_object_rcu(struct rcu_head *head)
4114 struct drm_i915_gem_object *obj =
4115 container_of(head, typeof(*obj), rcu);
4116 struct drm_i915_private *i915 = to_i915(obj->base.dev);
4118 /* We can't simply use call_rcu() from i915_gem_free_object()
4119 * as we need to block whilst unbinding, and the call_rcu
4120 * task may be called from softirq context. So we take a
4121 * detour through a worker.
4123 if (llist_add(&obj->freed, &i915->mm.free_list))
4124 schedule_work(&i915->mm.free_work);
4127 void i915_gem_free_object(struct drm_gem_object *gem_obj)
4129 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
4131 if (obj->mm.quirked)
4132 __i915_gem_object_unpin_pages(obj);
4134 if (discard_backing_storage(obj))
4135 obj->mm.madv = I915_MADV_DONTNEED;
4137 /* Before we free the object, make sure any pure RCU-only
4138 * read-side critical sections are complete, e.g.
4139 * i915_gem_busy_ioctl(). For the corresponding synchronized
4140 * lookup see i915_gem_object_lookup_rcu().
4142 call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
4145 void __i915_gem_object_release_unless_active(struct drm_i915_gem_object *obj)
4147 lockdep_assert_held(&obj->base.dev->struct_mutex);
4149 GEM_BUG_ON(i915_gem_object_has_active_reference(obj));
4150 if (i915_gem_object_is_active(obj))
4151 i915_gem_object_set_active_reference(obj);
4153 i915_gem_object_put(obj);
4156 static void assert_kernel_context_is_current(struct drm_i915_private *dev_priv)
4158 struct intel_engine_cs *engine;
4159 enum intel_engine_id id;
4161 for_each_engine(engine, dev_priv, id)
4162 GEM_BUG_ON(engine->last_context != dev_priv->kernel_context);
4165 int i915_gem_suspend(struct drm_device *dev)
4167 struct drm_i915_private *dev_priv = to_i915(dev);
4170 intel_suspend_gt_powersave(dev_priv);
4172 mutex_lock(&dev->struct_mutex);
4174 /* We have to flush all the executing contexts to main memory so
4175 * that they can saved in the hibernation image. To ensure the last
4176 * context image is coherent, we have to switch away from it. That
4177 * leaves the dev_priv->kernel_context still active when
4178 * we actually suspend, and its image in memory may not match the GPU
4179 * state. Fortunately, the kernel_context is disposable and we do
4180 * not rely on its state.
4182 ret = i915_gem_switch_to_kernel_context(dev_priv);
4186 ret = i915_gem_wait_for_idle(dev_priv,
4187 I915_WAIT_INTERRUPTIBLE |
4192 i915_gem_retire_requests(dev_priv);
4193 GEM_BUG_ON(dev_priv->gt.active_requests);
4195 assert_kernel_context_is_current(dev_priv);
4196 i915_gem_context_lost(dev_priv);
4197 mutex_unlock(&dev->struct_mutex);
4199 cancel_delayed_work_sync(&dev_priv->gpu_error.hangcheck_work);
4200 cancel_delayed_work_sync(&dev_priv->gt.retire_work);
4201 flush_delayed_work(&dev_priv->gt.idle_work);
4202 flush_work(&dev_priv->mm.free_work);
4204 /* Assert that we sucessfully flushed all the work and
4205 * reset the GPU back to its idle, low power state.
4207 WARN_ON(dev_priv->gt.awake);
4208 WARN_ON(!intel_execlists_idle(dev_priv));
4211 * Neither the BIOS, ourselves or any other kernel
4212 * expects the system to be in execlists mode on startup,
4213 * so we need to reset the GPU back to legacy mode. And the only
4214 * known way to disable logical contexts is through a GPU reset.
4216 * So in order to leave the system in a known default configuration,
4217 * always reset the GPU upon unload and suspend. Afterwards we then
4218 * clean up the GEM state tracking, flushing off the requests and
4219 * leaving the system in a known idle state.
4221 * Note that is of the upmost importance that the GPU is idle and
4222 * all stray writes are flushed *before* we dismantle the backing
4223 * storage for the pinned objects.
4225 * However, since we are uncertain that resetting the GPU on older
4226 * machines is a good idea, we don't - just in case it leaves the
4227 * machine in an unusable condition.
4229 if (HAS_HW_CONTEXTS(dev_priv)) {
4230 int reset = intel_gpu_reset(dev_priv, ALL_ENGINES);
4231 WARN_ON(reset && reset != -ENODEV);
4237 mutex_unlock(&dev->struct_mutex);
4241 void i915_gem_resume(struct drm_device *dev)
4243 struct drm_i915_private *dev_priv = to_i915(dev);
4245 WARN_ON(dev_priv->gt.awake);
4247 mutex_lock(&dev->struct_mutex);
4248 i915_gem_restore_gtt_mappings(dev_priv);
4250 /* As we didn't flush the kernel context before suspend, we cannot
4251 * guarantee that the context image is complete. So let's just reset
4252 * it and start again.
4254 dev_priv->gt.resume(dev_priv);
4256 mutex_unlock(&dev->struct_mutex);
4259 void i915_gem_init_swizzling(struct drm_i915_private *dev_priv)
4261 if (INTEL_GEN(dev_priv) < 5 ||
4262 dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
4265 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
4266 DISP_TILE_SURFACE_SWIZZLING);
4268 if (IS_GEN5(dev_priv))
4271 I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
4272 if (IS_GEN6(dev_priv))
4273 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
4274 else if (IS_GEN7(dev_priv))
4275 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
4276 else if (IS_GEN8(dev_priv))
4277 I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW));
4282 static void init_unused_ring(struct drm_i915_private *dev_priv, u32 base)
4284 I915_WRITE(RING_CTL(base), 0);
4285 I915_WRITE(RING_HEAD(base), 0);
4286 I915_WRITE(RING_TAIL(base), 0);
4287 I915_WRITE(RING_START(base), 0);
4290 static void init_unused_rings(struct drm_i915_private *dev_priv)
4292 if (IS_I830(dev_priv)) {
4293 init_unused_ring(dev_priv, PRB1_BASE);
4294 init_unused_ring(dev_priv, SRB0_BASE);
4295 init_unused_ring(dev_priv, SRB1_BASE);
4296 init_unused_ring(dev_priv, SRB2_BASE);
4297 init_unused_ring(dev_priv, SRB3_BASE);
4298 } else if (IS_GEN2(dev_priv)) {
4299 init_unused_ring(dev_priv, SRB0_BASE);
4300 init_unused_ring(dev_priv, SRB1_BASE);
4301 } else if (IS_GEN3(dev_priv)) {
4302 init_unused_ring(dev_priv, PRB1_BASE);
4303 init_unused_ring(dev_priv, PRB2_BASE);
4308 i915_gem_init_hw(struct drm_device *dev)
4310 struct drm_i915_private *dev_priv = to_i915(dev);
4311 struct intel_engine_cs *engine;
4312 enum intel_engine_id id;
4315 dev_priv->gt.last_init_time = ktime_get();
4317 /* Double layer security blanket, see i915_gem_init() */
4318 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4320 if (HAS_EDRAM(dev_priv) && INTEL_GEN(dev_priv) < 9)
4321 I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf));
4323 if (IS_HASWELL(dev_priv))
4324 I915_WRITE(MI_PREDICATE_RESULT_2, IS_HSW_GT3(dev_priv) ?
4325 LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED);
4327 if (HAS_PCH_NOP(dev_priv)) {
4328 if (IS_IVYBRIDGE(dev_priv)) {
4329 u32 temp = I915_READ(GEN7_MSG_CTL);
4330 temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK);
4331 I915_WRITE(GEN7_MSG_CTL, temp);
4332 } else if (INTEL_GEN(dev_priv) >= 7) {
4333 u32 temp = I915_READ(HSW_NDE_RSTWRN_OPT);
4334 temp &= ~RESET_PCH_HANDSHAKE_ENABLE;
4335 I915_WRITE(HSW_NDE_RSTWRN_OPT, temp);
4339 i915_gem_init_swizzling(dev_priv);
4342 * At least 830 can leave some of the unused rings
4343 * "active" (ie. head != tail) after resume which
4344 * will prevent c3 entry. Makes sure all unused rings
4347 init_unused_rings(dev_priv);
4349 BUG_ON(!dev_priv->kernel_context);
4351 ret = i915_ppgtt_init_hw(dev_priv);
4353 DRM_ERROR("PPGTT enable HW failed %d\n", ret);
4357 /* Need to do basic initialisation of all rings first: */
4358 for_each_engine(engine, dev_priv, id) {
4359 ret = engine->init_hw(engine);
4364 intel_mocs_init_l3cc_table(dev);
4366 /* We can't enable contexts until all firmware is loaded */
4367 ret = intel_guc_setup(dev);
4372 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4376 bool intel_sanitize_semaphores(struct drm_i915_private *dev_priv, int value)
4378 if (INTEL_INFO(dev_priv)->gen < 6)
4381 /* TODO: make semaphores and Execlists play nicely together */
4382 if (i915.enable_execlists)
4388 #ifdef CONFIG_INTEL_IOMMU
4389 /* Enable semaphores on SNB when IO remapping is off */
4390 if (INTEL_INFO(dev_priv)->gen == 6 && intel_iommu_gfx_mapped)
4397 int i915_gem_init(struct drm_device *dev)
4399 struct drm_i915_private *dev_priv = to_i915(dev);
4402 mutex_lock(&dev->struct_mutex);
4404 if (!i915.enable_execlists) {
4405 dev_priv->gt.resume = intel_legacy_submission_resume;
4406 dev_priv->gt.cleanup_engine = intel_engine_cleanup;
4408 dev_priv->gt.resume = intel_lr_context_resume;
4409 dev_priv->gt.cleanup_engine = intel_logical_ring_cleanup;
4412 /* This is just a security blanket to placate dragons.
4413 * On some systems, we very sporadically observe that the first TLBs
4414 * used by the CS may be stale, despite us poking the TLB reset. If
4415 * we hold the forcewake during initialisation these problems
4416 * just magically go away.
4418 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4420 i915_gem_init_userptr(dev_priv);
4422 ret = i915_gem_init_ggtt(dev_priv);
4426 ret = i915_gem_context_init(dev);
4430 ret = intel_engines_init(dev);
4434 ret = i915_gem_init_hw(dev);
4436 /* Allow engine initialisation to fail by marking the GPU as
4437 * wedged. But we only want to do this where the GPU is angry,
4438 * for all other failure, such as an allocation failure, bail.
4440 DRM_ERROR("Failed to initialize GPU, declaring it wedged\n");
4441 i915_gem_set_wedged(dev_priv);
4446 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4447 mutex_unlock(&dev->struct_mutex);
4453 i915_gem_cleanup_engines(struct drm_device *dev)
4455 struct drm_i915_private *dev_priv = to_i915(dev);
4456 struct intel_engine_cs *engine;
4457 enum intel_engine_id id;
4459 for_each_engine(engine, dev_priv, id)
4460 dev_priv->gt.cleanup_engine(engine);
4464 i915_gem_load_init_fences(struct drm_i915_private *dev_priv)
4468 if (INTEL_INFO(dev_priv)->gen >= 7 && !IS_VALLEYVIEW(dev_priv) &&
4469 !IS_CHERRYVIEW(dev_priv))
4470 dev_priv->num_fence_regs = 32;
4471 else if (INTEL_INFO(dev_priv)->gen >= 4 || IS_I945G(dev_priv) ||
4472 IS_I945GM(dev_priv) || IS_G33(dev_priv))
4473 dev_priv->num_fence_regs = 16;
4475 dev_priv->num_fence_regs = 8;
4477 if (intel_vgpu_active(dev_priv))
4478 dev_priv->num_fence_regs =
4479 I915_READ(vgtif_reg(avail_rs.fence_num));
4481 /* Initialize fence registers to zero */
4482 for (i = 0; i < dev_priv->num_fence_regs; i++) {
4483 struct drm_i915_fence_reg *fence = &dev_priv->fence_regs[i];
4485 fence->i915 = dev_priv;
4487 list_add_tail(&fence->link, &dev_priv->mm.fence_list);
4489 i915_gem_restore_fences(dev_priv);
4491 i915_gem_detect_bit_6_swizzle(dev_priv);
4495 i915_gem_load_init(struct drm_device *dev)
4497 struct drm_i915_private *dev_priv = to_i915(dev);
4500 dev_priv->objects = KMEM_CACHE(drm_i915_gem_object, SLAB_HWCACHE_ALIGN);
4501 if (!dev_priv->objects)
4504 dev_priv->vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
4505 if (!dev_priv->vmas)
4508 dev_priv->requests = KMEM_CACHE(drm_i915_gem_request,
4509 SLAB_HWCACHE_ALIGN |
4510 SLAB_RECLAIM_ACCOUNT |
4511 SLAB_DESTROY_BY_RCU);
4512 if (!dev_priv->requests)
4515 dev_priv->dependencies = KMEM_CACHE(i915_dependency,
4516 SLAB_HWCACHE_ALIGN |
4517 SLAB_RECLAIM_ACCOUNT);
4518 if (!dev_priv->dependencies)
4521 mutex_lock(&dev_priv->drm.struct_mutex);
4522 INIT_LIST_HEAD(&dev_priv->gt.timelines);
4523 err = i915_gem_timeline_init__global(dev_priv);
4524 mutex_unlock(&dev_priv->drm.struct_mutex);
4526 goto err_dependencies;
4528 INIT_LIST_HEAD(&dev_priv->context_list);
4529 INIT_WORK(&dev_priv->mm.free_work, __i915_gem_free_work);
4530 init_llist_head(&dev_priv->mm.free_list);
4531 INIT_LIST_HEAD(&dev_priv->mm.unbound_list);
4532 INIT_LIST_HEAD(&dev_priv->mm.bound_list);
4533 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4534 INIT_LIST_HEAD(&dev_priv->mm.userfault_list);
4535 INIT_DELAYED_WORK(&dev_priv->gt.retire_work,
4536 i915_gem_retire_work_handler);
4537 INIT_DELAYED_WORK(&dev_priv->gt.idle_work,
4538 i915_gem_idle_work_handler);
4539 init_waitqueue_head(&dev_priv->gpu_error.wait_queue);
4540 init_waitqueue_head(&dev_priv->gpu_error.reset_queue);
4542 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
4544 init_waitqueue_head(&dev_priv->pending_flip_queue);
4546 dev_priv->mm.interruptible = true;
4548 atomic_set(&dev_priv->mm.bsd_engine_dispatch_index, 0);
4550 spin_lock_init(&dev_priv->fb_tracking.lock);
4555 kmem_cache_destroy(dev_priv->dependencies);
4557 kmem_cache_destroy(dev_priv->requests);
4559 kmem_cache_destroy(dev_priv->vmas);
4561 kmem_cache_destroy(dev_priv->objects);
4566 void i915_gem_load_cleanup(struct drm_device *dev)
4568 struct drm_i915_private *dev_priv = to_i915(dev);
4570 WARN_ON(!llist_empty(&dev_priv->mm.free_list));
4572 mutex_lock(&dev_priv->drm.struct_mutex);
4573 i915_gem_timeline_fini(&dev_priv->gt.global_timeline);
4574 WARN_ON(!list_empty(&dev_priv->gt.timelines));
4575 mutex_unlock(&dev_priv->drm.struct_mutex);
4577 kmem_cache_destroy(dev_priv->dependencies);
4578 kmem_cache_destroy(dev_priv->requests);
4579 kmem_cache_destroy(dev_priv->vmas);
4580 kmem_cache_destroy(dev_priv->objects);
4582 /* And ensure that our DESTROY_BY_RCU slabs are truly destroyed */
4586 int i915_gem_freeze(struct drm_i915_private *dev_priv)
4588 intel_runtime_pm_get(dev_priv);
4590 mutex_lock(&dev_priv->drm.struct_mutex);
4591 i915_gem_shrink_all(dev_priv);
4592 mutex_unlock(&dev_priv->drm.struct_mutex);
4594 intel_runtime_pm_put(dev_priv);
4599 int i915_gem_freeze_late(struct drm_i915_private *dev_priv)
4601 struct drm_i915_gem_object *obj;
4602 struct list_head *phases[] = {
4603 &dev_priv->mm.unbound_list,
4604 &dev_priv->mm.bound_list,
4608 /* Called just before we write the hibernation image.
4610 * We need to update the domain tracking to reflect that the CPU
4611 * will be accessing all the pages to create and restore from the
4612 * hibernation, and so upon restoration those pages will be in the
4615 * To make sure the hibernation image contains the latest state,
4616 * we update that state just before writing out the image.
4618 * To try and reduce the hibernation image, we manually shrink
4619 * the objects as well.
4622 mutex_lock(&dev_priv->drm.struct_mutex);
4623 i915_gem_shrink(dev_priv, -1UL, I915_SHRINK_UNBOUND);
4625 for (p = phases; *p; p++) {
4626 list_for_each_entry(obj, *p, global_link) {
4627 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4628 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4631 mutex_unlock(&dev_priv->drm.struct_mutex);
4636 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
4638 struct drm_i915_file_private *file_priv = file->driver_priv;
4639 struct drm_i915_gem_request *request;
4641 /* Clean up our request list when the client is going away, so that
4642 * later retire_requests won't dereference our soon-to-be-gone
4645 spin_lock(&file_priv->mm.lock);
4646 list_for_each_entry(request, &file_priv->mm.request_list, client_list)
4647 request->file_priv = NULL;
4648 spin_unlock(&file_priv->mm.lock);
4650 if (!list_empty(&file_priv->rps.link)) {
4651 spin_lock(&to_i915(dev)->rps.client_lock);
4652 list_del(&file_priv->rps.link);
4653 spin_unlock(&to_i915(dev)->rps.client_lock);
4657 int i915_gem_open(struct drm_device *dev, struct drm_file *file)
4659 struct drm_i915_file_private *file_priv;
4664 file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL);
4668 file->driver_priv = file_priv;
4669 file_priv->dev_priv = to_i915(dev);
4670 file_priv->file = file;
4671 INIT_LIST_HEAD(&file_priv->rps.link);
4673 spin_lock_init(&file_priv->mm.lock);
4674 INIT_LIST_HEAD(&file_priv->mm.request_list);
4676 file_priv->bsd_engine = -1;
4678 ret = i915_gem_context_open(dev, file);
4686 * i915_gem_track_fb - update frontbuffer tracking
4687 * @old: current GEM buffer for the frontbuffer slots
4688 * @new: new GEM buffer for the frontbuffer slots
4689 * @frontbuffer_bits: bitmask of frontbuffer slots
4691 * This updates the frontbuffer tracking bits @frontbuffer_bits by clearing them
4692 * from @old and setting them in @new. Both @old and @new can be NULL.
4694 void i915_gem_track_fb(struct drm_i915_gem_object *old,
4695 struct drm_i915_gem_object *new,
4696 unsigned frontbuffer_bits)
4698 /* Control of individual bits within the mask are guarded by
4699 * the owning plane->mutex, i.e. we can never see concurrent
4700 * manipulation of individual bits. But since the bitfield as a whole
4701 * is updated using RMW, we need to use atomics in order to update
4704 BUILD_BUG_ON(INTEL_FRONTBUFFER_BITS_PER_PIPE * I915_MAX_PIPES >
4705 sizeof(atomic_t) * BITS_PER_BYTE);
4708 WARN_ON(!(atomic_read(&old->frontbuffer_bits) & frontbuffer_bits));
4709 atomic_andnot(frontbuffer_bits, &old->frontbuffer_bits);
4713 WARN_ON(atomic_read(&new->frontbuffer_bits) & frontbuffer_bits);
4714 atomic_or(frontbuffer_bits, &new->frontbuffer_bits);
4718 /* Allocate a new GEM object and fill it with the supplied data */
4719 struct drm_i915_gem_object *
4720 i915_gem_object_create_from_data(struct drm_device *dev,
4721 const void *data, size_t size)
4723 struct drm_i915_gem_object *obj;
4724 struct sg_table *sg;
4728 obj = i915_gem_object_create(dev, round_up(size, PAGE_SIZE));
4732 ret = i915_gem_object_set_to_cpu_domain(obj, true);
4736 ret = i915_gem_object_pin_pages(obj);
4741 bytes = sg_copy_from_buffer(sg->sgl, sg->nents, (void *)data, size);
4742 obj->mm.dirty = true; /* Backing store is now out of date */
4743 i915_gem_object_unpin_pages(obj);
4745 if (WARN_ON(bytes != size)) {
4746 DRM_ERROR("Incomplete copy, wrote %zu of %zu", bytes, size);
4754 i915_gem_object_put(obj);
4755 return ERR_PTR(ret);
4758 struct scatterlist *
4759 i915_gem_object_get_sg(struct drm_i915_gem_object *obj,
4761 unsigned int *offset)
4763 struct i915_gem_object_page_iter *iter = &obj->mm.get_page;
4764 struct scatterlist *sg;
4765 unsigned int idx, count;
4768 GEM_BUG_ON(n >= obj->base.size >> PAGE_SHIFT);
4769 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
4771 /* As we iterate forward through the sg, we record each entry in a
4772 * radixtree for quick repeated (backwards) lookups. If we have seen
4773 * this index previously, we will have an entry for it.
4775 * Initial lookup is O(N), but this is amortized to O(1) for
4776 * sequential page access (where each new request is consecutive
4777 * to the previous one). Repeated lookups are O(lg(obj->base.size)),
4778 * i.e. O(1) with a large constant!
4780 if (n < READ_ONCE(iter->sg_idx))
4783 mutex_lock(&iter->lock);
4785 /* We prefer to reuse the last sg so that repeated lookup of this
4786 * (or the subsequent) sg are fast - comparing against the last
4787 * sg is faster than going through the radixtree.
4792 count = __sg_page_count(sg);
4794 while (idx + count <= n) {
4795 unsigned long exception, i;
4798 /* If we cannot allocate and insert this entry, or the
4799 * individual pages from this range, cancel updating the
4800 * sg_idx so that on this lookup we are forced to linearly
4801 * scan onwards, but on future lookups we will try the
4802 * insertion again (in which case we need to be careful of
4803 * the error return reporting that we have already inserted
4806 ret = radix_tree_insert(&iter->radix, idx, sg);
4807 if (ret && ret != -EEXIST)
4811 RADIX_TREE_EXCEPTIONAL_ENTRY |
4812 idx << RADIX_TREE_EXCEPTIONAL_SHIFT;
4813 for (i = 1; i < count; i++) {
4814 ret = radix_tree_insert(&iter->radix, idx + i,
4816 if (ret && ret != -EEXIST)
4821 sg = ____sg_next(sg);
4822 count = __sg_page_count(sg);
4829 mutex_unlock(&iter->lock);
4831 if (unlikely(n < idx)) /* insertion completed by another thread */
4834 /* In case we failed to insert the entry into the radixtree, we need
4835 * to look beyond the current sg.
4837 while (idx + count <= n) {
4839 sg = ____sg_next(sg);
4840 count = __sg_page_count(sg);
4849 sg = radix_tree_lookup(&iter->radix, n);
4852 /* If this index is in the middle of multi-page sg entry,
4853 * the radixtree will contain an exceptional entry that points
4854 * to the start of that range. We will return the pointer to
4855 * the base page and the offset of this page within the
4859 if (unlikely(radix_tree_exception(sg))) {
4860 unsigned long base =
4861 (unsigned long)sg >> RADIX_TREE_EXCEPTIONAL_SHIFT;
4863 sg = radix_tree_lookup(&iter->radix, base);
4875 i915_gem_object_get_page(struct drm_i915_gem_object *obj, unsigned int n)
4877 struct scatterlist *sg;
4878 unsigned int offset;
4880 GEM_BUG_ON(!i915_gem_object_has_struct_page(obj));
4882 sg = i915_gem_object_get_sg(obj, n, &offset);
4883 return nth_page(sg_page(sg), offset);
4886 /* Like i915_gem_object_get_page(), but mark the returned page dirty */
4888 i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj,
4893 page = i915_gem_object_get_page(obj, n);
4895 set_page_dirty(page);
4901 i915_gem_object_get_dma_address(struct drm_i915_gem_object *obj,
4904 struct scatterlist *sg;
4905 unsigned int offset;
4907 sg = i915_gem_object_get_sg(obj, n, &offset);
4908 return sg_dma_address(sg) + (offset << PAGE_SHIFT);