2 * Copyright © 2008 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_trace.h"
33 #include "intel_drv.h"
34 #include <linux/oom.h>
35 #include <linux/shmem_fs.h>
36 #include <linux/slab.h>
37 #include <linux/swap.h>
38 #include <linux/pci.h>
39 #include <linux/dma-buf.h>
41 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);
42 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj,
44 static __must_check int
45 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
48 i915_gem_object_retire(struct drm_i915_gem_object *obj);
50 static void i915_gem_write_fence(struct drm_device *dev, int reg,
51 struct drm_i915_gem_object *obj);
52 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
53 struct drm_i915_fence_reg *fence,
56 static unsigned long i915_gem_shrinker_count(struct shrinker *shrinker,
57 struct shrink_control *sc);
58 static unsigned long i915_gem_shrinker_scan(struct shrinker *shrinker,
59 struct shrink_control *sc);
60 static int i915_gem_shrinker_oom(struct notifier_block *nb,
63 static unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv);
65 static bool cpu_cache_is_coherent(struct drm_device *dev,
66 enum i915_cache_level level)
68 return HAS_LLC(dev) || level != I915_CACHE_NONE;
71 static bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj)
73 if (!cpu_cache_is_coherent(obj->base.dev, obj->cache_level))
76 return obj->pin_display;
79 static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object *obj)
82 i915_gem_release_mmap(obj);
84 /* As we do not have an associated fence register, we will force
85 * a tiling change if we ever need to acquire one.
87 obj->fence_dirty = false;
88 obj->fence_reg = I915_FENCE_REG_NONE;
91 /* some bookkeeping */
92 static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
95 spin_lock(&dev_priv->mm.object_stat_lock);
96 dev_priv->mm.object_count++;
97 dev_priv->mm.object_memory += size;
98 spin_unlock(&dev_priv->mm.object_stat_lock);
101 static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
104 spin_lock(&dev_priv->mm.object_stat_lock);
105 dev_priv->mm.object_count--;
106 dev_priv->mm.object_memory -= size;
107 spin_unlock(&dev_priv->mm.object_stat_lock);
111 i915_gem_wait_for_error(struct i915_gpu_error *error)
115 #define EXIT_COND (!i915_reset_in_progress(error) || \
116 i915_terminally_wedged(error))
121 * Only wait 10 seconds for the gpu reset to complete to avoid hanging
122 * userspace. If it takes that long something really bad is going on and
123 * we should simply try to bail out and fail as gracefully as possible.
125 ret = wait_event_interruptible_timeout(error->reset_queue,
129 DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
131 } else if (ret < 0) {
139 int i915_mutex_lock_interruptible(struct drm_device *dev)
141 struct drm_i915_private *dev_priv = dev->dev_private;
144 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
148 ret = mutex_lock_interruptible(&dev->struct_mutex);
152 WARN_ON(i915_verify_lists(dev));
157 i915_gem_object_is_inactive(struct drm_i915_gem_object *obj)
159 return i915_gem_obj_bound_any(obj) && !obj->active;
163 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
164 struct drm_file *file)
166 struct drm_i915_private *dev_priv = dev->dev_private;
167 struct drm_i915_gem_get_aperture *args = data;
168 struct drm_i915_gem_object *obj;
172 mutex_lock(&dev->struct_mutex);
173 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list)
174 if (i915_gem_obj_is_pinned(obj))
175 pinned += i915_gem_obj_ggtt_size(obj);
176 mutex_unlock(&dev->struct_mutex);
178 args->aper_size = dev_priv->gtt.base.total;
179 args->aper_available_size = args->aper_size - pinned;
185 i915_gem_object_get_pages_phys(struct drm_i915_gem_object *obj)
187 struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
188 char *vaddr = obj->phys_handle->vaddr;
190 struct scatterlist *sg;
193 if (WARN_ON(i915_gem_object_needs_bit17_swizzle(obj)))
196 for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
200 page = shmem_read_mapping_page(mapping, i);
202 return PTR_ERR(page);
204 src = kmap_atomic(page);
205 memcpy(vaddr, src, PAGE_SIZE);
206 drm_clflush_virt_range(vaddr, PAGE_SIZE);
209 page_cache_release(page);
213 i915_gem_chipset_flush(obj->base.dev);
215 st = kmalloc(sizeof(*st), GFP_KERNEL);
219 if (sg_alloc_table(st, 1, GFP_KERNEL)) {
226 sg->length = obj->base.size;
228 sg_dma_address(sg) = obj->phys_handle->busaddr;
229 sg_dma_len(sg) = obj->base.size;
232 obj->has_dma_mapping = true;
237 i915_gem_object_put_pages_phys(struct drm_i915_gem_object *obj)
241 BUG_ON(obj->madv == __I915_MADV_PURGED);
243 ret = i915_gem_object_set_to_cpu_domain(obj, true);
245 /* In the event of a disaster, abandon all caches and
248 WARN_ON(ret != -EIO);
249 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
252 if (obj->madv == I915_MADV_DONTNEED)
256 struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
257 char *vaddr = obj->phys_handle->vaddr;
260 for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
264 page = shmem_read_mapping_page(mapping, i);
268 dst = kmap_atomic(page);
269 drm_clflush_virt_range(vaddr, PAGE_SIZE);
270 memcpy(dst, vaddr, PAGE_SIZE);
273 set_page_dirty(page);
274 if (obj->madv == I915_MADV_WILLNEED)
275 mark_page_accessed(page);
276 page_cache_release(page);
282 sg_free_table(obj->pages);
285 obj->has_dma_mapping = false;
289 i915_gem_object_release_phys(struct drm_i915_gem_object *obj)
291 drm_pci_free(obj->base.dev, obj->phys_handle);
294 static const struct drm_i915_gem_object_ops i915_gem_phys_ops = {
295 .get_pages = i915_gem_object_get_pages_phys,
296 .put_pages = i915_gem_object_put_pages_phys,
297 .release = i915_gem_object_release_phys,
301 drop_pages(struct drm_i915_gem_object *obj)
303 struct i915_vma *vma, *next;
306 drm_gem_object_reference(&obj->base);
307 list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link)
308 if (i915_vma_unbind(vma))
311 ret = i915_gem_object_put_pages(obj);
312 drm_gem_object_unreference(&obj->base);
318 i915_gem_object_attach_phys(struct drm_i915_gem_object *obj,
321 drm_dma_handle_t *phys;
324 if (obj->phys_handle) {
325 if ((unsigned long)obj->phys_handle->vaddr & (align -1))
331 if (obj->madv != I915_MADV_WILLNEED)
334 if (obj->base.filp == NULL)
337 ret = drop_pages(obj);
341 /* create a new object */
342 phys = drm_pci_alloc(obj->base.dev, obj->base.size, align);
346 obj->phys_handle = phys;
347 obj->ops = &i915_gem_phys_ops;
349 return i915_gem_object_get_pages(obj);
353 i915_gem_phys_pwrite(struct drm_i915_gem_object *obj,
354 struct drm_i915_gem_pwrite *args,
355 struct drm_file *file_priv)
357 struct drm_device *dev = obj->base.dev;
358 void *vaddr = obj->phys_handle->vaddr + args->offset;
359 char __user *user_data = to_user_ptr(args->data_ptr);
362 /* We manually control the domain here and pretend that it
363 * remains coherent i.e. in the GTT domain, like shmem_pwrite.
365 ret = i915_gem_object_wait_rendering(obj, false);
369 if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
370 unsigned long unwritten;
372 /* The physical object once assigned is fixed for the lifetime
373 * of the obj, so we can safely drop the lock and continue
376 mutex_unlock(&dev->struct_mutex);
377 unwritten = copy_from_user(vaddr, user_data, args->size);
378 mutex_lock(&dev->struct_mutex);
383 drm_clflush_virt_range(vaddr, args->size);
384 i915_gem_chipset_flush(dev);
388 void *i915_gem_object_alloc(struct drm_device *dev)
390 struct drm_i915_private *dev_priv = dev->dev_private;
391 return kmem_cache_zalloc(dev_priv->slab, GFP_KERNEL);
394 void i915_gem_object_free(struct drm_i915_gem_object *obj)
396 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
397 kmem_cache_free(dev_priv->slab, obj);
401 i915_gem_create(struct drm_file *file,
402 struct drm_device *dev,
407 struct drm_i915_gem_object *obj;
411 size = roundup(size, PAGE_SIZE);
415 /* Allocate the new object */
416 obj = i915_gem_alloc_object(dev, size);
420 obj->base.dumb = dumb;
421 ret = drm_gem_handle_create(file, &obj->base, &handle);
422 /* drop reference from allocate - handle holds it now */
423 drm_gem_object_unreference_unlocked(&obj->base);
432 i915_gem_dumb_create(struct drm_file *file,
433 struct drm_device *dev,
434 struct drm_mode_create_dumb *args)
436 /* have to work out size/pitch and return them */
437 args->pitch = ALIGN(args->width * DIV_ROUND_UP(args->bpp, 8), 64);
438 args->size = args->pitch * args->height;
439 return i915_gem_create(file, dev,
440 args->size, true, &args->handle);
444 * Creates a new mm object and returns a handle to it.
447 i915_gem_create_ioctl(struct drm_device *dev, void *data,
448 struct drm_file *file)
450 struct drm_i915_gem_create *args = data;
452 return i915_gem_create(file, dev,
453 args->size, false, &args->handle);
457 __copy_to_user_swizzled(char __user *cpu_vaddr,
458 const char *gpu_vaddr, int gpu_offset,
461 int ret, cpu_offset = 0;
464 int cacheline_end = ALIGN(gpu_offset + 1, 64);
465 int this_length = min(cacheline_end - gpu_offset, length);
466 int swizzled_gpu_offset = gpu_offset ^ 64;
468 ret = __copy_to_user(cpu_vaddr + cpu_offset,
469 gpu_vaddr + swizzled_gpu_offset,
474 cpu_offset += this_length;
475 gpu_offset += this_length;
476 length -= this_length;
483 __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset,
484 const char __user *cpu_vaddr,
487 int ret, cpu_offset = 0;
490 int cacheline_end = ALIGN(gpu_offset + 1, 64);
491 int this_length = min(cacheline_end - gpu_offset, length);
492 int swizzled_gpu_offset = gpu_offset ^ 64;
494 ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset,
495 cpu_vaddr + cpu_offset,
500 cpu_offset += this_length;
501 gpu_offset += this_length;
502 length -= this_length;
509 * Pins the specified object's pages and synchronizes the object with
510 * GPU accesses. Sets needs_clflush to non-zero if the caller should
511 * flush the object from the CPU cache.
513 int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj,
523 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) {
524 /* If we're not in the cpu read domain, set ourself into the gtt
525 * read domain and manually flush cachelines (if required). This
526 * optimizes for the case when the gpu will dirty the data
527 * anyway again before the next pread happens. */
528 *needs_clflush = !cpu_cache_is_coherent(obj->base.dev,
530 ret = i915_gem_object_wait_rendering(obj, true);
534 i915_gem_object_retire(obj);
537 ret = i915_gem_object_get_pages(obj);
541 i915_gem_object_pin_pages(obj);
546 /* Per-page copy function for the shmem pread fastpath.
547 * Flushes invalid cachelines before reading the target if
548 * needs_clflush is set. */
550 shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length,
551 char __user *user_data,
552 bool page_do_bit17_swizzling, bool needs_clflush)
557 if (unlikely(page_do_bit17_swizzling))
560 vaddr = kmap_atomic(page);
562 drm_clflush_virt_range(vaddr + shmem_page_offset,
564 ret = __copy_to_user_inatomic(user_data,
565 vaddr + shmem_page_offset,
567 kunmap_atomic(vaddr);
569 return ret ? -EFAULT : 0;
573 shmem_clflush_swizzled_range(char *addr, unsigned long length,
576 if (unlikely(swizzled)) {
577 unsigned long start = (unsigned long) addr;
578 unsigned long end = (unsigned long) addr + length;
580 /* For swizzling simply ensure that we always flush both
581 * channels. Lame, but simple and it works. Swizzled
582 * pwrite/pread is far from a hotpath - current userspace
583 * doesn't use it at all. */
584 start = round_down(start, 128);
585 end = round_up(end, 128);
587 drm_clflush_virt_range((void *)start, end - start);
589 drm_clflush_virt_range(addr, length);
594 /* Only difference to the fast-path function is that this can handle bit17
595 * and uses non-atomic copy and kmap functions. */
597 shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length,
598 char __user *user_data,
599 bool page_do_bit17_swizzling, bool needs_clflush)
606 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
608 page_do_bit17_swizzling);
610 if (page_do_bit17_swizzling)
611 ret = __copy_to_user_swizzled(user_data,
612 vaddr, shmem_page_offset,
615 ret = __copy_to_user(user_data,
616 vaddr + shmem_page_offset,
620 return ret ? - EFAULT : 0;
624 i915_gem_shmem_pread(struct drm_device *dev,
625 struct drm_i915_gem_object *obj,
626 struct drm_i915_gem_pread *args,
627 struct drm_file *file)
629 char __user *user_data;
632 int shmem_page_offset, page_length, ret = 0;
633 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
635 int needs_clflush = 0;
636 struct sg_page_iter sg_iter;
638 user_data = to_user_ptr(args->data_ptr);
641 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
643 ret = i915_gem_obj_prepare_shmem_read(obj, &needs_clflush);
647 offset = args->offset;
649 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
650 offset >> PAGE_SHIFT) {
651 struct page *page = sg_page_iter_page(&sg_iter);
656 /* Operation in this page
658 * shmem_page_offset = offset within page in shmem file
659 * page_length = bytes to copy for this page
661 shmem_page_offset = offset_in_page(offset);
662 page_length = remain;
663 if ((shmem_page_offset + page_length) > PAGE_SIZE)
664 page_length = PAGE_SIZE - shmem_page_offset;
666 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
667 (page_to_phys(page) & (1 << 17)) != 0;
669 ret = shmem_pread_fast(page, shmem_page_offset, page_length,
670 user_data, page_do_bit17_swizzling,
675 mutex_unlock(&dev->struct_mutex);
677 if (likely(!i915.prefault_disable) && !prefaulted) {
678 ret = fault_in_multipages_writeable(user_data, remain);
679 /* Userspace is tricking us, but we've already clobbered
680 * its pages with the prefault and promised to write the
681 * data up to the first fault. Hence ignore any errors
682 * and just continue. */
687 ret = shmem_pread_slow(page, shmem_page_offset, page_length,
688 user_data, page_do_bit17_swizzling,
691 mutex_lock(&dev->struct_mutex);
697 remain -= page_length;
698 user_data += page_length;
699 offset += page_length;
703 i915_gem_object_unpin_pages(obj);
709 * Reads data from the object referenced by handle.
711 * On error, the contents of *data are undefined.
714 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
715 struct drm_file *file)
717 struct drm_i915_gem_pread *args = data;
718 struct drm_i915_gem_object *obj;
724 if (!access_ok(VERIFY_WRITE,
725 to_user_ptr(args->data_ptr),
729 ret = i915_mutex_lock_interruptible(dev);
733 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
734 if (&obj->base == NULL) {
739 /* Bounds check source. */
740 if (args->offset > obj->base.size ||
741 args->size > obj->base.size - args->offset) {
746 /* prime objects have no backing filp to GEM pread/pwrite
749 if (!obj->base.filp) {
754 trace_i915_gem_object_pread(obj, args->offset, args->size);
756 ret = i915_gem_shmem_pread(dev, obj, args, file);
759 drm_gem_object_unreference(&obj->base);
761 mutex_unlock(&dev->struct_mutex);
765 /* This is the fast write path which cannot handle
766 * page faults in the source data
770 fast_user_write(struct io_mapping *mapping,
771 loff_t page_base, int page_offset,
772 char __user *user_data,
775 void __iomem *vaddr_atomic;
777 unsigned long unwritten;
779 vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
780 /* We can use the cpu mem copy function because this is X86. */
781 vaddr = (void __force*)vaddr_atomic + page_offset;
782 unwritten = __copy_from_user_inatomic_nocache(vaddr,
784 io_mapping_unmap_atomic(vaddr_atomic);
789 * This is the fast pwrite path, where we copy the data directly from the
790 * user into the GTT, uncached.
793 i915_gem_gtt_pwrite_fast(struct drm_device *dev,
794 struct drm_i915_gem_object *obj,
795 struct drm_i915_gem_pwrite *args,
796 struct drm_file *file)
798 struct drm_i915_private *dev_priv = dev->dev_private;
800 loff_t offset, page_base;
801 char __user *user_data;
802 int page_offset, page_length, ret;
804 ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE | PIN_NONBLOCK);
808 ret = i915_gem_object_set_to_gtt_domain(obj, true);
812 ret = i915_gem_object_put_fence(obj);
816 user_data = to_user_ptr(args->data_ptr);
819 offset = i915_gem_obj_ggtt_offset(obj) + args->offset;
822 /* Operation in this page
824 * page_base = page offset within aperture
825 * page_offset = offset within page
826 * page_length = bytes to copy for this page
828 page_base = offset & PAGE_MASK;
829 page_offset = offset_in_page(offset);
830 page_length = remain;
831 if ((page_offset + remain) > PAGE_SIZE)
832 page_length = PAGE_SIZE - page_offset;
834 /* If we get a fault while copying data, then (presumably) our
835 * source page isn't available. Return the error and we'll
836 * retry in the slow path.
838 if (fast_user_write(dev_priv->gtt.mappable, page_base,
839 page_offset, user_data, page_length)) {
844 remain -= page_length;
845 user_data += page_length;
846 offset += page_length;
850 i915_gem_object_ggtt_unpin(obj);
855 /* Per-page copy function for the shmem pwrite fastpath.
856 * Flushes invalid cachelines before writing to the target if
857 * needs_clflush_before is set and flushes out any written cachelines after
858 * writing if needs_clflush is set. */
860 shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
861 char __user *user_data,
862 bool page_do_bit17_swizzling,
863 bool needs_clflush_before,
864 bool needs_clflush_after)
869 if (unlikely(page_do_bit17_swizzling))
872 vaddr = kmap_atomic(page);
873 if (needs_clflush_before)
874 drm_clflush_virt_range(vaddr + shmem_page_offset,
876 ret = __copy_from_user_inatomic(vaddr + shmem_page_offset,
877 user_data, page_length);
878 if (needs_clflush_after)
879 drm_clflush_virt_range(vaddr + shmem_page_offset,
881 kunmap_atomic(vaddr);
883 return ret ? -EFAULT : 0;
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_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
890 char __user *user_data,
891 bool page_do_bit17_swizzling,
892 bool needs_clflush_before,
893 bool needs_clflush_after)
899 if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
900 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
902 page_do_bit17_swizzling);
903 if (page_do_bit17_swizzling)
904 ret = __copy_from_user_swizzled(vaddr, shmem_page_offset,
908 ret = __copy_from_user(vaddr + shmem_page_offset,
911 if (needs_clflush_after)
912 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
914 page_do_bit17_swizzling);
917 return ret ? -EFAULT : 0;
921 i915_gem_shmem_pwrite(struct drm_device *dev,
922 struct drm_i915_gem_object *obj,
923 struct drm_i915_gem_pwrite *args,
924 struct drm_file *file)
928 char __user *user_data;
929 int shmem_page_offset, page_length, ret = 0;
930 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
931 int hit_slowpath = 0;
932 int needs_clflush_after = 0;
933 int needs_clflush_before = 0;
934 struct sg_page_iter sg_iter;
936 user_data = to_user_ptr(args->data_ptr);
939 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
941 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
942 /* If we're not in the cpu write domain, set ourself into the gtt
943 * write domain and manually flush cachelines (if required). This
944 * optimizes for the case when the gpu will use the data
945 * right away and we therefore have to clflush anyway. */
946 needs_clflush_after = cpu_write_needs_clflush(obj);
947 ret = i915_gem_object_wait_rendering(obj, false);
951 i915_gem_object_retire(obj);
953 /* Same trick applies to invalidate partially written cachelines read
955 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0)
956 needs_clflush_before =
957 !cpu_cache_is_coherent(dev, obj->cache_level);
959 ret = i915_gem_object_get_pages(obj);
963 i915_gem_object_pin_pages(obj);
965 offset = args->offset;
968 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
969 offset >> PAGE_SHIFT) {
970 struct page *page = sg_page_iter_page(&sg_iter);
971 int partial_cacheline_write;
976 /* Operation in this page
978 * shmem_page_offset = offset within page in shmem file
979 * page_length = bytes to copy for this page
981 shmem_page_offset = offset_in_page(offset);
983 page_length = remain;
984 if ((shmem_page_offset + page_length) > PAGE_SIZE)
985 page_length = PAGE_SIZE - shmem_page_offset;
987 /* If we don't overwrite a cacheline completely we need to be
988 * careful to have up-to-date data by first clflushing. Don't
989 * overcomplicate things and flush the entire patch. */
990 partial_cacheline_write = needs_clflush_before &&
991 ((shmem_page_offset | page_length)
992 & (boot_cpu_data.x86_clflush_size - 1));
994 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
995 (page_to_phys(page) & (1 << 17)) != 0;
997 ret = shmem_pwrite_fast(page, shmem_page_offset, page_length,
998 user_data, page_do_bit17_swizzling,
999 partial_cacheline_write,
1000 needs_clflush_after);
1005 mutex_unlock(&dev->struct_mutex);
1006 ret = shmem_pwrite_slow(page, shmem_page_offset, page_length,
1007 user_data, page_do_bit17_swizzling,
1008 partial_cacheline_write,
1009 needs_clflush_after);
1011 mutex_lock(&dev->struct_mutex);
1017 remain -= page_length;
1018 user_data += page_length;
1019 offset += page_length;
1023 i915_gem_object_unpin_pages(obj);
1027 * Fixup: Flush cpu caches in case we didn't flush the dirty
1028 * cachelines in-line while writing and the object moved
1029 * out of the cpu write domain while we've dropped the lock.
1031 if (!needs_clflush_after &&
1032 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
1033 if (i915_gem_clflush_object(obj, obj->pin_display))
1034 i915_gem_chipset_flush(dev);
1038 if (needs_clflush_after)
1039 i915_gem_chipset_flush(dev);
1045 * Writes data to the object referenced by handle.
1047 * On error, the contents of the buffer that were to be modified are undefined.
1050 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
1051 struct drm_file *file)
1053 struct drm_i915_gem_pwrite *args = data;
1054 struct drm_i915_gem_object *obj;
1057 if (args->size == 0)
1060 if (!access_ok(VERIFY_READ,
1061 to_user_ptr(args->data_ptr),
1065 if (likely(!i915.prefault_disable)) {
1066 ret = fault_in_multipages_readable(to_user_ptr(args->data_ptr),
1072 ret = i915_mutex_lock_interruptible(dev);
1076 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1077 if (&obj->base == NULL) {
1082 /* Bounds check destination. */
1083 if (args->offset > obj->base.size ||
1084 args->size > obj->base.size - args->offset) {
1089 /* prime objects have no backing filp to GEM pread/pwrite
1092 if (!obj->base.filp) {
1097 trace_i915_gem_object_pwrite(obj, args->offset, args->size);
1100 /* We can only do the GTT pwrite on untiled buffers, as otherwise
1101 * it would end up going through the fenced access, and we'll get
1102 * different detiling behavior between reading and writing.
1103 * pread/pwrite currently are reading and writing from the CPU
1104 * perspective, requiring manual detiling by the client.
1106 if (obj->tiling_mode == I915_TILING_NONE &&
1107 obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
1108 cpu_write_needs_clflush(obj)) {
1109 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
1110 /* Note that the gtt paths might fail with non-page-backed user
1111 * pointers (e.g. gtt mappings when moving data between
1112 * textures). Fallback to the shmem path in that case. */
1115 if (ret == -EFAULT || ret == -ENOSPC) {
1116 if (obj->phys_handle)
1117 ret = i915_gem_phys_pwrite(obj, args, file);
1119 ret = i915_gem_shmem_pwrite(dev, obj, args, file);
1123 drm_gem_object_unreference(&obj->base);
1125 mutex_unlock(&dev->struct_mutex);
1130 i915_gem_check_wedge(struct i915_gpu_error *error,
1133 if (i915_reset_in_progress(error)) {
1134 /* Non-interruptible callers can't handle -EAGAIN, hence return
1135 * -EIO unconditionally for these. */
1139 /* Recovery complete, but the reset failed ... */
1140 if (i915_terminally_wedged(error))
1144 * Check if GPU Reset is in progress - we need intel_ring_begin
1145 * to work properly to reinit the hw state while the gpu is
1146 * still marked as reset-in-progress. Handle this with a flag.
1148 if (!error->reload_in_reset)
1156 * Compare seqno against outstanding lazy request. Emit a request if they are
1160 i915_gem_check_olr(struct intel_engine_cs *ring, u32 seqno)
1164 BUG_ON(!mutex_is_locked(&ring->dev->struct_mutex));
1167 if (seqno == ring->outstanding_lazy_seqno)
1168 ret = i915_add_request(ring, NULL);
1173 static void fake_irq(unsigned long data)
1175 wake_up_process((struct task_struct *)data);
1178 static bool missed_irq(struct drm_i915_private *dev_priv,
1179 struct intel_engine_cs *ring)
1181 return test_bit(ring->id, &dev_priv->gpu_error.missed_irq_rings);
1184 static bool can_wait_boost(struct drm_i915_file_private *file_priv)
1186 if (file_priv == NULL)
1189 return !atomic_xchg(&file_priv->rps_wait_boost, true);
1193 * __i915_wait_seqno - wait until execution of seqno has finished
1194 * @ring: the ring expected to report seqno
1196 * @reset_counter: reset sequence associated with the given seqno
1197 * @interruptible: do an interruptible wait (normally yes)
1198 * @timeout: in - how long to wait (NULL forever); out - how much time remaining
1200 * Note: It is of utmost importance that the passed in seqno and reset_counter
1201 * values have been read by the caller in an smp safe manner. Where read-side
1202 * locks are involved, it is sufficient to read the reset_counter before
1203 * unlocking the lock that protects the seqno. For lockless tricks, the
1204 * reset_counter _must_ be read before, and an appropriate smp_rmb must be
1207 * Returns 0 if the seqno was found within the alloted time. Else returns the
1208 * errno with remaining time filled in timeout argument.
1210 int __i915_wait_seqno(struct intel_engine_cs *ring, u32 seqno,
1211 unsigned reset_counter,
1214 struct drm_i915_file_private *file_priv)
1216 struct drm_device *dev = ring->dev;
1217 struct drm_i915_private *dev_priv = dev->dev_private;
1218 const bool irq_test_in_progress =
1219 ACCESS_ONCE(dev_priv->gpu_error.test_irq_rings) & intel_ring_flag(ring);
1221 unsigned long timeout_expire;
1225 WARN(!intel_irqs_enabled(dev_priv), "IRQs disabled");
1227 if (i915_seqno_passed(ring->get_seqno(ring, true), seqno))
1230 timeout_expire = timeout ? jiffies + nsecs_to_jiffies((u64)*timeout) : 0;
1232 if (INTEL_INFO(dev)->gen >= 6 && ring->id == RCS && can_wait_boost(file_priv)) {
1233 gen6_rps_boost(dev_priv);
1235 mod_delayed_work(dev_priv->wq,
1236 &file_priv->mm.idle_work,
1237 msecs_to_jiffies(100));
1240 if (!irq_test_in_progress && WARN_ON(!ring->irq_get(ring)))
1243 /* Record current time in case interrupted by signal, or wedged */
1244 trace_i915_gem_request_wait_begin(ring, seqno);
1245 before = ktime_get_raw_ns();
1247 struct timer_list timer;
1249 prepare_to_wait(&ring->irq_queue, &wait,
1250 interruptible ? TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE);
1252 /* We need to check whether any gpu reset happened in between
1253 * the caller grabbing the seqno and now ... */
1254 if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) {
1255 /* ... but upgrade the -EAGAIN to an -EIO if the gpu
1256 * is truely gone. */
1257 ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1263 if (i915_seqno_passed(ring->get_seqno(ring, false), seqno)) {
1268 if (interruptible && signal_pending(current)) {
1273 if (timeout && time_after_eq(jiffies, timeout_expire)) {
1278 timer.function = NULL;
1279 if (timeout || missed_irq(dev_priv, ring)) {
1280 unsigned long expire;
1282 setup_timer_on_stack(&timer, fake_irq, (unsigned long)current);
1283 expire = missed_irq(dev_priv, ring) ? jiffies + 1 : timeout_expire;
1284 mod_timer(&timer, expire);
1289 if (timer.function) {
1290 del_singleshot_timer_sync(&timer);
1291 destroy_timer_on_stack(&timer);
1294 now = ktime_get_raw_ns();
1295 trace_i915_gem_request_wait_end(ring, seqno);
1297 if (!irq_test_in_progress)
1298 ring->irq_put(ring);
1300 finish_wait(&ring->irq_queue, &wait);
1303 s64 tres = *timeout - (now - before);
1305 *timeout = tres < 0 ? 0 : tres;
1312 * Waits for a sequence number to be signaled, and cleans up the
1313 * request and object lists appropriately for that event.
1316 i915_wait_seqno(struct intel_engine_cs *ring, uint32_t seqno)
1318 struct drm_device *dev = ring->dev;
1319 struct drm_i915_private *dev_priv = dev->dev_private;
1320 bool interruptible = dev_priv->mm.interruptible;
1321 unsigned reset_counter;
1324 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1327 ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1331 ret = i915_gem_check_olr(ring, seqno);
1335 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
1336 return __i915_wait_seqno(ring, seqno, reset_counter, interruptible,
1341 i915_gem_object_wait_rendering__tail(struct drm_i915_gem_object *obj)
1346 /* Manually manage the write flush as we may have not yet
1347 * retired the buffer.
1349 * Note that the last_write_seqno is always the earlier of
1350 * the two (read/write) seqno, so if we haved successfully waited,
1351 * we know we have passed the last write.
1353 obj->last_write_seqno = 0;
1359 * Ensures that all rendering to the object has completed and the object is
1360 * safe to unbind from the GTT or access from the CPU.
1362 static __must_check int
1363 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
1366 struct intel_engine_cs *ring = obj->ring;
1370 seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno;
1374 ret = i915_wait_seqno(ring, seqno);
1378 return i915_gem_object_wait_rendering__tail(obj);
1381 /* A nonblocking variant of the above wait. This is a highly dangerous routine
1382 * as the object state may change during this call.
1384 static __must_check int
1385 i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
1386 struct drm_i915_file_private *file_priv,
1389 struct drm_device *dev = obj->base.dev;
1390 struct drm_i915_private *dev_priv = dev->dev_private;
1391 struct intel_engine_cs *ring = obj->ring;
1392 unsigned reset_counter;
1396 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1397 BUG_ON(!dev_priv->mm.interruptible);
1399 seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno;
1403 ret = i915_gem_check_wedge(&dev_priv->gpu_error, true);
1407 ret = i915_gem_check_olr(ring, seqno);
1411 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
1412 mutex_unlock(&dev->struct_mutex);
1413 ret = __i915_wait_seqno(ring, seqno, reset_counter, true, NULL,
1415 mutex_lock(&dev->struct_mutex);
1419 return i915_gem_object_wait_rendering__tail(obj);
1423 * Called when user space prepares to use an object with the CPU, either
1424 * through the mmap ioctl's mapping or a GTT mapping.
1427 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1428 struct drm_file *file)
1430 struct drm_i915_gem_set_domain *args = data;
1431 struct drm_i915_gem_object *obj;
1432 uint32_t read_domains = args->read_domains;
1433 uint32_t write_domain = args->write_domain;
1436 /* Only handle setting domains to types used by the CPU. */
1437 if (write_domain & I915_GEM_GPU_DOMAINS)
1440 if (read_domains & I915_GEM_GPU_DOMAINS)
1443 /* Having something in the write domain implies it's in the read
1444 * domain, and only that read domain. Enforce that in the request.
1446 if (write_domain != 0 && read_domains != write_domain)
1449 ret = i915_mutex_lock_interruptible(dev);
1453 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1454 if (&obj->base == NULL) {
1459 /* Try to flush the object off the GPU without holding the lock.
1460 * We will repeat the flush holding the lock in the normal manner
1461 * to catch cases where we are gazumped.
1463 ret = i915_gem_object_wait_rendering__nonblocking(obj,
1469 if (read_domains & I915_GEM_DOMAIN_GTT) {
1470 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1472 /* Silently promote "you're not bound, there was nothing to do"
1473 * to success, since the client was just asking us to
1474 * make sure everything was done.
1479 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1483 drm_gem_object_unreference(&obj->base);
1485 mutex_unlock(&dev->struct_mutex);
1490 * Called when user space has done writes to this buffer
1493 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1494 struct drm_file *file)
1496 struct drm_i915_gem_sw_finish *args = data;
1497 struct drm_i915_gem_object *obj;
1500 ret = i915_mutex_lock_interruptible(dev);
1504 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1505 if (&obj->base == NULL) {
1510 /* Pinned buffers may be scanout, so flush the cache */
1511 if (obj->pin_display)
1512 i915_gem_object_flush_cpu_write_domain(obj, true);
1514 drm_gem_object_unreference(&obj->base);
1516 mutex_unlock(&dev->struct_mutex);
1521 * Maps the contents of an object, returning the address it is mapped
1524 * While the mapping holds a reference on the contents of the object, it doesn't
1525 * imply a ref on the object itself.
1529 * DRM driver writers who look a this function as an example for how to do GEM
1530 * mmap support, please don't implement mmap support like here. The modern way
1531 * to implement DRM mmap support is with an mmap offset ioctl (like
1532 * i915_gem_mmap_gtt) and then using the mmap syscall on the DRM fd directly.
1533 * That way debug tooling like valgrind will understand what's going on, hiding
1534 * the mmap call in a driver private ioctl will break that. The i915 driver only
1535 * does cpu mmaps this way because we didn't know better.
1538 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1539 struct drm_file *file)
1541 struct drm_i915_gem_mmap *args = data;
1542 struct drm_gem_object *obj;
1545 obj = drm_gem_object_lookup(dev, file, args->handle);
1549 /* prime objects have no backing filp to GEM mmap
1553 drm_gem_object_unreference_unlocked(obj);
1557 addr = vm_mmap(obj->filp, 0, args->size,
1558 PROT_READ | PROT_WRITE, MAP_SHARED,
1560 drm_gem_object_unreference_unlocked(obj);
1561 if (IS_ERR((void *)addr))
1564 args->addr_ptr = (uint64_t) addr;
1570 * i915_gem_fault - fault a page into the GTT
1571 * vma: VMA in question
1574 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1575 * from userspace. The fault handler takes care of binding the object to
1576 * the GTT (if needed), allocating and programming a fence register (again,
1577 * only if needed based on whether the old reg is still valid or the object
1578 * is tiled) and inserting a new PTE into the faulting process.
1580 * Note that the faulting process may involve evicting existing objects
1581 * from the GTT and/or fence registers to make room. So performance may
1582 * suffer if the GTT working set is large or there are few fence registers
1585 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1587 struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1588 struct drm_device *dev = obj->base.dev;
1589 struct drm_i915_private *dev_priv = dev->dev_private;
1590 pgoff_t page_offset;
1593 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1595 intel_runtime_pm_get(dev_priv);
1597 /* We don't use vmf->pgoff since that has the fake offset */
1598 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1601 ret = i915_mutex_lock_interruptible(dev);
1605 trace_i915_gem_object_fault(obj, page_offset, true, write);
1607 /* Try to flush the object off the GPU first without holding the lock.
1608 * Upon reacquiring the lock, we will perform our sanity checks and then
1609 * repeat the flush holding the lock in the normal manner to catch cases
1610 * where we are gazumped.
1612 ret = i915_gem_object_wait_rendering__nonblocking(obj, NULL, !write);
1616 /* Access to snoopable pages through the GTT is incoherent. */
1617 if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) {
1622 /* Now bind it into the GTT if needed */
1623 ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE);
1627 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1631 ret = i915_gem_object_get_fence(obj);
1635 /* Finally, remap it using the new GTT offset */
1636 pfn = dev_priv->gtt.mappable_base + i915_gem_obj_ggtt_offset(obj);
1639 if (!obj->fault_mappable) {
1640 unsigned long size = min_t(unsigned long,
1641 vma->vm_end - vma->vm_start,
1645 for (i = 0; i < size >> PAGE_SHIFT; i++) {
1646 ret = vm_insert_pfn(vma,
1647 (unsigned long)vma->vm_start + i * PAGE_SIZE,
1653 obj->fault_mappable = true;
1655 ret = vm_insert_pfn(vma,
1656 (unsigned long)vmf->virtual_address,
1659 i915_gem_object_ggtt_unpin(obj);
1661 mutex_unlock(&dev->struct_mutex);
1666 * We eat errors when the gpu is terminally wedged to avoid
1667 * userspace unduly crashing (gl has no provisions for mmaps to
1668 * fail). But any other -EIO isn't ours (e.g. swap in failure)
1669 * and so needs to be reported.
1671 if (!i915_terminally_wedged(&dev_priv->gpu_error)) {
1672 ret = VM_FAULT_SIGBUS;
1677 * EAGAIN means the gpu is hung and we'll wait for the error
1678 * handler to reset everything when re-faulting in
1679 * i915_mutex_lock_interruptible.
1686 * EBUSY is ok: this just means that another thread
1687 * already did the job.
1689 ret = VM_FAULT_NOPAGE;
1696 ret = VM_FAULT_SIGBUS;
1699 WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret);
1700 ret = VM_FAULT_SIGBUS;
1704 intel_runtime_pm_put(dev_priv);
1709 * i915_gem_release_mmap - remove physical page mappings
1710 * @obj: obj in question
1712 * Preserve the reservation of the mmapping with the DRM core code, but
1713 * relinquish ownership of the pages back to the system.
1715 * It is vital that we remove the page mapping if we have mapped a tiled
1716 * object through the GTT and then lose the fence register due to
1717 * resource pressure. Similarly if the object has been moved out of the
1718 * aperture, than pages mapped into userspace must be revoked. Removing the
1719 * mapping will then trigger a page fault on the next user access, allowing
1720 * fixup by i915_gem_fault().
1723 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1725 if (!obj->fault_mappable)
1728 drm_vma_node_unmap(&obj->base.vma_node,
1729 obj->base.dev->anon_inode->i_mapping);
1730 obj->fault_mappable = false;
1734 i915_gem_release_all_mmaps(struct drm_i915_private *dev_priv)
1736 struct drm_i915_gem_object *obj;
1738 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list)
1739 i915_gem_release_mmap(obj);
1743 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1747 if (INTEL_INFO(dev)->gen >= 4 ||
1748 tiling_mode == I915_TILING_NONE)
1751 /* Previous chips need a power-of-two fence region when tiling */
1752 if (INTEL_INFO(dev)->gen == 3)
1753 gtt_size = 1024*1024;
1755 gtt_size = 512*1024;
1757 while (gtt_size < size)
1764 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1765 * @obj: object to check
1767 * Return the required GTT alignment for an object, taking into account
1768 * potential fence register mapping.
1771 i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size,
1772 int tiling_mode, bool fenced)
1775 * Minimum alignment is 4k (GTT page size), but might be greater
1776 * if a fence register is needed for the object.
1778 if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) ||
1779 tiling_mode == I915_TILING_NONE)
1783 * Previous chips need to be aligned to the size of the smallest
1784 * fence register that can contain the object.
1786 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1789 static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj)
1791 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1794 if (drm_vma_node_has_offset(&obj->base.vma_node))
1797 dev_priv->mm.shrinker_no_lock_stealing = true;
1799 ret = drm_gem_create_mmap_offset(&obj->base);
1803 /* Badly fragmented mmap space? The only way we can recover
1804 * space is by destroying unwanted objects. We can't randomly release
1805 * mmap_offsets as userspace expects them to be persistent for the
1806 * lifetime of the objects. The closest we can is to release the
1807 * offsets on purgeable objects by truncating it and marking it purged,
1808 * which prevents userspace from ever using that object again.
1810 i915_gem_shrink(dev_priv,
1811 obj->base.size >> PAGE_SHIFT,
1813 I915_SHRINK_UNBOUND |
1814 I915_SHRINK_PURGEABLE);
1815 ret = drm_gem_create_mmap_offset(&obj->base);
1819 i915_gem_shrink_all(dev_priv);
1820 ret = drm_gem_create_mmap_offset(&obj->base);
1822 dev_priv->mm.shrinker_no_lock_stealing = false;
1827 static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)
1829 drm_gem_free_mmap_offset(&obj->base);
1833 i915_gem_mmap_gtt(struct drm_file *file,
1834 struct drm_device *dev,
1835 uint32_t handle, bool dumb,
1838 struct drm_i915_private *dev_priv = dev->dev_private;
1839 struct drm_i915_gem_object *obj;
1842 ret = i915_mutex_lock_interruptible(dev);
1846 obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
1847 if (&obj->base == NULL) {
1853 * We don't allow dumb mmaps on objects created using another
1856 WARN_ONCE(dumb && !(obj->base.dumb || obj->base.import_attach),
1857 "Illegal dumb map of accelerated buffer.\n");
1859 if (obj->base.size > dev_priv->gtt.mappable_end) {
1864 if (obj->madv != I915_MADV_WILLNEED) {
1865 DRM_DEBUG("Attempting to mmap a purgeable buffer\n");
1870 ret = i915_gem_object_create_mmap_offset(obj);
1874 *offset = drm_vma_node_offset_addr(&obj->base.vma_node);
1877 drm_gem_object_unreference(&obj->base);
1879 mutex_unlock(&dev->struct_mutex);
1884 i915_gem_dumb_map_offset(struct drm_file *file,
1885 struct drm_device *dev,
1889 return i915_gem_mmap_gtt(file, dev, handle, true, offset);
1893 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1895 * @data: GTT mapping ioctl data
1896 * @file: GEM object info
1898 * Simply returns the fake offset to userspace so it can mmap it.
1899 * The mmap call will end up in drm_gem_mmap(), which will set things
1900 * up so we can get faults in the handler above.
1902 * The fault handler will take care of binding the object into the GTT
1903 * (since it may have been evicted to make room for something), allocating
1904 * a fence register, and mapping the appropriate aperture address into
1908 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1909 struct drm_file *file)
1911 struct drm_i915_gem_mmap_gtt *args = data;
1913 return i915_gem_mmap_gtt(file, dev, args->handle, false, &args->offset);
1917 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj)
1919 return obj->madv == I915_MADV_DONTNEED;
1922 /* Immediately discard the backing storage */
1924 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
1926 i915_gem_object_free_mmap_offset(obj);
1928 if (obj->base.filp == NULL)
1931 /* Our goal here is to return as much of the memory as
1932 * is possible back to the system as we are called from OOM.
1933 * To do this we must instruct the shmfs to drop all of its
1934 * backing pages, *now*.
1936 shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1);
1937 obj->madv = __I915_MADV_PURGED;
1940 /* Try to discard unwanted pages */
1942 i915_gem_object_invalidate(struct drm_i915_gem_object *obj)
1944 struct address_space *mapping;
1946 switch (obj->madv) {
1947 case I915_MADV_DONTNEED:
1948 i915_gem_object_truncate(obj);
1949 case __I915_MADV_PURGED:
1953 if (obj->base.filp == NULL)
1956 mapping = file_inode(obj->base.filp)->i_mapping,
1957 invalidate_mapping_pages(mapping, 0, (loff_t)-1);
1961 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
1963 struct sg_page_iter sg_iter;
1966 BUG_ON(obj->madv == __I915_MADV_PURGED);
1968 ret = i915_gem_object_set_to_cpu_domain(obj, true);
1970 /* In the event of a disaster, abandon all caches and
1971 * hope for the best.
1973 WARN_ON(ret != -EIO);
1974 i915_gem_clflush_object(obj, true);
1975 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
1978 if (i915_gem_object_needs_bit17_swizzle(obj))
1979 i915_gem_object_save_bit_17_swizzle(obj);
1981 if (obj->madv == I915_MADV_DONTNEED)
1984 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
1985 struct page *page = sg_page_iter_page(&sg_iter);
1988 set_page_dirty(page);
1990 if (obj->madv == I915_MADV_WILLNEED)
1991 mark_page_accessed(page);
1993 page_cache_release(page);
1997 sg_free_table(obj->pages);
2002 i915_gem_object_put_pages(struct drm_i915_gem_object *obj)
2004 const struct drm_i915_gem_object_ops *ops = obj->ops;
2006 if (obj->pages == NULL)
2009 if (obj->pages_pin_count)
2012 BUG_ON(i915_gem_obj_bound_any(obj));
2014 /* ->put_pages might need to allocate memory for the bit17 swizzle
2015 * array, hence protect them from being reaped by removing them from gtt
2017 list_del(&obj->global_list);
2019 ops->put_pages(obj);
2022 i915_gem_object_invalidate(obj);
2028 i915_gem_shrink(struct drm_i915_private *dev_priv,
2029 long target, unsigned flags)
2032 struct list_head *list;
2035 { &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND },
2036 { &dev_priv->mm.bound_list, I915_SHRINK_BOUND },
2039 unsigned long count = 0;
2042 * As we may completely rewrite the (un)bound list whilst unbinding
2043 * (due to retiring requests) we have to strictly process only
2044 * one element of the list at the time, and recheck the list
2045 * on every iteration.
2047 * In particular, we must hold a reference whilst removing the
2048 * object as we may end up waiting for and/or retiring the objects.
2049 * This might release the final reference (held by the active list)
2050 * and result in the object being freed from under us. This is
2051 * similar to the precautions the eviction code must take whilst
2054 * Also note that although these lists do not hold a reference to
2055 * the object we can safely grab one here: The final object
2056 * unreferencing and the bound_list are both protected by the
2057 * dev->struct_mutex and so we won't ever be able to observe an
2058 * object on the bound_list with a reference count equals 0.
2060 for (phase = phases; phase->list; phase++) {
2061 struct list_head still_in_list;
2063 if ((flags & phase->bit) == 0)
2066 INIT_LIST_HEAD(&still_in_list);
2067 while (count < target && !list_empty(phase->list)) {
2068 struct drm_i915_gem_object *obj;
2069 struct i915_vma *vma, *v;
2071 obj = list_first_entry(phase->list,
2072 typeof(*obj), global_list);
2073 list_move_tail(&obj->global_list, &still_in_list);
2075 if (flags & I915_SHRINK_PURGEABLE &&
2076 !i915_gem_object_is_purgeable(obj))
2079 drm_gem_object_reference(&obj->base);
2081 /* For the unbound phase, this should be a no-op! */
2082 list_for_each_entry_safe(vma, v,
2083 &obj->vma_list, vma_link)
2084 if (i915_vma_unbind(vma))
2087 if (i915_gem_object_put_pages(obj) == 0)
2088 count += obj->base.size >> PAGE_SHIFT;
2090 drm_gem_object_unreference(&obj->base);
2092 list_splice(&still_in_list, phase->list);
2098 static unsigned long
2099 i915_gem_shrink_all(struct drm_i915_private *dev_priv)
2101 i915_gem_evict_everything(dev_priv->dev);
2102 return i915_gem_shrink(dev_priv, LONG_MAX,
2103 I915_SHRINK_BOUND | I915_SHRINK_UNBOUND);
2107 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
2109 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2111 struct address_space *mapping;
2112 struct sg_table *st;
2113 struct scatterlist *sg;
2114 struct sg_page_iter sg_iter;
2116 unsigned long last_pfn = 0; /* suppress gcc warning */
2119 /* Assert that the object is not currently in any GPU domain. As it
2120 * wasn't in the GTT, there shouldn't be any way it could have been in
2123 BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
2124 BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
2126 st = kmalloc(sizeof(*st), GFP_KERNEL);
2130 page_count = obj->base.size / PAGE_SIZE;
2131 if (sg_alloc_table(st, page_count, GFP_KERNEL)) {
2136 /* Get the list of pages out of our struct file. They'll be pinned
2137 * at this point until we release them.
2139 * Fail silently without starting the shrinker
2141 mapping = file_inode(obj->base.filp)->i_mapping;
2142 gfp = mapping_gfp_mask(mapping);
2143 gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD;
2144 gfp &= ~(__GFP_IO | __GFP_WAIT);
2147 for (i = 0; i < page_count; i++) {
2148 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2150 i915_gem_shrink(dev_priv,
2153 I915_SHRINK_UNBOUND |
2154 I915_SHRINK_PURGEABLE);
2155 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2158 /* We've tried hard to allocate the memory by reaping
2159 * our own buffer, now let the real VM do its job and
2160 * go down in flames if truly OOM.
2162 i915_gem_shrink_all(dev_priv);
2163 page = shmem_read_mapping_page(mapping, i);
2167 #ifdef CONFIG_SWIOTLB
2168 if (swiotlb_nr_tbl()) {
2170 sg_set_page(sg, page, PAGE_SIZE, 0);
2175 if (!i || page_to_pfn(page) != last_pfn + 1) {
2179 sg_set_page(sg, page, PAGE_SIZE, 0);
2181 sg->length += PAGE_SIZE;
2183 last_pfn = page_to_pfn(page);
2185 /* Check that the i965g/gm workaround works. */
2186 WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
2188 #ifdef CONFIG_SWIOTLB
2189 if (!swiotlb_nr_tbl())
2194 if (i915_gem_object_needs_bit17_swizzle(obj))
2195 i915_gem_object_do_bit_17_swizzle(obj);
2197 if (obj->tiling_mode != I915_TILING_NONE &&
2198 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES)
2199 i915_gem_object_pin_pages(obj);
2205 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0)
2206 page_cache_release(sg_page_iter_page(&sg_iter));
2210 /* shmemfs first checks if there is enough memory to allocate the page
2211 * and reports ENOSPC should there be insufficient, along with the usual
2212 * ENOMEM for a genuine allocation failure.
2214 * We use ENOSPC in our driver to mean that we have run out of aperture
2215 * space and so want to translate the error from shmemfs back to our
2216 * usual understanding of ENOMEM.
2218 if (PTR_ERR(page) == -ENOSPC)
2221 return PTR_ERR(page);
2224 /* Ensure that the associated pages are gathered from the backing storage
2225 * and pinned into our object. i915_gem_object_get_pages() may be called
2226 * multiple times before they are released by a single call to
2227 * i915_gem_object_put_pages() - once the pages are no longer referenced
2228 * either as a result of memory pressure (reaping pages under the shrinker)
2229 * or as the object is itself released.
2232 i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
2234 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2235 const struct drm_i915_gem_object_ops *ops = obj->ops;
2241 if (obj->madv != I915_MADV_WILLNEED) {
2242 DRM_DEBUG("Attempting to obtain a purgeable object\n");
2246 BUG_ON(obj->pages_pin_count);
2248 ret = ops->get_pages(obj);
2252 list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list);
2257 i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
2258 struct intel_engine_cs *ring)
2260 u32 seqno = intel_ring_get_seqno(ring);
2262 BUG_ON(ring == NULL);
2263 if (obj->ring != ring && obj->last_write_seqno) {
2264 /* Keep the seqno relative to the current ring */
2265 obj->last_write_seqno = seqno;
2269 /* Add a reference if we're newly entering the active list. */
2271 drm_gem_object_reference(&obj->base);
2275 list_move_tail(&obj->ring_list, &ring->active_list);
2277 obj->last_read_seqno = seqno;
2280 void i915_vma_move_to_active(struct i915_vma *vma,
2281 struct intel_engine_cs *ring)
2283 list_move_tail(&vma->mm_list, &vma->vm->active_list);
2284 return i915_gem_object_move_to_active(vma->obj, ring);
2288 i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
2290 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2291 struct i915_address_space *vm;
2292 struct i915_vma *vma;
2294 BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS);
2295 BUG_ON(!obj->active);
2297 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
2298 vma = i915_gem_obj_to_vma(obj, vm);
2299 if (vma && !list_empty(&vma->mm_list))
2300 list_move_tail(&vma->mm_list, &vm->inactive_list);
2303 intel_fb_obj_flush(obj, true);
2305 list_del_init(&obj->ring_list);
2308 obj->last_read_seqno = 0;
2309 obj->last_write_seqno = 0;
2310 obj->base.write_domain = 0;
2312 obj->last_fenced_seqno = 0;
2315 drm_gem_object_unreference(&obj->base);
2317 WARN_ON(i915_verify_lists(dev));
2321 i915_gem_object_retire(struct drm_i915_gem_object *obj)
2323 struct intel_engine_cs *ring = obj->ring;
2328 if (i915_seqno_passed(ring->get_seqno(ring, true),
2329 obj->last_read_seqno))
2330 i915_gem_object_move_to_inactive(obj);
2334 i915_gem_init_seqno(struct drm_device *dev, u32 seqno)
2336 struct drm_i915_private *dev_priv = dev->dev_private;
2337 struct intel_engine_cs *ring;
2340 /* Carefully retire all requests without writing to the rings */
2341 for_each_ring(ring, dev_priv, i) {
2342 ret = intel_ring_idle(ring);
2346 i915_gem_retire_requests(dev);
2348 /* Finally reset hw state */
2349 for_each_ring(ring, dev_priv, i) {
2350 intel_ring_init_seqno(ring, seqno);
2352 for (j = 0; j < ARRAY_SIZE(ring->semaphore.sync_seqno); j++)
2353 ring->semaphore.sync_seqno[j] = 0;
2359 int i915_gem_set_seqno(struct drm_device *dev, u32 seqno)
2361 struct drm_i915_private *dev_priv = dev->dev_private;
2367 /* HWS page needs to be set less than what we
2368 * will inject to ring
2370 ret = i915_gem_init_seqno(dev, seqno - 1);
2374 /* Carefully set the last_seqno value so that wrap
2375 * detection still works
2377 dev_priv->next_seqno = seqno;
2378 dev_priv->last_seqno = seqno - 1;
2379 if (dev_priv->last_seqno == 0)
2380 dev_priv->last_seqno--;
2386 i915_gem_get_seqno(struct drm_device *dev, u32 *seqno)
2388 struct drm_i915_private *dev_priv = dev->dev_private;
2390 /* reserve 0 for non-seqno */
2391 if (dev_priv->next_seqno == 0) {
2392 int ret = i915_gem_init_seqno(dev, 0);
2396 dev_priv->next_seqno = 1;
2399 *seqno = dev_priv->last_seqno = dev_priv->next_seqno++;
2403 int __i915_add_request(struct intel_engine_cs *ring,
2404 struct drm_file *file,
2405 struct drm_i915_gem_object *obj,
2408 struct drm_i915_private *dev_priv = ring->dev->dev_private;
2409 struct drm_i915_gem_request *request;
2410 struct intel_ringbuffer *ringbuf;
2411 u32 request_ring_position, request_start;
2414 request = ring->preallocated_lazy_request;
2415 if (WARN_ON(request == NULL))
2418 if (i915.enable_execlists) {
2419 struct intel_context *ctx = request->ctx;
2420 ringbuf = ctx->engine[ring->id].ringbuf;
2422 ringbuf = ring->buffer;
2424 request_start = intel_ring_get_tail(ringbuf);
2426 * Emit any outstanding flushes - execbuf can fail to emit the flush
2427 * after having emitted the batchbuffer command. Hence we need to fix
2428 * things up similar to emitting the lazy request. The difference here
2429 * is that the flush _must_ happen before the next request, no matter
2432 if (i915.enable_execlists) {
2433 ret = logical_ring_flush_all_caches(ringbuf);
2437 ret = intel_ring_flush_all_caches(ring);
2442 /* Record the position of the start of the request so that
2443 * should we detect the updated seqno part-way through the
2444 * GPU processing the request, we never over-estimate the
2445 * position of the head.
2447 request_ring_position = intel_ring_get_tail(ringbuf);
2449 if (i915.enable_execlists) {
2450 ret = ring->emit_request(ringbuf);
2454 ret = ring->add_request(ring);
2459 request->seqno = intel_ring_get_seqno(ring);
2460 request->ring = ring;
2461 request->head = request_start;
2462 request->tail = request_ring_position;
2464 /* Whilst this request exists, batch_obj will be on the
2465 * active_list, and so will hold the active reference. Only when this
2466 * request is retired will the the batch_obj be moved onto the
2467 * inactive_list and lose its active reference. Hence we do not need
2468 * to explicitly hold another reference here.
2470 request->batch_obj = obj;
2472 if (!i915.enable_execlists) {
2473 /* Hold a reference to the current context so that we can inspect
2474 * it later in case a hangcheck error event fires.
2476 request->ctx = ring->last_context;
2478 i915_gem_context_reference(request->ctx);
2481 request->emitted_jiffies = jiffies;
2482 list_add_tail(&request->list, &ring->request_list);
2483 request->file_priv = NULL;
2486 struct drm_i915_file_private *file_priv = file->driver_priv;
2488 spin_lock(&file_priv->mm.lock);
2489 request->file_priv = file_priv;
2490 list_add_tail(&request->client_list,
2491 &file_priv->mm.request_list);
2492 spin_unlock(&file_priv->mm.lock);
2495 trace_i915_gem_request_add(ring, request->seqno);
2496 ring->outstanding_lazy_seqno = 0;
2497 ring->preallocated_lazy_request = NULL;
2499 i915_queue_hangcheck(ring->dev);
2501 cancel_delayed_work_sync(&dev_priv->mm.idle_work);
2502 queue_delayed_work(dev_priv->wq,
2503 &dev_priv->mm.retire_work,
2504 round_jiffies_up_relative(HZ));
2505 intel_mark_busy(dev_priv->dev);
2508 *out_seqno = request->seqno;
2513 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
2515 struct drm_i915_file_private *file_priv = request->file_priv;
2520 spin_lock(&file_priv->mm.lock);
2521 list_del(&request->client_list);
2522 request->file_priv = NULL;
2523 spin_unlock(&file_priv->mm.lock);
2526 static bool i915_context_is_banned(struct drm_i915_private *dev_priv,
2527 const struct intel_context *ctx)
2529 unsigned long elapsed;
2531 elapsed = get_seconds() - ctx->hang_stats.guilty_ts;
2533 if (ctx->hang_stats.banned)
2536 if (elapsed <= DRM_I915_CTX_BAN_PERIOD) {
2537 if (!i915_gem_context_is_default(ctx)) {
2538 DRM_DEBUG("context hanging too fast, banning!\n");
2540 } else if (i915_stop_ring_allow_ban(dev_priv)) {
2541 if (i915_stop_ring_allow_warn(dev_priv))
2542 DRM_ERROR("gpu hanging too fast, banning!\n");
2550 static void i915_set_reset_status(struct drm_i915_private *dev_priv,
2551 struct intel_context *ctx,
2554 struct i915_ctx_hang_stats *hs;
2559 hs = &ctx->hang_stats;
2562 hs->banned = i915_context_is_banned(dev_priv, ctx);
2564 hs->guilty_ts = get_seconds();
2566 hs->batch_pending++;
2570 static void i915_gem_free_request(struct drm_i915_gem_request *request)
2572 struct intel_context *ctx = request->ctx;
2574 list_del(&request->list);
2575 i915_gem_request_remove_from_client(request);
2578 if (i915.enable_execlists) {
2579 struct intel_engine_cs *ring = request->ring;
2581 if (ctx != ring->default_context)
2582 intel_lr_context_unpin(ring, ctx);
2584 i915_gem_context_unreference(ctx);
2589 struct drm_i915_gem_request *
2590 i915_gem_find_active_request(struct intel_engine_cs *ring)
2592 struct drm_i915_gem_request *request;
2593 u32 completed_seqno;
2595 completed_seqno = ring->get_seqno(ring, false);
2597 list_for_each_entry(request, &ring->request_list, list) {
2598 if (i915_seqno_passed(completed_seqno, request->seqno))
2607 static void i915_gem_reset_ring_status(struct drm_i915_private *dev_priv,
2608 struct intel_engine_cs *ring)
2610 struct drm_i915_gem_request *request;
2613 request = i915_gem_find_active_request(ring);
2615 if (request == NULL)
2618 ring_hung = ring->hangcheck.score >= HANGCHECK_SCORE_RING_HUNG;
2620 i915_set_reset_status(dev_priv, request->ctx, ring_hung);
2622 list_for_each_entry_continue(request, &ring->request_list, list)
2623 i915_set_reset_status(dev_priv, request->ctx, false);
2626 static void i915_gem_reset_ring_cleanup(struct drm_i915_private *dev_priv,
2627 struct intel_engine_cs *ring)
2629 while (!list_empty(&ring->active_list)) {
2630 struct drm_i915_gem_object *obj;
2632 obj = list_first_entry(&ring->active_list,
2633 struct drm_i915_gem_object,
2636 i915_gem_object_move_to_inactive(obj);
2640 * Clear the execlists queue up before freeing the requests, as those
2641 * are the ones that keep the context and ringbuffer backing objects
2644 while (!list_empty(&ring->execlist_queue)) {
2645 struct intel_ctx_submit_request *submit_req;
2647 submit_req = list_first_entry(&ring->execlist_queue,
2648 struct intel_ctx_submit_request,
2650 list_del(&submit_req->execlist_link);
2651 intel_runtime_pm_put(dev_priv);
2652 i915_gem_context_unreference(submit_req->ctx);
2657 * We must free the requests after all the corresponding objects have
2658 * been moved off active lists. Which is the same order as the normal
2659 * retire_requests function does. This is important if object hold
2660 * implicit references on things like e.g. ppgtt address spaces through
2663 while (!list_empty(&ring->request_list)) {
2664 struct drm_i915_gem_request *request;
2666 request = list_first_entry(&ring->request_list,
2667 struct drm_i915_gem_request,
2670 i915_gem_free_request(request);
2673 /* These may not have been flush before the reset, do so now */
2674 kfree(ring->preallocated_lazy_request);
2675 ring->preallocated_lazy_request = NULL;
2676 ring->outstanding_lazy_seqno = 0;
2679 void i915_gem_restore_fences(struct drm_device *dev)
2681 struct drm_i915_private *dev_priv = dev->dev_private;
2684 for (i = 0; i < dev_priv->num_fence_regs; i++) {
2685 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
2688 * Commit delayed tiling changes if we have an object still
2689 * attached to the fence, otherwise just clear the fence.
2692 i915_gem_object_update_fence(reg->obj, reg,
2693 reg->obj->tiling_mode);
2695 i915_gem_write_fence(dev, i, NULL);
2700 void i915_gem_reset(struct drm_device *dev)
2702 struct drm_i915_private *dev_priv = dev->dev_private;
2703 struct intel_engine_cs *ring;
2707 * Before we free the objects from the requests, we need to inspect
2708 * them for finding the guilty party. As the requests only borrow
2709 * their reference to the objects, the inspection must be done first.
2711 for_each_ring(ring, dev_priv, i)
2712 i915_gem_reset_ring_status(dev_priv, ring);
2714 for_each_ring(ring, dev_priv, i)
2715 i915_gem_reset_ring_cleanup(dev_priv, ring);
2717 i915_gem_context_reset(dev);
2719 i915_gem_restore_fences(dev);
2723 * This function clears the request list as sequence numbers are passed.
2726 i915_gem_retire_requests_ring(struct intel_engine_cs *ring)
2730 if (list_empty(&ring->request_list))
2733 WARN_ON(i915_verify_lists(ring->dev));
2735 seqno = ring->get_seqno(ring, true);
2737 /* Move any buffers on the active list that are no longer referenced
2738 * by the ringbuffer to the flushing/inactive lists as appropriate,
2739 * before we free the context associated with the requests.
2741 while (!list_empty(&ring->active_list)) {
2742 struct drm_i915_gem_object *obj;
2744 obj = list_first_entry(&ring->active_list,
2745 struct drm_i915_gem_object,
2748 if (!i915_seqno_passed(seqno, obj->last_read_seqno))
2751 i915_gem_object_move_to_inactive(obj);
2755 while (!list_empty(&ring->request_list)) {
2756 struct drm_i915_gem_request *request;
2757 struct intel_ringbuffer *ringbuf;
2759 request = list_first_entry(&ring->request_list,
2760 struct drm_i915_gem_request,
2763 if (!i915_seqno_passed(seqno, request->seqno))
2766 trace_i915_gem_request_retire(ring, request->seqno);
2768 /* This is one of the few common intersection points
2769 * between legacy ringbuffer submission and execlists:
2770 * we need to tell them apart in order to find the correct
2771 * ringbuffer to which the request belongs to.
2773 if (i915.enable_execlists) {
2774 struct intel_context *ctx = request->ctx;
2775 ringbuf = ctx->engine[ring->id].ringbuf;
2777 ringbuf = ring->buffer;
2779 /* We know the GPU must have read the request to have
2780 * sent us the seqno + interrupt, so use the position
2781 * of tail of the request to update the last known position
2784 ringbuf->last_retired_head = request->tail;
2786 i915_gem_free_request(request);
2789 if (unlikely(ring->trace_irq_seqno &&
2790 i915_seqno_passed(seqno, ring->trace_irq_seqno))) {
2791 ring->irq_put(ring);
2792 ring->trace_irq_seqno = 0;
2795 WARN_ON(i915_verify_lists(ring->dev));
2799 i915_gem_retire_requests(struct drm_device *dev)
2801 struct drm_i915_private *dev_priv = dev->dev_private;
2802 struct intel_engine_cs *ring;
2806 for_each_ring(ring, dev_priv, i) {
2807 i915_gem_retire_requests_ring(ring);
2808 idle &= list_empty(&ring->request_list);
2809 if (i915.enable_execlists) {
2810 unsigned long flags;
2812 spin_lock_irqsave(&ring->execlist_lock, flags);
2813 idle &= list_empty(&ring->execlist_queue);
2814 spin_unlock_irqrestore(&ring->execlist_lock, flags);
2816 intel_execlists_retire_requests(ring);
2821 mod_delayed_work(dev_priv->wq,
2822 &dev_priv->mm.idle_work,
2823 msecs_to_jiffies(100));
2829 i915_gem_retire_work_handler(struct work_struct *work)
2831 struct drm_i915_private *dev_priv =
2832 container_of(work, typeof(*dev_priv), mm.retire_work.work);
2833 struct drm_device *dev = dev_priv->dev;
2836 /* Come back later if the device is busy... */
2838 if (mutex_trylock(&dev->struct_mutex)) {
2839 idle = i915_gem_retire_requests(dev);
2840 mutex_unlock(&dev->struct_mutex);
2843 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
2844 round_jiffies_up_relative(HZ));
2848 i915_gem_idle_work_handler(struct work_struct *work)
2850 struct drm_i915_private *dev_priv =
2851 container_of(work, typeof(*dev_priv), mm.idle_work.work);
2853 intel_mark_idle(dev_priv->dev);
2857 * Ensures that an object will eventually get non-busy by flushing any required
2858 * write domains, emitting any outstanding lazy request and retiring and
2859 * completed requests.
2862 i915_gem_object_flush_active(struct drm_i915_gem_object *obj)
2867 ret = i915_gem_check_olr(obj->ring, obj->last_read_seqno);
2871 i915_gem_retire_requests_ring(obj->ring);
2878 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
2879 * @DRM_IOCTL_ARGS: standard ioctl arguments
2881 * Returns 0 if successful, else an error is returned with the remaining time in
2882 * the timeout parameter.
2883 * -ETIME: object is still busy after timeout
2884 * -ERESTARTSYS: signal interrupted the wait
2885 * -ENONENT: object doesn't exist
2886 * Also possible, but rare:
2887 * -EAGAIN: GPU wedged
2889 * -ENODEV: Internal IRQ fail
2890 * -E?: The add request failed
2892 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
2893 * non-zero timeout parameter the wait ioctl will wait for the given number of
2894 * nanoseconds on an object becoming unbusy. Since the wait itself does so
2895 * without holding struct_mutex the object may become re-busied before this
2896 * function completes. A similar but shorter * race condition exists in the busy
2900 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
2902 struct drm_i915_private *dev_priv = dev->dev_private;
2903 struct drm_i915_gem_wait *args = data;
2904 struct drm_i915_gem_object *obj;
2905 struct intel_engine_cs *ring = NULL;
2906 unsigned reset_counter;
2910 if (args->flags != 0)
2913 ret = i915_mutex_lock_interruptible(dev);
2917 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle));
2918 if (&obj->base == NULL) {
2919 mutex_unlock(&dev->struct_mutex);
2923 /* Need to make sure the object gets inactive eventually. */
2924 ret = i915_gem_object_flush_active(obj);
2929 seqno = obj->last_read_seqno;
2936 /* Do this after OLR check to make sure we make forward progress polling
2937 * on this IOCTL with a timeout <=0 (like busy ioctl)
2939 if (args->timeout_ns <= 0) {
2944 drm_gem_object_unreference(&obj->base);
2945 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
2946 mutex_unlock(&dev->struct_mutex);
2948 return __i915_wait_seqno(ring, seqno, reset_counter, true,
2949 &args->timeout_ns, file->driver_priv);
2952 drm_gem_object_unreference(&obj->base);
2953 mutex_unlock(&dev->struct_mutex);
2958 * i915_gem_object_sync - sync an object to a ring.
2960 * @obj: object which may be in use on another ring.
2961 * @to: ring we wish to use the object on. May be NULL.
2963 * This code is meant to abstract object synchronization with the GPU.
2964 * Calling with NULL implies synchronizing the object with the CPU
2965 * rather than a particular GPU ring.
2967 * Returns 0 if successful, else propagates up the lower layer error.
2970 i915_gem_object_sync(struct drm_i915_gem_object *obj,
2971 struct intel_engine_cs *to)
2973 struct intel_engine_cs *from = obj->ring;
2977 if (from == NULL || to == from)
2980 if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev))
2981 return i915_gem_object_wait_rendering(obj, false);
2983 idx = intel_ring_sync_index(from, to);
2985 seqno = obj->last_read_seqno;
2986 /* Optimization: Avoid semaphore sync when we are sure we already
2987 * waited for an object with higher seqno */
2988 if (seqno <= from->semaphore.sync_seqno[idx])
2991 ret = i915_gem_check_olr(obj->ring, seqno);
2995 trace_i915_gem_ring_sync_to(from, to, seqno);
2996 ret = to->semaphore.sync_to(to, from, seqno);
2998 /* We use last_read_seqno because sync_to()
2999 * might have just caused seqno wrap under
3002 from->semaphore.sync_seqno[idx] = obj->last_read_seqno;
3007 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
3009 u32 old_write_domain, old_read_domains;
3011 /* Force a pagefault for domain tracking on next user access */
3012 i915_gem_release_mmap(obj);
3014 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
3017 /* Wait for any direct GTT access to complete */
3020 old_read_domains = obj->base.read_domains;
3021 old_write_domain = obj->base.write_domain;
3023 obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
3024 obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
3026 trace_i915_gem_object_change_domain(obj,
3031 int i915_vma_unbind(struct i915_vma *vma)
3033 struct drm_i915_gem_object *obj = vma->obj;
3034 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3037 if (list_empty(&vma->vma_link))
3040 if (!drm_mm_node_allocated(&vma->node)) {
3041 i915_gem_vma_destroy(vma);
3048 BUG_ON(obj->pages == NULL);
3050 ret = i915_gem_object_finish_gpu(obj);
3053 /* Continue on if we fail due to EIO, the GPU is hung so we
3054 * should be safe and we need to cleanup or else we might
3055 * cause memory corruption through use-after-free.
3058 /* Throw away the active reference before moving to the unbound list */
3059 i915_gem_object_retire(obj);
3061 if (i915_is_ggtt(vma->vm)) {
3062 i915_gem_object_finish_gtt(obj);
3064 /* release the fence reg _after_ flushing */
3065 ret = i915_gem_object_put_fence(obj);
3070 trace_i915_vma_unbind(vma);
3072 vma->unbind_vma(vma);
3074 list_del_init(&vma->mm_list);
3075 if (i915_is_ggtt(vma->vm))
3076 obj->map_and_fenceable = false;
3078 drm_mm_remove_node(&vma->node);
3079 i915_gem_vma_destroy(vma);
3081 /* Since the unbound list is global, only move to that list if
3082 * no more VMAs exist. */
3083 if (list_empty(&obj->vma_list)) {
3084 i915_gem_gtt_finish_object(obj);
3085 list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list);
3088 /* And finally now the object is completely decoupled from this vma,
3089 * we can drop its hold on the backing storage and allow it to be
3090 * reaped by the shrinker.
3092 i915_gem_object_unpin_pages(obj);
3097 int i915_gpu_idle(struct drm_device *dev)
3099 struct drm_i915_private *dev_priv = dev->dev_private;
3100 struct intel_engine_cs *ring;
3103 /* Flush everything onto the inactive list. */
3104 for_each_ring(ring, dev_priv, i) {
3105 if (!i915.enable_execlists) {
3106 ret = i915_switch_context(ring, ring->default_context);
3111 ret = intel_ring_idle(ring);
3119 static void i965_write_fence_reg(struct drm_device *dev, int reg,
3120 struct drm_i915_gem_object *obj)
3122 struct drm_i915_private *dev_priv = dev->dev_private;
3124 int fence_pitch_shift;
3126 if (INTEL_INFO(dev)->gen >= 6) {
3127 fence_reg = FENCE_REG_SANDYBRIDGE_0;
3128 fence_pitch_shift = SANDYBRIDGE_FENCE_PITCH_SHIFT;
3130 fence_reg = FENCE_REG_965_0;
3131 fence_pitch_shift = I965_FENCE_PITCH_SHIFT;
3134 fence_reg += reg * 8;
3136 /* To w/a incoherency with non-atomic 64-bit register updates,
3137 * we split the 64-bit update into two 32-bit writes. In order
3138 * for a partial fence not to be evaluated between writes, we
3139 * precede the update with write to turn off the fence register,
3140 * and only enable the fence as the last step.
3142 * For extra levels of paranoia, we make sure each step lands
3143 * before applying the next step.
3145 I915_WRITE(fence_reg, 0);
3146 POSTING_READ(fence_reg);
3149 u32 size = i915_gem_obj_ggtt_size(obj);
3152 val = (uint64_t)((i915_gem_obj_ggtt_offset(obj) + size - 4096) &
3154 val |= i915_gem_obj_ggtt_offset(obj) & 0xfffff000;
3155 val |= (uint64_t)((obj->stride / 128) - 1) << fence_pitch_shift;
3156 if (obj->tiling_mode == I915_TILING_Y)
3157 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
3158 val |= I965_FENCE_REG_VALID;
3160 I915_WRITE(fence_reg + 4, val >> 32);
3161 POSTING_READ(fence_reg + 4);
3163 I915_WRITE(fence_reg + 0, val);
3164 POSTING_READ(fence_reg);
3166 I915_WRITE(fence_reg + 4, 0);
3167 POSTING_READ(fence_reg + 4);
3171 static void i915_write_fence_reg(struct drm_device *dev, int reg,
3172 struct drm_i915_gem_object *obj)
3174 struct drm_i915_private *dev_priv = dev->dev_private;
3178 u32 size = i915_gem_obj_ggtt_size(obj);
3182 WARN((i915_gem_obj_ggtt_offset(obj) & ~I915_FENCE_START_MASK) ||
3183 (size & -size) != size ||
3184 (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
3185 "object 0x%08lx [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
3186 i915_gem_obj_ggtt_offset(obj), obj->map_and_fenceable, size);
3188 if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
3193 /* Note: pitch better be a power of two tile widths */
3194 pitch_val = obj->stride / tile_width;
3195 pitch_val = ffs(pitch_val) - 1;
3197 val = i915_gem_obj_ggtt_offset(obj);
3198 if (obj->tiling_mode == I915_TILING_Y)
3199 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
3200 val |= I915_FENCE_SIZE_BITS(size);
3201 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
3202 val |= I830_FENCE_REG_VALID;
3207 reg = FENCE_REG_830_0 + reg * 4;
3209 reg = FENCE_REG_945_8 + (reg - 8) * 4;
3211 I915_WRITE(reg, val);
3215 static void i830_write_fence_reg(struct drm_device *dev, int reg,
3216 struct drm_i915_gem_object *obj)
3218 struct drm_i915_private *dev_priv = dev->dev_private;
3222 u32 size = i915_gem_obj_ggtt_size(obj);
3225 WARN((i915_gem_obj_ggtt_offset(obj) & ~I830_FENCE_START_MASK) ||
3226 (size & -size) != size ||
3227 (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
3228 "object 0x%08lx not 512K or pot-size 0x%08x aligned\n",
3229 i915_gem_obj_ggtt_offset(obj), size);
3231 pitch_val = obj->stride / 128;
3232 pitch_val = ffs(pitch_val) - 1;
3234 val = i915_gem_obj_ggtt_offset(obj);
3235 if (obj->tiling_mode == I915_TILING_Y)
3236 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
3237 val |= I830_FENCE_SIZE_BITS(size);
3238 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
3239 val |= I830_FENCE_REG_VALID;
3243 I915_WRITE(FENCE_REG_830_0 + reg * 4, val);
3244 POSTING_READ(FENCE_REG_830_0 + reg * 4);
3247 inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object *obj)
3249 return obj && obj->base.read_domains & I915_GEM_DOMAIN_GTT;
3252 static void i915_gem_write_fence(struct drm_device *dev, int reg,
3253 struct drm_i915_gem_object *obj)
3255 struct drm_i915_private *dev_priv = dev->dev_private;
3257 /* Ensure that all CPU reads are completed before installing a fence
3258 * and all writes before removing the fence.
3260 if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj))
3263 WARN(obj && (!obj->stride || !obj->tiling_mode),
3264 "bogus fence setup with stride: 0x%x, tiling mode: %i\n",
3265 obj->stride, obj->tiling_mode);
3267 switch (INTEL_INFO(dev)->gen) {
3273 case 4: i965_write_fence_reg(dev, reg, obj); break;
3274 case 3: i915_write_fence_reg(dev, reg, obj); break;
3275 case 2: i830_write_fence_reg(dev, reg, obj); break;
3279 /* And similarly be paranoid that no direct access to this region
3280 * is reordered to before the fence is installed.
3282 if (i915_gem_object_needs_mb(obj))
3286 static inline int fence_number(struct drm_i915_private *dev_priv,
3287 struct drm_i915_fence_reg *fence)
3289 return fence - dev_priv->fence_regs;
3292 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
3293 struct drm_i915_fence_reg *fence,
3296 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3297 int reg = fence_number(dev_priv, fence);
3299 i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL);
3302 obj->fence_reg = reg;
3304 list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list);
3306 obj->fence_reg = I915_FENCE_REG_NONE;
3308 list_del_init(&fence->lru_list);
3310 obj->fence_dirty = false;
3314 i915_gem_object_wait_fence(struct drm_i915_gem_object *obj)
3316 if (obj->last_fenced_seqno) {
3317 int ret = i915_wait_seqno(obj->ring, obj->last_fenced_seqno);
3321 obj->last_fenced_seqno = 0;
3328 i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
3330 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3331 struct drm_i915_fence_reg *fence;
3334 ret = i915_gem_object_wait_fence(obj);
3338 if (obj->fence_reg == I915_FENCE_REG_NONE)
3341 fence = &dev_priv->fence_regs[obj->fence_reg];
3343 if (WARN_ON(fence->pin_count))
3346 i915_gem_object_fence_lost(obj);
3347 i915_gem_object_update_fence(obj, fence, false);
3352 static struct drm_i915_fence_reg *
3353 i915_find_fence_reg(struct drm_device *dev)
3355 struct drm_i915_private *dev_priv = dev->dev_private;
3356 struct drm_i915_fence_reg *reg, *avail;
3359 /* First try to find a free reg */
3361 for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
3362 reg = &dev_priv->fence_regs[i];
3366 if (!reg->pin_count)
3373 /* None available, try to steal one or wait for a user to finish */
3374 list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
3382 /* Wait for completion of pending flips which consume fences */
3383 if (intel_has_pending_fb_unpin(dev))
3384 return ERR_PTR(-EAGAIN);
3386 return ERR_PTR(-EDEADLK);
3390 * i915_gem_object_get_fence - set up fencing for an object
3391 * @obj: object to map through a fence reg
3393 * When mapping objects through the GTT, userspace wants to be able to write
3394 * to them without having to worry about swizzling if the object is tiled.
3395 * This function walks the fence regs looking for a free one for @obj,
3396 * stealing one if it can't find any.
3398 * It then sets up the reg based on the object's properties: address, pitch
3399 * and tiling format.
3401 * For an untiled surface, this removes any existing fence.
3404 i915_gem_object_get_fence(struct drm_i915_gem_object *obj)
3406 struct drm_device *dev = obj->base.dev;
3407 struct drm_i915_private *dev_priv = dev->dev_private;
3408 bool enable = obj->tiling_mode != I915_TILING_NONE;
3409 struct drm_i915_fence_reg *reg;
3412 /* Have we updated the tiling parameters upon the object and so
3413 * will need to serialise the write to the associated fence register?
3415 if (obj->fence_dirty) {
3416 ret = i915_gem_object_wait_fence(obj);
3421 /* Just update our place in the LRU if our fence is getting reused. */
3422 if (obj->fence_reg != I915_FENCE_REG_NONE) {
3423 reg = &dev_priv->fence_regs[obj->fence_reg];
3424 if (!obj->fence_dirty) {
3425 list_move_tail(®->lru_list,
3426 &dev_priv->mm.fence_list);
3429 } else if (enable) {
3430 if (WARN_ON(!obj->map_and_fenceable))
3433 reg = i915_find_fence_reg(dev);
3435 return PTR_ERR(reg);
3438 struct drm_i915_gem_object *old = reg->obj;
3440 ret = i915_gem_object_wait_fence(old);
3444 i915_gem_object_fence_lost(old);
3449 i915_gem_object_update_fence(obj, reg, enable);
3454 static bool i915_gem_valid_gtt_space(struct i915_vma *vma,
3455 unsigned long cache_level)
3457 struct drm_mm_node *gtt_space = &vma->node;
3458 struct drm_mm_node *other;
3461 * On some machines we have to be careful when putting differing types
3462 * of snoopable memory together to avoid the prefetcher crossing memory
3463 * domains and dying. During vm initialisation, we decide whether or not
3464 * these constraints apply and set the drm_mm.color_adjust
3467 if (vma->vm->mm.color_adjust == NULL)
3470 if (!drm_mm_node_allocated(gtt_space))
3473 if (list_empty(>t_space->node_list))
3476 other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list);
3477 if (other->allocated && !other->hole_follows && other->color != cache_level)
3480 other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list);
3481 if (other->allocated && !gtt_space->hole_follows && other->color != cache_level)
3488 * Finds free space in the GTT aperture and binds the object there.
3490 static struct i915_vma *
3491 i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj,
3492 struct i915_address_space *vm,
3496 struct drm_device *dev = obj->base.dev;
3497 struct drm_i915_private *dev_priv = dev->dev_private;
3498 u32 size, fence_size, fence_alignment, unfenced_alignment;
3499 unsigned long start =
3500 flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
3502 flags & PIN_MAPPABLE ? dev_priv->gtt.mappable_end : vm->total;
3503 struct i915_vma *vma;
3506 fence_size = i915_gem_get_gtt_size(dev,
3509 fence_alignment = i915_gem_get_gtt_alignment(dev,
3511 obj->tiling_mode, true);
3512 unfenced_alignment =
3513 i915_gem_get_gtt_alignment(dev,
3515 obj->tiling_mode, false);
3518 alignment = flags & PIN_MAPPABLE ? fence_alignment :
3520 if (flags & PIN_MAPPABLE && alignment & (fence_alignment - 1)) {
3521 DRM_DEBUG("Invalid object alignment requested %u\n", alignment);
3522 return ERR_PTR(-EINVAL);
3525 size = flags & PIN_MAPPABLE ? fence_size : obj->base.size;
3527 /* If the object is bigger than the entire aperture, reject it early
3528 * before evicting everything in a vain attempt to find space.
3530 if (obj->base.size > end) {
3531 DRM_DEBUG("Attempting to bind an object larger than the aperture: object=%zd > %s aperture=%lu\n",
3533 flags & PIN_MAPPABLE ? "mappable" : "total",
3535 return ERR_PTR(-E2BIG);
3538 ret = i915_gem_object_get_pages(obj);
3540 return ERR_PTR(ret);
3542 i915_gem_object_pin_pages(obj);
3544 vma = i915_gem_obj_lookup_or_create_vma(obj, vm);
3549 ret = drm_mm_insert_node_in_range_generic(&vm->mm, &vma->node,
3553 DRM_MM_SEARCH_DEFAULT,
3554 DRM_MM_CREATE_DEFAULT);
3556 ret = i915_gem_evict_something(dev, vm, size, alignment,
3565 if (WARN_ON(!i915_gem_valid_gtt_space(vma, obj->cache_level))) {
3567 goto err_remove_node;
3570 ret = i915_gem_gtt_prepare_object(obj);
3572 goto err_remove_node;
3574 list_move_tail(&obj->global_list, &dev_priv->mm.bound_list);
3575 list_add_tail(&vma->mm_list, &vm->inactive_list);
3577 trace_i915_vma_bind(vma, flags);
3578 vma->bind_vma(vma, obj->cache_level,
3579 flags & PIN_GLOBAL ? GLOBAL_BIND : 0);
3584 drm_mm_remove_node(&vma->node);
3586 i915_gem_vma_destroy(vma);
3589 i915_gem_object_unpin_pages(obj);
3594 i915_gem_clflush_object(struct drm_i915_gem_object *obj,
3597 /* If we don't have a page list set up, then we're not pinned
3598 * to GPU, and we can ignore the cache flush because it'll happen
3599 * again at bind time.
3601 if (obj->pages == NULL)
3605 * Stolen memory is always coherent with the GPU as it is explicitly
3606 * marked as wc by the system, or the system is cache-coherent.
3608 if (obj->stolen || obj->phys_handle)
3611 /* If the GPU is snooping the contents of the CPU cache,
3612 * we do not need to manually clear the CPU cache lines. However,
3613 * the caches are only snooped when the render cache is
3614 * flushed/invalidated. As we always have to emit invalidations
3615 * and flushes when moving into and out of the RENDER domain, correct
3616 * snooping behaviour occurs naturally as the result of our domain
3619 if (!force && cpu_cache_is_coherent(obj->base.dev, obj->cache_level))
3622 trace_i915_gem_object_clflush(obj);
3623 drm_clflush_sg(obj->pages);
3628 /** Flushes the GTT write domain for the object if it's dirty. */
3630 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
3632 uint32_t old_write_domain;
3634 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
3637 /* No actual flushing is required for the GTT write domain. Writes
3638 * to it immediately go to main memory as far as we know, so there's
3639 * no chipset flush. It also doesn't land in render cache.
3641 * However, we do have to enforce the order so that all writes through
3642 * the GTT land before any writes to the device, such as updates to
3647 old_write_domain = obj->base.write_domain;
3648 obj->base.write_domain = 0;
3650 intel_fb_obj_flush(obj, false);
3652 trace_i915_gem_object_change_domain(obj,
3653 obj->base.read_domains,
3657 /** Flushes the CPU write domain for the object if it's dirty. */
3659 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj,
3662 uint32_t old_write_domain;
3664 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
3667 if (i915_gem_clflush_object(obj, force))
3668 i915_gem_chipset_flush(obj->base.dev);
3670 old_write_domain = obj->base.write_domain;
3671 obj->base.write_domain = 0;
3673 intel_fb_obj_flush(obj, false);
3675 trace_i915_gem_object_change_domain(obj,
3676 obj->base.read_domains,
3681 * Moves a single object to the GTT read, and possibly write domain.
3683 * This function returns when the move is complete, including waiting on
3687 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
3689 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3690 struct i915_vma *vma = i915_gem_obj_to_ggtt(obj);
3691 uint32_t old_write_domain, old_read_domains;
3694 /* Not valid to be called on unbound objects. */
3698 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
3701 ret = i915_gem_object_wait_rendering(obj, !write);
3705 i915_gem_object_retire(obj);
3706 i915_gem_object_flush_cpu_write_domain(obj, false);
3708 /* Serialise direct access to this object with the barriers for
3709 * coherent writes from the GPU, by effectively invalidating the
3710 * GTT domain upon first access.
3712 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
3715 old_write_domain = obj->base.write_domain;
3716 old_read_domains = obj->base.read_domains;
3718 /* It should now be out of any other write domains, and we can update
3719 * the domain values for our changes.
3721 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3722 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3724 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
3725 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
3730 intel_fb_obj_invalidate(obj, NULL);
3732 trace_i915_gem_object_change_domain(obj,
3736 /* And bump the LRU for this access */
3737 if (i915_gem_object_is_inactive(obj))
3738 list_move_tail(&vma->mm_list,
3739 &dev_priv->gtt.base.inactive_list);
3744 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
3745 enum i915_cache_level cache_level)
3747 struct drm_device *dev = obj->base.dev;
3748 struct i915_vma *vma, *next;
3751 if (obj->cache_level == cache_level)
3754 if (i915_gem_obj_is_pinned(obj)) {
3755 DRM_DEBUG("can not change the cache level of pinned objects\n");
3759 list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) {
3760 if (!i915_gem_valid_gtt_space(vma, cache_level)) {
3761 ret = i915_vma_unbind(vma);
3767 if (i915_gem_obj_bound_any(obj)) {
3768 ret = i915_gem_object_finish_gpu(obj);
3772 i915_gem_object_finish_gtt(obj);
3774 /* Before SandyBridge, you could not use tiling or fence
3775 * registers with snooped memory, so relinquish any fences
3776 * currently pointing to our region in the aperture.
3778 if (INTEL_INFO(dev)->gen < 6) {
3779 ret = i915_gem_object_put_fence(obj);
3784 list_for_each_entry(vma, &obj->vma_list, vma_link)
3785 if (drm_mm_node_allocated(&vma->node))
3786 vma->bind_vma(vma, cache_level,
3787 vma->bound & GLOBAL_BIND);
3790 list_for_each_entry(vma, &obj->vma_list, vma_link)
3791 vma->node.color = cache_level;
3792 obj->cache_level = cache_level;
3794 if (cpu_write_needs_clflush(obj)) {
3795 u32 old_read_domains, old_write_domain;
3797 /* If we're coming from LLC cached, then we haven't
3798 * actually been tracking whether the data is in the
3799 * CPU cache or not, since we only allow one bit set
3800 * in obj->write_domain and have been skipping the clflushes.
3801 * Just set it to the CPU cache for now.
3803 i915_gem_object_retire(obj);
3804 WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU);
3806 old_read_domains = obj->base.read_domains;
3807 old_write_domain = obj->base.write_domain;
3809 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3810 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3812 trace_i915_gem_object_change_domain(obj,
3820 int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data,
3821 struct drm_file *file)
3823 struct drm_i915_gem_caching *args = data;
3824 struct drm_i915_gem_object *obj;
3827 ret = i915_mutex_lock_interruptible(dev);
3831 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3832 if (&obj->base == NULL) {
3837 switch (obj->cache_level) {
3838 case I915_CACHE_LLC:
3839 case I915_CACHE_L3_LLC:
3840 args->caching = I915_CACHING_CACHED;
3844 args->caching = I915_CACHING_DISPLAY;
3848 args->caching = I915_CACHING_NONE;
3852 drm_gem_object_unreference(&obj->base);
3854 mutex_unlock(&dev->struct_mutex);
3858 int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
3859 struct drm_file *file)
3861 struct drm_i915_gem_caching *args = data;
3862 struct drm_i915_gem_object *obj;
3863 enum i915_cache_level level;
3866 switch (args->caching) {
3867 case I915_CACHING_NONE:
3868 level = I915_CACHE_NONE;
3870 case I915_CACHING_CACHED:
3871 level = I915_CACHE_LLC;
3873 case I915_CACHING_DISPLAY:
3874 level = HAS_WT(dev) ? I915_CACHE_WT : I915_CACHE_NONE;
3880 ret = i915_mutex_lock_interruptible(dev);
3884 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3885 if (&obj->base == NULL) {
3890 ret = i915_gem_object_set_cache_level(obj, level);
3892 drm_gem_object_unreference(&obj->base);
3894 mutex_unlock(&dev->struct_mutex);
3898 static bool is_pin_display(struct drm_i915_gem_object *obj)
3900 struct i915_vma *vma;
3902 vma = i915_gem_obj_to_ggtt(obj);
3906 /* There are 3 sources that pin objects:
3907 * 1. The display engine (scanouts, sprites, cursors);
3908 * 2. Reservations for execbuffer;
3911 * We can ignore reservations as we hold the struct_mutex and
3912 * are only called outside of the reservation path. The user
3913 * can only increment pin_count once, and so if after
3914 * subtracting the potential reference by the user, any pin_count
3915 * remains, it must be due to another use by the display engine.
3917 return vma->pin_count - !!obj->user_pin_count;
3921 * Prepare buffer for display plane (scanout, cursors, etc).
3922 * Can be called from an uninterruptible phase (modesetting) and allows
3923 * any flushes to be pipelined (for pageflips).
3926 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
3928 struct intel_engine_cs *pipelined)
3930 u32 old_read_domains, old_write_domain;
3931 bool was_pin_display;
3934 if (pipelined != obj->ring) {
3935 ret = i915_gem_object_sync(obj, pipelined);
3940 /* Mark the pin_display early so that we account for the
3941 * display coherency whilst setting up the cache domains.
3943 was_pin_display = obj->pin_display;
3944 obj->pin_display = true;
3946 /* The display engine is not coherent with the LLC cache on gen6. As
3947 * a result, we make sure that the pinning that is about to occur is
3948 * done with uncached PTEs. This is lowest common denominator for all
3951 * However for gen6+, we could do better by using the GFDT bit instead
3952 * of uncaching, which would allow us to flush all the LLC-cached data
3953 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3955 ret = i915_gem_object_set_cache_level(obj,
3956 HAS_WT(obj->base.dev) ? I915_CACHE_WT : I915_CACHE_NONE);
3958 goto err_unpin_display;
3960 /* As the user may map the buffer once pinned in the display plane
3961 * (e.g. libkms for the bootup splash), we have to ensure that we
3962 * always use map_and_fenceable for all scanout buffers.
3964 ret = i915_gem_obj_ggtt_pin(obj, alignment, PIN_MAPPABLE);
3966 goto err_unpin_display;
3968 i915_gem_object_flush_cpu_write_domain(obj, true);
3970 old_write_domain = obj->base.write_domain;
3971 old_read_domains = obj->base.read_domains;
3973 /* It should now be out of any other write domains, and we can update
3974 * the domain values for our changes.
3976 obj->base.write_domain = 0;
3977 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3979 trace_i915_gem_object_change_domain(obj,
3986 WARN_ON(was_pin_display != is_pin_display(obj));
3987 obj->pin_display = was_pin_display;
3992 i915_gem_object_unpin_from_display_plane(struct drm_i915_gem_object *obj)
3994 i915_gem_object_ggtt_unpin(obj);
3995 obj->pin_display = is_pin_display(obj);
3999 i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
4003 if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
4006 ret = i915_gem_object_wait_rendering(obj, false);
4010 /* Ensure that we invalidate the GPU's caches and TLBs. */
4011 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
4016 * Moves a single object to the CPU read, and possibly write domain.
4018 * This function returns when the move is complete, including waiting on
4022 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
4024 uint32_t old_write_domain, old_read_domains;
4027 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
4030 ret = i915_gem_object_wait_rendering(obj, !write);
4034 i915_gem_object_retire(obj);
4035 i915_gem_object_flush_gtt_write_domain(obj);
4037 old_write_domain = obj->base.write_domain;
4038 old_read_domains = obj->base.read_domains;
4040 /* Flush the CPU cache if it's still invalid. */
4041 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
4042 i915_gem_clflush_object(obj, false);
4044 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
4047 /* It should now be out of any other write domains, and we can update
4048 * the domain values for our changes.
4050 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
4052 /* If we're writing through the CPU, then the GPU read domains will
4053 * need to be invalidated at next use.
4056 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4057 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4061 intel_fb_obj_invalidate(obj, NULL);
4063 trace_i915_gem_object_change_domain(obj,
4070 /* Throttle our rendering by waiting until the ring has completed our requests
4071 * emitted over 20 msec ago.
4073 * Note that if we were to use the current jiffies each time around the loop,
4074 * we wouldn't escape the function with any frames outstanding if the time to
4075 * render a frame was over 20ms.
4077 * This should get us reasonable parallelism between CPU and GPU but also
4078 * relatively low latency when blocking on a particular request to finish.
4081 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
4083 struct drm_i915_private *dev_priv = dev->dev_private;
4084 struct drm_i915_file_private *file_priv = file->driver_priv;
4085 unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
4086 struct drm_i915_gem_request *request;
4087 struct intel_engine_cs *ring = NULL;
4088 unsigned reset_counter;
4092 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
4096 ret = i915_gem_check_wedge(&dev_priv->gpu_error, false);
4100 spin_lock(&file_priv->mm.lock);
4101 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
4102 if (time_after_eq(request->emitted_jiffies, recent_enough))
4105 ring = request->ring;
4106 seqno = request->seqno;
4108 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
4109 spin_unlock(&file_priv->mm.lock);
4114 ret = __i915_wait_seqno(ring, seqno, reset_counter, true, NULL, NULL);
4116 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
4122 i915_vma_misplaced(struct i915_vma *vma, uint32_t alignment, uint64_t flags)
4124 struct drm_i915_gem_object *obj = vma->obj;
4127 vma->node.start & (alignment - 1))
4130 if (flags & PIN_MAPPABLE && !obj->map_and_fenceable)
4133 if (flags & PIN_OFFSET_BIAS &&
4134 vma->node.start < (flags & PIN_OFFSET_MASK))
4141 i915_gem_object_pin(struct drm_i915_gem_object *obj,
4142 struct i915_address_space *vm,
4146 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
4147 struct i915_vma *vma;
4151 if (WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base))
4154 if (WARN_ON(flags & (PIN_GLOBAL | PIN_MAPPABLE) && !i915_is_ggtt(vm)))
4157 if (WARN_ON((flags & (PIN_MAPPABLE | PIN_GLOBAL)) == PIN_MAPPABLE))
4160 vma = i915_gem_obj_to_vma(obj, vm);
4162 if (WARN_ON(vma->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT))
4165 if (i915_vma_misplaced(vma, alignment, flags)) {
4166 WARN(vma->pin_count,
4167 "bo is already pinned with incorrect alignment:"
4168 " offset=%lx, req.alignment=%x, req.map_and_fenceable=%d,"
4169 " obj->map_and_fenceable=%d\n",
4170 i915_gem_obj_offset(obj, vm), alignment,
4171 !!(flags & PIN_MAPPABLE),
4172 obj->map_and_fenceable);
4173 ret = i915_vma_unbind(vma);
4181 bound = vma ? vma->bound : 0;
4182 if (vma == NULL || !drm_mm_node_allocated(&vma->node)) {
4183 vma = i915_gem_object_bind_to_vm(obj, vm, alignment, flags);
4185 return PTR_ERR(vma);
4188 if (flags & PIN_GLOBAL && !(vma->bound & GLOBAL_BIND))
4189 vma->bind_vma(vma, obj->cache_level, GLOBAL_BIND);
4191 if ((bound ^ vma->bound) & GLOBAL_BIND) {
4192 bool mappable, fenceable;
4193 u32 fence_size, fence_alignment;
4195 fence_size = i915_gem_get_gtt_size(obj->base.dev,
4198 fence_alignment = i915_gem_get_gtt_alignment(obj->base.dev,
4203 fenceable = (vma->node.size == fence_size &&
4204 (vma->node.start & (fence_alignment - 1)) == 0);
4206 mappable = (vma->node.start + obj->base.size <=
4207 dev_priv->gtt.mappable_end);
4209 obj->map_and_fenceable = mappable && fenceable;
4212 WARN_ON(flags & PIN_MAPPABLE && !obj->map_and_fenceable);
4215 if (flags & PIN_MAPPABLE)
4216 obj->pin_mappable |= true;
4222 i915_gem_object_ggtt_unpin(struct drm_i915_gem_object *obj)
4224 struct i915_vma *vma = i915_gem_obj_to_ggtt(obj);
4227 BUG_ON(vma->pin_count == 0);
4228 BUG_ON(!i915_gem_obj_ggtt_bound(obj));
4230 if (--vma->pin_count == 0)
4231 obj->pin_mappable = false;
4235 i915_gem_object_pin_fence(struct drm_i915_gem_object *obj)
4237 if (obj->fence_reg != I915_FENCE_REG_NONE) {
4238 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
4239 struct i915_vma *ggtt_vma = i915_gem_obj_to_ggtt(obj);
4241 WARN_ON(!ggtt_vma ||
4242 dev_priv->fence_regs[obj->fence_reg].pin_count >
4243 ggtt_vma->pin_count);
4244 dev_priv->fence_regs[obj->fence_reg].pin_count++;
4251 i915_gem_object_unpin_fence(struct drm_i915_gem_object *obj)
4253 if (obj->fence_reg != I915_FENCE_REG_NONE) {
4254 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
4255 WARN_ON(dev_priv->fence_regs[obj->fence_reg].pin_count <= 0);
4256 dev_priv->fence_regs[obj->fence_reg].pin_count--;
4261 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
4262 struct drm_file *file)
4264 struct drm_i915_gem_pin *args = data;
4265 struct drm_i915_gem_object *obj;
4268 if (drm_core_check_feature(dev, DRIVER_MODESET))
4271 ret = i915_mutex_lock_interruptible(dev);
4275 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
4276 if (&obj->base == NULL) {
4281 if (obj->madv != I915_MADV_WILLNEED) {
4282 DRM_DEBUG("Attempting to pin a purgeable buffer\n");
4287 if (obj->pin_filp != NULL && obj->pin_filp != file) {
4288 DRM_DEBUG("Already pinned in i915_gem_pin_ioctl(): %d\n",
4294 if (obj->user_pin_count == ULONG_MAX) {
4299 if (obj->user_pin_count == 0) {
4300 ret = i915_gem_obj_ggtt_pin(obj, args->alignment, PIN_MAPPABLE);
4305 obj->user_pin_count++;
4306 obj->pin_filp = file;
4308 args->offset = i915_gem_obj_ggtt_offset(obj);
4310 drm_gem_object_unreference(&obj->base);
4312 mutex_unlock(&dev->struct_mutex);
4317 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
4318 struct drm_file *file)
4320 struct drm_i915_gem_pin *args = data;
4321 struct drm_i915_gem_object *obj;
4324 if (drm_core_check_feature(dev, DRIVER_MODESET))
4327 ret = i915_mutex_lock_interruptible(dev);
4331 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
4332 if (&obj->base == NULL) {
4337 if (obj->pin_filp != file) {
4338 DRM_DEBUG("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
4343 obj->user_pin_count--;
4344 if (obj->user_pin_count == 0) {
4345 obj->pin_filp = NULL;
4346 i915_gem_object_ggtt_unpin(obj);
4350 drm_gem_object_unreference(&obj->base);
4352 mutex_unlock(&dev->struct_mutex);
4357 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
4358 struct drm_file *file)
4360 struct drm_i915_gem_busy *args = data;
4361 struct drm_i915_gem_object *obj;
4364 ret = i915_mutex_lock_interruptible(dev);
4368 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
4369 if (&obj->base == NULL) {
4374 /* Count all active objects as busy, even if they are currently not used
4375 * by the gpu. Users of this interface expect objects to eventually
4376 * become non-busy without any further actions, therefore emit any
4377 * necessary flushes here.
4379 ret = i915_gem_object_flush_active(obj);
4381 args->busy = obj->active;
4383 BUILD_BUG_ON(I915_NUM_RINGS > 16);
4384 args->busy |= intel_ring_flag(obj->ring) << 16;
4387 drm_gem_object_unreference(&obj->base);
4389 mutex_unlock(&dev->struct_mutex);
4394 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
4395 struct drm_file *file_priv)
4397 return i915_gem_ring_throttle(dev, file_priv);
4401 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
4402 struct drm_file *file_priv)
4404 struct drm_i915_private *dev_priv = dev->dev_private;
4405 struct drm_i915_gem_madvise *args = data;
4406 struct drm_i915_gem_object *obj;
4409 switch (args->madv) {
4410 case I915_MADV_DONTNEED:
4411 case I915_MADV_WILLNEED:
4417 ret = i915_mutex_lock_interruptible(dev);
4421 obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
4422 if (&obj->base == NULL) {
4427 if (i915_gem_obj_is_pinned(obj)) {
4433 obj->tiling_mode != I915_TILING_NONE &&
4434 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
4435 if (obj->madv == I915_MADV_WILLNEED)
4436 i915_gem_object_unpin_pages(obj);
4437 if (args->madv == I915_MADV_WILLNEED)
4438 i915_gem_object_pin_pages(obj);
4441 if (obj->madv != __I915_MADV_PURGED)
4442 obj->madv = args->madv;
4444 /* if the object is no longer attached, discard its backing storage */
4445 if (i915_gem_object_is_purgeable(obj) && obj->pages == NULL)
4446 i915_gem_object_truncate(obj);
4448 args->retained = obj->madv != __I915_MADV_PURGED;
4451 drm_gem_object_unreference(&obj->base);
4453 mutex_unlock(&dev->struct_mutex);
4457 void i915_gem_object_init(struct drm_i915_gem_object *obj,
4458 const struct drm_i915_gem_object_ops *ops)
4460 INIT_LIST_HEAD(&obj->global_list);
4461 INIT_LIST_HEAD(&obj->ring_list);
4462 INIT_LIST_HEAD(&obj->obj_exec_link);
4463 INIT_LIST_HEAD(&obj->vma_list);
4467 obj->fence_reg = I915_FENCE_REG_NONE;
4468 obj->madv = I915_MADV_WILLNEED;
4470 i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size);
4473 static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
4474 .get_pages = i915_gem_object_get_pages_gtt,
4475 .put_pages = i915_gem_object_put_pages_gtt,
4478 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
4481 struct drm_i915_gem_object *obj;
4482 struct address_space *mapping;
4485 obj = i915_gem_object_alloc(dev);
4489 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
4490 i915_gem_object_free(obj);
4494 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
4495 if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) {
4496 /* 965gm cannot relocate objects above 4GiB. */
4497 mask &= ~__GFP_HIGHMEM;
4498 mask |= __GFP_DMA32;
4501 mapping = file_inode(obj->base.filp)->i_mapping;
4502 mapping_set_gfp_mask(mapping, mask);
4504 i915_gem_object_init(obj, &i915_gem_object_ops);
4506 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4507 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4510 /* On some devices, we can have the GPU use the LLC (the CPU
4511 * cache) for about a 10% performance improvement
4512 * compared to uncached. Graphics requests other than
4513 * display scanout are coherent with the CPU in
4514 * accessing this cache. This means in this mode we
4515 * don't need to clflush on the CPU side, and on the
4516 * GPU side we only need to flush internal caches to
4517 * get data visible to the CPU.
4519 * However, we maintain the display planes as UC, and so
4520 * need to rebind when first used as such.
4522 obj->cache_level = I915_CACHE_LLC;
4524 obj->cache_level = I915_CACHE_NONE;
4526 trace_i915_gem_object_create(obj);
4531 static bool discard_backing_storage(struct drm_i915_gem_object *obj)
4533 /* If we are the last user of the backing storage (be it shmemfs
4534 * pages or stolen etc), we know that the pages are going to be
4535 * immediately released. In this case, we can then skip copying
4536 * back the contents from the GPU.
4539 if (obj->madv != I915_MADV_WILLNEED)
4542 if (obj->base.filp == NULL)
4545 /* At first glance, this looks racy, but then again so would be
4546 * userspace racing mmap against close. However, the first external
4547 * reference to the filp can only be obtained through the
4548 * i915_gem_mmap_ioctl() which safeguards us against the user
4549 * acquiring such a reference whilst we are in the middle of
4550 * freeing the object.
4552 return atomic_long_read(&obj->base.filp->f_count) == 1;
4555 void i915_gem_free_object(struct drm_gem_object *gem_obj)
4557 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
4558 struct drm_device *dev = obj->base.dev;
4559 struct drm_i915_private *dev_priv = dev->dev_private;
4560 struct i915_vma *vma, *next;
4562 intel_runtime_pm_get(dev_priv);
4564 trace_i915_gem_object_destroy(obj);
4566 list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) {
4570 ret = i915_vma_unbind(vma);
4571 if (WARN_ON(ret == -ERESTARTSYS)) {
4572 bool was_interruptible;
4574 was_interruptible = dev_priv->mm.interruptible;
4575 dev_priv->mm.interruptible = false;
4577 WARN_ON(i915_vma_unbind(vma));
4579 dev_priv->mm.interruptible = was_interruptible;
4583 /* Stolen objects don't hold a ref, but do hold pin count. Fix that up
4584 * before progressing. */
4586 i915_gem_object_unpin_pages(obj);
4588 WARN_ON(obj->frontbuffer_bits);
4590 if (obj->pages && obj->madv == I915_MADV_WILLNEED &&
4591 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES &&
4592 obj->tiling_mode != I915_TILING_NONE)
4593 i915_gem_object_unpin_pages(obj);
4595 if (WARN_ON(obj->pages_pin_count))
4596 obj->pages_pin_count = 0;
4597 if (discard_backing_storage(obj))
4598 obj->madv = I915_MADV_DONTNEED;
4599 i915_gem_object_put_pages(obj);
4600 i915_gem_object_free_mmap_offset(obj);
4604 if (obj->base.import_attach)
4605 drm_prime_gem_destroy(&obj->base, NULL);
4607 if (obj->ops->release)
4608 obj->ops->release(obj);
4610 drm_gem_object_release(&obj->base);
4611 i915_gem_info_remove_obj(dev_priv, obj->base.size);
4614 i915_gem_object_free(obj);
4616 intel_runtime_pm_put(dev_priv);
4619 struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj,
4620 struct i915_address_space *vm)
4622 struct i915_vma *vma;
4623 list_for_each_entry(vma, &obj->vma_list, vma_link)
4630 void i915_gem_vma_destroy(struct i915_vma *vma)
4632 struct i915_address_space *vm = NULL;
4633 WARN_ON(vma->node.allocated);
4635 /* Keep the vma as a placeholder in the execbuffer reservation lists */
4636 if (!list_empty(&vma->exec_list))
4641 if (!i915_is_ggtt(vm))
4642 i915_ppgtt_put(i915_vm_to_ppgtt(vm));
4644 list_del(&vma->vma_link);
4650 i915_gem_stop_ringbuffers(struct drm_device *dev)
4652 struct drm_i915_private *dev_priv = dev->dev_private;
4653 struct intel_engine_cs *ring;
4656 for_each_ring(ring, dev_priv, i)
4657 dev_priv->gt.stop_ring(ring);
4661 i915_gem_suspend(struct drm_device *dev)
4663 struct drm_i915_private *dev_priv = dev->dev_private;
4666 mutex_lock(&dev->struct_mutex);
4667 ret = i915_gpu_idle(dev);
4671 i915_gem_retire_requests(dev);
4673 /* Under UMS, be paranoid and evict. */
4674 if (!drm_core_check_feature(dev, DRIVER_MODESET))
4675 i915_gem_evict_everything(dev);
4677 i915_gem_stop_ringbuffers(dev);
4678 mutex_unlock(&dev->struct_mutex);
4680 del_timer_sync(&dev_priv->gpu_error.hangcheck_timer);
4681 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
4682 flush_delayed_work(&dev_priv->mm.idle_work);
4687 mutex_unlock(&dev->struct_mutex);
4691 int i915_gem_l3_remap(struct intel_engine_cs *ring, int slice)
4693 struct drm_device *dev = ring->dev;
4694 struct drm_i915_private *dev_priv = dev->dev_private;
4695 u32 reg_base = GEN7_L3LOG_BASE + (slice * 0x200);
4696 u32 *remap_info = dev_priv->l3_parity.remap_info[slice];
4699 if (!HAS_L3_DPF(dev) || !remap_info)
4702 ret = intel_ring_begin(ring, GEN7_L3LOG_SIZE / 4 * 3);
4707 * Note: We do not worry about the concurrent register cacheline hang
4708 * here because no other code should access these registers other than
4709 * at initialization time.
4711 for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) {
4712 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
4713 intel_ring_emit(ring, reg_base + i);
4714 intel_ring_emit(ring, remap_info[i/4]);
4717 intel_ring_advance(ring);
4722 void i915_gem_init_swizzling(struct drm_device *dev)
4724 struct drm_i915_private *dev_priv = dev->dev_private;
4726 if (INTEL_INFO(dev)->gen < 5 ||
4727 dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
4730 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
4731 DISP_TILE_SURFACE_SWIZZLING);
4736 I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
4738 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
4739 else if (IS_GEN7(dev))
4740 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
4741 else if (IS_GEN8(dev))
4742 I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW));
4748 intel_enable_blt(struct drm_device *dev)
4753 /* The blitter was dysfunctional on early prototypes */
4754 if (IS_GEN6(dev) && dev->pdev->revision < 8) {
4755 DRM_INFO("BLT not supported on this pre-production hardware;"
4756 " graphics performance will be degraded.\n");
4763 static void init_unused_ring(struct drm_device *dev, u32 base)
4765 struct drm_i915_private *dev_priv = dev->dev_private;
4767 I915_WRITE(RING_CTL(base), 0);
4768 I915_WRITE(RING_HEAD(base), 0);
4769 I915_WRITE(RING_TAIL(base), 0);
4770 I915_WRITE(RING_START(base), 0);
4773 static void init_unused_rings(struct drm_device *dev)
4776 init_unused_ring(dev, PRB1_BASE);
4777 init_unused_ring(dev, SRB0_BASE);
4778 init_unused_ring(dev, SRB1_BASE);
4779 init_unused_ring(dev, SRB2_BASE);
4780 init_unused_ring(dev, SRB3_BASE);
4781 } else if (IS_GEN2(dev)) {
4782 init_unused_ring(dev, SRB0_BASE);
4783 init_unused_ring(dev, SRB1_BASE);
4784 } else if (IS_GEN3(dev)) {
4785 init_unused_ring(dev, PRB1_BASE);
4786 init_unused_ring(dev, PRB2_BASE);
4790 int i915_gem_init_rings(struct drm_device *dev)
4792 struct drm_i915_private *dev_priv = dev->dev_private;
4796 * At least 830 can leave some of the unused rings
4797 * "active" (ie. head != tail) after resume which
4798 * will prevent c3 entry. Makes sure all unused rings
4801 init_unused_rings(dev);
4803 ret = intel_init_render_ring_buffer(dev);
4808 ret = intel_init_bsd_ring_buffer(dev);
4810 goto cleanup_render_ring;
4813 if (intel_enable_blt(dev)) {
4814 ret = intel_init_blt_ring_buffer(dev);
4816 goto cleanup_bsd_ring;
4819 if (HAS_VEBOX(dev)) {
4820 ret = intel_init_vebox_ring_buffer(dev);
4822 goto cleanup_blt_ring;
4825 if (HAS_BSD2(dev)) {
4826 ret = intel_init_bsd2_ring_buffer(dev);
4828 goto cleanup_vebox_ring;
4831 ret = i915_gem_set_seqno(dev, ((u32)~0 - 0x1000));
4833 goto cleanup_bsd2_ring;
4838 intel_cleanup_ring_buffer(&dev_priv->ring[VCS2]);
4840 intel_cleanup_ring_buffer(&dev_priv->ring[VECS]);
4842 intel_cleanup_ring_buffer(&dev_priv->ring[BCS]);
4844 intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
4845 cleanup_render_ring:
4846 intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
4852 i915_gem_init_hw(struct drm_device *dev)
4854 struct drm_i915_private *dev_priv = dev->dev_private;
4857 if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt())
4860 if (dev_priv->ellc_size)
4861 I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf));
4863 if (IS_HASWELL(dev))
4864 I915_WRITE(MI_PREDICATE_RESULT_2, IS_HSW_GT3(dev) ?
4865 LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED);
4867 if (HAS_PCH_NOP(dev)) {
4868 if (IS_IVYBRIDGE(dev)) {
4869 u32 temp = I915_READ(GEN7_MSG_CTL);
4870 temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK);
4871 I915_WRITE(GEN7_MSG_CTL, temp);
4872 } else if (INTEL_INFO(dev)->gen >= 7) {
4873 u32 temp = I915_READ(HSW_NDE_RSTWRN_OPT);
4874 temp &= ~RESET_PCH_HANDSHAKE_ENABLE;
4875 I915_WRITE(HSW_NDE_RSTWRN_OPT, temp);
4879 i915_gem_init_swizzling(dev);
4881 ret = dev_priv->gt.init_rings(dev);
4885 for (i = 0; i < NUM_L3_SLICES(dev); i++)
4886 i915_gem_l3_remap(&dev_priv->ring[RCS], i);
4889 * XXX: Contexts should only be initialized once. Doing a switch to the
4890 * default context switch however is something we'd like to do after
4891 * reset or thaw (the latter may not actually be necessary for HW, but
4892 * goes with our code better). Context switching requires rings (for
4893 * the do_switch), but before enabling PPGTT. So don't move this.
4895 ret = i915_gem_context_enable(dev_priv);
4896 if (ret && ret != -EIO) {
4897 DRM_ERROR("Context enable failed %d\n", ret);
4898 i915_gem_cleanup_ringbuffer(dev);
4903 ret = i915_ppgtt_init_hw(dev);
4904 if (ret && ret != -EIO) {
4905 DRM_ERROR("PPGTT enable failed %d\n", ret);
4906 i915_gem_cleanup_ringbuffer(dev);
4912 int i915_gem_init(struct drm_device *dev)
4914 struct drm_i915_private *dev_priv = dev->dev_private;
4917 i915.enable_execlists = intel_sanitize_enable_execlists(dev,
4918 i915.enable_execlists);
4920 mutex_lock(&dev->struct_mutex);
4922 if (IS_VALLEYVIEW(dev)) {
4923 /* VLVA0 (potential hack), BIOS isn't actually waking us */
4924 I915_WRITE(VLV_GTLC_WAKE_CTRL, VLV_GTLC_ALLOWWAKEREQ);
4925 if (wait_for((I915_READ(VLV_GTLC_PW_STATUS) &
4926 VLV_GTLC_ALLOWWAKEACK), 10))
4927 DRM_DEBUG_DRIVER("allow wake ack timed out\n");
4930 if (!i915.enable_execlists) {
4931 dev_priv->gt.do_execbuf = i915_gem_ringbuffer_submission;
4932 dev_priv->gt.init_rings = i915_gem_init_rings;
4933 dev_priv->gt.cleanup_ring = intel_cleanup_ring_buffer;
4934 dev_priv->gt.stop_ring = intel_stop_ring_buffer;
4936 dev_priv->gt.do_execbuf = intel_execlists_submission;
4937 dev_priv->gt.init_rings = intel_logical_rings_init;
4938 dev_priv->gt.cleanup_ring = intel_logical_ring_cleanup;
4939 dev_priv->gt.stop_ring = intel_logical_ring_stop;
4942 ret = i915_gem_init_userptr(dev);
4944 mutex_unlock(&dev->struct_mutex);
4948 i915_gem_init_global_gtt(dev);
4950 ret = i915_gem_context_init(dev);
4952 mutex_unlock(&dev->struct_mutex);
4956 ret = i915_gem_init_hw(dev);
4958 /* Allow ring initialisation to fail by marking the GPU as
4959 * wedged. But we only want to do this where the GPU is angry,
4960 * for all other failure, such as an allocation failure, bail.
4962 DRM_ERROR("Failed to initialize GPU, declaring it wedged\n");
4963 atomic_set_mask(I915_WEDGED, &dev_priv->gpu_error.reset_counter);
4966 mutex_unlock(&dev->struct_mutex);
4972 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
4974 struct drm_i915_private *dev_priv = dev->dev_private;
4975 struct intel_engine_cs *ring;
4978 for_each_ring(ring, dev_priv, i)
4979 dev_priv->gt.cleanup_ring(ring);
4983 init_ring_lists(struct intel_engine_cs *ring)
4985 INIT_LIST_HEAD(&ring->active_list);
4986 INIT_LIST_HEAD(&ring->request_list);
4989 void i915_init_vm(struct drm_i915_private *dev_priv,
4990 struct i915_address_space *vm)
4992 if (!i915_is_ggtt(vm))
4993 drm_mm_init(&vm->mm, vm->start, vm->total);
4994 vm->dev = dev_priv->dev;
4995 INIT_LIST_HEAD(&vm->active_list);
4996 INIT_LIST_HEAD(&vm->inactive_list);
4997 INIT_LIST_HEAD(&vm->global_link);
4998 list_add_tail(&vm->global_link, &dev_priv->vm_list);
5002 i915_gem_load(struct drm_device *dev)
5004 struct drm_i915_private *dev_priv = dev->dev_private;
5008 kmem_cache_create("i915_gem_object",
5009 sizeof(struct drm_i915_gem_object), 0,
5013 INIT_LIST_HEAD(&dev_priv->vm_list);
5014 i915_init_vm(dev_priv, &dev_priv->gtt.base);
5016 INIT_LIST_HEAD(&dev_priv->context_list);
5017 INIT_LIST_HEAD(&dev_priv->mm.unbound_list);
5018 INIT_LIST_HEAD(&dev_priv->mm.bound_list);
5019 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
5020 for (i = 0; i < I915_NUM_RINGS; i++)
5021 init_ring_lists(&dev_priv->ring[i]);
5022 for (i = 0; i < I915_MAX_NUM_FENCES; i++)
5023 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
5024 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
5025 i915_gem_retire_work_handler);
5026 INIT_DELAYED_WORK(&dev_priv->mm.idle_work,
5027 i915_gem_idle_work_handler);
5028 init_waitqueue_head(&dev_priv->gpu_error.reset_queue);
5030 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
5031 if (!drm_core_check_feature(dev, DRIVER_MODESET) && IS_GEN3(dev)) {
5032 I915_WRITE(MI_ARB_STATE,
5033 _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
5036 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
5038 /* Old X drivers will take 0-2 for front, back, depth buffers */
5039 if (!drm_core_check_feature(dev, DRIVER_MODESET))
5040 dev_priv->fence_reg_start = 3;
5042 if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev))
5043 dev_priv->num_fence_regs = 32;
5044 else if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
5045 dev_priv->num_fence_regs = 16;
5047 dev_priv->num_fence_regs = 8;
5049 /* Initialize fence registers to zero */
5050 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
5051 i915_gem_restore_fences(dev);
5053 i915_gem_detect_bit_6_swizzle(dev);
5054 init_waitqueue_head(&dev_priv->pending_flip_queue);
5056 dev_priv->mm.interruptible = true;
5058 dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
5059 dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count;
5060 dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS;
5061 register_shrinker(&dev_priv->mm.shrinker);
5063 dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
5064 register_oom_notifier(&dev_priv->mm.oom_notifier);
5066 mutex_init(&dev_priv->fb_tracking.lock);
5069 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
5071 struct drm_i915_file_private *file_priv = file->driver_priv;
5073 cancel_delayed_work_sync(&file_priv->mm.idle_work);
5075 /* Clean up our request list when the client is going away, so that
5076 * later retire_requests won't dereference our soon-to-be-gone
5079 spin_lock(&file_priv->mm.lock);
5080 while (!list_empty(&file_priv->mm.request_list)) {
5081 struct drm_i915_gem_request *request;
5083 request = list_first_entry(&file_priv->mm.request_list,
5084 struct drm_i915_gem_request,
5086 list_del(&request->client_list);
5087 request->file_priv = NULL;
5089 spin_unlock(&file_priv->mm.lock);
5093 i915_gem_file_idle_work_handler(struct work_struct *work)
5095 struct drm_i915_file_private *file_priv =
5096 container_of(work, typeof(*file_priv), mm.idle_work.work);
5098 atomic_set(&file_priv->rps_wait_boost, false);
5101 int i915_gem_open(struct drm_device *dev, struct drm_file *file)
5103 struct drm_i915_file_private *file_priv;
5106 DRM_DEBUG_DRIVER("\n");
5108 file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL);
5112 file->driver_priv = file_priv;
5113 file_priv->dev_priv = dev->dev_private;
5114 file_priv->file = file;
5116 spin_lock_init(&file_priv->mm.lock);
5117 INIT_LIST_HEAD(&file_priv->mm.request_list);
5118 INIT_DELAYED_WORK(&file_priv->mm.idle_work,
5119 i915_gem_file_idle_work_handler);
5121 ret = i915_gem_context_open(dev, file);
5129 * i915_gem_track_fb - update frontbuffer tracking
5130 * old: current GEM buffer for the frontbuffer slots
5131 * new: new GEM buffer for the frontbuffer slots
5132 * frontbuffer_bits: bitmask of frontbuffer slots
5134 * This updates the frontbuffer tracking bits @frontbuffer_bits by clearing them
5135 * from @old and setting them in @new. Both @old and @new can be NULL.
5137 void i915_gem_track_fb(struct drm_i915_gem_object *old,
5138 struct drm_i915_gem_object *new,
5139 unsigned frontbuffer_bits)
5142 WARN_ON(!mutex_is_locked(&old->base.dev->struct_mutex));
5143 WARN_ON(!(old->frontbuffer_bits & frontbuffer_bits));
5144 old->frontbuffer_bits &= ~frontbuffer_bits;
5148 WARN_ON(!mutex_is_locked(&new->base.dev->struct_mutex));
5149 WARN_ON(new->frontbuffer_bits & frontbuffer_bits);
5150 new->frontbuffer_bits |= frontbuffer_bits;
5154 static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task)
5156 if (!mutex_is_locked(mutex))
5159 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES)
5160 return mutex->owner == task;
5162 /* Since UP may be pre-empted, we cannot assume that we own the lock */
5167 static bool i915_gem_shrinker_lock(struct drm_device *dev, bool *unlock)
5169 if (!mutex_trylock(&dev->struct_mutex)) {
5170 if (!mutex_is_locked_by(&dev->struct_mutex, current))
5173 if (to_i915(dev)->mm.shrinker_no_lock_stealing)
5183 static int num_vma_bound(struct drm_i915_gem_object *obj)
5185 struct i915_vma *vma;
5188 list_for_each_entry(vma, &obj->vma_list, vma_link)
5189 if (drm_mm_node_allocated(&vma->node))
5195 static unsigned long
5196 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
5198 struct drm_i915_private *dev_priv =
5199 container_of(shrinker, struct drm_i915_private, mm.shrinker);
5200 struct drm_device *dev = dev_priv->dev;
5201 struct drm_i915_gem_object *obj;
5202 unsigned long count;
5205 if (!i915_gem_shrinker_lock(dev, &unlock))
5209 list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list)
5210 if (obj->pages_pin_count == 0)
5211 count += obj->base.size >> PAGE_SHIFT;
5213 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
5214 if (!i915_gem_obj_is_pinned(obj) &&
5215 obj->pages_pin_count == num_vma_bound(obj))
5216 count += obj->base.size >> PAGE_SHIFT;
5220 mutex_unlock(&dev->struct_mutex);
5225 /* All the new VM stuff */
5226 unsigned long i915_gem_obj_offset(struct drm_i915_gem_object *o,
5227 struct i915_address_space *vm)
5229 struct drm_i915_private *dev_priv = o->base.dev->dev_private;
5230 struct i915_vma *vma;
5232 WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base);
5234 list_for_each_entry(vma, &o->vma_list, vma_link) {
5236 return vma->node.start;
5239 WARN(1, "%s vma for this object not found.\n",
5240 i915_is_ggtt(vm) ? "global" : "ppgtt");
5244 bool i915_gem_obj_bound(struct drm_i915_gem_object *o,
5245 struct i915_address_space *vm)
5247 struct i915_vma *vma;
5249 list_for_each_entry(vma, &o->vma_list, vma_link)
5250 if (vma->vm == vm && drm_mm_node_allocated(&vma->node))
5256 bool i915_gem_obj_bound_any(struct drm_i915_gem_object *o)
5258 struct i915_vma *vma;
5260 list_for_each_entry(vma, &o->vma_list, vma_link)
5261 if (drm_mm_node_allocated(&vma->node))
5267 unsigned long i915_gem_obj_size(struct drm_i915_gem_object *o,
5268 struct i915_address_space *vm)
5270 struct drm_i915_private *dev_priv = o->base.dev->dev_private;
5271 struct i915_vma *vma;
5273 WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base);
5275 BUG_ON(list_empty(&o->vma_list));
5277 list_for_each_entry(vma, &o->vma_list, vma_link)
5279 return vma->node.size;
5284 static unsigned long
5285 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
5287 struct drm_i915_private *dev_priv =
5288 container_of(shrinker, struct drm_i915_private, mm.shrinker);
5289 struct drm_device *dev = dev_priv->dev;
5290 unsigned long freed;
5293 if (!i915_gem_shrinker_lock(dev, &unlock))
5296 freed = i915_gem_shrink(dev_priv,
5299 I915_SHRINK_UNBOUND |
5300 I915_SHRINK_PURGEABLE);
5301 if (freed < sc->nr_to_scan)
5302 freed += i915_gem_shrink(dev_priv,
5303 sc->nr_to_scan - freed,
5305 I915_SHRINK_UNBOUND);
5307 mutex_unlock(&dev->struct_mutex);
5313 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
5315 struct drm_i915_private *dev_priv =
5316 container_of(nb, struct drm_i915_private, mm.oom_notifier);
5317 struct drm_device *dev = dev_priv->dev;
5318 struct drm_i915_gem_object *obj;
5319 unsigned long timeout = msecs_to_jiffies(5000) + 1;
5320 unsigned long pinned, bound, unbound, freed_pages;
5321 bool was_interruptible;
5324 while (!i915_gem_shrinker_lock(dev, &unlock) && --timeout) {
5325 schedule_timeout_killable(1);
5326 if (fatal_signal_pending(current))
5330 pr_err("Unable to purge GPU memory due lock contention.\n");
5334 was_interruptible = dev_priv->mm.interruptible;
5335 dev_priv->mm.interruptible = false;
5337 freed_pages = i915_gem_shrink_all(dev_priv);
5339 dev_priv->mm.interruptible = was_interruptible;
5341 /* Because we may be allocating inside our own driver, we cannot
5342 * assert that there are no objects with pinned pages that are not
5343 * being pointed to by hardware.
5345 unbound = bound = pinned = 0;
5346 list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) {
5347 if (!obj->base.filp) /* not backed by a freeable object */
5350 if (obj->pages_pin_count)
5351 pinned += obj->base.size;
5353 unbound += obj->base.size;
5355 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
5356 if (!obj->base.filp)
5359 if (obj->pages_pin_count)
5360 pinned += obj->base.size;
5362 bound += obj->base.size;
5366 mutex_unlock(&dev->struct_mutex);
5368 if (freed_pages || unbound || bound)
5369 pr_info("Purging GPU memory, %lu bytes freed, %lu bytes still pinned.\n",
5370 freed_pages << PAGE_SHIFT, pinned);
5371 if (unbound || bound)
5372 pr_err("%lu and %lu bytes still available in the "
5373 "bound and unbound GPU page lists.\n",
5376 *(unsigned long *)ptr += freed_pages;
5380 struct i915_vma *i915_gem_obj_to_ggtt(struct drm_i915_gem_object *obj)
5382 struct i915_vma *vma;
5384 vma = list_first_entry(&obj->vma_list, typeof(*vma), vma_link);
5385 if (vma->vm != i915_obj_to_ggtt(obj))