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_private *dev_priv = dev->dev_private;
1054 struct drm_i915_gem_pwrite *args = data;
1055 struct drm_i915_gem_object *obj;
1058 if (args->size == 0)
1061 if (!access_ok(VERIFY_READ,
1062 to_user_ptr(args->data_ptr),
1066 if (likely(!i915.prefault_disable)) {
1067 ret = fault_in_multipages_readable(to_user_ptr(args->data_ptr),
1073 intel_runtime_pm_get(dev_priv);
1075 ret = i915_mutex_lock_interruptible(dev);
1079 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1080 if (&obj->base == NULL) {
1085 /* Bounds check destination. */
1086 if (args->offset > obj->base.size ||
1087 args->size > obj->base.size - args->offset) {
1092 /* prime objects have no backing filp to GEM pread/pwrite
1095 if (!obj->base.filp) {
1100 trace_i915_gem_object_pwrite(obj, args->offset, args->size);
1103 /* We can only do the GTT pwrite on untiled buffers, as otherwise
1104 * it would end up going through the fenced access, and we'll get
1105 * different detiling behavior between reading and writing.
1106 * pread/pwrite currently are reading and writing from the CPU
1107 * perspective, requiring manual detiling by the client.
1109 if (obj->tiling_mode == I915_TILING_NONE &&
1110 obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
1111 cpu_write_needs_clflush(obj)) {
1112 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
1113 /* Note that the gtt paths might fail with non-page-backed user
1114 * pointers (e.g. gtt mappings when moving data between
1115 * textures). Fallback to the shmem path in that case. */
1118 if (ret == -EFAULT || ret == -ENOSPC) {
1119 if (obj->phys_handle)
1120 ret = i915_gem_phys_pwrite(obj, args, file);
1122 ret = i915_gem_shmem_pwrite(dev, obj, args, file);
1126 drm_gem_object_unreference(&obj->base);
1128 mutex_unlock(&dev->struct_mutex);
1130 intel_runtime_pm_put(dev_priv);
1136 i915_gem_check_wedge(struct i915_gpu_error *error,
1139 if (i915_reset_in_progress(error)) {
1140 /* Non-interruptible callers can't handle -EAGAIN, hence return
1141 * -EIO unconditionally for these. */
1145 /* Recovery complete, but the reset failed ... */
1146 if (i915_terminally_wedged(error))
1150 * Check if GPU Reset is in progress - we need intel_ring_begin
1151 * to work properly to reinit the hw state while the gpu is
1152 * still marked as reset-in-progress. Handle this with a flag.
1154 if (!error->reload_in_reset)
1162 * Compare seqno against outstanding lazy request. Emit a request if they are
1166 i915_gem_check_olr(struct intel_engine_cs *ring, u32 seqno)
1170 BUG_ON(!mutex_is_locked(&ring->dev->struct_mutex));
1173 if (seqno == ring->outstanding_lazy_seqno)
1174 ret = i915_add_request(ring, NULL);
1179 static void fake_irq(unsigned long data)
1181 wake_up_process((struct task_struct *)data);
1184 static bool missed_irq(struct drm_i915_private *dev_priv,
1185 struct intel_engine_cs *ring)
1187 return test_bit(ring->id, &dev_priv->gpu_error.missed_irq_rings);
1190 static bool can_wait_boost(struct drm_i915_file_private *file_priv)
1192 if (file_priv == NULL)
1195 return !atomic_xchg(&file_priv->rps_wait_boost, true);
1199 * __i915_wait_seqno - wait until execution of seqno has finished
1200 * @ring: the ring expected to report seqno
1202 * @reset_counter: reset sequence associated with the given seqno
1203 * @interruptible: do an interruptible wait (normally yes)
1204 * @timeout: in - how long to wait (NULL forever); out - how much time remaining
1206 * Note: It is of utmost importance that the passed in seqno and reset_counter
1207 * values have been read by the caller in an smp safe manner. Where read-side
1208 * locks are involved, it is sufficient to read the reset_counter before
1209 * unlocking the lock that protects the seqno. For lockless tricks, the
1210 * reset_counter _must_ be read before, and an appropriate smp_rmb must be
1213 * Returns 0 if the seqno was found within the alloted time. Else returns the
1214 * errno with remaining time filled in timeout argument.
1216 int __i915_wait_seqno(struct intel_engine_cs *ring, u32 seqno,
1217 unsigned reset_counter,
1220 struct drm_i915_file_private *file_priv)
1222 struct drm_device *dev = ring->dev;
1223 struct drm_i915_private *dev_priv = dev->dev_private;
1224 const bool irq_test_in_progress =
1225 ACCESS_ONCE(dev_priv->gpu_error.test_irq_rings) & intel_ring_flag(ring);
1227 unsigned long timeout_expire;
1231 WARN(!intel_irqs_enabled(dev_priv), "IRQs disabled");
1233 if (i915_seqno_passed(ring->get_seqno(ring, true), seqno))
1236 timeout_expire = timeout ?
1237 jiffies + nsecs_to_jiffies_timeout((u64)*timeout) : 0;
1239 if (INTEL_INFO(dev)->gen >= 6 && ring->id == RCS && can_wait_boost(file_priv)) {
1240 gen6_rps_boost(dev_priv);
1242 mod_delayed_work(dev_priv->wq,
1243 &file_priv->mm.idle_work,
1244 msecs_to_jiffies(100));
1247 if (!irq_test_in_progress && WARN_ON(!ring->irq_get(ring)))
1250 /* Record current time in case interrupted by signal, or wedged */
1251 trace_i915_gem_request_wait_begin(ring, seqno);
1252 before = ktime_get_raw_ns();
1254 struct timer_list timer;
1256 prepare_to_wait(&ring->irq_queue, &wait,
1257 interruptible ? TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE);
1259 /* We need to check whether any gpu reset happened in between
1260 * the caller grabbing the seqno and now ... */
1261 if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) {
1262 /* ... but upgrade the -EAGAIN to an -EIO if the gpu
1263 * is truely gone. */
1264 ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1270 if (i915_seqno_passed(ring->get_seqno(ring, false), seqno)) {
1275 if (interruptible && signal_pending(current)) {
1280 if (timeout && time_after_eq(jiffies, timeout_expire)) {
1285 timer.function = NULL;
1286 if (timeout || missed_irq(dev_priv, ring)) {
1287 unsigned long expire;
1289 setup_timer_on_stack(&timer, fake_irq, (unsigned long)current);
1290 expire = missed_irq(dev_priv, ring) ? jiffies + 1 : timeout_expire;
1291 mod_timer(&timer, expire);
1296 if (timer.function) {
1297 del_singleshot_timer_sync(&timer);
1298 destroy_timer_on_stack(&timer);
1301 now = ktime_get_raw_ns();
1302 trace_i915_gem_request_wait_end(ring, seqno);
1304 if (!irq_test_in_progress)
1305 ring->irq_put(ring);
1307 finish_wait(&ring->irq_queue, &wait);
1310 s64 tres = *timeout - (now - before);
1312 *timeout = tres < 0 ? 0 : tres;
1315 * Apparently ktime isn't accurate enough and occasionally has a
1316 * bit of mismatch in the jiffies<->nsecs<->ktime loop. So patch
1317 * things up to make the test happy. We allow up to 1 jiffy.
1319 * This is a regrssion from the timespec->ktime conversion.
1321 if (ret == -ETIME && *timeout < jiffies_to_usecs(1)*1000)
1329 * Waits for a sequence number to be signaled, and cleans up the
1330 * request and object lists appropriately for that event.
1333 i915_wait_seqno(struct intel_engine_cs *ring, uint32_t seqno)
1335 struct drm_device *dev = ring->dev;
1336 struct drm_i915_private *dev_priv = dev->dev_private;
1337 bool interruptible = dev_priv->mm.interruptible;
1338 unsigned reset_counter;
1341 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1344 ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1348 ret = i915_gem_check_olr(ring, seqno);
1352 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
1353 return __i915_wait_seqno(ring, seqno, reset_counter, interruptible,
1358 i915_gem_object_wait_rendering__tail(struct drm_i915_gem_object *obj)
1363 /* Manually manage the write flush as we may have not yet
1364 * retired the buffer.
1366 * Note that the last_write_seqno is always the earlier of
1367 * the two (read/write) seqno, so if we haved successfully waited,
1368 * we know we have passed the last write.
1370 obj->last_write_seqno = 0;
1376 * Ensures that all rendering to the object has completed and the object is
1377 * safe to unbind from the GTT or access from the CPU.
1379 static __must_check int
1380 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
1383 struct intel_engine_cs *ring = obj->ring;
1387 seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno;
1391 ret = i915_wait_seqno(ring, seqno);
1395 return i915_gem_object_wait_rendering__tail(obj);
1398 /* A nonblocking variant of the above wait. This is a highly dangerous routine
1399 * as the object state may change during this call.
1401 static __must_check int
1402 i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
1403 struct drm_i915_file_private *file_priv,
1406 struct drm_device *dev = obj->base.dev;
1407 struct drm_i915_private *dev_priv = dev->dev_private;
1408 struct intel_engine_cs *ring = obj->ring;
1409 unsigned reset_counter;
1413 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1414 BUG_ON(!dev_priv->mm.interruptible);
1416 seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno;
1420 ret = i915_gem_check_wedge(&dev_priv->gpu_error, true);
1424 ret = i915_gem_check_olr(ring, seqno);
1428 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
1429 mutex_unlock(&dev->struct_mutex);
1430 ret = __i915_wait_seqno(ring, seqno, reset_counter, true, NULL,
1432 mutex_lock(&dev->struct_mutex);
1436 return i915_gem_object_wait_rendering__tail(obj);
1440 * Called when user space prepares to use an object with the CPU, either
1441 * through the mmap ioctl's mapping or a GTT mapping.
1444 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1445 struct drm_file *file)
1447 struct drm_i915_gem_set_domain *args = data;
1448 struct drm_i915_gem_object *obj;
1449 uint32_t read_domains = args->read_domains;
1450 uint32_t write_domain = args->write_domain;
1453 /* Only handle setting domains to types used by the CPU. */
1454 if (write_domain & I915_GEM_GPU_DOMAINS)
1457 if (read_domains & I915_GEM_GPU_DOMAINS)
1460 /* Having something in the write domain implies it's in the read
1461 * domain, and only that read domain. Enforce that in the request.
1463 if (write_domain != 0 && read_domains != write_domain)
1466 ret = i915_mutex_lock_interruptible(dev);
1470 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1471 if (&obj->base == NULL) {
1476 /* Try to flush the object off the GPU without holding the lock.
1477 * We will repeat the flush holding the lock in the normal manner
1478 * to catch cases where we are gazumped.
1480 ret = i915_gem_object_wait_rendering__nonblocking(obj,
1486 if (read_domains & I915_GEM_DOMAIN_GTT) {
1487 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1489 /* Silently promote "you're not bound, there was nothing to do"
1490 * to success, since the client was just asking us to
1491 * make sure everything was done.
1496 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1500 drm_gem_object_unreference(&obj->base);
1502 mutex_unlock(&dev->struct_mutex);
1507 * Called when user space has done writes to this buffer
1510 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1511 struct drm_file *file)
1513 struct drm_i915_gem_sw_finish *args = data;
1514 struct drm_i915_gem_object *obj;
1517 ret = i915_mutex_lock_interruptible(dev);
1521 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1522 if (&obj->base == NULL) {
1527 /* Pinned buffers may be scanout, so flush the cache */
1528 if (obj->pin_display)
1529 i915_gem_object_flush_cpu_write_domain(obj, true);
1531 drm_gem_object_unreference(&obj->base);
1533 mutex_unlock(&dev->struct_mutex);
1538 * Maps the contents of an object, returning the address it is mapped
1541 * While the mapping holds a reference on the contents of the object, it doesn't
1542 * imply a ref on the object itself.
1546 * DRM driver writers who look a this function as an example for how to do GEM
1547 * mmap support, please don't implement mmap support like here. The modern way
1548 * to implement DRM mmap support is with an mmap offset ioctl (like
1549 * i915_gem_mmap_gtt) and then using the mmap syscall on the DRM fd directly.
1550 * That way debug tooling like valgrind will understand what's going on, hiding
1551 * the mmap call in a driver private ioctl will break that. The i915 driver only
1552 * does cpu mmaps this way because we didn't know better.
1555 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1556 struct drm_file *file)
1558 struct drm_i915_gem_mmap *args = data;
1559 struct drm_gem_object *obj;
1562 obj = drm_gem_object_lookup(dev, file, args->handle);
1566 /* prime objects have no backing filp to GEM mmap
1570 drm_gem_object_unreference_unlocked(obj);
1574 addr = vm_mmap(obj->filp, 0, args->size,
1575 PROT_READ | PROT_WRITE, MAP_SHARED,
1577 drm_gem_object_unreference_unlocked(obj);
1578 if (IS_ERR((void *)addr))
1581 args->addr_ptr = (uint64_t) addr;
1587 * i915_gem_fault - fault a page into the GTT
1588 * vma: VMA in question
1591 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1592 * from userspace. The fault handler takes care of binding the object to
1593 * the GTT (if needed), allocating and programming a fence register (again,
1594 * only if needed based on whether the old reg is still valid or the object
1595 * is tiled) and inserting a new PTE into the faulting process.
1597 * Note that the faulting process may involve evicting existing objects
1598 * from the GTT and/or fence registers to make room. So performance may
1599 * suffer if the GTT working set is large or there are few fence registers
1602 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1604 struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1605 struct drm_device *dev = obj->base.dev;
1606 struct drm_i915_private *dev_priv = dev->dev_private;
1607 pgoff_t page_offset;
1610 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1612 intel_runtime_pm_get(dev_priv);
1614 /* We don't use vmf->pgoff since that has the fake offset */
1615 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1618 ret = i915_mutex_lock_interruptible(dev);
1622 trace_i915_gem_object_fault(obj, page_offset, true, write);
1624 /* Try to flush the object off the GPU first without holding the lock.
1625 * Upon reacquiring the lock, we will perform our sanity checks and then
1626 * repeat the flush holding the lock in the normal manner to catch cases
1627 * where we are gazumped.
1629 ret = i915_gem_object_wait_rendering__nonblocking(obj, NULL, !write);
1633 /* Access to snoopable pages through the GTT is incoherent. */
1634 if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) {
1639 /* Now bind it into the GTT if needed */
1640 ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE);
1644 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1648 ret = i915_gem_object_get_fence(obj);
1652 /* Finally, remap it using the new GTT offset */
1653 pfn = dev_priv->gtt.mappable_base + i915_gem_obj_ggtt_offset(obj);
1656 if (!obj->fault_mappable) {
1657 unsigned long size = min_t(unsigned long,
1658 vma->vm_end - vma->vm_start,
1662 for (i = 0; i < size >> PAGE_SHIFT; i++) {
1663 ret = vm_insert_pfn(vma,
1664 (unsigned long)vma->vm_start + i * PAGE_SIZE,
1670 obj->fault_mappable = true;
1672 ret = vm_insert_pfn(vma,
1673 (unsigned long)vmf->virtual_address,
1676 i915_gem_object_ggtt_unpin(obj);
1678 mutex_unlock(&dev->struct_mutex);
1683 * We eat errors when the gpu is terminally wedged to avoid
1684 * userspace unduly crashing (gl has no provisions for mmaps to
1685 * fail). But any other -EIO isn't ours (e.g. swap in failure)
1686 * and so needs to be reported.
1688 if (!i915_terminally_wedged(&dev_priv->gpu_error)) {
1689 ret = VM_FAULT_SIGBUS;
1694 * EAGAIN means the gpu is hung and we'll wait for the error
1695 * handler to reset everything when re-faulting in
1696 * i915_mutex_lock_interruptible.
1703 * EBUSY is ok: this just means that another thread
1704 * already did the job.
1706 ret = VM_FAULT_NOPAGE;
1713 ret = VM_FAULT_SIGBUS;
1716 WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret);
1717 ret = VM_FAULT_SIGBUS;
1721 intel_runtime_pm_put(dev_priv);
1726 * i915_gem_release_mmap - remove physical page mappings
1727 * @obj: obj in question
1729 * Preserve the reservation of the mmapping with the DRM core code, but
1730 * relinquish ownership of the pages back to the system.
1732 * It is vital that we remove the page mapping if we have mapped a tiled
1733 * object through the GTT and then lose the fence register due to
1734 * resource pressure. Similarly if the object has been moved out of the
1735 * aperture, than pages mapped into userspace must be revoked. Removing the
1736 * mapping will then trigger a page fault on the next user access, allowing
1737 * fixup by i915_gem_fault().
1740 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1742 if (!obj->fault_mappable)
1745 drm_vma_node_unmap(&obj->base.vma_node,
1746 obj->base.dev->anon_inode->i_mapping);
1747 obj->fault_mappable = false;
1751 i915_gem_release_all_mmaps(struct drm_i915_private *dev_priv)
1753 struct drm_i915_gem_object *obj;
1755 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list)
1756 i915_gem_release_mmap(obj);
1760 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1764 if (INTEL_INFO(dev)->gen >= 4 ||
1765 tiling_mode == I915_TILING_NONE)
1768 /* Previous chips need a power-of-two fence region when tiling */
1769 if (INTEL_INFO(dev)->gen == 3)
1770 gtt_size = 1024*1024;
1772 gtt_size = 512*1024;
1774 while (gtt_size < size)
1781 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1782 * @obj: object to check
1784 * Return the required GTT alignment for an object, taking into account
1785 * potential fence register mapping.
1788 i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size,
1789 int tiling_mode, bool fenced)
1792 * Minimum alignment is 4k (GTT page size), but might be greater
1793 * if a fence register is needed for the object.
1795 if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) ||
1796 tiling_mode == I915_TILING_NONE)
1800 * Previous chips need to be aligned to the size of the smallest
1801 * fence register that can contain the object.
1803 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1806 static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj)
1808 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1811 if (drm_vma_node_has_offset(&obj->base.vma_node))
1814 dev_priv->mm.shrinker_no_lock_stealing = true;
1816 ret = drm_gem_create_mmap_offset(&obj->base);
1820 /* Badly fragmented mmap space? The only way we can recover
1821 * space is by destroying unwanted objects. We can't randomly release
1822 * mmap_offsets as userspace expects them to be persistent for the
1823 * lifetime of the objects. The closest we can is to release the
1824 * offsets on purgeable objects by truncating it and marking it purged,
1825 * which prevents userspace from ever using that object again.
1827 i915_gem_shrink(dev_priv,
1828 obj->base.size >> PAGE_SHIFT,
1830 I915_SHRINK_UNBOUND |
1831 I915_SHRINK_PURGEABLE);
1832 ret = drm_gem_create_mmap_offset(&obj->base);
1836 i915_gem_shrink_all(dev_priv);
1837 ret = drm_gem_create_mmap_offset(&obj->base);
1839 dev_priv->mm.shrinker_no_lock_stealing = false;
1844 static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)
1846 drm_gem_free_mmap_offset(&obj->base);
1850 i915_gem_mmap_gtt(struct drm_file *file,
1851 struct drm_device *dev,
1852 uint32_t handle, bool dumb,
1855 struct drm_i915_private *dev_priv = dev->dev_private;
1856 struct drm_i915_gem_object *obj;
1859 ret = i915_mutex_lock_interruptible(dev);
1863 obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
1864 if (&obj->base == NULL) {
1870 * We don't allow dumb mmaps on objects created using another
1873 WARN_ONCE(dumb && !(obj->base.dumb || obj->base.import_attach),
1874 "Illegal dumb map of accelerated buffer.\n");
1876 if (obj->base.size > dev_priv->gtt.mappable_end) {
1881 if (obj->madv != I915_MADV_WILLNEED) {
1882 DRM_DEBUG("Attempting to mmap a purgeable buffer\n");
1887 ret = i915_gem_object_create_mmap_offset(obj);
1891 *offset = drm_vma_node_offset_addr(&obj->base.vma_node);
1894 drm_gem_object_unreference(&obj->base);
1896 mutex_unlock(&dev->struct_mutex);
1901 i915_gem_dumb_map_offset(struct drm_file *file,
1902 struct drm_device *dev,
1906 return i915_gem_mmap_gtt(file, dev, handle, true, offset);
1910 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1912 * @data: GTT mapping ioctl data
1913 * @file: GEM object info
1915 * Simply returns the fake offset to userspace so it can mmap it.
1916 * The mmap call will end up in drm_gem_mmap(), which will set things
1917 * up so we can get faults in the handler above.
1919 * The fault handler will take care of binding the object into the GTT
1920 * (since it may have been evicted to make room for something), allocating
1921 * a fence register, and mapping the appropriate aperture address into
1925 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1926 struct drm_file *file)
1928 struct drm_i915_gem_mmap_gtt *args = data;
1930 return i915_gem_mmap_gtt(file, dev, args->handle, false, &args->offset);
1934 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj)
1936 return obj->madv == I915_MADV_DONTNEED;
1939 /* Immediately discard the backing storage */
1941 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
1943 i915_gem_object_free_mmap_offset(obj);
1945 if (obj->base.filp == NULL)
1948 /* Our goal here is to return as much of the memory as
1949 * is possible back to the system as we are called from OOM.
1950 * To do this we must instruct the shmfs to drop all of its
1951 * backing pages, *now*.
1953 shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1);
1954 obj->madv = __I915_MADV_PURGED;
1957 /* Try to discard unwanted pages */
1959 i915_gem_object_invalidate(struct drm_i915_gem_object *obj)
1961 struct address_space *mapping;
1963 switch (obj->madv) {
1964 case I915_MADV_DONTNEED:
1965 i915_gem_object_truncate(obj);
1966 case __I915_MADV_PURGED:
1970 if (obj->base.filp == NULL)
1973 mapping = file_inode(obj->base.filp)->i_mapping,
1974 invalidate_mapping_pages(mapping, 0, (loff_t)-1);
1978 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
1980 struct sg_page_iter sg_iter;
1983 BUG_ON(obj->madv == __I915_MADV_PURGED);
1985 ret = i915_gem_object_set_to_cpu_domain(obj, true);
1987 /* In the event of a disaster, abandon all caches and
1988 * hope for the best.
1990 WARN_ON(ret != -EIO);
1991 i915_gem_clflush_object(obj, true);
1992 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
1995 if (i915_gem_object_needs_bit17_swizzle(obj))
1996 i915_gem_object_save_bit_17_swizzle(obj);
1998 if (obj->madv == I915_MADV_DONTNEED)
2001 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
2002 struct page *page = sg_page_iter_page(&sg_iter);
2005 set_page_dirty(page);
2007 if (obj->madv == I915_MADV_WILLNEED)
2008 mark_page_accessed(page);
2010 page_cache_release(page);
2014 sg_free_table(obj->pages);
2019 i915_gem_object_put_pages(struct drm_i915_gem_object *obj)
2021 const struct drm_i915_gem_object_ops *ops = obj->ops;
2023 if (obj->pages == NULL)
2026 if (obj->pages_pin_count)
2029 BUG_ON(i915_gem_obj_bound_any(obj));
2031 /* ->put_pages might need to allocate memory for the bit17 swizzle
2032 * array, hence protect them from being reaped by removing them from gtt
2034 list_del(&obj->global_list);
2036 ops->put_pages(obj);
2039 i915_gem_object_invalidate(obj);
2045 i915_gem_shrink(struct drm_i915_private *dev_priv,
2046 long target, unsigned flags)
2049 struct list_head *list;
2052 { &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND },
2053 { &dev_priv->mm.bound_list, I915_SHRINK_BOUND },
2056 unsigned long count = 0;
2059 * As we may completely rewrite the (un)bound list whilst unbinding
2060 * (due to retiring requests) we have to strictly process only
2061 * one element of the list at the time, and recheck the list
2062 * on every iteration.
2064 * In particular, we must hold a reference whilst removing the
2065 * object as we may end up waiting for and/or retiring the objects.
2066 * This might release the final reference (held by the active list)
2067 * and result in the object being freed from under us. This is
2068 * similar to the precautions the eviction code must take whilst
2071 * Also note that although these lists do not hold a reference to
2072 * the object we can safely grab one here: The final object
2073 * unreferencing and the bound_list are both protected by the
2074 * dev->struct_mutex and so we won't ever be able to observe an
2075 * object on the bound_list with a reference count equals 0.
2077 for (phase = phases; phase->list; phase++) {
2078 struct list_head still_in_list;
2080 if ((flags & phase->bit) == 0)
2083 INIT_LIST_HEAD(&still_in_list);
2084 while (count < target && !list_empty(phase->list)) {
2085 struct drm_i915_gem_object *obj;
2086 struct i915_vma *vma, *v;
2088 obj = list_first_entry(phase->list,
2089 typeof(*obj), global_list);
2090 list_move_tail(&obj->global_list, &still_in_list);
2092 if (flags & I915_SHRINK_PURGEABLE &&
2093 !i915_gem_object_is_purgeable(obj))
2096 drm_gem_object_reference(&obj->base);
2098 /* For the unbound phase, this should be a no-op! */
2099 list_for_each_entry_safe(vma, v,
2100 &obj->vma_list, vma_link)
2101 if (i915_vma_unbind(vma))
2104 if (i915_gem_object_put_pages(obj) == 0)
2105 count += obj->base.size >> PAGE_SHIFT;
2107 drm_gem_object_unreference(&obj->base);
2109 list_splice(&still_in_list, phase->list);
2115 static unsigned long
2116 i915_gem_shrink_all(struct drm_i915_private *dev_priv)
2118 i915_gem_evict_everything(dev_priv->dev);
2119 return i915_gem_shrink(dev_priv, LONG_MAX,
2120 I915_SHRINK_BOUND | I915_SHRINK_UNBOUND);
2124 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
2126 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2128 struct address_space *mapping;
2129 struct sg_table *st;
2130 struct scatterlist *sg;
2131 struct sg_page_iter sg_iter;
2133 unsigned long last_pfn = 0; /* suppress gcc warning */
2136 /* Assert that the object is not currently in any GPU domain. As it
2137 * wasn't in the GTT, there shouldn't be any way it could have been in
2140 BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
2141 BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
2143 st = kmalloc(sizeof(*st), GFP_KERNEL);
2147 page_count = obj->base.size / PAGE_SIZE;
2148 if (sg_alloc_table(st, page_count, GFP_KERNEL)) {
2153 /* Get the list of pages out of our struct file. They'll be pinned
2154 * at this point until we release them.
2156 * Fail silently without starting the shrinker
2158 mapping = file_inode(obj->base.filp)->i_mapping;
2159 gfp = mapping_gfp_mask(mapping);
2160 gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD;
2161 gfp &= ~(__GFP_IO | __GFP_WAIT);
2164 for (i = 0; i < page_count; i++) {
2165 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2167 i915_gem_shrink(dev_priv,
2170 I915_SHRINK_UNBOUND |
2171 I915_SHRINK_PURGEABLE);
2172 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2175 /* We've tried hard to allocate the memory by reaping
2176 * our own buffer, now let the real VM do its job and
2177 * go down in flames if truly OOM.
2179 i915_gem_shrink_all(dev_priv);
2180 page = shmem_read_mapping_page(mapping, i);
2184 #ifdef CONFIG_SWIOTLB
2185 if (swiotlb_nr_tbl()) {
2187 sg_set_page(sg, page, PAGE_SIZE, 0);
2192 if (!i || page_to_pfn(page) != last_pfn + 1) {
2196 sg_set_page(sg, page, PAGE_SIZE, 0);
2198 sg->length += PAGE_SIZE;
2200 last_pfn = page_to_pfn(page);
2202 /* Check that the i965g/gm workaround works. */
2203 WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
2205 #ifdef CONFIG_SWIOTLB
2206 if (!swiotlb_nr_tbl())
2211 if (i915_gem_object_needs_bit17_swizzle(obj))
2212 i915_gem_object_do_bit_17_swizzle(obj);
2214 if (obj->tiling_mode != I915_TILING_NONE &&
2215 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES)
2216 i915_gem_object_pin_pages(obj);
2222 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0)
2223 page_cache_release(sg_page_iter_page(&sg_iter));
2227 /* shmemfs first checks if there is enough memory to allocate the page
2228 * and reports ENOSPC should there be insufficient, along with the usual
2229 * ENOMEM for a genuine allocation failure.
2231 * We use ENOSPC in our driver to mean that we have run out of aperture
2232 * space and so want to translate the error from shmemfs back to our
2233 * usual understanding of ENOMEM.
2235 if (PTR_ERR(page) == -ENOSPC)
2238 return PTR_ERR(page);
2241 /* Ensure that the associated pages are gathered from the backing storage
2242 * and pinned into our object. i915_gem_object_get_pages() may be called
2243 * multiple times before they are released by a single call to
2244 * i915_gem_object_put_pages() - once the pages are no longer referenced
2245 * either as a result of memory pressure (reaping pages under the shrinker)
2246 * or as the object is itself released.
2249 i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
2251 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2252 const struct drm_i915_gem_object_ops *ops = obj->ops;
2258 if (obj->madv != I915_MADV_WILLNEED) {
2259 DRM_DEBUG("Attempting to obtain a purgeable object\n");
2263 BUG_ON(obj->pages_pin_count);
2265 ret = ops->get_pages(obj);
2269 list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list);
2274 i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
2275 struct intel_engine_cs *ring)
2277 u32 seqno = intel_ring_get_seqno(ring);
2279 BUG_ON(ring == NULL);
2280 if (obj->ring != ring && obj->last_write_seqno) {
2281 /* Keep the seqno relative to the current ring */
2282 obj->last_write_seqno = seqno;
2286 /* Add a reference if we're newly entering the active list. */
2288 drm_gem_object_reference(&obj->base);
2292 list_move_tail(&obj->ring_list, &ring->active_list);
2294 obj->last_read_seqno = seqno;
2297 void i915_vma_move_to_active(struct i915_vma *vma,
2298 struct intel_engine_cs *ring)
2300 list_move_tail(&vma->mm_list, &vma->vm->active_list);
2301 return i915_gem_object_move_to_active(vma->obj, ring);
2305 i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
2307 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2308 struct i915_address_space *vm;
2309 struct i915_vma *vma;
2311 BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS);
2312 BUG_ON(!obj->active);
2314 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
2315 vma = i915_gem_obj_to_vma(obj, vm);
2316 if (vma && !list_empty(&vma->mm_list))
2317 list_move_tail(&vma->mm_list, &vm->inactive_list);
2320 intel_fb_obj_flush(obj, true);
2322 list_del_init(&obj->ring_list);
2325 obj->last_read_seqno = 0;
2326 obj->last_write_seqno = 0;
2327 obj->base.write_domain = 0;
2329 obj->last_fenced_seqno = 0;
2332 drm_gem_object_unreference(&obj->base);
2334 WARN_ON(i915_verify_lists(dev));
2338 i915_gem_object_retire(struct drm_i915_gem_object *obj)
2340 struct intel_engine_cs *ring = obj->ring;
2345 if (i915_seqno_passed(ring->get_seqno(ring, true),
2346 obj->last_read_seqno))
2347 i915_gem_object_move_to_inactive(obj);
2351 i915_gem_init_seqno(struct drm_device *dev, u32 seqno)
2353 struct drm_i915_private *dev_priv = dev->dev_private;
2354 struct intel_engine_cs *ring;
2357 /* Carefully retire all requests without writing to the rings */
2358 for_each_ring(ring, dev_priv, i) {
2359 ret = intel_ring_idle(ring);
2363 i915_gem_retire_requests(dev);
2365 /* Finally reset hw state */
2366 for_each_ring(ring, dev_priv, i) {
2367 intel_ring_init_seqno(ring, seqno);
2369 for (j = 0; j < ARRAY_SIZE(ring->semaphore.sync_seqno); j++)
2370 ring->semaphore.sync_seqno[j] = 0;
2376 int i915_gem_set_seqno(struct drm_device *dev, u32 seqno)
2378 struct drm_i915_private *dev_priv = dev->dev_private;
2384 /* HWS page needs to be set less than what we
2385 * will inject to ring
2387 ret = i915_gem_init_seqno(dev, seqno - 1);
2391 /* Carefully set the last_seqno value so that wrap
2392 * detection still works
2394 dev_priv->next_seqno = seqno;
2395 dev_priv->last_seqno = seqno - 1;
2396 if (dev_priv->last_seqno == 0)
2397 dev_priv->last_seqno--;
2403 i915_gem_get_seqno(struct drm_device *dev, u32 *seqno)
2405 struct drm_i915_private *dev_priv = dev->dev_private;
2407 /* reserve 0 for non-seqno */
2408 if (dev_priv->next_seqno == 0) {
2409 int ret = i915_gem_init_seqno(dev, 0);
2413 dev_priv->next_seqno = 1;
2416 *seqno = dev_priv->last_seqno = dev_priv->next_seqno++;
2420 int __i915_add_request(struct intel_engine_cs *ring,
2421 struct drm_file *file,
2422 struct drm_i915_gem_object *obj,
2425 struct drm_i915_private *dev_priv = ring->dev->dev_private;
2426 struct drm_i915_gem_request *request;
2427 struct intel_ringbuffer *ringbuf;
2428 u32 request_ring_position, request_start;
2431 request = ring->preallocated_lazy_request;
2432 if (WARN_ON(request == NULL))
2435 if (i915.enable_execlists) {
2436 struct intel_context *ctx = request->ctx;
2437 ringbuf = ctx->engine[ring->id].ringbuf;
2439 ringbuf = ring->buffer;
2441 request_start = intel_ring_get_tail(ringbuf);
2443 * Emit any outstanding flushes - execbuf can fail to emit the flush
2444 * after having emitted the batchbuffer command. Hence we need to fix
2445 * things up similar to emitting the lazy request. The difference here
2446 * is that the flush _must_ happen before the next request, no matter
2449 if (i915.enable_execlists) {
2450 ret = logical_ring_flush_all_caches(ringbuf);
2454 ret = intel_ring_flush_all_caches(ring);
2459 /* Record the position of the start of the request so that
2460 * should we detect the updated seqno part-way through the
2461 * GPU processing the request, we never over-estimate the
2462 * position of the head.
2464 request_ring_position = intel_ring_get_tail(ringbuf);
2466 if (i915.enable_execlists) {
2467 ret = ring->emit_request(ringbuf);
2471 ret = ring->add_request(ring);
2476 request->seqno = intel_ring_get_seqno(ring);
2477 request->ring = ring;
2478 request->head = request_start;
2479 request->tail = request_ring_position;
2481 /* Whilst this request exists, batch_obj will be on the
2482 * active_list, and so will hold the active reference. Only when this
2483 * request is retired will the the batch_obj be moved onto the
2484 * inactive_list and lose its active reference. Hence we do not need
2485 * to explicitly hold another reference here.
2487 request->batch_obj = obj;
2489 if (!i915.enable_execlists) {
2490 /* Hold a reference to the current context so that we can inspect
2491 * it later in case a hangcheck error event fires.
2493 request->ctx = ring->last_context;
2495 i915_gem_context_reference(request->ctx);
2498 request->emitted_jiffies = jiffies;
2499 list_add_tail(&request->list, &ring->request_list);
2500 request->file_priv = NULL;
2503 struct drm_i915_file_private *file_priv = file->driver_priv;
2505 spin_lock(&file_priv->mm.lock);
2506 request->file_priv = file_priv;
2507 list_add_tail(&request->client_list,
2508 &file_priv->mm.request_list);
2509 spin_unlock(&file_priv->mm.lock);
2512 trace_i915_gem_request_add(ring, request->seqno);
2513 ring->outstanding_lazy_seqno = 0;
2514 ring->preallocated_lazy_request = NULL;
2516 i915_queue_hangcheck(ring->dev);
2518 cancel_delayed_work_sync(&dev_priv->mm.idle_work);
2519 queue_delayed_work(dev_priv->wq,
2520 &dev_priv->mm.retire_work,
2521 round_jiffies_up_relative(HZ));
2522 intel_mark_busy(dev_priv->dev);
2525 *out_seqno = request->seqno;
2530 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
2532 struct drm_i915_file_private *file_priv = request->file_priv;
2537 spin_lock(&file_priv->mm.lock);
2538 list_del(&request->client_list);
2539 request->file_priv = NULL;
2540 spin_unlock(&file_priv->mm.lock);
2543 static bool i915_context_is_banned(struct drm_i915_private *dev_priv,
2544 const struct intel_context *ctx)
2546 unsigned long elapsed;
2548 elapsed = get_seconds() - ctx->hang_stats.guilty_ts;
2550 if (ctx->hang_stats.banned)
2553 if (elapsed <= DRM_I915_CTX_BAN_PERIOD) {
2554 if (!i915_gem_context_is_default(ctx)) {
2555 DRM_DEBUG("context hanging too fast, banning!\n");
2557 } else if (i915_stop_ring_allow_ban(dev_priv)) {
2558 if (i915_stop_ring_allow_warn(dev_priv))
2559 DRM_ERROR("gpu hanging too fast, banning!\n");
2567 static void i915_set_reset_status(struct drm_i915_private *dev_priv,
2568 struct intel_context *ctx,
2571 struct i915_ctx_hang_stats *hs;
2576 hs = &ctx->hang_stats;
2579 hs->banned = i915_context_is_banned(dev_priv, ctx);
2581 hs->guilty_ts = get_seconds();
2583 hs->batch_pending++;
2587 static void i915_gem_free_request(struct drm_i915_gem_request *request)
2589 struct intel_context *ctx = request->ctx;
2591 list_del(&request->list);
2592 i915_gem_request_remove_from_client(request);
2595 if (i915.enable_execlists) {
2596 struct intel_engine_cs *ring = request->ring;
2598 if (ctx != ring->default_context)
2599 intel_lr_context_unpin(ring, ctx);
2601 i915_gem_context_unreference(ctx);
2606 struct drm_i915_gem_request *
2607 i915_gem_find_active_request(struct intel_engine_cs *ring)
2609 struct drm_i915_gem_request *request;
2610 u32 completed_seqno;
2612 completed_seqno = ring->get_seqno(ring, false);
2614 list_for_each_entry(request, &ring->request_list, list) {
2615 if (i915_seqno_passed(completed_seqno, request->seqno))
2624 static void i915_gem_reset_ring_status(struct drm_i915_private *dev_priv,
2625 struct intel_engine_cs *ring)
2627 struct drm_i915_gem_request *request;
2630 request = i915_gem_find_active_request(ring);
2632 if (request == NULL)
2635 ring_hung = ring->hangcheck.score >= HANGCHECK_SCORE_RING_HUNG;
2637 i915_set_reset_status(dev_priv, request->ctx, ring_hung);
2639 list_for_each_entry_continue(request, &ring->request_list, list)
2640 i915_set_reset_status(dev_priv, request->ctx, false);
2643 static void i915_gem_reset_ring_cleanup(struct drm_i915_private *dev_priv,
2644 struct intel_engine_cs *ring)
2646 while (!list_empty(&ring->active_list)) {
2647 struct drm_i915_gem_object *obj;
2649 obj = list_first_entry(&ring->active_list,
2650 struct drm_i915_gem_object,
2653 i915_gem_object_move_to_inactive(obj);
2657 * Clear the execlists queue up before freeing the requests, as those
2658 * are the ones that keep the context and ringbuffer backing objects
2661 while (!list_empty(&ring->execlist_queue)) {
2662 struct intel_ctx_submit_request *submit_req;
2664 submit_req = list_first_entry(&ring->execlist_queue,
2665 struct intel_ctx_submit_request,
2667 list_del(&submit_req->execlist_link);
2668 intel_runtime_pm_put(dev_priv);
2669 i915_gem_context_unreference(submit_req->ctx);
2674 * We must free the requests after all the corresponding objects have
2675 * been moved off active lists. Which is the same order as the normal
2676 * retire_requests function does. This is important if object hold
2677 * implicit references on things like e.g. ppgtt address spaces through
2680 while (!list_empty(&ring->request_list)) {
2681 struct drm_i915_gem_request *request;
2683 request = list_first_entry(&ring->request_list,
2684 struct drm_i915_gem_request,
2687 i915_gem_free_request(request);
2690 /* These may not have been flush before the reset, do so now */
2691 kfree(ring->preallocated_lazy_request);
2692 ring->preallocated_lazy_request = NULL;
2693 ring->outstanding_lazy_seqno = 0;
2696 void i915_gem_restore_fences(struct drm_device *dev)
2698 struct drm_i915_private *dev_priv = dev->dev_private;
2701 for (i = 0; i < dev_priv->num_fence_regs; i++) {
2702 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
2705 * Commit delayed tiling changes if we have an object still
2706 * attached to the fence, otherwise just clear the fence.
2709 i915_gem_object_update_fence(reg->obj, reg,
2710 reg->obj->tiling_mode);
2712 i915_gem_write_fence(dev, i, NULL);
2717 void i915_gem_reset(struct drm_device *dev)
2719 struct drm_i915_private *dev_priv = dev->dev_private;
2720 struct intel_engine_cs *ring;
2724 * Before we free the objects from the requests, we need to inspect
2725 * them for finding the guilty party. As the requests only borrow
2726 * their reference to the objects, the inspection must be done first.
2728 for_each_ring(ring, dev_priv, i)
2729 i915_gem_reset_ring_status(dev_priv, ring);
2731 for_each_ring(ring, dev_priv, i)
2732 i915_gem_reset_ring_cleanup(dev_priv, ring);
2734 i915_gem_context_reset(dev);
2736 i915_gem_restore_fences(dev);
2740 * This function clears the request list as sequence numbers are passed.
2743 i915_gem_retire_requests_ring(struct intel_engine_cs *ring)
2747 if (list_empty(&ring->request_list))
2750 WARN_ON(i915_verify_lists(ring->dev));
2752 seqno = ring->get_seqno(ring, true);
2754 /* Move any buffers on the active list that are no longer referenced
2755 * by the ringbuffer to the flushing/inactive lists as appropriate,
2756 * before we free the context associated with the requests.
2758 while (!list_empty(&ring->active_list)) {
2759 struct drm_i915_gem_object *obj;
2761 obj = list_first_entry(&ring->active_list,
2762 struct drm_i915_gem_object,
2765 if (!i915_seqno_passed(seqno, obj->last_read_seqno))
2768 i915_gem_object_move_to_inactive(obj);
2772 while (!list_empty(&ring->request_list)) {
2773 struct drm_i915_gem_request *request;
2774 struct intel_ringbuffer *ringbuf;
2776 request = list_first_entry(&ring->request_list,
2777 struct drm_i915_gem_request,
2780 if (!i915_seqno_passed(seqno, request->seqno))
2783 trace_i915_gem_request_retire(ring, request->seqno);
2785 /* This is one of the few common intersection points
2786 * between legacy ringbuffer submission and execlists:
2787 * we need to tell them apart in order to find the correct
2788 * ringbuffer to which the request belongs to.
2790 if (i915.enable_execlists) {
2791 struct intel_context *ctx = request->ctx;
2792 ringbuf = ctx->engine[ring->id].ringbuf;
2794 ringbuf = ring->buffer;
2796 /* We know the GPU must have read the request to have
2797 * sent us the seqno + interrupt, so use the position
2798 * of tail of the request to update the last known position
2801 ringbuf->last_retired_head = request->tail;
2803 i915_gem_free_request(request);
2806 if (unlikely(ring->trace_irq_seqno &&
2807 i915_seqno_passed(seqno, ring->trace_irq_seqno))) {
2808 ring->irq_put(ring);
2809 ring->trace_irq_seqno = 0;
2812 WARN_ON(i915_verify_lists(ring->dev));
2816 i915_gem_retire_requests(struct drm_device *dev)
2818 struct drm_i915_private *dev_priv = dev->dev_private;
2819 struct intel_engine_cs *ring;
2823 for_each_ring(ring, dev_priv, i) {
2824 i915_gem_retire_requests_ring(ring);
2825 idle &= list_empty(&ring->request_list);
2826 if (i915.enable_execlists) {
2827 unsigned long flags;
2829 spin_lock_irqsave(&ring->execlist_lock, flags);
2830 idle &= list_empty(&ring->execlist_queue);
2831 spin_unlock_irqrestore(&ring->execlist_lock, flags);
2833 intel_execlists_retire_requests(ring);
2838 mod_delayed_work(dev_priv->wq,
2839 &dev_priv->mm.idle_work,
2840 msecs_to_jiffies(100));
2846 i915_gem_retire_work_handler(struct work_struct *work)
2848 struct drm_i915_private *dev_priv =
2849 container_of(work, typeof(*dev_priv), mm.retire_work.work);
2850 struct drm_device *dev = dev_priv->dev;
2853 /* Come back later if the device is busy... */
2855 if (mutex_trylock(&dev->struct_mutex)) {
2856 idle = i915_gem_retire_requests(dev);
2857 mutex_unlock(&dev->struct_mutex);
2860 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
2861 round_jiffies_up_relative(HZ));
2865 i915_gem_idle_work_handler(struct work_struct *work)
2867 struct drm_i915_private *dev_priv =
2868 container_of(work, typeof(*dev_priv), mm.idle_work.work);
2870 intel_mark_idle(dev_priv->dev);
2874 * Ensures that an object will eventually get non-busy by flushing any required
2875 * write domains, emitting any outstanding lazy request and retiring and
2876 * completed requests.
2879 i915_gem_object_flush_active(struct drm_i915_gem_object *obj)
2884 ret = i915_gem_check_olr(obj->ring, obj->last_read_seqno);
2888 i915_gem_retire_requests_ring(obj->ring);
2895 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
2896 * @DRM_IOCTL_ARGS: standard ioctl arguments
2898 * Returns 0 if successful, else an error is returned with the remaining time in
2899 * the timeout parameter.
2900 * -ETIME: object is still busy after timeout
2901 * -ERESTARTSYS: signal interrupted the wait
2902 * -ENONENT: object doesn't exist
2903 * Also possible, but rare:
2904 * -EAGAIN: GPU wedged
2906 * -ENODEV: Internal IRQ fail
2907 * -E?: The add request failed
2909 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
2910 * non-zero timeout parameter the wait ioctl will wait for the given number of
2911 * nanoseconds on an object becoming unbusy. Since the wait itself does so
2912 * without holding struct_mutex the object may become re-busied before this
2913 * function completes. A similar but shorter * race condition exists in the busy
2917 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
2919 struct drm_i915_private *dev_priv = dev->dev_private;
2920 struct drm_i915_gem_wait *args = data;
2921 struct drm_i915_gem_object *obj;
2922 struct intel_engine_cs *ring = NULL;
2923 unsigned reset_counter;
2927 if (args->flags != 0)
2930 ret = i915_mutex_lock_interruptible(dev);
2934 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle));
2935 if (&obj->base == NULL) {
2936 mutex_unlock(&dev->struct_mutex);
2940 /* Need to make sure the object gets inactive eventually. */
2941 ret = i915_gem_object_flush_active(obj);
2946 seqno = obj->last_read_seqno;
2953 /* Do this after OLR check to make sure we make forward progress polling
2954 * on this IOCTL with a timeout <=0 (like busy ioctl)
2956 if (args->timeout_ns <= 0) {
2961 drm_gem_object_unreference(&obj->base);
2962 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
2963 mutex_unlock(&dev->struct_mutex);
2965 return __i915_wait_seqno(ring, seqno, reset_counter, true,
2966 &args->timeout_ns, file->driver_priv);
2969 drm_gem_object_unreference(&obj->base);
2970 mutex_unlock(&dev->struct_mutex);
2975 * i915_gem_object_sync - sync an object to a ring.
2977 * @obj: object which may be in use on another ring.
2978 * @to: ring we wish to use the object on. May be NULL.
2980 * This code is meant to abstract object synchronization with the GPU.
2981 * Calling with NULL implies synchronizing the object with the CPU
2982 * rather than a particular GPU ring.
2984 * Returns 0 if successful, else propagates up the lower layer error.
2987 i915_gem_object_sync(struct drm_i915_gem_object *obj,
2988 struct intel_engine_cs *to)
2990 struct intel_engine_cs *from = obj->ring;
2994 if (from == NULL || to == from)
2997 if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev))
2998 return i915_gem_object_wait_rendering(obj, false);
3000 idx = intel_ring_sync_index(from, to);
3002 seqno = obj->last_read_seqno;
3003 /* Optimization: Avoid semaphore sync when we are sure we already
3004 * waited for an object with higher seqno */
3005 if (seqno <= from->semaphore.sync_seqno[idx])
3008 ret = i915_gem_check_olr(obj->ring, seqno);
3012 trace_i915_gem_ring_sync_to(from, to, seqno);
3013 ret = to->semaphore.sync_to(to, from, seqno);
3015 /* We use last_read_seqno because sync_to()
3016 * might have just caused seqno wrap under
3019 from->semaphore.sync_seqno[idx] = obj->last_read_seqno;
3024 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
3026 u32 old_write_domain, old_read_domains;
3028 /* Force a pagefault for domain tracking on next user access */
3029 i915_gem_release_mmap(obj);
3031 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
3034 /* Wait for any direct GTT access to complete */
3037 old_read_domains = obj->base.read_domains;
3038 old_write_domain = obj->base.write_domain;
3040 obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
3041 obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
3043 trace_i915_gem_object_change_domain(obj,
3048 int i915_vma_unbind(struct i915_vma *vma)
3050 struct drm_i915_gem_object *obj = vma->obj;
3051 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3054 if (list_empty(&vma->vma_link))
3057 if (!drm_mm_node_allocated(&vma->node)) {
3058 i915_gem_vma_destroy(vma);
3065 BUG_ON(obj->pages == NULL);
3067 ret = i915_gem_object_finish_gpu(obj);
3070 /* Continue on if we fail due to EIO, the GPU is hung so we
3071 * should be safe and we need to cleanup or else we might
3072 * cause memory corruption through use-after-free.
3075 /* Throw away the active reference before moving to the unbound list */
3076 i915_gem_object_retire(obj);
3078 if (i915_is_ggtt(vma->vm)) {
3079 i915_gem_object_finish_gtt(obj);
3081 /* release the fence reg _after_ flushing */
3082 ret = i915_gem_object_put_fence(obj);
3087 trace_i915_vma_unbind(vma);
3089 vma->unbind_vma(vma);
3091 list_del_init(&vma->mm_list);
3092 if (i915_is_ggtt(vma->vm))
3093 obj->map_and_fenceable = false;
3095 drm_mm_remove_node(&vma->node);
3096 i915_gem_vma_destroy(vma);
3098 /* Since the unbound list is global, only move to that list if
3099 * no more VMAs exist. */
3100 if (list_empty(&obj->vma_list)) {
3101 i915_gem_gtt_finish_object(obj);
3102 list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list);
3105 /* And finally now the object is completely decoupled from this vma,
3106 * we can drop its hold on the backing storage and allow it to be
3107 * reaped by the shrinker.
3109 i915_gem_object_unpin_pages(obj);
3114 int i915_gpu_idle(struct drm_device *dev)
3116 struct drm_i915_private *dev_priv = dev->dev_private;
3117 struct intel_engine_cs *ring;
3120 /* Flush everything onto the inactive list. */
3121 for_each_ring(ring, dev_priv, i) {
3122 if (!i915.enable_execlists) {
3123 ret = i915_switch_context(ring, ring->default_context);
3128 ret = intel_ring_idle(ring);
3136 static void i965_write_fence_reg(struct drm_device *dev, int reg,
3137 struct drm_i915_gem_object *obj)
3139 struct drm_i915_private *dev_priv = dev->dev_private;
3141 int fence_pitch_shift;
3143 if (INTEL_INFO(dev)->gen >= 6) {
3144 fence_reg = FENCE_REG_SANDYBRIDGE_0;
3145 fence_pitch_shift = SANDYBRIDGE_FENCE_PITCH_SHIFT;
3147 fence_reg = FENCE_REG_965_0;
3148 fence_pitch_shift = I965_FENCE_PITCH_SHIFT;
3151 fence_reg += reg * 8;
3153 /* To w/a incoherency with non-atomic 64-bit register updates,
3154 * we split the 64-bit update into two 32-bit writes. In order
3155 * for a partial fence not to be evaluated between writes, we
3156 * precede the update with write to turn off the fence register,
3157 * and only enable the fence as the last step.
3159 * For extra levels of paranoia, we make sure each step lands
3160 * before applying the next step.
3162 I915_WRITE(fence_reg, 0);
3163 POSTING_READ(fence_reg);
3166 u32 size = i915_gem_obj_ggtt_size(obj);
3169 val = (uint64_t)((i915_gem_obj_ggtt_offset(obj) + size - 4096) &
3171 val |= i915_gem_obj_ggtt_offset(obj) & 0xfffff000;
3172 val |= (uint64_t)((obj->stride / 128) - 1) << fence_pitch_shift;
3173 if (obj->tiling_mode == I915_TILING_Y)
3174 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
3175 val |= I965_FENCE_REG_VALID;
3177 I915_WRITE(fence_reg + 4, val >> 32);
3178 POSTING_READ(fence_reg + 4);
3180 I915_WRITE(fence_reg + 0, val);
3181 POSTING_READ(fence_reg);
3183 I915_WRITE(fence_reg + 4, 0);
3184 POSTING_READ(fence_reg + 4);
3188 static void i915_write_fence_reg(struct drm_device *dev, int reg,
3189 struct drm_i915_gem_object *obj)
3191 struct drm_i915_private *dev_priv = dev->dev_private;
3195 u32 size = i915_gem_obj_ggtt_size(obj);
3199 WARN((i915_gem_obj_ggtt_offset(obj) & ~I915_FENCE_START_MASK) ||
3200 (size & -size) != size ||
3201 (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
3202 "object 0x%08lx [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
3203 i915_gem_obj_ggtt_offset(obj), obj->map_and_fenceable, size);
3205 if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
3210 /* Note: pitch better be a power of two tile widths */
3211 pitch_val = obj->stride / tile_width;
3212 pitch_val = ffs(pitch_val) - 1;
3214 val = i915_gem_obj_ggtt_offset(obj);
3215 if (obj->tiling_mode == I915_TILING_Y)
3216 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
3217 val |= I915_FENCE_SIZE_BITS(size);
3218 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
3219 val |= I830_FENCE_REG_VALID;
3224 reg = FENCE_REG_830_0 + reg * 4;
3226 reg = FENCE_REG_945_8 + (reg - 8) * 4;
3228 I915_WRITE(reg, val);
3232 static void i830_write_fence_reg(struct drm_device *dev, int reg,
3233 struct drm_i915_gem_object *obj)
3235 struct drm_i915_private *dev_priv = dev->dev_private;
3239 u32 size = i915_gem_obj_ggtt_size(obj);
3242 WARN((i915_gem_obj_ggtt_offset(obj) & ~I830_FENCE_START_MASK) ||
3243 (size & -size) != size ||
3244 (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
3245 "object 0x%08lx not 512K or pot-size 0x%08x aligned\n",
3246 i915_gem_obj_ggtt_offset(obj), size);
3248 pitch_val = obj->stride / 128;
3249 pitch_val = ffs(pitch_val) - 1;
3251 val = i915_gem_obj_ggtt_offset(obj);
3252 if (obj->tiling_mode == I915_TILING_Y)
3253 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
3254 val |= I830_FENCE_SIZE_BITS(size);
3255 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
3256 val |= I830_FENCE_REG_VALID;
3260 I915_WRITE(FENCE_REG_830_0 + reg * 4, val);
3261 POSTING_READ(FENCE_REG_830_0 + reg * 4);
3264 inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object *obj)
3266 return obj && obj->base.read_domains & I915_GEM_DOMAIN_GTT;
3269 static void i915_gem_write_fence(struct drm_device *dev, int reg,
3270 struct drm_i915_gem_object *obj)
3272 struct drm_i915_private *dev_priv = dev->dev_private;
3274 /* Ensure that all CPU reads are completed before installing a fence
3275 * and all writes before removing the fence.
3277 if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj))
3280 WARN(obj && (!obj->stride || !obj->tiling_mode),
3281 "bogus fence setup with stride: 0x%x, tiling mode: %i\n",
3282 obj->stride, obj->tiling_mode);
3284 switch (INTEL_INFO(dev)->gen) {
3290 case 4: i965_write_fence_reg(dev, reg, obj); break;
3291 case 3: i915_write_fence_reg(dev, reg, obj); break;
3292 case 2: i830_write_fence_reg(dev, reg, obj); break;
3296 /* And similarly be paranoid that no direct access to this region
3297 * is reordered to before the fence is installed.
3299 if (i915_gem_object_needs_mb(obj))
3303 static inline int fence_number(struct drm_i915_private *dev_priv,
3304 struct drm_i915_fence_reg *fence)
3306 return fence - dev_priv->fence_regs;
3309 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
3310 struct drm_i915_fence_reg *fence,
3313 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3314 int reg = fence_number(dev_priv, fence);
3316 i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL);
3319 obj->fence_reg = reg;
3321 list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list);
3323 obj->fence_reg = I915_FENCE_REG_NONE;
3325 list_del_init(&fence->lru_list);
3327 obj->fence_dirty = false;
3331 i915_gem_object_wait_fence(struct drm_i915_gem_object *obj)
3333 if (obj->last_fenced_seqno) {
3334 int ret = i915_wait_seqno(obj->ring, obj->last_fenced_seqno);
3338 obj->last_fenced_seqno = 0;
3345 i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
3347 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3348 struct drm_i915_fence_reg *fence;
3351 ret = i915_gem_object_wait_fence(obj);
3355 if (obj->fence_reg == I915_FENCE_REG_NONE)
3358 fence = &dev_priv->fence_regs[obj->fence_reg];
3360 if (WARN_ON(fence->pin_count))
3363 i915_gem_object_fence_lost(obj);
3364 i915_gem_object_update_fence(obj, fence, false);
3369 static struct drm_i915_fence_reg *
3370 i915_find_fence_reg(struct drm_device *dev)
3372 struct drm_i915_private *dev_priv = dev->dev_private;
3373 struct drm_i915_fence_reg *reg, *avail;
3376 /* First try to find a free reg */
3378 for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
3379 reg = &dev_priv->fence_regs[i];
3383 if (!reg->pin_count)
3390 /* None available, try to steal one or wait for a user to finish */
3391 list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
3399 /* Wait for completion of pending flips which consume fences */
3400 if (intel_has_pending_fb_unpin(dev))
3401 return ERR_PTR(-EAGAIN);
3403 return ERR_PTR(-EDEADLK);
3407 * i915_gem_object_get_fence - set up fencing for an object
3408 * @obj: object to map through a fence reg
3410 * When mapping objects through the GTT, userspace wants to be able to write
3411 * to them without having to worry about swizzling if the object is tiled.
3412 * This function walks the fence regs looking for a free one for @obj,
3413 * stealing one if it can't find any.
3415 * It then sets up the reg based on the object's properties: address, pitch
3416 * and tiling format.
3418 * For an untiled surface, this removes any existing fence.
3421 i915_gem_object_get_fence(struct drm_i915_gem_object *obj)
3423 struct drm_device *dev = obj->base.dev;
3424 struct drm_i915_private *dev_priv = dev->dev_private;
3425 bool enable = obj->tiling_mode != I915_TILING_NONE;
3426 struct drm_i915_fence_reg *reg;
3429 /* Have we updated the tiling parameters upon the object and so
3430 * will need to serialise the write to the associated fence register?
3432 if (obj->fence_dirty) {
3433 ret = i915_gem_object_wait_fence(obj);
3438 /* Just update our place in the LRU if our fence is getting reused. */
3439 if (obj->fence_reg != I915_FENCE_REG_NONE) {
3440 reg = &dev_priv->fence_regs[obj->fence_reg];
3441 if (!obj->fence_dirty) {
3442 list_move_tail(®->lru_list,
3443 &dev_priv->mm.fence_list);
3446 } else if (enable) {
3447 if (WARN_ON(!obj->map_and_fenceable))
3450 reg = i915_find_fence_reg(dev);
3452 return PTR_ERR(reg);
3455 struct drm_i915_gem_object *old = reg->obj;
3457 ret = i915_gem_object_wait_fence(old);
3461 i915_gem_object_fence_lost(old);
3466 i915_gem_object_update_fence(obj, reg, enable);
3471 static bool i915_gem_valid_gtt_space(struct i915_vma *vma,
3472 unsigned long cache_level)
3474 struct drm_mm_node *gtt_space = &vma->node;
3475 struct drm_mm_node *other;
3478 * On some machines we have to be careful when putting differing types
3479 * of snoopable memory together to avoid the prefetcher crossing memory
3480 * domains and dying. During vm initialisation, we decide whether or not
3481 * these constraints apply and set the drm_mm.color_adjust
3484 if (vma->vm->mm.color_adjust == NULL)
3487 if (!drm_mm_node_allocated(gtt_space))
3490 if (list_empty(>t_space->node_list))
3493 other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list);
3494 if (other->allocated && !other->hole_follows && other->color != cache_level)
3497 other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list);
3498 if (other->allocated && !gtt_space->hole_follows && other->color != cache_level)
3505 * Finds free space in the GTT aperture and binds the object there.
3507 static struct i915_vma *
3508 i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj,
3509 struct i915_address_space *vm,
3513 struct drm_device *dev = obj->base.dev;
3514 struct drm_i915_private *dev_priv = dev->dev_private;
3515 u32 size, fence_size, fence_alignment, unfenced_alignment;
3516 unsigned long start =
3517 flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
3519 flags & PIN_MAPPABLE ? dev_priv->gtt.mappable_end : vm->total;
3520 struct i915_vma *vma;
3523 fence_size = i915_gem_get_gtt_size(dev,
3526 fence_alignment = i915_gem_get_gtt_alignment(dev,
3528 obj->tiling_mode, true);
3529 unfenced_alignment =
3530 i915_gem_get_gtt_alignment(dev,
3532 obj->tiling_mode, false);
3535 alignment = flags & PIN_MAPPABLE ? fence_alignment :
3537 if (flags & PIN_MAPPABLE && alignment & (fence_alignment - 1)) {
3538 DRM_DEBUG("Invalid object alignment requested %u\n", alignment);
3539 return ERR_PTR(-EINVAL);
3542 size = flags & PIN_MAPPABLE ? fence_size : obj->base.size;
3544 /* If the object is bigger than the entire aperture, reject it early
3545 * before evicting everything in a vain attempt to find space.
3547 if (obj->base.size > end) {
3548 DRM_DEBUG("Attempting to bind an object larger than the aperture: object=%zd > %s aperture=%lu\n",
3550 flags & PIN_MAPPABLE ? "mappable" : "total",
3552 return ERR_PTR(-E2BIG);
3555 ret = i915_gem_object_get_pages(obj);
3557 return ERR_PTR(ret);
3559 i915_gem_object_pin_pages(obj);
3561 vma = i915_gem_obj_lookup_or_create_vma(obj, vm);
3566 ret = drm_mm_insert_node_in_range_generic(&vm->mm, &vma->node,
3570 DRM_MM_SEARCH_DEFAULT,
3571 DRM_MM_CREATE_DEFAULT);
3573 ret = i915_gem_evict_something(dev, vm, size, alignment,
3582 if (WARN_ON(!i915_gem_valid_gtt_space(vma, obj->cache_level))) {
3584 goto err_remove_node;
3587 ret = i915_gem_gtt_prepare_object(obj);
3589 goto err_remove_node;
3591 list_move_tail(&obj->global_list, &dev_priv->mm.bound_list);
3592 list_add_tail(&vma->mm_list, &vm->inactive_list);
3594 trace_i915_vma_bind(vma, flags);
3595 vma->bind_vma(vma, obj->cache_level,
3596 flags & PIN_GLOBAL ? GLOBAL_BIND : 0);
3601 drm_mm_remove_node(&vma->node);
3603 i915_gem_vma_destroy(vma);
3606 i915_gem_object_unpin_pages(obj);
3611 i915_gem_clflush_object(struct drm_i915_gem_object *obj,
3614 /* If we don't have a page list set up, then we're not pinned
3615 * to GPU, and we can ignore the cache flush because it'll happen
3616 * again at bind time.
3618 if (obj->pages == NULL)
3622 * Stolen memory is always coherent with the GPU as it is explicitly
3623 * marked as wc by the system, or the system is cache-coherent.
3625 if (obj->stolen || obj->phys_handle)
3628 /* If the GPU is snooping the contents of the CPU cache,
3629 * we do not need to manually clear the CPU cache lines. However,
3630 * the caches are only snooped when the render cache is
3631 * flushed/invalidated. As we always have to emit invalidations
3632 * and flushes when moving into and out of the RENDER domain, correct
3633 * snooping behaviour occurs naturally as the result of our domain
3636 if (!force && cpu_cache_is_coherent(obj->base.dev, obj->cache_level))
3639 trace_i915_gem_object_clflush(obj);
3640 drm_clflush_sg(obj->pages);
3645 /** Flushes the GTT write domain for the object if it's dirty. */
3647 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
3649 uint32_t old_write_domain;
3651 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
3654 /* No actual flushing is required for the GTT write domain. Writes
3655 * to it immediately go to main memory as far as we know, so there's
3656 * no chipset flush. It also doesn't land in render cache.
3658 * However, we do have to enforce the order so that all writes through
3659 * the GTT land before any writes to the device, such as updates to
3664 old_write_domain = obj->base.write_domain;
3665 obj->base.write_domain = 0;
3667 intel_fb_obj_flush(obj, false);
3669 trace_i915_gem_object_change_domain(obj,
3670 obj->base.read_domains,
3674 /** Flushes the CPU write domain for the object if it's dirty. */
3676 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj,
3679 uint32_t old_write_domain;
3681 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
3684 if (i915_gem_clflush_object(obj, force))
3685 i915_gem_chipset_flush(obj->base.dev);
3687 old_write_domain = obj->base.write_domain;
3688 obj->base.write_domain = 0;
3690 intel_fb_obj_flush(obj, false);
3692 trace_i915_gem_object_change_domain(obj,
3693 obj->base.read_domains,
3698 * Moves a single object to the GTT read, and possibly write domain.
3700 * This function returns when the move is complete, including waiting on
3704 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
3706 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3707 struct i915_vma *vma = i915_gem_obj_to_ggtt(obj);
3708 uint32_t old_write_domain, old_read_domains;
3711 /* Not valid to be called on unbound objects. */
3715 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
3718 ret = i915_gem_object_wait_rendering(obj, !write);
3722 i915_gem_object_retire(obj);
3723 i915_gem_object_flush_cpu_write_domain(obj, false);
3725 /* Serialise direct access to this object with the barriers for
3726 * coherent writes from the GPU, by effectively invalidating the
3727 * GTT domain upon first access.
3729 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
3732 old_write_domain = obj->base.write_domain;
3733 old_read_domains = obj->base.read_domains;
3735 /* It should now be out of any other write domains, and we can update
3736 * the domain values for our changes.
3738 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3739 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3741 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
3742 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
3747 intel_fb_obj_invalidate(obj, NULL);
3749 trace_i915_gem_object_change_domain(obj,
3753 /* And bump the LRU for this access */
3754 if (i915_gem_object_is_inactive(obj))
3755 list_move_tail(&vma->mm_list,
3756 &dev_priv->gtt.base.inactive_list);
3761 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
3762 enum i915_cache_level cache_level)
3764 struct drm_device *dev = obj->base.dev;
3765 struct i915_vma *vma, *next;
3768 if (obj->cache_level == cache_level)
3771 if (i915_gem_obj_is_pinned(obj)) {
3772 DRM_DEBUG("can not change the cache level of pinned objects\n");
3776 list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) {
3777 if (!i915_gem_valid_gtt_space(vma, cache_level)) {
3778 ret = i915_vma_unbind(vma);
3784 if (i915_gem_obj_bound_any(obj)) {
3785 ret = i915_gem_object_finish_gpu(obj);
3789 i915_gem_object_finish_gtt(obj);
3791 /* Before SandyBridge, you could not use tiling or fence
3792 * registers with snooped memory, so relinquish any fences
3793 * currently pointing to our region in the aperture.
3795 if (INTEL_INFO(dev)->gen < 6) {
3796 ret = i915_gem_object_put_fence(obj);
3801 list_for_each_entry(vma, &obj->vma_list, vma_link)
3802 if (drm_mm_node_allocated(&vma->node))
3803 vma->bind_vma(vma, cache_level,
3804 vma->bound & GLOBAL_BIND);
3807 list_for_each_entry(vma, &obj->vma_list, vma_link)
3808 vma->node.color = cache_level;
3809 obj->cache_level = cache_level;
3811 if (cpu_write_needs_clflush(obj)) {
3812 u32 old_read_domains, old_write_domain;
3814 /* If we're coming from LLC cached, then we haven't
3815 * actually been tracking whether the data is in the
3816 * CPU cache or not, since we only allow one bit set
3817 * in obj->write_domain and have been skipping the clflushes.
3818 * Just set it to the CPU cache for now.
3820 i915_gem_object_retire(obj);
3821 WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU);
3823 old_read_domains = obj->base.read_domains;
3824 old_write_domain = obj->base.write_domain;
3826 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3827 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3829 trace_i915_gem_object_change_domain(obj,
3837 int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data,
3838 struct drm_file *file)
3840 struct drm_i915_gem_caching *args = data;
3841 struct drm_i915_gem_object *obj;
3844 ret = i915_mutex_lock_interruptible(dev);
3848 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3849 if (&obj->base == NULL) {
3854 switch (obj->cache_level) {
3855 case I915_CACHE_LLC:
3856 case I915_CACHE_L3_LLC:
3857 args->caching = I915_CACHING_CACHED;
3861 args->caching = I915_CACHING_DISPLAY;
3865 args->caching = I915_CACHING_NONE;
3869 drm_gem_object_unreference(&obj->base);
3871 mutex_unlock(&dev->struct_mutex);
3875 int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
3876 struct drm_file *file)
3878 struct drm_i915_gem_caching *args = data;
3879 struct drm_i915_gem_object *obj;
3880 enum i915_cache_level level;
3883 switch (args->caching) {
3884 case I915_CACHING_NONE:
3885 level = I915_CACHE_NONE;
3887 case I915_CACHING_CACHED:
3888 level = I915_CACHE_LLC;
3890 case I915_CACHING_DISPLAY:
3891 level = HAS_WT(dev) ? I915_CACHE_WT : I915_CACHE_NONE;
3897 ret = i915_mutex_lock_interruptible(dev);
3901 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3902 if (&obj->base == NULL) {
3907 ret = i915_gem_object_set_cache_level(obj, level);
3909 drm_gem_object_unreference(&obj->base);
3911 mutex_unlock(&dev->struct_mutex);
3915 static bool is_pin_display(struct drm_i915_gem_object *obj)
3917 struct i915_vma *vma;
3919 vma = i915_gem_obj_to_ggtt(obj);
3923 /* There are 3 sources that pin objects:
3924 * 1. The display engine (scanouts, sprites, cursors);
3925 * 2. Reservations for execbuffer;
3928 * We can ignore reservations as we hold the struct_mutex and
3929 * are only called outside of the reservation path. The user
3930 * can only increment pin_count once, and so if after
3931 * subtracting the potential reference by the user, any pin_count
3932 * remains, it must be due to another use by the display engine.
3934 return vma->pin_count - !!obj->user_pin_count;
3938 * Prepare buffer for display plane (scanout, cursors, etc).
3939 * Can be called from an uninterruptible phase (modesetting) and allows
3940 * any flushes to be pipelined (for pageflips).
3943 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
3945 struct intel_engine_cs *pipelined)
3947 u32 old_read_domains, old_write_domain;
3948 bool was_pin_display;
3951 if (pipelined != obj->ring) {
3952 ret = i915_gem_object_sync(obj, pipelined);
3957 /* Mark the pin_display early so that we account for the
3958 * display coherency whilst setting up the cache domains.
3960 was_pin_display = obj->pin_display;
3961 obj->pin_display = true;
3963 /* The display engine is not coherent with the LLC cache on gen6. As
3964 * a result, we make sure that the pinning that is about to occur is
3965 * done with uncached PTEs. This is lowest common denominator for all
3968 * However for gen6+, we could do better by using the GFDT bit instead
3969 * of uncaching, which would allow us to flush all the LLC-cached data
3970 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3972 ret = i915_gem_object_set_cache_level(obj,
3973 HAS_WT(obj->base.dev) ? I915_CACHE_WT : I915_CACHE_NONE);
3975 goto err_unpin_display;
3977 /* As the user may map the buffer once pinned in the display plane
3978 * (e.g. libkms for the bootup splash), we have to ensure that we
3979 * always use map_and_fenceable for all scanout buffers.
3981 ret = i915_gem_obj_ggtt_pin(obj, alignment, PIN_MAPPABLE);
3983 goto err_unpin_display;
3985 i915_gem_object_flush_cpu_write_domain(obj, true);
3987 old_write_domain = obj->base.write_domain;
3988 old_read_domains = obj->base.read_domains;
3990 /* It should now be out of any other write domains, and we can update
3991 * the domain values for our changes.
3993 obj->base.write_domain = 0;
3994 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3996 trace_i915_gem_object_change_domain(obj,
4003 WARN_ON(was_pin_display != is_pin_display(obj));
4004 obj->pin_display = was_pin_display;
4009 i915_gem_object_unpin_from_display_plane(struct drm_i915_gem_object *obj)
4011 i915_gem_object_ggtt_unpin(obj);
4012 obj->pin_display = is_pin_display(obj);
4016 i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
4020 if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
4023 ret = i915_gem_object_wait_rendering(obj, false);
4027 /* Ensure that we invalidate the GPU's caches and TLBs. */
4028 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
4033 * Moves a single object to the CPU read, and possibly write domain.
4035 * This function returns when the move is complete, including waiting on
4039 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
4041 uint32_t old_write_domain, old_read_domains;
4044 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
4047 ret = i915_gem_object_wait_rendering(obj, !write);
4051 i915_gem_object_retire(obj);
4052 i915_gem_object_flush_gtt_write_domain(obj);
4054 old_write_domain = obj->base.write_domain;
4055 old_read_domains = obj->base.read_domains;
4057 /* Flush the CPU cache if it's still invalid. */
4058 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
4059 i915_gem_clflush_object(obj, false);
4061 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
4064 /* It should now be out of any other write domains, and we can update
4065 * the domain values for our changes.
4067 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
4069 /* If we're writing through the CPU, then the GPU read domains will
4070 * need to be invalidated at next use.
4073 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4074 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4078 intel_fb_obj_invalidate(obj, NULL);
4080 trace_i915_gem_object_change_domain(obj,
4087 /* Throttle our rendering by waiting until the ring has completed our requests
4088 * emitted over 20 msec ago.
4090 * Note that if we were to use the current jiffies each time around the loop,
4091 * we wouldn't escape the function with any frames outstanding if the time to
4092 * render a frame was over 20ms.
4094 * This should get us reasonable parallelism between CPU and GPU but also
4095 * relatively low latency when blocking on a particular request to finish.
4098 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
4100 struct drm_i915_private *dev_priv = dev->dev_private;
4101 struct drm_i915_file_private *file_priv = file->driver_priv;
4102 unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
4103 struct drm_i915_gem_request *request;
4104 struct intel_engine_cs *ring = NULL;
4105 unsigned reset_counter;
4109 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
4113 ret = i915_gem_check_wedge(&dev_priv->gpu_error, false);
4117 spin_lock(&file_priv->mm.lock);
4118 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
4119 if (time_after_eq(request->emitted_jiffies, recent_enough))
4122 ring = request->ring;
4123 seqno = request->seqno;
4125 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
4126 spin_unlock(&file_priv->mm.lock);
4131 ret = __i915_wait_seqno(ring, seqno, reset_counter, true, NULL, NULL);
4133 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
4139 i915_vma_misplaced(struct i915_vma *vma, uint32_t alignment, uint64_t flags)
4141 struct drm_i915_gem_object *obj = vma->obj;
4144 vma->node.start & (alignment - 1))
4147 if (flags & PIN_MAPPABLE && !obj->map_and_fenceable)
4150 if (flags & PIN_OFFSET_BIAS &&
4151 vma->node.start < (flags & PIN_OFFSET_MASK))
4158 i915_gem_object_pin(struct drm_i915_gem_object *obj,
4159 struct i915_address_space *vm,
4163 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
4164 struct i915_vma *vma;
4168 if (WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base))
4171 if (WARN_ON(flags & (PIN_GLOBAL | PIN_MAPPABLE) && !i915_is_ggtt(vm)))
4174 if (WARN_ON((flags & (PIN_MAPPABLE | PIN_GLOBAL)) == PIN_MAPPABLE))
4177 vma = i915_gem_obj_to_vma(obj, vm);
4179 if (WARN_ON(vma->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT))
4182 if (i915_vma_misplaced(vma, alignment, flags)) {
4183 WARN(vma->pin_count,
4184 "bo is already pinned with incorrect alignment:"
4185 " offset=%lx, req.alignment=%x, req.map_and_fenceable=%d,"
4186 " obj->map_and_fenceable=%d\n",
4187 i915_gem_obj_offset(obj, vm), alignment,
4188 !!(flags & PIN_MAPPABLE),
4189 obj->map_and_fenceable);
4190 ret = i915_vma_unbind(vma);
4198 bound = vma ? vma->bound : 0;
4199 if (vma == NULL || !drm_mm_node_allocated(&vma->node)) {
4200 vma = i915_gem_object_bind_to_vm(obj, vm, alignment, flags);
4202 return PTR_ERR(vma);
4205 if (flags & PIN_GLOBAL && !(vma->bound & GLOBAL_BIND))
4206 vma->bind_vma(vma, obj->cache_level, GLOBAL_BIND);
4208 if ((bound ^ vma->bound) & GLOBAL_BIND) {
4209 bool mappable, fenceable;
4210 u32 fence_size, fence_alignment;
4212 fence_size = i915_gem_get_gtt_size(obj->base.dev,
4215 fence_alignment = i915_gem_get_gtt_alignment(obj->base.dev,
4220 fenceable = (vma->node.size == fence_size &&
4221 (vma->node.start & (fence_alignment - 1)) == 0);
4223 mappable = (vma->node.start + obj->base.size <=
4224 dev_priv->gtt.mappable_end);
4226 obj->map_and_fenceable = mappable && fenceable;
4229 WARN_ON(flags & PIN_MAPPABLE && !obj->map_and_fenceable);
4232 if (flags & PIN_MAPPABLE)
4233 obj->pin_mappable |= true;
4239 i915_gem_object_ggtt_unpin(struct drm_i915_gem_object *obj)
4241 struct i915_vma *vma = i915_gem_obj_to_ggtt(obj);
4244 BUG_ON(vma->pin_count == 0);
4245 BUG_ON(!i915_gem_obj_ggtt_bound(obj));
4247 if (--vma->pin_count == 0)
4248 obj->pin_mappable = false;
4252 i915_gem_object_pin_fence(struct drm_i915_gem_object *obj)
4254 if (obj->fence_reg != I915_FENCE_REG_NONE) {
4255 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
4256 struct i915_vma *ggtt_vma = i915_gem_obj_to_ggtt(obj);
4258 WARN_ON(!ggtt_vma ||
4259 dev_priv->fence_regs[obj->fence_reg].pin_count >
4260 ggtt_vma->pin_count);
4261 dev_priv->fence_regs[obj->fence_reg].pin_count++;
4268 i915_gem_object_unpin_fence(struct drm_i915_gem_object *obj)
4270 if (obj->fence_reg != I915_FENCE_REG_NONE) {
4271 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
4272 WARN_ON(dev_priv->fence_regs[obj->fence_reg].pin_count <= 0);
4273 dev_priv->fence_regs[obj->fence_reg].pin_count--;
4278 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
4279 struct drm_file *file)
4281 struct drm_i915_gem_pin *args = data;
4282 struct drm_i915_gem_object *obj;
4285 if (drm_core_check_feature(dev, DRIVER_MODESET))
4288 ret = i915_mutex_lock_interruptible(dev);
4292 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
4293 if (&obj->base == NULL) {
4298 if (obj->madv != I915_MADV_WILLNEED) {
4299 DRM_DEBUG("Attempting to pin a purgeable buffer\n");
4304 if (obj->pin_filp != NULL && obj->pin_filp != file) {
4305 DRM_DEBUG("Already pinned in i915_gem_pin_ioctl(): %d\n",
4311 if (obj->user_pin_count == ULONG_MAX) {
4316 if (obj->user_pin_count == 0) {
4317 ret = i915_gem_obj_ggtt_pin(obj, args->alignment, PIN_MAPPABLE);
4322 obj->user_pin_count++;
4323 obj->pin_filp = file;
4325 args->offset = i915_gem_obj_ggtt_offset(obj);
4327 drm_gem_object_unreference(&obj->base);
4329 mutex_unlock(&dev->struct_mutex);
4334 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
4335 struct drm_file *file)
4337 struct drm_i915_gem_pin *args = data;
4338 struct drm_i915_gem_object *obj;
4341 if (drm_core_check_feature(dev, DRIVER_MODESET))
4344 ret = i915_mutex_lock_interruptible(dev);
4348 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
4349 if (&obj->base == NULL) {
4354 if (obj->pin_filp != file) {
4355 DRM_DEBUG("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
4360 obj->user_pin_count--;
4361 if (obj->user_pin_count == 0) {
4362 obj->pin_filp = NULL;
4363 i915_gem_object_ggtt_unpin(obj);
4367 drm_gem_object_unreference(&obj->base);
4369 mutex_unlock(&dev->struct_mutex);
4374 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
4375 struct drm_file *file)
4377 struct drm_i915_gem_busy *args = data;
4378 struct drm_i915_gem_object *obj;
4381 ret = i915_mutex_lock_interruptible(dev);
4385 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
4386 if (&obj->base == NULL) {
4391 /* Count all active objects as busy, even if they are currently not used
4392 * by the gpu. Users of this interface expect objects to eventually
4393 * become non-busy without any further actions, therefore emit any
4394 * necessary flushes here.
4396 ret = i915_gem_object_flush_active(obj);
4398 args->busy = obj->active;
4400 BUILD_BUG_ON(I915_NUM_RINGS > 16);
4401 args->busy |= intel_ring_flag(obj->ring) << 16;
4404 drm_gem_object_unreference(&obj->base);
4406 mutex_unlock(&dev->struct_mutex);
4411 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
4412 struct drm_file *file_priv)
4414 return i915_gem_ring_throttle(dev, file_priv);
4418 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
4419 struct drm_file *file_priv)
4421 struct drm_i915_private *dev_priv = dev->dev_private;
4422 struct drm_i915_gem_madvise *args = data;
4423 struct drm_i915_gem_object *obj;
4426 switch (args->madv) {
4427 case I915_MADV_DONTNEED:
4428 case I915_MADV_WILLNEED:
4434 ret = i915_mutex_lock_interruptible(dev);
4438 obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
4439 if (&obj->base == NULL) {
4444 if (i915_gem_obj_is_pinned(obj)) {
4450 obj->tiling_mode != I915_TILING_NONE &&
4451 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
4452 if (obj->madv == I915_MADV_WILLNEED)
4453 i915_gem_object_unpin_pages(obj);
4454 if (args->madv == I915_MADV_WILLNEED)
4455 i915_gem_object_pin_pages(obj);
4458 if (obj->madv != __I915_MADV_PURGED)
4459 obj->madv = args->madv;
4461 /* if the object is no longer attached, discard its backing storage */
4462 if (i915_gem_object_is_purgeable(obj) && obj->pages == NULL)
4463 i915_gem_object_truncate(obj);
4465 args->retained = obj->madv != __I915_MADV_PURGED;
4468 drm_gem_object_unreference(&obj->base);
4470 mutex_unlock(&dev->struct_mutex);
4474 void i915_gem_object_init(struct drm_i915_gem_object *obj,
4475 const struct drm_i915_gem_object_ops *ops)
4477 INIT_LIST_HEAD(&obj->global_list);
4478 INIT_LIST_HEAD(&obj->ring_list);
4479 INIT_LIST_HEAD(&obj->obj_exec_link);
4480 INIT_LIST_HEAD(&obj->vma_list);
4484 obj->fence_reg = I915_FENCE_REG_NONE;
4485 obj->madv = I915_MADV_WILLNEED;
4487 i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size);
4490 static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
4491 .get_pages = i915_gem_object_get_pages_gtt,
4492 .put_pages = i915_gem_object_put_pages_gtt,
4495 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
4498 struct drm_i915_gem_object *obj;
4499 struct address_space *mapping;
4502 obj = i915_gem_object_alloc(dev);
4506 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
4507 i915_gem_object_free(obj);
4511 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
4512 if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) {
4513 /* 965gm cannot relocate objects above 4GiB. */
4514 mask &= ~__GFP_HIGHMEM;
4515 mask |= __GFP_DMA32;
4518 mapping = file_inode(obj->base.filp)->i_mapping;
4519 mapping_set_gfp_mask(mapping, mask);
4521 i915_gem_object_init(obj, &i915_gem_object_ops);
4523 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4524 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4527 /* On some devices, we can have the GPU use the LLC (the CPU
4528 * cache) for about a 10% performance improvement
4529 * compared to uncached. Graphics requests other than
4530 * display scanout are coherent with the CPU in
4531 * accessing this cache. This means in this mode we
4532 * don't need to clflush on the CPU side, and on the
4533 * GPU side we only need to flush internal caches to
4534 * get data visible to the CPU.
4536 * However, we maintain the display planes as UC, and so
4537 * need to rebind when first used as such.
4539 obj->cache_level = I915_CACHE_LLC;
4541 obj->cache_level = I915_CACHE_NONE;
4543 trace_i915_gem_object_create(obj);
4548 static bool discard_backing_storage(struct drm_i915_gem_object *obj)
4550 /* If we are the last user of the backing storage (be it shmemfs
4551 * pages or stolen etc), we know that the pages are going to be
4552 * immediately released. In this case, we can then skip copying
4553 * back the contents from the GPU.
4556 if (obj->madv != I915_MADV_WILLNEED)
4559 if (obj->base.filp == NULL)
4562 /* At first glance, this looks racy, but then again so would be
4563 * userspace racing mmap against close. However, the first external
4564 * reference to the filp can only be obtained through the
4565 * i915_gem_mmap_ioctl() which safeguards us against the user
4566 * acquiring such a reference whilst we are in the middle of
4567 * freeing the object.
4569 return atomic_long_read(&obj->base.filp->f_count) == 1;
4572 void i915_gem_free_object(struct drm_gem_object *gem_obj)
4574 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
4575 struct drm_device *dev = obj->base.dev;
4576 struct drm_i915_private *dev_priv = dev->dev_private;
4577 struct i915_vma *vma, *next;
4579 intel_runtime_pm_get(dev_priv);
4581 trace_i915_gem_object_destroy(obj);
4583 list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) {
4587 ret = i915_vma_unbind(vma);
4588 if (WARN_ON(ret == -ERESTARTSYS)) {
4589 bool was_interruptible;
4591 was_interruptible = dev_priv->mm.interruptible;
4592 dev_priv->mm.interruptible = false;
4594 WARN_ON(i915_vma_unbind(vma));
4596 dev_priv->mm.interruptible = was_interruptible;
4600 /* Stolen objects don't hold a ref, but do hold pin count. Fix that up
4601 * before progressing. */
4603 i915_gem_object_unpin_pages(obj);
4605 WARN_ON(obj->frontbuffer_bits);
4607 if (obj->pages && obj->madv == I915_MADV_WILLNEED &&
4608 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES &&
4609 obj->tiling_mode != I915_TILING_NONE)
4610 i915_gem_object_unpin_pages(obj);
4612 if (WARN_ON(obj->pages_pin_count))
4613 obj->pages_pin_count = 0;
4614 if (discard_backing_storage(obj))
4615 obj->madv = I915_MADV_DONTNEED;
4616 i915_gem_object_put_pages(obj);
4617 i915_gem_object_free_mmap_offset(obj);
4621 if (obj->base.import_attach)
4622 drm_prime_gem_destroy(&obj->base, NULL);
4624 if (obj->ops->release)
4625 obj->ops->release(obj);
4627 drm_gem_object_release(&obj->base);
4628 i915_gem_info_remove_obj(dev_priv, obj->base.size);
4631 i915_gem_object_free(obj);
4633 intel_runtime_pm_put(dev_priv);
4636 struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj,
4637 struct i915_address_space *vm)
4639 struct i915_vma *vma;
4640 list_for_each_entry(vma, &obj->vma_list, vma_link)
4647 void i915_gem_vma_destroy(struct i915_vma *vma)
4649 struct i915_address_space *vm = NULL;
4650 WARN_ON(vma->node.allocated);
4652 /* Keep the vma as a placeholder in the execbuffer reservation lists */
4653 if (!list_empty(&vma->exec_list))
4658 if (!i915_is_ggtt(vm))
4659 i915_ppgtt_put(i915_vm_to_ppgtt(vm));
4661 list_del(&vma->vma_link);
4667 i915_gem_stop_ringbuffers(struct drm_device *dev)
4669 struct drm_i915_private *dev_priv = dev->dev_private;
4670 struct intel_engine_cs *ring;
4673 for_each_ring(ring, dev_priv, i)
4674 dev_priv->gt.stop_ring(ring);
4678 i915_gem_suspend(struct drm_device *dev)
4680 struct drm_i915_private *dev_priv = dev->dev_private;
4683 mutex_lock(&dev->struct_mutex);
4684 ret = i915_gpu_idle(dev);
4688 i915_gem_retire_requests(dev);
4690 /* Under UMS, be paranoid and evict. */
4691 if (!drm_core_check_feature(dev, DRIVER_MODESET))
4692 i915_gem_evict_everything(dev);
4694 i915_gem_stop_ringbuffers(dev);
4695 mutex_unlock(&dev->struct_mutex);
4697 del_timer_sync(&dev_priv->gpu_error.hangcheck_timer);
4698 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
4699 flush_delayed_work(&dev_priv->mm.idle_work);
4704 mutex_unlock(&dev->struct_mutex);
4708 int i915_gem_l3_remap(struct intel_engine_cs *ring, int slice)
4710 struct drm_device *dev = ring->dev;
4711 struct drm_i915_private *dev_priv = dev->dev_private;
4712 u32 reg_base = GEN7_L3LOG_BASE + (slice * 0x200);
4713 u32 *remap_info = dev_priv->l3_parity.remap_info[slice];
4716 if (!HAS_L3_DPF(dev) || !remap_info)
4719 ret = intel_ring_begin(ring, GEN7_L3LOG_SIZE / 4 * 3);
4724 * Note: We do not worry about the concurrent register cacheline hang
4725 * here because no other code should access these registers other than
4726 * at initialization time.
4728 for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) {
4729 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
4730 intel_ring_emit(ring, reg_base + i);
4731 intel_ring_emit(ring, remap_info[i/4]);
4734 intel_ring_advance(ring);
4739 void i915_gem_init_swizzling(struct drm_device *dev)
4741 struct drm_i915_private *dev_priv = dev->dev_private;
4743 if (INTEL_INFO(dev)->gen < 5 ||
4744 dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
4747 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
4748 DISP_TILE_SURFACE_SWIZZLING);
4753 I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
4755 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
4756 else if (IS_GEN7(dev))
4757 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
4758 else if (IS_GEN8(dev))
4759 I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW));
4765 intel_enable_blt(struct drm_device *dev)
4770 /* The blitter was dysfunctional on early prototypes */
4771 if (IS_GEN6(dev) && dev->pdev->revision < 8) {
4772 DRM_INFO("BLT not supported on this pre-production hardware;"
4773 " graphics performance will be degraded.\n");
4780 static void init_unused_ring(struct drm_device *dev, u32 base)
4782 struct drm_i915_private *dev_priv = dev->dev_private;
4784 I915_WRITE(RING_CTL(base), 0);
4785 I915_WRITE(RING_HEAD(base), 0);
4786 I915_WRITE(RING_TAIL(base), 0);
4787 I915_WRITE(RING_START(base), 0);
4790 static void init_unused_rings(struct drm_device *dev)
4793 init_unused_ring(dev, PRB1_BASE);
4794 init_unused_ring(dev, SRB0_BASE);
4795 init_unused_ring(dev, SRB1_BASE);
4796 init_unused_ring(dev, SRB2_BASE);
4797 init_unused_ring(dev, SRB3_BASE);
4798 } else if (IS_GEN2(dev)) {
4799 init_unused_ring(dev, SRB0_BASE);
4800 init_unused_ring(dev, SRB1_BASE);
4801 } else if (IS_GEN3(dev)) {
4802 init_unused_ring(dev, PRB1_BASE);
4803 init_unused_ring(dev, PRB2_BASE);
4807 int i915_gem_init_rings(struct drm_device *dev)
4809 struct drm_i915_private *dev_priv = dev->dev_private;
4813 * At least 830 can leave some of the unused rings
4814 * "active" (ie. head != tail) after resume which
4815 * will prevent c3 entry. Makes sure all unused rings
4818 init_unused_rings(dev);
4820 ret = intel_init_render_ring_buffer(dev);
4825 ret = intel_init_bsd_ring_buffer(dev);
4827 goto cleanup_render_ring;
4830 if (intel_enable_blt(dev)) {
4831 ret = intel_init_blt_ring_buffer(dev);
4833 goto cleanup_bsd_ring;
4836 if (HAS_VEBOX(dev)) {
4837 ret = intel_init_vebox_ring_buffer(dev);
4839 goto cleanup_blt_ring;
4842 if (HAS_BSD2(dev)) {
4843 ret = intel_init_bsd2_ring_buffer(dev);
4845 goto cleanup_vebox_ring;
4848 ret = i915_gem_set_seqno(dev, ((u32)~0 - 0x1000));
4850 goto cleanup_bsd2_ring;
4855 intel_cleanup_ring_buffer(&dev_priv->ring[VCS2]);
4857 intel_cleanup_ring_buffer(&dev_priv->ring[VECS]);
4859 intel_cleanup_ring_buffer(&dev_priv->ring[BCS]);
4861 intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
4862 cleanup_render_ring:
4863 intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
4869 i915_gem_init_hw(struct drm_device *dev)
4871 struct drm_i915_private *dev_priv = dev->dev_private;
4874 if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt())
4877 if (dev_priv->ellc_size)
4878 I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf));
4880 if (IS_HASWELL(dev))
4881 I915_WRITE(MI_PREDICATE_RESULT_2, IS_HSW_GT3(dev) ?
4882 LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED);
4884 if (HAS_PCH_NOP(dev)) {
4885 if (IS_IVYBRIDGE(dev)) {
4886 u32 temp = I915_READ(GEN7_MSG_CTL);
4887 temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK);
4888 I915_WRITE(GEN7_MSG_CTL, temp);
4889 } else if (INTEL_INFO(dev)->gen >= 7) {
4890 u32 temp = I915_READ(HSW_NDE_RSTWRN_OPT);
4891 temp &= ~RESET_PCH_HANDSHAKE_ENABLE;
4892 I915_WRITE(HSW_NDE_RSTWRN_OPT, temp);
4896 i915_gem_init_swizzling(dev);
4898 ret = dev_priv->gt.init_rings(dev);
4902 for (i = 0; i < NUM_L3_SLICES(dev); i++)
4903 i915_gem_l3_remap(&dev_priv->ring[RCS], i);
4906 * XXX: Contexts should only be initialized once. Doing a switch to the
4907 * default context switch however is something we'd like to do after
4908 * reset or thaw (the latter may not actually be necessary for HW, but
4909 * goes with our code better). Context switching requires rings (for
4910 * the do_switch), but before enabling PPGTT. So don't move this.
4912 ret = i915_gem_context_enable(dev_priv);
4913 if (ret && ret != -EIO) {
4914 DRM_ERROR("Context enable failed %d\n", ret);
4915 i915_gem_cleanup_ringbuffer(dev);
4920 ret = i915_ppgtt_init_hw(dev);
4921 if (ret && ret != -EIO) {
4922 DRM_ERROR("PPGTT enable failed %d\n", ret);
4923 i915_gem_cleanup_ringbuffer(dev);
4929 int i915_gem_init(struct drm_device *dev)
4931 struct drm_i915_private *dev_priv = dev->dev_private;
4934 i915.enable_execlists = intel_sanitize_enable_execlists(dev,
4935 i915.enable_execlists);
4937 mutex_lock(&dev->struct_mutex);
4939 if (IS_VALLEYVIEW(dev)) {
4940 /* VLVA0 (potential hack), BIOS isn't actually waking us */
4941 I915_WRITE(VLV_GTLC_WAKE_CTRL, VLV_GTLC_ALLOWWAKEREQ);
4942 if (wait_for((I915_READ(VLV_GTLC_PW_STATUS) &
4943 VLV_GTLC_ALLOWWAKEACK), 10))
4944 DRM_DEBUG_DRIVER("allow wake ack timed out\n");
4947 if (!i915.enable_execlists) {
4948 dev_priv->gt.do_execbuf = i915_gem_ringbuffer_submission;
4949 dev_priv->gt.init_rings = i915_gem_init_rings;
4950 dev_priv->gt.cleanup_ring = intel_cleanup_ring_buffer;
4951 dev_priv->gt.stop_ring = intel_stop_ring_buffer;
4953 dev_priv->gt.do_execbuf = intel_execlists_submission;
4954 dev_priv->gt.init_rings = intel_logical_rings_init;
4955 dev_priv->gt.cleanup_ring = intel_logical_ring_cleanup;
4956 dev_priv->gt.stop_ring = intel_logical_ring_stop;
4959 ret = i915_gem_init_userptr(dev);
4961 mutex_unlock(&dev->struct_mutex);
4965 i915_gem_init_global_gtt(dev);
4967 ret = i915_gem_context_init(dev);
4969 mutex_unlock(&dev->struct_mutex);
4973 ret = i915_gem_init_hw(dev);
4975 /* Allow ring initialisation to fail by marking the GPU as
4976 * wedged. But we only want to do this where the GPU is angry,
4977 * for all other failure, such as an allocation failure, bail.
4979 DRM_ERROR("Failed to initialize GPU, declaring it wedged\n");
4980 atomic_set_mask(I915_WEDGED, &dev_priv->gpu_error.reset_counter);
4983 mutex_unlock(&dev->struct_mutex);
4989 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
4991 struct drm_i915_private *dev_priv = dev->dev_private;
4992 struct intel_engine_cs *ring;
4995 for_each_ring(ring, dev_priv, i)
4996 dev_priv->gt.cleanup_ring(ring);
5000 init_ring_lists(struct intel_engine_cs *ring)
5002 INIT_LIST_HEAD(&ring->active_list);
5003 INIT_LIST_HEAD(&ring->request_list);
5006 void i915_init_vm(struct drm_i915_private *dev_priv,
5007 struct i915_address_space *vm)
5009 if (!i915_is_ggtt(vm))
5010 drm_mm_init(&vm->mm, vm->start, vm->total);
5011 vm->dev = dev_priv->dev;
5012 INIT_LIST_HEAD(&vm->active_list);
5013 INIT_LIST_HEAD(&vm->inactive_list);
5014 INIT_LIST_HEAD(&vm->global_link);
5015 list_add_tail(&vm->global_link, &dev_priv->vm_list);
5019 i915_gem_load(struct drm_device *dev)
5021 struct drm_i915_private *dev_priv = dev->dev_private;
5025 kmem_cache_create("i915_gem_object",
5026 sizeof(struct drm_i915_gem_object), 0,
5030 INIT_LIST_HEAD(&dev_priv->vm_list);
5031 i915_init_vm(dev_priv, &dev_priv->gtt.base);
5033 INIT_LIST_HEAD(&dev_priv->context_list);
5034 INIT_LIST_HEAD(&dev_priv->mm.unbound_list);
5035 INIT_LIST_HEAD(&dev_priv->mm.bound_list);
5036 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
5037 for (i = 0; i < I915_NUM_RINGS; i++)
5038 init_ring_lists(&dev_priv->ring[i]);
5039 for (i = 0; i < I915_MAX_NUM_FENCES; i++)
5040 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
5041 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
5042 i915_gem_retire_work_handler);
5043 INIT_DELAYED_WORK(&dev_priv->mm.idle_work,
5044 i915_gem_idle_work_handler);
5045 init_waitqueue_head(&dev_priv->gpu_error.reset_queue);
5047 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
5048 if (!drm_core_check_feature(dev, DRIVER_MODESET) && IS_GEN3(dev)) {
5049 I915_WRITE(MI_ARB_STATE,
5050 _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
5053 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
5055 /* Old X drivers will take 0-2 for front, back, depth buffers */
5056 if (!drm_core_check_feature(dev, DRIVER_MODESET))
5057 dev_priv->fence_reg_start = 3;
5059 if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev))
5060 dev_priv->num_fence_regs = 32;
5061 else if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
5062 dev_priv->num_fence_regs = 16;
5064 dev_priv->num_fence_regs = 8;
5066 /* Initialize fence registers to zero */
5067 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
5068 i915_gem_restore_fences(dev);
5070 i915_gem_detect_bit_6_swizzle(dev);
5071 init_waitqueue_head(&dev_priv->pending_flip_queue);
5073 dev_priv->mm.interruptible = true;
5075 dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
5076 dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count;
5077 dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS;
5078 register_shrinker(&dev_priv->mm.shrinker);
5080 dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
5081 register_oom_notifier(&dev_priv->mm.oom_notifier);
5083 mutex_init(&dev_priv->fb_tracking.lock);
5086 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
5088 struct drm_i915_file_private *file_priv = file->driver_priv;
5090 cancel_delayed_work_sync(&file_priv->mm.idle_work);
5092 /* Clean up our request list when the client is going away, so that
5093 * later retire_requests won't dereference our soon-to-be-gone
5096 spin_lock(&file_priv->mm.lock);
5097 while (!list_empty(&file_priv->mm.request_list)) {
5098 struct drm_i915_gem_request *request;
5100 request = list_first_entry(&file_priv->mm.request_list,
5101 struct drm_i915_gem_request,
5103 list_del(&request->client_list);
5104 request->file_priv = NULL;
5106 spin_unlock(&file_priv->mm.lock);
5110 i915_gem_file_idle_work_handler(struct work_struct *work)
5112 struct drm_i915_file_private *file_priv =
5113 container_of(work, typeof(*file_priv), mm.idle_work.work);
5115 atomic_set(&file_priv->rps_wait_boost, false);
5118 int i915_gem_open(struct drm_device *dev, struct drm_file *file)
5120 struct drm_i915_file_private *file_priv;
5123 DRM_DEBUG_DRIVER("\n");
5125 file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL);
5129 file->driver_priv = file_priv;
5130 file_priv->dev_priv = dev->dev_private;
5131 file_priv->file = file;
5133 spin_lock_init(&file_priv->mm.lock);
5134 INIT_LIST_HEAD(&file_priv->mm.request_list);
5135 INIT_DELAYED_WORK(&file_priv->mm.idle_work,
5136 i915_gem_file_idle_work_handler);
5138 ret = i915_gem_context_open(dev, file);
5146 * i915_gem_track_fb - update frontbuffer tracking
5147 * old: current GEM buffer for the frontbuffer slots
5148 * new: new GEM buffer for the frontbuffer slots
5149 * frontbuffer_bits: bitmask of frontbuffer slots
5151 * This updates the frontbuffer tracking bits @frontbuffer_bits by clearing them
5152 * from @old and setting them in @new. Both @old and @new can be NULL.
5154 void i915_gem_track_fb(struct drm_i915_gem_object *old,
5155 struct drm_i915_gem_object *new,
5156 unsigned frontbuffer_bits)
5159 WARN_ON(!mutex_is_locked(&old->base.dev->struct_mutex));
5160 WARN_ON(!(old->frontbuffer_bits & frontbuffer_bits));
5161 old->frontbuffer_bits &= ~frontbuffer_bits;
5165 WARN_ON(!mutex_is_locked(&new->base.dev->struct_mutex));
5166 WARN_ON(new->frontbuffer_bits & frontbuffer_bits);
5167 new->frontbuffer_bits |= frontbuffer_bits;
5171 static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task)
5173 if (!mutex_is_locked(mutex))
5176 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES)
5177 return mutex->owner == task;
5179 /* Since UP may be pre-empted, we cannot assume that we own the lock */
5184 static bool i915_gem_shrinker_lock(struct drm_device *dev, bool *unlock)
5186 if (!mutex_trylock(&dev->struct_mutex)) {
5187 if (!mutex_is_locked_by(&dev->struct_mutex, current))
5190 if (to_i915(dev)->mm.shrinker_no_lock_stealing)
5200 static int num_vma_bound(struct drm_i915_gem_object *obj)
5202 struct i915_vma *vma;
5205 list_for_each_entry(vma, &obj->vma_list, vma_link)
5206 if (drm_mm_node_allocated(&vma->node))
5212 static unsigned long
5213 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
5215 struct drm_i915_private *dev_priv =
5216 container_of(shrinker, struct drm_i915_private, mm.shrinker);
5217 struct drm_device *dev = dev_priv->dev;
5218 struct drm_i915_gem_object *obj;
5219 unsigned long count;
5222 if (!i915_gem_shrinker_lock(dev, &unlock))
5226 list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list)
5227 if (obj->pages_pin_count == 0)
5228 count += obj->base.size >> PAGE_SHIFT;
5230 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
5231 if (!i915_gem_obj_is_pinned(obj) &&
5232 obj->pages_pin_count == num_vma_bound(obj))
5233 count += obj->base.size >> PAGE_SHIFT;
5237 mutex_unlock(&dev->struct_mutex);
5242 /* All the new VM stuff */
5243 unsigned long i915_gem_obj_offset(struct drm_i915_gem_object *o,
5244 struct i915_address_space *vm)
5246 struct drm_i915_private *dev_priv = o->base.dev->dev_private;
5247 struct i915_vma *vma;
5249 WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base);
5251 list_for_each_entry(vma, &o->vma_list, vma_link) {
5253 return vma->node.start;
5256 WARN(1, "%s vma for this object not found.\n",
5257 i915_is_ggtt(vm) ? "global" : "ppgtt");
5261 bool i915_gem_obj_bound(struct drm_i915_gem_object *o,
5262 struct i915_address_space *vm)
5264 struct i915_vma *vma;
5266 list_for_each_entry(vma, &o->vma_list, vma_link)
5267 if (vma->vm == vm && drm_mm_node_allocated(&vma->node))
5273 bool i915_gem_obj_bound_any(struct drm_i915_gem_object *o)
5275 struct i915_vma *vma;
5277 list_for_each_entry(vma, &o->vma_list, vma_link)
5278 if (drm_mm_node_allocated(&vma->node))
5284 unsigned long i915_gem_obj_size(struct drm_i915_gem_object *o,
5285 struct i915_address_space *vm)
5287 struct drm_i915_private *dev_priv = o->base.dev->dev_private;
5288 struct i915_vma *vma;
5290 WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base);
5292 BUG_ON(list_empty(&o->vma_list));
5294 list_for_each_entry(vma, &o->vma_list, vma_link)
5296 return vma->node.size;
5301 static unsigned long
5302 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
5304 struct drm_i915_private *dev_priv =
5305 container_of(shrinker, struct drm_i915_private, mm.shrinker);
5306 struct drm_device *dev = dev_priv->dev;
5307 unsigned long freed;
5310 if (!i915_gem_shrinker_lock(dev, &unlock))
5313 freed = i915_gem_shrink(dev_priv,
5316 I915_SHRINK_UNBOUND |
5317 I915_SHRINK_PURGEABLE);
5318 if (freed < sc->nr_to_scan)
5319 freed += i915_gem_shrink(dev_priv,
5320 sc->nr_to_scan - freed,
5322 I915_SHRINK_UNBOUND);
5324 mutex_unlock(&dev->struct_mutex);
5330 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
5332 struct drm_i915_private *dev_priv =
5333 container_of(nb, struct drm_i915_private, mm.oom_notifier);
5334 struct drm_device *dev = dev_priv->dev;
5335 struct drm_i915_gem_object *obj;
5336 unsigned long timeout = msecs_to_jiffies(5000) + 1;
5337 unsigned long pinned, bound, unbound, freed_pages;
5338 bool was_interruptible;
5341 while (!i915_gem_shrinker_lock(dev, &unlock) && --timeout) {
5342 schedule_timeout_killable(1);
5343 if (fatal_signal_pending(current))
5347 pr_err("Unable to purge GPU memory due lock contention.\n");
5351 was_interruptible = dev_priv->mm.interruptible;
5352 dev_priv->mm.interruptible = false;
5354 freed_pages = i915_gem_shrink_all(dev_priv);
5356 dev_priv->mm.interruptible = was_interruptible;
5358 /* Because we may be allocating inside our own driver, we cannot
5359 * assert that there are no objects with pinned pages that are not
5360 * being pointed to by hardware.
5362 unbound = bound = pinned = 0;
5363 list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) {
5364 if (!obj->base.filp) /* not backed by a freeable object */
5367 if (obj->pages_pin_count)
5368 pinned += obj->base.size;
5370 unbound += obj->base.size;
5372 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
5373 if (!obj->base.filp)
5376 if (obj->pages_pin_count)
5377 pinned += obj->base.size;
5379 bound += obj->base.size;
5383 mutex_unlock(&dev->struct_mutex);
5385 if (freed_pages || unbound || bound)
5386 pr_info("Purging GPU memory, %lu bytes freed, %lu bytes still pinned.\n",
5387 freed_pages << PAGE_SHIFT, pinned);
5388 if (unbound || bound)
5389 pr_err("%lu and %lu bytes still available in the "
5390 "bound and unbound GPU page lists.\n",
5393 *(unsigned long *)ptr += freed_pages;
5397 struct i915_vma *i915_gem_obj_to_ggtt(struct drm_i915_gem_object *obj)
5399 struct i915_vma *vma;
5401 vma = list_first_entry(&obj->vma_list, typeof(*vma), vma_link);
5402 if (vma->vm != i915_obj_to_ggtt(obj))