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_get_aperture_ioctl(struct drm_device *dev, void *data,
158 struct drm_file *file)
160 struct drm_i915_private *dev_priv = dev->dev_private;
161 struct drm_i915_gem_get_aperture *args = data;
162 struct drm_i915_gem_object *obj;
166 mutex_lock(&dev->struct_mutex);
167 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list)
168 if (i915_gem_obj_is_pinned(obj))
169 pinned += i915_gem_obj_ggtt_size(obj);
170 mutex_unlock(&dev->struct_mutex);
172 args->aper_size = dev_priv->gtt.base.total;
173 args->aper_available_size = args->aper_size - pinned;
179 i915_gem_object_get_pages_phys(struct drm_i915_gem_object *obj)
181 struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
182 char *vaddr = obj->phys_handle->vaddr;
184 struct scatterlist *sg;
187 if (WARN_ON(i915_gem_object_needs_bit17_swizzle(obj)))
190 for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
194 page = shmem_read_mapping_page(mapping, i);
196 return PTR_ERR(page);
198 src = kmap_atomic(page);
199 memcpy(vaddr, src, PAGE_SIZE);
200 drm_clflush_virt_range(vaddr, PAGE_SIZE);
203 page_cache_release(page);
207 i915_gem_chipset_flush(obj->base.dev);
209 st = kmalloc(sizeof(*st), GFP_KERNEL);
213 if (sg_alloc_table(st, 1, GFP_KERNEL)) {
220 sg->length = obj->base.size;
222 sg_dma_address(sg) = obj->phys_handle->busaddr;
223 sg_dma_len(sg) = obj->base.size;
226 obj->has_dma_mapping = true;
231 i915_gem_object_put_pages_phys(struct drm_i915_gem_object *obj)
235 BUG_ON(obj->madv == __I915_MADV_PURGED);
237 ret = i915_gem_object_set_to_cpu_domain(obj, true);
239 /* In the event of a disaster, abandon all caches and
242 WARN_ON(ret != -EIO);
243 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
246 if (obj->madv == I915_MADV_DONTNEED)
250 struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
251 char *vaddr = obj->phys_handle->vaddr;
254 for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
258 page = shmem_read_mapping_page(mapping, i);
262 dst = kmap_atomic(page);
263 drm_clflush_virt_range(vaddr, PAGE_SIZE);
264 memcpy(dst, vaddr, PAGE_SIZE);
267 set_page_dirty(page);
268 if (obj->madv == I915_MADV_WILLNEED)
269 mark_page_accessed(page);
270 page_cache_release(page);
276 sg_free_table(obj->pages);
279 obj->has_dma_mapping = false;
283 i915_gem_object_release_phys(struct drm_i915_gem_object *obj)
285 drm_pci_free(obj->base.dev, obj->phys_handle);
288 static const struct drm_i915_gem_object_ops i915_gem_phys_ops = {
289 .get_pages = i915_gem_object_get_pages_phys,
290 .put_pages = i915_gem_object_put_pages_phys,
291 .release = i915_gem_object_release_phys,
295 drop_pages(struct drm_i915_gem_object *obj)
297 struct i915_vma *vma, *next;
300 drm_gem_object_reference(&obj->base);
301 list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link)
302 if (i915_vma_unbind(vma))
305 ret = i915_gem_object_put_pages(obj);
306 drm_gem_object_unreference(&obj->base);
312 i915_gem_object_attach_phys(struct drm_i915_gem_object *obj,
315 drm_dma_handle_t *phys;
318 if (obj->phys_handle) {
319 if ((unsigned long)obj->phys_handle->vaddr & (align -1))
325 if (obj->madv != I915_MADV_WILLNEED)
328 if (obj->base.filp == NULL)
331 ret = drop_pages(obj);
335 /* create a new object */
336 phys = drm_pci_alloc(obj->base.dev, obj->base.size, align);
340 obj->phys_handle = phys;
341 obj->ops = &i915_gem_phys_ops;
343 return i915_gem_object_get_pages(obj);
347 i915_gem_phys_pwrite(struct drm_i915_gem_object *obj,
348 struct drm_i915_gem_pwrite *args,
349 struct drm_file *file_priv)
351 struct drm_device *dev = obj->base.dev;
352 void *vaddr = obj->phys_handle->vaddr + args->offset;
353 char __user *user_data = to_user_ptr(args->data_ptr);
356 /* We manually control the domain here and pretend that it
357 * remains coherent i.e. in the GTT domain, like shmem_pwrite.
359 ret = i915_gem_object_wait_rendering(obj, false);
363 if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
364 unsigned long unwritten;
366 /* The physical object once assigned is fixed for the lifetime
367 * of the obj, so we can safely drop the lock and continue
370 mutex_unlock(&dev->struct_mutex);
371 unwritten = copy_from_user(vaddr, user_data, args->size);
372 mutex_lock(&dev->struct_mutex);
377 drm_clflush_virt_range(vaddr, args->size);
378 i915_gem_chipset_flush(dev);
382 void *i915_gem_object_alloc(struct drm_device *dev)
384 struct drm_i915_private *dev_priv = dev->dev_private;
385 return kmem_cache_zalloc(dev_priv->slab, GFP_KERNEL);
388 void i915_gem_object_free(struct drm_i915_gem_object *obj)
390 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
391 kmem_cache_free(dev_priv->slab, obj);
395 i915_gem_create(struct drm_file *file,
396 struct drm_device *dev,
401 struct drm_i915_gem_object *obj;
405 size = roundup(size, PAGE_SIZE);
409 /* Allocate the new object */
410 obj = i915_gem_alloc_object(dev, size);
414 obj->base.dumb = dumb;
415 ret = drm_gem_handle_create(file, &obj->base, &handle);
416 /* drop reference from allocate - handle holds it now */
417 drm_gem_object_unreference_unlocked(&obj->base);
426 i915_gem_dumb_create(struct drm_file *file,
427 struct drm_device *dev,
428 struct drm_mode_create_dumb *args)
430 /* have to work out size/pitch and return them */
431 args->pitch = ALIGN(args->width * DIV_ROUND_UP(args->bpp, 8), 64);
432 args->size = args->pitch * args->height;
433 return i915_gem_create(file, dev,
434 args->size, true, &args->handle);
438 * Creates a new mm object and returns a handle to it.
441 i915_gem_create_ioctl(struct drm_device *dev, void *data,
442 struct drm_file *file)
444 struct drm_i915_gem_create *args = data;
446 return i915_gem_create(file, dev,
447 args->size, false, &args->handle);
451 __copy_to_user_swizzled(char __user *cpu_vaddr,
452 const char *gpu_vaddr, int gpu_offset,
455 int ret, cpu_offset = 0;
458 int cacheline_end = ALIGN(gpu_offset + 1, 64);
459 int this_length = min(cacheline_end - gpu_offset, length);
460 int swizzled_gpu_offset = gpu_offset ^ 64;
462 ret = __copy_to_user(cpu_vaddr + cpu_offset,
463 gpu_vaddr + swizzled_gpu_offset,
468 cpu_offset += this_length;
469 gpu_offset += this_length;
470 length -= this_length;
477 __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset,
478 const char __user *cpu_vaddr,
481 int ret, cpu_offset = 0;
484 int cacheline_end = ALIGN(gpu_offset + 1, 64);
485 int this_length = min(cacheline_end - gpu_offset, length);
486 int swizzled_gpu_offset = gpu_offset ^ 64;
488 ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset,
489 cpu_vaddr + cpu_offset,
494 cpu_offset += this_length;
495 gpu_offset += this_length;
496 length -= this_length;
503 * Pins the specified object's pages and synchronizes the object with
504 * GPU accesses. Sets needs_clflush to non-zero if the caller should
505 * flush the object from the CPU cache.
507 int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj,
517 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) {
518 /* If we're not in the cpu read domain, set ourself into the gtt
519 * read domain and manually flush cachelines (if required). This
520 * optimizes for the case when the gpu will dirty the data
521 * anyway again before the next pread happens. */
522 *needs_clflush = !cpu_cache_is_coherent(obj->base.dev,
524 ret = i915_gem_object_wait_rendering(obj, true);
528 i915_gem_object_retire(obj);
531 ret = i915_gem_object_get_pages(obj);
535 i915_gem_object_pin_pages(obj);
540 /* Per-page copy function for the shmem pread fastpath.
541 * Flushes invalid cachelines before reading the target if
542 * needs_clflush is set. */
544 shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length,
545 char __user *user_data,
546 bool page_do_bit17_swizzling, bool needs_clflush)
551 if (unlikely(page_do_bit17_swizzling))
554 vaddr = kmap_atomic(page);
556 drm_clflush_virt_range(vaddr + shmem_page_offset,
558 ret = __copy_to_user_inatomic(user_data,
559 vaddr + shmem_page_offset,
561 kunmap_atomic(vaddr);
563 return ret ? -EFAULT : 0;
567 shmem_clflush_swizzled_range(char *addr, unsigned long length,
570 if (unlikely(swizzled)) {
571 unsigned long start = (unsigned long) addr;
572 unsigned long end = (unsigned long) addr + length;
574 /* For swizzling simply ensure that we always flush both
575 * channels. Lame, but simple and it works. Swizzled
576 * pwrite/pread is far from a hotpath - current userspace
577 * doesn't use it at all. */
578 start = round_down(start, 128);
579 end = round_up(end, 128);
581 drm_clflush_virt_range((void *)start, end - start);
583 drm_clflush_virt_range(addr, length);
588 /* Only difference to the fast-path function is that this can handle bit17
589 * and uses non-atomic copy and kmap functions. */
591 shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length,
592 char __user *user_data,
593 bool page_do_bit17_swizzling, bool needs_clflush)
600 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
602 page_do_bit17_swizzling);
604 if (page_do_bit17_swizzling)
605 ret = __copy_to_user_swizzled(user_data,
606 vaddr, shmem_page_offset,
609 ret = __copy_to_user(user_data,
610 vaddr + shmem_page_offset,
614 return ret ? - EFAULT : 0;
618 i915_gem_shmem_pread(struct drm_device *dev,
619 struct drm_i915_gem_object *obj,
620 struct drm_i915_gem_pread *args,
621 struct drm_file *file)
623 char __user *user_data;
626 int shmem_page_offset, page_length, ret = 0;
627 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
629 int needs_clflush = 0;
630 struct sg_page_iter sg_iter;
632 user_data = to_user_ptr(args->data_ptr);
635 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
637 ret = i915_gem_obj_prepare_shmem_read(obj, &needs_clflush);
641 offset = args->offset;
643 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
644 offset >> PAGE_SHIFT) {
645 struct page *page = sg_page_iter_page(&sg_iter);
650 /* Operation in this page
652 * shmem_page_offset = offset within page in shmem file
653 * page_length = bytes to copy for this page
655 shmem_page_offset = offset_in_page(offset);
656 page_length = remain;
657 if ((shmem_page_offset + page_length) > PAGE_SIZE)
658 page_length = PAGE_SIZE - shmem_page_offset;
660 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
661 (page_to_phys(page) & (1 << 17)) != 0;
663 ret = shmem_pread_fast(page, shmem_page_offset, page_length,
664 user_data, page_do_bit17_swizzling,
669 mutex_unlock(&dev->struct_mutex);
671 if (likely(!i915.prefault_disable) && !prefaulted) {
672 ret = fault_in_multipages_writeable(user_data, remain);
673 /* Userspace is tricking us, but we've already clobbered
674 * its pages with the prefault and promised to write the
675 * data up to the first fault. Hence ignore any errors
676 * and just continue. */
681 ret = shmem_pread_slow(page, shmem_page_offset, page_length,
682 user_data, page_do_bit17_swizzling,
685 mutex_lock(&dev->struct_mutex);
691 remain -= page_length;
692 user_data += page_length;
693 offset += page_length;
697 i915_gem_object_unpin_pages(obj);
703 * Reads data from the object referenced by handle.
705 * On error, the contents of *data are undefined.
708 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
709 struct drm_file *file)
711 struct drm_i915_gem_pread *args = data;
712 struct drm_i915_gem_object *obj;
718 if (!access_ok(VERIFY_WRITE,
719 to_user_ptr(args->data_ptr),
723 ret = i915_mutex_lock_interruptible(dev);
727 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
728 if (&obj->base == NULL) {
733 /* Bounds check source. */
734 if (args->offset > obj->base.size ||
735 args->size > obj->base.size - args->offset) {
740 /* prime objects have no backing filp to GEM pread/pwrite
743 if (!obj->base.filp) {
748 trace_i915_gem_object_pread(obj, args->offset, args->size);
750 ret = i915_gem_shmem_pread(dev, obj, args, file);
753 drm_gem_object_unreference(&obj->base);
755 mutex_unlock(&dev->struct_mutex);
759 /* This is the fast write path which cannot handle
760 * page faults in the source data
764 fast_user_write(struct io_mapping *mapping,
765 loff_t page_base, int page_offset,
766 char __user *user_data,
769 void __iomem *vaddr_atomic;
771 unsigned long unwritten;
773 vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
774 /* We can use the cpu mem copy function because this is X86. */
775 vaddr = (void __force*)vaddr_atomic + page_offset;
776 unwritten = __copy_from_user_inatomic_nocache(vaddr,
778 io_mapping_unmap_atomic(vaddr_atomic);
783 * This is the fast pwrite path, where we copy the data directly from the
784 * user into the GTT, uncached.
787 i915_gem_gtt_pwrite_fast(struct drm_device *dev,
788 struct drm_i915_gem_object *obj,
789 struct drm_i915_gem_pwrite *args,
790 struct drm_file *file)
792 struct drm_i915_private *dev_priv = dev->dev_private;
794 loff_t offset, page_base;
795 char __user *user_data;
796 int page_offset, page_length, ret;
798 ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE | PIN_NONBLOCK);
802 ret = i915_gem_object_set_to_gtt_domain(obj, true);
806 ret = i915_gem_object_put_fence(obj);
810 user_data = to_user_ptr(args->data_ptr);
813 offset = i915_gem_obj_ggtt_offset(obj) + args->offset;
816 /* Operation in this page
818 * page_base = page offset within aperture
819 * page_offset = offset within page
820 * page_length = bytes to copy for this page
822 page_base = offset & PAGE_MASK;
823 page_offset = offset_in_page(offset);
824 page_length = remain;
825 if ((page_offset + remain) > PAGE_SIZE)
826 page_length = PAGE_SIZE - page_offset;
828 /* If we get a fault while copying data, then (presumably) our
829 * source page isn't available. Return the error and we'll
830 * retry in the slow path.
832 if (fast_user_write(dev_priv->gtt.mappable, page_base,
833 page_offset, user_data, page_length)) {
838 remain -= page_length;
839 user_data += page_length;
840 offset += page_length;
844 i915_gem_object_ggtt_unpin(obj);
849 /* Per-page copy function for the shmem pwrite fastpath.
850 * Flushes invalid cachelines before writing to the target if
851 * needs_clflush_before is set and flushes out any written cachelines after
852 * writing if needs_clflush is set. */
854 shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
855 char __user *user_data,
856 bool page_do_bit17_swizzling,
857 bool needs_clflush_before,
858 bool needs_clflush_after)
863 if (unlikely(page_do_bit17_swizzling))
866 vaddr = kmap_atomic(page);
867 if (needs_clflush_before)
868 drm_clflush_virt_range(vaddr + shmem_page_offset,
870 ret = __copy_from_user_inatomic(vaddr + shmem_page_offset,
871 user_data, page_length);
872 if (needs_clflush_after)
873 drm_clflush_virt_range(vaddr + shmem_page_offset,
875 kunmap_atomic(vaddr);
877 return ret ? -EFAULT : 0;
880 /* Only difference to the fast-path function is that this can handle bit17
881 * and uses non-atomic copy and kmap functions. */
883 shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
884 char __user *user_data,
885 bool page_do_bit17_swizzling,
886 bool needs_clflush_before,
887 bool needs_clflush_after)
893 if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
894 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
896 page_do_bit17_swizzling);
897 if (page_do_bit17_swizzling)
898 ret = __copy_from_user_swizzled(vaddr, shmem_page_offset,
902 ret = __copy_from_user(vaddr + shmem_page_offset,
905 if (needs_clflush_after)
906 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
908 page_do_bit17_swizzling);
911 return ret ? -EFAULT : 0;
915 i915_gem_shmem_pwrite(struct drm_device *dev,
916 struct drm_i915_gem_object *obj,
917 struct drm_i915_gem_pwrite *args,
918 struct drm_file *file)
922 char __user *user_data;
923 int shmem_page_offset, page_length, ret = 0;
924 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
925 int hit_slowpath = 0;
926 int needs_clflush_after = 0;
927 int needs_clflush_before = 0;
928 struct sg_page_iter sg_iter;
930 user_data = to_user_ptr(args->data_ptr);
933 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
935 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
936 /* If we're not in the cpu write domain, set ourself into the gtt
937 * write domain and manually flush cachelines (if required). This
938 * optimizes for the case when the gpu will use the data
939 * right away and we therefore have to clflush anyway. */
940 needs_clflush_after = cpu_write_needs_clflush(obj);
941 ret = i915_gem_object_wait_rendering(obj, false);
945 i915_gem_object_retire(obj);
947 /* Same trick applies to invalidate partially written cachelines read
949 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0)
950 needs_clflush_before =
951 !cpu_cache_is_coherent(dev, obj->cache_level);
953 ret = i915_gem_object_get_pages(obj);
957 i915_gem_object_pin_pages(obj);
959 offset = args->offset;
962 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
963 offset >> PAGE_SHIFT) {
964 struct page *page = sg_page_iter_page(&sg_iter);
965 int partial_cacheline_write;
970 /* Operation in this page
972 * shmem_page_offset = offset within page in shmem file
973 * page_length = bytes to copy for this page
975 shmem_page_offset = offset_in_page(offset);
977 page_length = remain;
978 if ((shmem_page_offset + page_length) > PAGE_SIZE)
979 page_length = PAGE_SIZE - shmem_page_offset;
981 /* If we don't overwrite a cacheline completely we need to be
982 * careful to have up-to-date data by first clflushing. Don't
983 * overcomplicate things and flush the entire patch. */
984 partial_cacheline_write = needs_clflush_before &&
985 ((shmem_page_offset | page_length)
986 & (boot_cpu_data.x86_clflush_size - 1));
988 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
989 (page_to_phys(page) & (1 << 17)) != 0;
991 ret = shmem_pwrite_fast(page, shmem_page_offset, page_length,
992 user_data, page_do_bit17_swizzling,
993 partial_cacheline_write,
994 needs_clflush_after);
999 mutex_unlock(&dev->struct_mutex);
1000 ret = shmem_pwrite_slow(page, shmem_page_offset, page_length,
1001 user_data, page_do_bit17_swizzling,
1002 partial_cacheline_write,
1003 needs_clflush_after);
1005 mutex_lock(&dev->struct_mutex);
1011 remain -= page_length;
1012 user_data += page_length;
1013 offset += page_length;
1017 i915_gem_object_unpin_pages(obj);
1021 * Fixup: Flush cpu caches in case we didn't flush the dirty
1022 * cachelines in-line while writing and the object moved
1023 * out of the cpu write domain while we've dropped the lock.
1025 if (!needs_clflush_after &&
1026 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
1027 if (i915_gem_clflush_object(obj, obj->pin_display))
1028 i915_gem_chipset_flush(dev);
1032 if (needs_clflush_after)
1033 i915_gem_chipset_flush(dev);
1039 * Writes data to the object referenced by handle.
1041 * On error, the contents of the buffer that were to be modified are undefined.
1044 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
1045 struct drm_file *file)
1047 struct drm_i915_gem_pwrite *args = data;
1048 struct drm_i915_gem_object *obj;
1051 if (args->size == 0)
1054 if (!access_ok(VERIFY_READ,
1055 to_user_ptr(args->data_ptr),
1059 if (likely(!i915.prefault_disable)) {
1060 ret = fault_in_multipages_readable(to_user_ptr(args->data_ptr),
1066 ret = i915_mutex_lock_interruptible(dev);
1070 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1071 if (&obj->base == NULL) {
1076 /* Bounds check destination. */
1077 if (args->offset > obj->base.size ||
1078 args->size > obj->base.size - args->offset) {
1083 /* prime objects have no backing filp to GEM pread/pwrite
1086 if (!obj->base.filp) {
1091 trace_i915_gem_object_pwrite(obj, args->offset, args->size);
1094 /* We can only do the GTT pwrite on untiled buffers, as otherwise
1095 * it would end up going through the fenced access, and we'll get
1096 * different detiling behavior between reading and writing.
1097 * pread/pwrite currently are reading and writing from the CPU
1098 * perspective, requiring manual detiling by the client.
1100 if (obj->tiling_mode == I915_TILING_NONE &&
1101 obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
1102 cpu_write_needs_clflush(obj)) {
1103 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
1104 /* Note that the gtt paths might fail with non-page-backed user
1105 * pointers (e.g. gtt mappings when moving data between
1106 * textures). Fallback to the shmem path in that case. */
1109 if (ret == -EFAULT || ret == -ENOSPC) {
1110 if (obj->phys_handle)
1111 ret = i915_gem_phys_pwrite(obj, args, file);
1113 ret = i915_gem_shmem_pwrite(dev, obj, args, file);
1117 drm_gem_object_unreference(&obj->base);
1119 mutex_unlock(&dev->struct_mutex);
1124 i915_gem_check_wedge(struct i915_gpu_error *error,
1127 if (i915_reset_in_progress(error)) {
1128 /* Non-interruptible callers can't handle -EAGAIN, hence return
1129 * -EIO unconditionally for these. */
1133 /* Recovery complete, but the reset failed ... */
1134 if (i915_terminally_wedged(error))
1138 * Check if GPU Reset is in progress - we need intel_ring_begin
1139 * to work properly to reinit the hw state while the gpu is
1140 * still marked as reset-in-progress. Handle this with a flag.
1142 if (!error->reload_in_reset)
1150 * Compare arbitrary request against outstanding lazy request. Emit on match.
1153 i915_gem_check_olr(struct drm_i915_gem_request *req)
1157 WARN_ON(!mutex_is_locked(&req->ring->dev->struct_mutex));
1160 if (req == req->ring->outstanding_lazy_request)
1161 ret = i915_add_request(req->ring);
1166 static void fake_irq(unsigned long data)
1168 wake_up_process((struct task_struct *)data);
1171 static bool missed_irq(struct drm_i915_private *dev_priv,
1172 struct intel_engine_cs *ring)
1174 return test_bit(ring->id, &dev_priv->gpu_error.missed_irq_rings);
1177 static bool can_wait_boost(struct drm_i915_file_private *file_priv)
1179 if (file_priv == NULL)
1182 return !atomic_xchg(&file_priv->rps_wait_boost, true);
1186 * __i915_wait_request - wait until execution of request has finished
1188 * @reset_counter: reset sequence associated with the given request
1189 * @interruptible: do an interruptible wait (normally yes)
1190 * @timeout: in - how long to wait (NULL forever); out - how much time remaining
1192 * Note: It is of utmost importance that the passed in seqno and reset_counter
1193 * values have been read by the caller in an smp safe manner. Where read-side
1194 * locks are involved, it is sufficient to read the reset_counter before
1195 * unlocking the lock that protects the seqno. For lockless tricks, the
1196 * reset_counter _must_ be read before, and an appropriate smp_rmb must be
1199 * Returns 0 if the request was found within the alloted time. Else returns the
1200 * errno with remaining time filled in timeout argument.
1202 int __i915_wait_request(struct drm_i915_gem_request *req,
1203 unsigned reset_counter,
1206 struct drm_i915_file_private *file_priv)
1208 struct intel_engine_cs *ring = i915_gem_request_get_ring(req);
1209 struct drm_device *dev = ring->dev;
1210 struct drm_i915_private *dev_priv = dev->dev_private;
1211 const bool irq_test_in_progress =
1212 ACCESS_ONCE(dev_priv->gpu_error.test_irq_rings) & intel_ring_flag(ring);
1214 unsigned long timeout_expire;
1218 WARN(!intel_irqs_enabled(dev_priv), "IRQs disabled");
1220 if (i915_gem_request_completed(req, true))
1223 timeout_expire = timeout ? jiffies + nsecs_to_jiffies((u64)*timeout) : 0;
1225 if (INTEL_INFO(dev)->gen >= 6 && ring->id == RCS && can_wait_boost(file_priv)) {
1226 gen6_rps_boost(dev_priv);
1228 mod_delayed_work(dev_priv->wq,
1229 &file_priv->mm.idle_work,
1230 msecs_to_jiffies(100));
1233 if (!irq_test_in_progress && WARN_ON(!ring->irq_get(ring)))
1236 /* Record current time in case interrupted by signal, or wedged */
1237 trace_i915_gem_request_wait_begin(req);
1238 before = ktime_get_raw_ns();
1240 struct timer_list timer;
1242 prepare_to_wait(&ring->irq_queue, &wait,
1243 interruptible ? TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE);
1245 /* We need to check whether any gpu reset happened in between
1246 * the caller grabbing the seqno and now ... */
1247 if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) {
1248 /* ... but upgrade the -EAGAIN to an -EIO if the gpu
1249 * is truely gone. */
1250 ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1256 if (i915_gem_request_completed(req, false)) {
1261 if (interruptible && signal_pending(current)) {
1266 if (timeout && time_after_eq(jiffies, timeout_expire)) {
1271 timer.function = NULL;
1272 if (timeout || missed_irq(dev_priv, ring)) {
1273 unsigned long expire;
1275 setup_timer_on_stack(&timer, fake_irq, (unsigned long)current);
1276 expire = missed_irq(dev_priv, ring) ? jiffies + 1 : timeout_expire;
1277 mod_timer(&timer, expire);
1282 if (timer.function) {
1283 del_singleshot_timer_sync(&timer);
1284 destroy_timer_on_stack(&timer);
1287 now = ktime_get_raw_ns();
1288 trace_i915_gem_request_wait_end(req);
1290 if (!irq_test_in_progress)
1291 ring->irq_put(ring);
1293 finish_wait(&ring->irq_queue, &wait);
1296 s64 tres = *timeout - (now - before);
1298 *timeout = tres < 0 ? 0 : tres;
1305 * Waits for a request to be signaled, and cleans up the
1306 * request and object lists appropriately for that event.
1309 i915_wait_request(struct drm_i915_gem_request *req)
1311 struct drm_device *dev;
1312 struct drm_i915_private *dev_priv;
1314 unsigned reset_counter;
1317 BUG_ON(req == NULL);
1319 dev = req->ring->dev;
1320 dev_priv = dev->dev_private;
1321 interruptible = dev_priv->mm.interruptible;
1323 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1325 ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1329 ret = i915_gem_check_olr(req);
1333 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
1334 i915_gem_request_reference(req);
1335 ret = __i915_wait_request(req, reset_counter,
1336 interruptible, NULL, NULL);
1337 i915_gem_request_unreference(req);
1342 i915_gem_object_wait_rendering__tail(struct drm_i915_gem_object *obj)
1347 /* Manually manage the write flush as we may have not yet
1348 * retired the buffer.
1350 * Note that the last_write_req is always the earlier of
1351 * the two (read/write) requests, so if we haved successfully waited,
1352 * we know we have passed the last write.
1354 i915_gem_request_assign(&obj->last_write_req, NULL);
1360 * Ensures that all rendering to the object has completed and the object is
1361 * safe to unbind from the GTT or access from the CPU.
1363 static __must_check int
1364 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
1367 struct drm_i915_gem_request *req;
1370 req = readonly ? obj->last_write_req : obj->last_read_req;
1374 ret = i915_wait_request(req);
1378 return i915_gem_object_wait_rendering__tail(obj);
1381 /* A nonblocking variant of the above wait. This is a highly dangerous routine
1382 * as the object state may change during this call.
1384 static __must_check int
1385 i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
1386 struct drm_i915_file_private *file_priv,
1389 struct drm_i915_gem_request *req;
1390 struct drm_device *dev = obj->base.dev;
1391 struct drm_i915_private *dev_priv = dev->dev_private;
1392 unsigned reset_counter;
1395 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1396 BUG_ON(!dev_priv->mm.interruptible);
1398 req = readonly ? obj->last_write_req : obj->last_read_req;
1402 ret = i915_gem_check_wedge(&dev_priv->gpu_error, true);
1406 ret = i915_gem_check_olr(req);
1410 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
1411 i915_gem_request_reference(req);
1412 mutex_unlock(&dev->struct_mutex);
1413 ret = __i915_wait_request(req, reset_counter, true, NULL, file_priv);
1414 mutex_lock(&dev->struct_mutex);
1415 i915_gem_request_unreference(req);
1419 return i915_gem_object_wait_rendering__tail(obj);
1423 * Called when user space prepares to use an object with the CPU, either
1424 * through the mmap ioctl's mapping or a GTT mapping.
1427 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1428 struct drm_file *file)
1430 struct drm_i915_gem_set_domain *args = data;
1431 struct drm_i915_gem_object *obj;
1432 uint32_t read_domains = args->read_domains;
1433 uint32_t write_domain = args->write_domain;
1436 /* Only handle setting domains to types used by the CPU. */
1437 if (write_domain & I915_GEM_GPU_DOMAINS)
1440 if (read_domains & I915_GEM_GPU_DOMAINS)
1443 /* Having something in the write domain implies it's in the read
1444 * domain, and only that read domain. Enforce that in the request.
1446 if (write_domain != 0 && read_domains != write_domain)
1449 ret = i915_mutex_lock_interruptible(dev);
1453 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1454 if (&obj->base == NULL) {
1459 /* Try to flush the object off the GPU without holding the lock.
1460 * We will repeat the flush holding the lock in the normal manner
1461 * to catch cases where we are gazumped.
1463 ret = i915_gem_object_wait_rendering__nonblocking(obj,
1469 if (read_domains & I915_GEM_DOMAIN_GTT)
1470 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1472 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1475 drm_gem_object_unreference(&obj->base);
1477 mutex_unlock(&dev->struct_mutex);
1482 * Called when user space has done writes to this buffer
1485 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1486 struct drm_file *file)
1488 struct drm_i915_gem_sw_finish *args = data;
1489 struct drm_i915_gem_object *obj;
1492 ret = i915_mutex_lock_interruptible(dev);
1496 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1497 if (&obj->base == NULL) {
1502 /* Pinned buffers may be scanout, so flush the cache */
1503 if (obj->pin_display)
1504 i915_gem_object_flush_cpu_write_domain(obj, true);
1506 drm_gem_object_unreference(&obj->base);
1508 mutex_unlock(&dev->struct_mutex);
1513 * Maps the contents of an object, returning the address it is mapped
1516 * While the mapping holds a reference on the contents of the object, it doesn't
1517 * imply a ref on the object itself.
1521 * DRM driver writers who look a this function as an example for how to do GEM
1522 * mmap support, please don't implement mmap support like here. The modern way
1523 * to implement DRM mmap support is with an mmap offset ioctl (like
1524 * i915_gem_mmap_gtt) and then using the mmap syscall on the DRM fd directly.
1525 * That way debug tooling like valgrind will understand what's going on, hiding
1526 * the mmap call in a driver private ioctl will break that. The i915 driver only
1527 * does cpu mmaps this way because we didn't know better.
1530 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1531 struct drm_file *file)
1533 struct drm_i915_gem_mmap *args = data;
1534 struct drm_gem_object *obj;
1537 obj = drm_gem_object_lookup(dev, file, args->handle);
1541 /* prime objects have no backing filp to GEM mmap
1545 drm_gem_object_unreference_unlocked(obj);
1549 addr = vm_mmap(obj->filp, 0, args->size,
1550 PROT_READ | PROT_WRITE, MAP_SHARED,
1552 drm_gem_object_unreference_unlocked(obj);
1553 if (IS_ERR((void *)addr))
1556 args->addr_ptr = (uint64_t) addr;
1562 * i915_gem_fault - fault a page into the GTT
1563 * vma: VMA in question
1566 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1567 * from userspace. The fault handler takes care of binding the object to
1568 * the GTT (if needed), allocating and programming a fence register (again,
1569 * only if needed based on whether the old reg is still valid or the object
1570 * is tiled) and inserting a new PTE into the faulting process.
1572 * Note that the faulting process may involve evicting existing objects
1573 * from the GTT and/or fence registers to make room. So performance may
1574 * suffer if the GTT working set is large or there are few fence registers
1577 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1579 struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1580 struct drm_device *dev = obj->base.dev;
1581 struct drm_i915_private *dev_priv = dev->dev_private;
1582 pgoff_t page_offset;
1585 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1587 intel_runtime_pm_get(dev_priv);
1589 /* We don't use vmf->pgoff since that has the fake offset */
1590 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1593 ret = i915_mutex_lock_interruptible(dev);
1597 trace_i915_gem_object_fault(obj, page_offset, true, write);
1599 /* Try to flush the object off the GPU first without holding the lock.
1600 * Upon reacquiring the lock, we will perform our sanity checks and then
1601 * repeat the flush holding the lock in the normal manner to catch cases
1602 * where we are gazumped.
1604 ret = i915_gem_object_wait_rendering__nonblocking(obj, NULL, !write);
1608 /* Access to snoopable pages through the GTT is incoherent. */
1609 if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) {
1614 /* Now bind it into the GTT if needed */
1615 ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE);
1619 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1623 ret = i915_gem_object_get_fence(obj);
1627 /* Finally, remap it using the new GTT offset */
1628 pfn = dev_priv->gtt.mappable_base + i915_gem_obj_ggtt_offset(obj);
1631 if (!obj->fault_mappable) {
1632 unsigned long size = min_t(unsigned long,
1633 vma->vm_end - vma->vm_start,
1637 for (i = 0; i < size >> PAGE_SHIFT; i++) {
1638 ret = vm_insert_pfn(vma,
1639 (unsigned long)vma->vm_start + i * PAGE_SIZE,
1645 obj->fault_mappable = true;
1647 ret = vm_insert_pfn(vma,
1648 (unsigned long)vmf->virtual_address,
1651 i915_gem_object_ggtt_unpin(obj);
1653 mutex_unlock(&dev->struct_mutex);
1658 * We eat errors when the gpu is terminally wedged to avoid
1659 * userspace unduly crashing (gl has no provisions for mmaps to
1660 * fail). But any other -EIO isn't ours (e.g. swap in failure)
1661 * and so needs to be reported.
1663 if (!i915_terminally_wedged(&dev_priv->gpu_error)) {
1664 ret = VM_FAULT_SIGBUS;
1669 * EAGAIN means the gpu is hung and we'll wait for the error
1670 * handler to reset everything when re-faulting in
1671 * i915_mutex_lock_interruptible.
1678 * EBUSY is ok: this just means that another thread
1679 * already did the job.
1681 ret = VM_FAULT_NOPAGE;
1688 ret = VM_FAULT_SIGBUS;
1691 WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret);
1692 ret = VM_FAULT_SIGBUS;
1696 intel_runtime_pm_put(dev_priv);
1701 * i915_gem_release_mmap - remove physical page mappings
1702 * @obj: obj in question
1704 * Preserve the reservation of the mmapping with the DRM core code, but
1705 * relinquish ownership of the pages back to the system.
1707 * It is vital that we remove the page mapping if we have mapped a tiled
1708 * object through the GTT and then lose the fence register due to
1709 * resource pressure. Similarly if the object has been moved out of the
1710 * aperture, than pages mapped into userspace must be revoked. Removing the
1711 * mapping will then trigger a page fault on the next user access, allowing
1712 * fixup by i915_gem_fault().
1715 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1717 if (!obj->fault_mappable)
1720 drm_vma_node_unmap(&obj->base.vma_node,
1721 obj->base.dev->anon_inode->i_mapping);
1722 obj->fault_mappable = false;
1726 i915_gem_release_all_mmaps(struct drm_i915_private *dev_priv)
1728 struct drm_i915_gem_object *obj;
1730 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list)
1731 i915_gem_release_mmap(obj);
1735 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1739 if (INTEL_INFO(dev)->gen >= 4 ||
1740 tiling_mode == I915_TILING_NONE)
1743 /* Previous chips need a power-of-two fence region when tiling */
1744 if (INTEL_INFO(dev)->gen == 3)
1745 gtt_size = 1024*1024;
1747 gtt_size = 512*1024;
1749 while (gtt_size < size)
1756 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1757 * @obj: object to check
1759 * Return the required GTT alignment for an object, taking into account
1760 * potential fence register mapping.
1763 i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size,
1764 int tiling_mode, bool fenced)
1767 * Minimum alignment is 4k (GTT page size), but might be greater
1768 * if a fence register is needed for the object.
1770 if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) ||
1771 tiling_mode == I915_TILING_NONE)
1775 * Previous chips need to be aligned to the size of the smallest
1776 * fence register that can contain the object.
1778 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1781 static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj)
1783 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1786 if (drm_vma_node_has_offset(&obj->base.vma_node))
1789 dev_priv->mm.shrinker_no_lock_stealing = true;
1791 ret = drm_gem_create_mmap_offset(&obj->base);
1795 /* Badly fragmented mmap space? The only way we can recover
1796 * space is by destroying unwanted objects. We can't randomly release
1797 * mmap_offsets as userspace expects them to be persistent for the
1798 * lifetime of the objects. The closest we can is to release the
1799 * offsets on purgeable objects by truncating it and marking it purged,
1800 * which prevents userspace from ever using that object again.
1802 i915_gem_shrink(dev_priv,
1803 obj->base.size >> PAGE_SHIFT,
1805 I915_SHRINK_UNBOUND |
1806 I915_SHRINK_PURGEABLE);
1807 ret = drm_gem_create_mmap_offset(&obj->base);
1811 i915_gem_shrink_all(dev_priv);
1812 ret = drm_gem_create_mmap_offset(&obj->base);
1814 dev_priv->mm.shrinker_no_lock_stealing = false;
1819 static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)
1821 drm_gem_free_mmap_offset(&obj->base);
1825 i915_gem_mmap_gtt(struct drm_file *file,
1826 struct drm_device *dev,
1827 uint32_t handle, bool dumb,
1830 struct drm_i915_private *dev_priv = dev->dev_private;
1831 struct drm_i915_gem_object *obj;
1834 ret = i915_mutex_lock_interruptible(dev);
1838 obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
1839 if (&obj->base == NULL) {
1845 * We don't allow dumb mmaps on objects created using another
1848 WARN_ONCE(dumb && !(obj->base.dumb || obj->base.import_attach),
1849 "Illegal dumb map of accelerated buffer.\n");
1851 if (obj->base.size > dev_priv->gtt.mappable_end) {
1856 if (obj->madv != I915_MADV_WILLNEED) {
1857 DRM_DEBUG("Attempting to mmap a purgeable buffer\n");
1862 ret = i915_gem_object_create_mmap_offset(obj);
1866 *offset = drm_vma_node_offset_addr(&obj->base.vma_node);
1869 drm_gem_object_unreference(&obj->base);
1871 mutex_unlock(&dev->struct_mutex);
1876 i915_gem_dumb_map_offset(struct drm_file *file,
1877 struct drm_device *dev,
1881 return i915_gem_mmap_gtt(file, dev, handle, true, offset);
1885 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1887 * @data: GTT mapping ioctl data
1888 * @file: GEM object info
1890 * Simply returns the fake offset to userspace so it can mmap it.
1891 * The mmap call will end up in drm_gem_mmap(), which will set things
1892 * up so we can get faults in the handler above.
1894 * The fault handler will take care of binding the object into the GTT
1895 * (since it may have been evicted to make room for something), allocating
1896 * a fence register, and mapping the appropriate aperture address into
1900 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1901 struct drm_file *file)
1903 struct drm_i915_gem_mmap_gtt *args = data;
1905 return i915_gem_mmap_gtt(file, dev, args->handle, false, &args->offset);
1909 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj)
1911 return obj->madv == I915_MADV_DONTNEED;
1914 /* Immediately discard the backing storage */
1916 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
1918 i915_gem_object_free_mmap_offset(obj);
1920 if (obj->base.filp == NULL)
1923 /* Our goal here is to return as much of the memory as
1924 * is possible back to the system as we are called from OOM.
1925 * To do this we must instruct the shmfs to drop all of its
1926 * backing pages, *now*.
1928 shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1);
1929 obj->madv = __I915_MADV_PURGED;
1932 /* Try to discard unwanted pages */
1934 i915_gem_object_invalidate(struct drm_i915_gem_object *obj)
1936 struct address_space *mapping;
1938 switch (obj->madv) {
1939 case I915_MADV_DONTNEED:
1940 i915_gem_object_truncate(obj);
1941 case __I915_MADV_PURGED:
1945 if (obj->base.filp == NULL)
1948 mapping = file_inode(obj->base.filp)->i_mapping,
1949 invalidate_mapping_pages(mapping, 0, (loff_t)-1);
1953 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
1955 struct sg_page_iter sg_iter;
1958 BUG_ON(obj->madv == __I915_MADV_PURGED);
1960 ret = i915_gem_object_set_to_cpu_domain(obj, true);
1962 /* In the event of a disaster, abandon all caches and
1963 * hope for the best.
1965 WARN_ON(ret != -EIO);
1966 i915_gem_clflush_object(obj, true);
1967 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
1970 if (i915_gem_object_needs_bit17_swizzle(obj))
1971 i915_gem_object_save_bit_17_swizzle(obj);
1973 if (obj->madv == I915_MADV_DONTNEED)
1976 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
1977 struct page *page = sg_page_iter_page(&sg_iter);
1980 set_page_dirty(page);
1982 if (obj->madv == I915_MADV_WILLNEED)
1983 mark_page_accessed(page);
1985 page_cache_release(page);
1989 sg_free_table(obj->pages);
1994 i915_gem_object_put_pages(struct drm_i915_gem_object *obj)
1996 const struct drm_i915_gem_object_ops *ops = obj->ops;
1998 if (obj->pages == NULL)
2001 if (obj->pages_pin_count)
2004 BUG_ON(i915_gem_obj_bound_any(obj));
2006 /* ->put_pages might need to allocate memory for the bit17 swizzle
2007 * array, hence protect them from being reaped by removing them from gtt
2009 list_del(&obj->global_list);
2011 ops->put_pages(obj);
2014 i915_gem_object_invalidate(obj);
2020 i915_gem_shrink(struct drm_i915_private *dev_priv,
2021 long target, unsigned flags)
2024 struct list_head *list;
2027 { &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND },
2028 { &dev_priv->mm.bound_list, I915_SHRINK_BOUND },
2031 unsigned long count = 0;
2034 * As we may completely rewrite the (un)bound list whilst unbinding
2035 * (due to retiring requests) we have to strictly process only
2036 * one element of the list at the time, and recheck the list
2037 * on every iteration.
2039 * In particular, we must hold a reference whilst removing the
2040 * object as we may end up waiting for and/or retiring the objects.
2041 * This might release the final reference (held by the active list)
2042 * and result in the object being freed from under us. This is
2043 * similar to the precautions the eviction code must take whilst
2046 * Also note that although these lists do not hold a reference to
2047 * the object we can safely grab one here: The final object
2048 * unreferencing and the bound_list are both protected by the
2049 * dev->struct_mutex and so we won't ever be able to observe an
2050 * object on the bound_list with a reference count equals 0.
2052 for (phase = phases; phase->list; phase++) {
2053 struct list_head still_in_list;
2055 if ((flags & phase->bit) == 0)
2058 INIT_LIST_HEAD(&still_in_list);
2059 while (count < target && !list_empty(phase->list)) {
2060 struct drm_i915_gem_object *obj;
2061 struct i915_vma *vma, *v;
2063 obj = list_first_entry(phase->list,
2064 typeof(*obj), global_list);
2065 list_move_tail(&obj->global_list, &still_in_list);
2067 if (flags & I915_SHRINK_PURGEABLE &&
2068 !i915_gem_object_is_purgeable(obj))
2071 drm_gem_object_reference(&obj->base);
2073 /* For the unbound phase, this should be a no-op! */
2074 list_for_each_entry_safe(vma, v,
2075 &obj->vma_list, vma_link)
2076 if (i915_vma_unbind(vma))
2079 if (i915_gem_object_put_pages(obj) == 0)
2080 count += obj->base.size >> PAGE_SHIFT;
2082 drm_gem_object_unreference(&obj->base);
2084 list_splice(&still_in_list, phase->list);
2090 static unsigned long
2091 i915_gem_shrink_all(struct drm_i915_private *dev_priv)
2093 i915_gem_evict_everything(dev_priv->dev);
2094 return i915_gem_shrink(dev_priv, LONG_MAX,
2095 I915_SHRINK_BOUND | I915_SHRINK_UNBOUND);
2099 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
2101 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2103 struct address_space *mapping;
2104 struct sg_table *st;
2105 struct scatterlist *sg;
2106 struct sg_page_iter sg_iter;
2108 unsigned long last_pfn = 0; /* suppress gcc warning */
2111 /* Assert that the object is not currently in any GPU domain. As it
2112 * wasn't in the GTT, there shouldn't be any way it could have been in
2115 BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
2116 BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
2118 st = kmalloc(sizeof(*st), GFP_KERNEL);
2122 page_count = obj->base.size / PAGE_SIZE;
2123 if (sg_alloc_table(st, page_count, GFP_KERNEL)) {
2128 /* Get the list of pages out of our struct file. They'll be pinned
2129 * at this point until we release them.
2131 * Fail silently without starting the shrinker
2133 mapping = file_inode(obj->base.filp)->i_mapping;
2134 gfp = mapping_gfp_mask(mapping);
2135 gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD;
2136 gfp &= ~(__GFP_IO | __GFP_WAIT);
2139 for (i = 0; i < page_count; i++) {
2140 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2142 i915_gem_shrink(dev_priv,
2145 I915_SHRINK_UNBOUND |
2146 I915_SHRINK_PURGEABLE);
2147 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2150 /* We've tried hard to allocate the memory by reaping
2151 * our own buffer, now let the real VM do its job and
2152 * go down in flames if truly OOM.
2154 i915_gem_shrink_all(dev_priv);
2155 page = shmem_read_mapping_page(mapping, i);
2159 #ifdef CONFIG_SWIOTLB
2160 if (swiotlb_nr_tbl()) {
2162 sg_set_page(sg, page, PAGE_SIZE, 0);
2167 if (!i || page_to_pfn(page) != last_pfn + 1) {
2171 sg_set_page(sg, page, PAGE_SIZE, 0);
2173 sg->length += PAGE_SIZE;
2175 last_pfn = page_to_pfn(page);
2177 /* Check that the i965g/gm workaround works. */
2178 WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
2180 #ifdef CONFIG_SWIOTLB
2181 if (!swiotlb_nr_tbl())
2186 if (i915_gem_object_needs_bit17_swizzle(obj))
2187 i915_gem_object_do_bit_17_swizzle(obj);
2189 if (obj->tiling_mode != I915_TILING_NONE &&
2190 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES)
2191 i915_gem_object_pin_pages(obj);
2197 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0)
2198 page_cache_release(sg_page_iter_page(&sg_iter));
2202 /* shmemfs first checks if there is enough memory to allocate the page
2203 * and reports ENOSPC should there be insufficient, along with the usual
2204 * ENOMEM for a genuine allocation failure.
2206 * We use ENOSPC in our driver to mean that we have run out of aperture
2207 * space and so want to translate the error from shmemfs back to our
2208 * usual understanding of ENOMEM.
2210 if (PTR_ERR(page) == -ENOSPC)
2213 return PTR_ERR(page);
2216 /* Ensure that the associated pages are gathered from the backing storage
2217 * and pinned into our object. i915_gem_object_get_pages() may be called
2218 * multiple times before they are released by a single call to
2219 * i915_gem_object_put_pages() - once the pages are no longer referenced
2220 * either as a result of memory pressure (reaping pages under the shrinker)
2221 * or as the object is itself released.
2224 i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
2226 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2227 const struct drm_i915_gem_object_ops *ops = obj->ops;
2233 if (obj->madv != I915_MADV_WILLNEED) {
2234 DRM_DEBUG("Attempting to obtain a purgeable object\n");
2238 BUG_ON(obj->pages_pin_count);
2240 ret = ops->get_pages(obj);
2244 list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list);
2249 i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
2250 struct intel_engine_cs *ring)
2252 struct drm_i915_gem_request *req;
2253 struct intel_engine_cs *old_ring;
2255 BUG_ON(ring == NULL);
2257 req = intel_ring_get_request(ring);
2258 old_ring = i915_gem_request_get_ring(obj->last_read_req);
2260 if (old_ring != ring && obj->last_write_req) {
2261 /* Keep the request relative to the current ring */
2262 i915_gem_request_assign(&obj->last_write_req, req);
2265 /* Add a reference if we're newly entering the active list. */
2267 drm_gem_object_reference(&obj->base);
2271 list_move_tail(&obj->ring_list, &ring->active_list);
2273 i915_gem_request_assign(&obj->last_read_req, req);
2276 void i915_vma_move_to_active(struct i915_vma *vma,
2277 struct intel_engine_cs *ring)
2279 list_move_tail(&vma->mm_list, &vma->vm->active_list);
2280 return i915_gem_object_move_to_active(vma->obj, ring);
2284 i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
2286 struct i915_vma *vma;
2288 BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS);
2289 BUG_ON(!obj->active);
2291 list_for_each_entry(vma, &obj->vma_list, vma_link) {
2292 if (!list_empty(&vma->mm_list))
2293 list_move_tail(&vma->mm_list, &vma->vm->inactive_list);
2296 intel_fb_obj_flush(obj, true);
2298 list_del_init(&obj->ring_list);
2300 i915_gem_request_assign(&obj->last_read_req, NULL);
2301 i915_gem_request_assign(&obj->last_write_req, NULL);
2302 obj->base.write_domain = 0;
2304 i915_gem_request_assign(&obj->last_fenced_req, NULL);
2307 drm_gem_object_unreference(&obj->base);
2309 WARN_ON(i915_verify_lists(dev));
2313 i915_gem_object_retire(struct drm_i915_gem_object *obj)
2315 if (obj->last_read_req == NULL)
2318 if (i915_gem_request_completed(obj->last_read_req, true))
2319 i915_gem_object_move_to_inactive(obj);
2323 i915_gem_init_seqno(struct drm_device *dev, u32 seqno)
2325 struct drm_i915_private *dev_priv = dev->dev_private;
2326 struct intel_engine_cs *ring;
2329 /* Carefully retire all requests without writing to the rings */
2330 for_each_ring(ring, dev_priv, i) {
2331 ret = intel_ring_idle(ring);
2335 i915_gem_retire_requests(dev);
2337 /* Finally reset hw state */
2338 for_each_ring(ring, dev_priv, i) {
2339 intel_ring_init_seqno(ring, seqno);
2341 for (j = 0; j < ARRAY_SIZE(ring->semaphore.sync_seqno); j++)
2342 ring->semaphore.sync_seqno[j] = 0;
2348 int i915_gem_set_seqno(struct drm_device *dev, u32 seqno)
2350 struct drm_i915_private *dev_priv = dev->dev_private;
2356 /* HWS page needs to be set less than what we
2357 * will inject to ring
2359 ret = i915_gem_init_seqno(dev, seqno - 1);
2363 /* Carefully set the last_seqno value so that wrap
2364 * detection still works
2366 dev_priv->next_seqno = seqno;
2367 dev_priv->last_seqno = seqno - 1;
2368 if (dev_priv->last_seqno == 0)
2369 dev_priv->last_seqno--;
2375 i915_gem_get_seqno(struct drm_device *dev, u32 *seqno)
2377 struct drm_i915_private *dev_priv = dev->dev_private;
2379 /* reserve 0 for non-seqno */
2380 if (dev_priv->next_seqno == 0) {
2381 int ret = i915_gem_init_seqno(dev, 0);
2385 dev_priv->next_seqno = 1;
2388 *seqno = dev_priv->last_seqno = dev_priv->next_seqno++;
2392 int __i915_add_request(struct intel_engine_cs *ring,
2393 struct drm_file *file,
2394 struct drm_i915_gem_object *obj)
2396 struct drm_i915_private *dev_priv = ring->dev->dev_private;
2397 struct drm_i915_gem_request *request;
2398 struct intel_ringbuffer *ringbuf;
2399 u32 request_ring_position, request_start;
2402 request = ring->outstanding_lazy_request;
2403 if (WARN_ON(request == NULL))
2406 if (i915.enable_execlists) {
2407 struct intel_context *ctx = request->ctx;
2408 ringbuf = ctx->engine[ring->id].ringbuf;
2410 ringbuf = ring->buffer;
2412 request_start = intel_ring_get_tail(ringbuf);
2414 * Emit any outstanding flushes - execbuf can fail to emit the flush
2415 * after having emitted the batchbuffer command. Hence we need to fix
2416 * things up similar to emitting the lazy request. The difference here
2417 * is that the flush _must_ happen before the next request, no matter
2420 if (i915.enable_execlists) {
2421 ret = logical_ring_flush_all_caches(ringbuf);
2425 ret = intel_ring_flush_all_caches(ring);
2430 /* Record the position of the start of the request so that
2431 * should we detect the updated seqno part-way through the
2432 * GPU processing the request, we never over-estimate the
2433 * position of the head.
2435 request_ring_position = intel_ring_get_tail(ringbuf);
2437 if (i915.enable_execlists) {
2438 ret = ring->emit_request(ringbuf);
2442 ret = ring->add_request(ring);
2447 request->head = request_start;
2448 request->tail = request_ring_position;
2450 /* Whilst this request exists, batch_obj will be on the
2451 * active_list, and so will hold the active reference. Only when this
2452 * request is retired will the the batch_obj be moved onto the
2453 * inactive_list and lose its active reference. Hence we do not need
2454 * to explicitly hold another reference here.
2456 request->batch_obj = obj;
2458 if (!i915.enable_execlists) {
2459 /* Hold a reference to the current context so that we can inspect
2460 * it later in case a hangcheck error event fires.
2462 request->ctx = ring->last_context;
2464 i915_gem_context_reference(request->ctx);
2467 request->emitted_jiffies = jiffies;
2468 list_add_tail(&request->list, &ring->request_list);
2469 request->file_priv = NULL;
2472 struct drm_i915_file_private *file_priv = file->driver_priv;
2474 spin_lock(&file_priv->mm.lock);
2475 request->file_priv = file_priv;
2476 list_add_tail(&request->client_list,
2477 &file_priv->mm.request_list);
2478 spin_unlock(&file_priv->mm.lock);
2481 trace_i915_gem_request_add(request);
2482 ring->outstanding_lazy_request = NULL;
2484 i915_queue_hangcheck(ring->dev);
2486 cancel_delayed_work_sync(&dev_priv->mm.idle_work);
2487 queue_delayed_work(dev_priv->wq,
2488 &dev_priv->mm.retire_work,
2489 round_jiffies_up_relative(HZ));
2490 intel_mark_busy(dev_priv->dev);
2496 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
2498 struct drm_i915_file_private *file_priv = request->file_priv;
2503 spin_lock(&file_priv->mm.lock);
2504 list_del(&request->client_list);
2505 request->file_priv = NULL;
2506 spin_unlock(&file_priv->mm.lock);
2509 static bool i915_context_is_banned(struct drm_i915_private *dev_priv,
2510 const struct intel_context *ctx)
2512 unsigned long elapsed;
2514 elapsed = get_seconds() - ctx->hang_stats.guilty_ts;
2516 if (ctx->hang_stats.banned)
2519 if (elapsed <= DRM_I915_CTX_BAN_PERIOD) {
2520 if (!i915_gem_context_is_default(ctx)) {
2521 DRM_DEBUG("context hanging too fast, banning!\n");
2523 } else if (i915_stop_ring_allow_ban(dev_priv)) {
2524 if (i915_stop_ring_allow_warn(dev_priv))
2525 DRM_ERROR("gpu hanging too fast, banning!\n");
2533 static void i915_set_reset_status(struct drm_i915_private *dev_priv,
2534 struct intel_context *ctx,
2537 struct i915_ctx_hang_stats *hs;
2542 hs = &ctx->hang_stats;
2545 hs->banned = i915_context_is_banned(dev_priv, ctx);
2547 hs->guilty_ts = get_seconds();
2549 hs->batch_pending++;
2553 static void i915_gem_free_request(struct drm_i915_gem_request *request)
2555 list_del(&request->list);
2556 i915_gem_request_remove_from_client(request);
2558 i915_gem_request_unreference(request);
2561 void i915_gem_request_free(struct kref *req_ref)
2563 struct drm_i915_gem_request *req = container_of(req_ref,
2565 struct intel_context *ctx = req->ctx;
2568 if (i915.enable_execlists) {
2569 struct intel_engine_cs *ring = req->ring;
2571 if (ctx != ring->default_context)
2572 intel_lr_context_unpin(ring, ctx);
2575 i915_gem_context_unreference(ctx);
2581 struct drm_i915_gem_request *
2582 i915_gem_find_active_request(struct intel_engine_cs *ring)
2584 struct drm_i915_gem_request *request;
2586 list_for_each_entry(request, &ring->request_list, list) {
2587 if (i915_gem_request_completed(request, false))
2596 static void i915_gem_reset_ring_status(struct drm_i915_private *dev_priv,
2597 struct intel_engine_cs *ring)
2599 struct drm_i915_gem_request *request;
2602 request = i915_gem_find_active_request(ring);
2604 if (request == NULL)
2607 ring_hung = ring->hangcheck.score >= HANGCHECK_SCORE_RING_HUNG;
2609 i915_set_reset_status(dev_priv, request->ctx, ring_hung);
2611 list_for_each_entry_continue(request, &ring->request_list, list)
2612 i915_set_reset_status(dev_priv, request->ctx, false);
2615 static void i915_gem_reset_ring_cleanup(struct drm_i915_private *dev_priv,
2616 struct intel_engine_cs *ring)
2618 while (!list_empty(&ring->active_list)) {
2619 struct drm_i915_gem_object *obj;
2621 obj = list_first_entry(&ring->active_list,
2622 struct drm_i915_gem_object,
2625 i915_gem_object_move_to_inactive(obj);
2629 * Clear the execlists queue up before freeing the requests, as those
2630 * are the ones that keep the context and ringbuffer backing objects
2633 while (!list_empty(&ring->execlist_queue)) {
2634 struct intel_ctx_submit_request *submit_req;
2636 submit_req = list_first_entry(&ring->execlist_queue,
2637 struct intel_ctx_submit_request,
2639 list_del(&submit_req->execlist_link);
2640 intel_runtime_pm_put(dev_priv);
2641 i915_gem_context_unreference(submit_req->ctx);
2646 * We must free the requests after all the corresponding objects have
2647 * been moved off active lists. Which is the same order as the normal
2648 * retire_requests function does. This is important if object hold
2649 * implicit references on things like e.g. ppgtt address spaces through
2652 while (!list_empty(&ring->request_list)) {
2653 struct drm_i915_gem_request *request;
2655 request = list_first_entry(&ring->request_list,
2656 struct drm_i915_gem_request,
2659 i915_gem_free_request(request);
2662 /* This may not have been flushed before the reset, so clean it now */
2663 i915_gem_request_assign(&ring->outstanding_lazy_request, NULL);
2666 void i915_gem_restore_fences(struct drm_device *dev)
2668 struct drm_i915_private *dev_priv = dev->dev_private;
2671 for (i = 0; i < dev_priv->num_fence_regs; i++) {
2672 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
2675 * Commit delayed tiling changes if we have an object still
2676 * attached to the fence, otherwise just clear the fence.
2679 i915_gem_object_update_fence(reg->obj, reg,
2680 reg->obj->tiling_mode);
2682 i915_gem_write_fence(dev, i, NULL);
2687 void i915_gem_reset(struct drm_device *dev)
2689 struct drm_i915_private *dev_priv = dev->dev_private;
2690 struct intel_engine_cs *ring;
2694 * Before we free the objects from the requests, we need to inspect
2695 * them for finding the guilty party. As the requests only borrow
2696 * their reference to the objects, the inspection must be done first.
2698 for_each_ring(ring, dev_priv, i)
2699 i915_gem_reset_ring_status(dev_priv, ring);
2701 for_each_ring(ring, dev_priv, i)
2702 i915_gem_reset_ring_cleanup(dev_priv, ring);
2704 i915_gem_context_reset(dev);
2706 i915_gem_restore_fences(dev);
2710 * This function clears the request list as sequence numbers are passed.
2713 i915_gem_retire_requests_ring(struct intel_engine_cs *ring)
2715 if (list_empty(&ring->request_list))
2718 WARN_ON(i915_verify_lists(ring->dev));
2720 /* Move any buffers on the active list that are no longer referenced
2721 * by the ringbuffer to the flushing/inactive lists as appropriate,
2722 * before we free the context associated with the requests.
2724 while (!list_empty(&ring->active_list)) {
2725 struct drm_i915_gem_object *obj;
2727 obj = list_first_entry(&ring->active_list,
2728 struct drm_i915_gem_object,
2731 if (!i915_gem_request_completed(obj->last_read_req, true))
2734 i915_gem_object_move_to_inactive(obj);
2738 while (!list_empty(&ring->request_list)) {
2739 struct drm_i915_gem_request *request;
2740 struct intel_ringbuffer *ringbuf;
2742 request = list_first_entry(&ring->request_list,
2743 struct drm_i915_gem_request,
2746 if (!i915_gem_request_completed(request, true))
2749 trace_i915_gem_request_retire(request);
2751 /* This is one of the few common intersection points
2752 * between legacy ringbuffer submission and execlists:
2753 * we need to tell them apart in order to find the correct
2754 * ringbuffer to which the request belongs to.
2756 if (i915.enable_execlists) {
2757 struct intel_context *ctx = request->ctx;
2758 ringbuf = ctx->engine[ring->id].ringbuf;
2760 ringbuf = ring->buffer;
2762 /* We know the GPU must have read the request to have
2763 * sent us the seqno + interrupt, so use the position
2764 * of tail of the request to update the last known position
2767 ringbuf->last_retired_head = request->tail;
2769 i915_gem_free_request(request);
2772 if (unlikely(ring->trace_irq_req &&
2773 i915_gem_request_completed(ring->trace_irq_req, true))) {
2774 ring->irq_put(ring);
2775 i915_gem_request_assign(&ring->trace_irq_req, NULL);
2778 WARN_ON(i915_verify_lists(ring->dev));
2782 i915_gem_retire_requests(struct drm_device *dev)
2784 struct drm_i915_private *dev_priv = dev->dev_private;
2785 struct intel_engine_cs *ring;
2789 for_each_ring(ring, dev_priv, i) {
2790 i915_gem_retire_requests_ring(ring);
2791 idle &= list_empty(&ring->request_list);
2792 if (i915.enable_execlists) {
2793 unsigned long flags;
2795 spin_lock_irqsave(&ring->execlist_lock, flags);
2796 idle &= list_empty(&ring->execlist_queue);
2797 spin_unlock_irqrestore(&ring->execlist_lock, flags);
2799 intel_execlists_retire_requests(ring);
2804 mod_delayed_work(dev_priv->wq,
2805 &dev_priv->mm.idle_work,
2806 msecs_to_jiffies(100));
2812 i915_gem_retire_work_handler(struct work_struct *work)
2814 struct drm_i915_private *dev_priv =
2815 container_of(work, typeof(*dev_priv), mm.retire_work.work);
2816 struct drm_device *dev = dev_priv->dev;
2819 /* Come back later if the device is busy... */
2821 if (mutex_trylock(&dev->struct_mutex)) {
2822 idle = i915_gem_retire_requests(dev);
2823 mutex_unlock(&dev->struct_mutex);
2826 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
2827 round_jiffies_up_relative(HZ));
2831 i915_gem_idle_work_handler(struct work_struct *work)
2833 struct drm_i915_private *dev_priv =
2834 container_of(work, typeof(*dev_priv), mm.idle_work.work);
2836 intel_mark_idle(dev_priv->dev);
2840 * Ensures that an object will eventually get non-busy by flushing any required
2841 * write domains, emitting any outstanding lazy request and retiring and
2842 * completed requests.
2845 i915_gem_object_flush_active(struct drm_i915_gem_object *obj)
2847 struct intel_engine_cs *ring;
2851 ring = i915_gem_request_get_ring(obj->last_read_req);
2853 ret = i915_gem_check_olr(obj->last_read_req);
2857 i915_gem_retire_requests_ring(ring);
2864 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
2865 * @DRM_IOCTL_ARGS: standard ioctl arguments
2867 * Returns 0 if successful, else an error is returned with the remaining time in
2868 * the timeout parameter.
2869 * -ETIME: object is still busy after timeout
2870 * -ERESTARTSYS: signal interrupted the wait
2871 * -ENONENT: object doesn't exist
2872 * Also possible, but rare:
2873 * -EAGAIN: GPU wedged
2875 * -ENODEV: Internal IRQ fail
2876 * -E?: The add request failed
2878 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
2879 * non-zero timeout parameter the wait ioctl will wait for the given number of
2880 * nanoseconds on an object becoming unbusy. Since the wait itself does so
2881 * without holding struct_mutex the object may become re-busied before this
2882 * function completes. A similar but shorter * race condition exists in the busy
2886 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
2888 struct drm_i915_private *dev_priv = dev->dev_private;
2889 struct drm_i915_gem_wait *args = data;
2890 struct drm_i915_gem_object *obj;
2891 struct drm_i915_gem_request *req;
2892 unsigned reset_counter;
2895 if (args->flags != 0)
2898 ret = i915_mutex_lock_interruptible(dev);
2902 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle));
2903 if (&obj->base == NULL) {
2904 mutex_unlock(&dev->struct_mutex);
2908 /* Need to make sure the object gets inactive eventually. */
2909 ret = i915_gem_object_flush_active(obj);
2913 if (!obj->active || !obj->last_read_req)
2916 req = obj->last_read_req;
2918 /* Do this after OLR check to make sure we make forward progress polling
2919 * on this IOCTL with a timeout <=0 (like busy ioctl)
2921 if (args->timeout_ns <= 0) {
2926 drm_gem_object_unreference(&obj->base);
2927 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
2928 i915_gem_request_reference(req);
2929 mutex_unlock(&dev->struct_mutex);
2931 ret = __i915_wait_request(req, reset_counter, true, &args->timeout_ns,
2933 mutex_lock(&dev->struct_mutex);
2934 i915_gem_request_unreference(req);
2935 mutex_unlock(&dev->struct_mutex);
2939 drm_gem_object_unreference(&obj->base);
2940 mutex_unlock(&dev->struct_mutex);
2945 * i915_gem_object_sync - sync an object to a ring.
2947 * @obj: object which may be in use on another ring.
2948 * @to: ring we wish to use the object on. May be NULL.
2950 * This code is meant to abstract object synchronization with the GPU.
2951 * Calling with NULL implies synchronizing the object with the CPU
2952 * rather than a particular GPU ring.
2954 * Returns 0 if successful, else propagates up the lower layer error.
2957 i915_gem_object_sync(struct drm_i915_gem_object *obj,
2958 struct intel_engine_cs *to)
2960 struct intel_engine_cs *from;
2964 from = i915_gem_request_get_ring(obj->last_read_req);
2966 if (from == NULL || to == from)
2969 if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev))
2970 return i915_gem_object_wait_rendering(obj, false);
2972 idx = intel_ring_sync_index(from, to);
2974 seqno = i915_gem_request_get_seqno(obj->last_read_req);
2975 /* Optimization: Avoid semaphore sync when we are sure we already
2976 * waited for an object with higher seqno */
2977 if (seqno <= from->semaphore.sync_seqno[idx])
2980 ret = i915_gem_check_olr(obj->last_read_req);
2984 trace_i915_gem_ring_sync_to(from, to, obj->last_read_req);
2985 ret = to->semaphore.sync_to(to, from, seqno);
2987 /* We use last_read_req because sync_to()
2988 * might have just caused seqno wrap under
2991 from->semaphore.sync_seqno[idx] =
2992 i915_gem_request_get_seqno(obj->last_read_req);
2997 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
2999 u32 old_write_domain, old_read_domains;
3001 /* Force a pagefault for domain tracking on next user access */
3002 i915_gem_release_mmap(obj);
3004 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
3007 /* Wait for any direct GTT access to complete */
3010 old_read_domains = obj->base.read_domains;
3011 old_write_domain = obj->base.write_domain;
3013 obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
3014 obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
3016 trace_i915_gem_object_change_domain(obj,
3021 int i915_vma_unbind(struct i915_vma *vma)
3023 struct drm_i915_gem_object *obj = vma->obj;
3024 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3027 if (list_empty(&vma->vma_link))
3030 if (!drm_mm_node_allocated(&vma->node)) {
3031 i915_gem_vma_destroy(vma);
3038 BUG_ON(obj->pages == NULL);
3040 ret = i915_gem_object_finish_gpu(obj);
3043 /* Continue on if we fail due to EIO, the GPU is hung so we
3044 * should be safe and we need to cleanup or else we might
3045 * cause memory corruption through use-after-free.
3048 if (i915_is_ggtt(vma->vm) &&
3049 vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL) {
3050 i915_gem_object_finish_gtt(obj);
3052 /* release the fence reg _after_ flushing */
3053 ret = i915_gem_object_put_fence(obj);
3058 trace_i915_vma_unbind(vma);
3060 vma->unbind_vma(vma);
3062 list_del_init(&vma->mm_list);
3063 if (i915_is_ggtt(vma->vm)) {
3064 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL) {
3065 obj->map_and_fenceable = false;
3066 } else if (vma->ggtt_view.pages) {
3067 sg_free_table(vma->ggtt_view.pages);
3068 kfree(vma->ggtt_view.pages);
3069 vma->ggtt_view.pages = NULL;
3073 drm_mm_remove_node(&vma->node);
3074 i915_gem_vma_destroy(vma);
3076 /* Since the unbound list is global, only move to that list if
3077 * no more VMAs exist. */
3078 if (list_empty(&obj->vma_list)) {
3079 /* Throw away the active reference before
3080 * moving to the unbound list. */
3081 i915_gem_object_retire(obj);
3083 i915_gem_gtt_finish_object(obj);
3084 list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list);
3087 /* And finally now the object is completely decoupled from this vma,
3088 * we can drop its hold on the backing storage and allow it to be
3089 * reaped by the shrinker.
3091 i915_gem_object_unpin_pages(obj);
3096 int i915_gpu_idle(struct drm_device *dev)
3098 struct drm_i915_private *dev_priv = dev->dev_private;
3099 struct intel_engine_cs *ring;
3102 /* Flush everything onto the inactive list. */
3103 for_each_ring(ring, dev_priv, i) {
3104 if (!i915.enable_execlists) {
3105 ret = i915_switch_context(ring, ring->default_context);
3110 ret = intel_ring_idle(ring);
3118 static void i965_write_fence_reg(struct drm_device *dev, int reg,
3119 struct drm_i915_gem_object *obj)
3121 struct drm_i915_private *dev_priv = dev->dev_private;
3123 int fence_pitch_shift;
3125 if (INTEL_INFO(dev)->gen >= 6) {
3126 fence_reg = FENCE_REG_SANDYBRIDGE_0;
3127 fence_pitch_shift = SANDYBRIDGE_FENCE_PITCH_SHIFT;
3129 fence_reg = FENCE_REG_965_0;
3130 fence_pitch_shift = I965_FENCE_PITCH_SHIFT;
3133 fence_reg += reg * 8;
3135 /* To w/a incoherency with non-atomic 64-bit register updates,
3136 * we split the 64-bit update into two 32-bit writes. In order
3137 * for a partial fence not to be evaluated between writes, we
3138 * precede the update with write to turn off the fence register,
3139 * and only enable the fence as the last step.
3141 * For extra levels of paranoia, we make sure each step lands
3142 * before applying the next step.
3144 I915_WRITE(fence_reg, 0);
3145 POSTING_READ(fence_reg);
3148 u32 size = i915_gem_obj_ggtt_size(obj);
3151 val = (uint64_t)((i915_gem_obj_ggtt_offset(obj) + size - 4096) &
3153 val |= i915_gem_obj_ggtt_offset(obj) & 0xfffff000;
3154 val |= (uint64_t)((obj->stride / 128) - 1) << fence_pitch_shift;
3155 if (obj->tiling_mode == I915_TILING_Y)
3156 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
3157 val |= I965_FENCE_REG_VALID;
3159 I915_WRITE(fence_reg + 4, val >> 32);
3160 POSTING_READ(fence_reg + 4);
3162 I915_WRITE(fence_reg + 0, val);
3163 POSTING_READ(fence_reg);
3165 I915_WRITE(fence_reg + 4, 0);
3166 POSTING_READ(fence_reg + 4);
3170 static void i915_write_fence_reg(struct drm_device *dev, int reg,
3171 struct drm_i915_gem_object *obj)
3173 struct drm_i915_private *dev_priv = dev->dev_private;
3177 u32 size = i915_gem_obj_ggtt_size(obj);
3181 WARN((i915_gem_obj_ggtt_offset(obj) & ~I915_FENCE_START_MASK) ||
3182 (size & -size) != size ||
3183 (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
3184 "object 0x%08lx [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
3185 i915_gem_obj_ggtt_offset(obj), obj->map_and_fenceable, size);
3187 if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
3192 /* Note: pitch better be a power of two tile widths */
3193 pitch_val = obj->stride / tile_width;
3194 pitch_val = ffs(pitch_val) - 1;
3196 val = i915_gem_obj_ggtt_offset(obj);
3197 if (obj->tiling_mode == I915_TILING_Y)
3198 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
3199 val |= I915_FENCE_SIZE_BITS(size);
3200 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
3201 val |= I830_FENCE_REG_VALID;
3206 reg = FENCE_REG_830_0 + reg * 4;
3208 reg = FENCE_REG_945_8 + (reg - 8) * 4;
3210 I915_WRITE(reg, val);
3214 static void i830_write_fence_reg(struct drm_device *dev, int reg,
3215 struct drm_i915_gem_object *obj)
3217 struct drm_i915_private *dev_priv = dev->dev_private;
3221 u32 size = i915_gem_obj_ggtt_size(obj);
3224 WARN((i915_gem_obj_ggtt_offset(obj) & ~I830_FENCE_START_MASK) ||
3225 (size & -size) != size ||
3226 (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
3227 "object 0x%08lx not 512K or pot-size 0x%08x aligned\n",
3228 i915_gem_obj_ggtt_offset(obj), size);
3230 pitch_val = obj->stride / 128;
3231 pitch_val = ffs(pitch_val) - 1;
3233 val = i915_gem_obj_ggtt_offset(obj);
3234 if (obj->tiling_mode == I915_TILING_Y)
3235 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
3236 val |= I830_FENCE_SIZE_BITS(size);
3237 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
3238 val |= I830_FENCE_REG_VALID;
3242 I915_WRITE(FENCE_REG_830_0 + reg * 4, val);
3243 POSTING_READ(FENCE_REG_830_0 + reg * 4);
3246 inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object *obj)
3248 return obj && obj->base.read_domains & I915_GEM_DOMAIN_GTT;
3251 static void i915_gem_write_fence(struct drm_device *dev, int reg,
3252 struct drm_i915_gem_object *obj)
3254 struct drm_i915_private *dev_priv = dev->dev_private;
3256 /* Ensure that all CPU reads are completed before installing a fence
3257 * and all writes before removing the fence.
3259 if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj))
3262 WARN(obj && (!obj->stride || !obj->tiling_mode),
3263 "bogus fence setup with stride: 0x%x, tiling mode: %i\n",
3264 obj->stride, obj->tiling_mode);
3267 i830_write_fence_reg(dev, reg, obj);
3268 else if (IS_GEN3(dev))
3269 i915_write_fence_reg(dev, reg, obj);
3270 else if (INTEL_INFO(dev)->gen >= 4)
3271 i965_write_fence_reg(dev, reg, obj);
3273 /* And similarly be paranoid that no direct access to this region
3274 * is reordered to before the fence is installed.
3276 if (i915_gem_object_needs_mb(obj))
3280 static inline int fence_number(struct drm_i915_private *dev_priv,
3281 struct drm_i915_fence_reg *fence)
3283 return fence - dev_priv->fence_regs;
3286 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
3287 struct drm_i915_fence_reg *fence,
3290 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3291 int reg = fence_number(dev_priv, fence);
3293 i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL);
3296 obj->fence_reg = reg;
3298 list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list);
3300 obj->fence_reg = I915_FENCE_REG_NONE;
3302 list_del_init(&fence->lru_list);
3304 obj->fence_dirty = false;
3308 i915_gem_object_wait_fence(struct drm_i915_gem_object *obj)
3310 if (obj->last_fenced_req) {
3311 int ret = i915_wait_request(obj->last_fenced_req);
3315 i915_gem_request_assign(&obj->last_fenced_req, NULL);
3322 i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
3324 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3325 struct drm_i915_fence_reg *fence;
3328 ret = i915_gem_object_wait_fence(obj);
3332 if (obj->fence_reg == I915_FENCE_REG_NONE)
3335 fence = &dev_priv->fence_regs[obj->fence_reg];
3337 if (WARN_ON(fence->pin_count))
3340 i915_gem_object_fence_lost(obj);
3341 i915_gem_object_update_fence(obj, fence, false);
3346 static struct drm_i915_fence_reg *
3347 i915_find_fence_reg(struct drm_device *dev)
3349 struct drm_i915_private *dev_priv = dev->dev_private;
3350 struct drm_i915_fence_reg *reg, *avail;
3353 /* First try to find a free reg */
3355 for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
3356 reg = &dev_priv->fence_regs[i];
3360 if (!reg->pin_count)
3367 /* None available, try to steal one or wait for a user to finish */
3368 list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
3376 /* Wait for completion of pending flips which consume fences */
3377 if (intel_has_pending_fb_unpin(dev))
3378 return ERR_PTR(-EAGAIN);
3380 return ERR_PTR(-EDEADLK);
3384 * i915_gem_object_get_fence - set up fencing for an object
3385 * @obj: object to map through a fence reg
3387 * When mapping objects through the GTT, userspace wants to be able to write
3388 * to them without having to worry about swizzling if the object is tiled.
3389 * This function walks the fence regs looking for a free one for @obj,
3390 * stealing one if it can't find any.
3392 * It then sets up the reg based on the object's properties: address, pitch
3393 * and tiling format.
3395 * For an untiled surface, this removes any existing fence.
3398 i915_gem_object_get_fence(struct drm_i915_gem_object *obj)
3400 struct drm_device *dev = obj->base.dev;
3401 struct drm_i915_private *dev_priv = dev->dev_private;
3402 bool enable = obj->tiling_mode != I915_TILING_NONE;
3403 struct drm_i915_fence_reg *reg;
3406 /* Have we updated the tiling parameters upon the object and so
3407 * will need to serialise the write to the associated fence register?
3409 if (obj->fence_dirty) {
3410 ret = i915_gem_object_wait_fence(obj);
3415 /* Just update our place in the LRU if our fence is getting reused. */
3416 if (obj->fence_reg != I915_FENCE_REG_NONE) {
3417 reg = &dev_priv->fence_regs[obj->fence_reg];
3418 if (!obj->fence_dirty) {
3419 list_move_tail(®->lru_list,
3420 &dev_priv->mm.fence_list);
3423 } else if (enable) {
3424 if (WARN_ON(!obj->map_and_fenceable))
3427 reg = i915_find_fence_reg(dev);
3429 return PTR_ERR(reg);
3432 struct drm_i915_gem_object *old = reg->obj;
3434 ret = i915_gem_object_wait_fence(old);
3438 i915_gem_object_fence_lost(old);
3443 i915_gem_object_update_fence(obj, reg, enable);
3448 static bool i915_gem_valid_gtt_space(struct i915_vma *vma,
3449 unsigned long cache_level)
3451 struct drm_mm_node *gtt_space = &vma->node;
3452 struct drm_mm_node *other;
3455 * On some machines we have to be careful when putting differing types
3456 * of snoopable memory together to avoid the prefetcher crossing memory
3457 * domains and dying. During vm initialisation, we decide whether or not
3458 * these constraints apply and set the drm_mm.color_adjust
3461 if (vma->vm->mm.color_adjust == NULL)
3464 if (!drm_mm_node_allocated(gtt_space))
3467 if (list_empty(>t_space->node_list))
3470 other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list);
3471 if (other->allocated && !other->hole_follows && other->color != cache_level)
3474 other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list);
3475 if (other->allocated && !gtt_space->hole_follows && other->color != cache_level)
3482 * Finds free space in the GTT aperture and binds the object there.
3484 static struct i915_vma *
3485 i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj,
3486 struct i915_address_space *vm,
3489 const struct i915_ggtt_view *view)
3491 struct drm_device *dev = obj->base.dev;
3492 struct drm_i915_private *dev_priv = dev->dev_private;
3493 u32 size, fence_size, fence_alignment, unfenced_alignment;
3494 unsigned long start =
3495 flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
3497 flags & PIN_MAPPABLE ? dev_priv->gtt.mappable_end : vm->total;
3498 struct i915_vma *vma;
3501 fence_size = i915_gem_get_gtt_size(dev,
3504 fence_alignment = i915_gem_get_gtt_alignment(dev,
3506 obj->tiling_mode, true);
3507 unfenced_alignment =
3508 i915_gem_get_gtt_alignment(dev,
3510 obj->tiling_mode, false);
3513 alignment = flags & PIN_MAPPABLE ? fence_alignment :
3515 if (flags & PIN_MAPPABLE && alignment & (fence_alignment - 1)) {
3516 DRM_DEBUG("Invalid object alignment requested %u\n", alignment);
3517 return ERR_PTR(-EINVAL);
3520 size = flags & PIN_MAPPABLE ? fence_size : obj->base.size;
3522 /* If the object is bigger than the entire aperture, reject it early
3523 * before evicting everything in a vain attempt to find space.
3525 if (obj->base.size > end) {
3526 DRM_DEBUG("Attempting to bind an object larger than the aperture: object=%zd > %s aperture=%lu\n",
3528 flags & PIN_MAPPABLE ? "mappable" : "total",
3530 return ERR_PTR(-E2BIG);
3533 ret = i915_gem_object_get_pages(obj);
3535 return ERR_PTR(ret);
3537 i915_gem_object_pin_pages(obj);
3539 vma = i915_gem_obj_lookup_or_create_vma_view(obj, vm, view);
3544 ret = drm_mm_insert_node_in_range_generic(&vm->mm, &vma->node,
3548 DRM_MM_SEARCH_DEFAULT,
3549 DRM_MM_CREATE_DEFAULT);
3551 ret = i915_gem_evict_something(dev, vm, size, alignment,
3560 if (WARN_ON(!i915_gem_valid_gtt_space(vma, obj->cache_level))) {
3562 goto err_remove_node;
3565 ret = i915_gem_gtt_prepare_object(obj);
3567 goto err_remove_node;
3569 trace_i915_vma_bind(vma, flags);
3570 ret = i915_vma_bind(vma, obj->cache_level,
3571 flags & PIN_GLOBAL ? GLOBAL_BIND : 0);
3573 goto err_finish_gtt;
3575 list_move_tail(&obj->global_list, &dev_priv->mm.bound_list);
3576 list_add_tail(&vma->mm_list, &vm->inactive_list);
3581 i915_gem_gtt_finish_object(obj);
3583 drm_mm_remove_node(&vma->node);
3585 i915_gem_vma_destroy(vma);
3588 i915_gem_object_unpin_pages(obj);
3593 i915_gem_clflush_object(struct drm_i915_gem_object *obj,
3596 /* If we don't have a page list set up, then we're not pinned
3597 * to GPU, and we can ignore the cache flush because it'll happen
3598 * again at bind time.
3600 if (obj->pages == NULL)
3604 * Stolen memory is always coherent with the GPU as it is explicitly
3605 * marked as wc by the system, or the system is cache-coherent.
3607 if (obj->stolen || obj->phys_handle)
3610 /* If the GPU is snooping the contents of the CPU cache,
3611 * we do not need to manually clear the CPU cache lines. However,
3612 * the caches are only snooped when the render cache is
3613 * flushed/invalidated. As we always have to emit invalidations
3614 * and flushes when moving into and out of the RENDER domain, correct
3615 * snooping behaviour occurs naturally as the result of our domain
3618 if (!force && cpu_cache_is_coherent(obj->base.dev, obj->cache_level))
3621 trace_i915_gem_object_clflush(obj);
3622 drm_clflush_sg(obj->pages);
3627 /** Flushes the GTT write domain for the object if it's dirty. */
3629 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
3631 uint32_t old_write_domain;
3633 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
3636 /* No actual flushing is required for the GTT write domain. Writes
3637 * to it immediately go to main memory as far as we know, so there's
3638 * no chipset flush. It also doesn't land in render cache.
3640 * However, we do have to enforce the order so that all writes through
3641 * the GTT land before any writes to the device, such as updates to
3646 old_write_domain = obj->base.write_domain;
3647 obj->base.write_domain = 0;
3649 intel_fb_obj_flush(obj, false);
3651 trace_i915_gem_object_change_domain(obj,
3652 obj->base.read_domains,
3656 /** Flushes the CPU write domain for the object if it's dirty. */
3658 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj,
3661 uint32_t old_write_domain;
3663 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
3666 if (i915_gem_clflush_object(obj, force))
3667 i915_gem_chipset_flush(obj->base.dev);
3669 old_write_domain = obj->base.write_domain;
3670 obj->base.write_domain = 0;
3672 intel_fb_obj_flush(obj, false);
3674 trace_i915_gem_object_change_domain(obj,
3675 obj->base.read_domains,
3680 * Moves a single object to the GTT read, and possibly write domain.
3682 * This function returns when the move is complete, including waiting on
3686 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
3688 uint32_t old_write_domain, old_read_domains;
3689 struct i915_vma *vma;
3692 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
3695 ret = i915_gem_object_wait_rendering(obj, !write);
3699 i915_gem_object_retire(obj);
3701 /* Flush and acquire obj->pages so that we are coherent through
3702 * direct access in memory with previous cached writes through
3703 * shmemfs and that our cache domain tracking remains valid.
3704 * For example, if the obj->filp was moved to swap without us
3705 * being notified and releasing the pages, we would mistakenly
3706 * continue to assume that the obj remained out of the CPU cached
3709 ret = i915_gem_object_get_pages(obj);
3713 i915_gem_object_flush_cpu_write_domain(obj, false);
3715 /* Serialise direct access to this object with the barriers for
3716 * coherent writes from the GPU, by effectively invalidating the
3717 * GTT domain upon first access.
3719 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
3722 old_write_domain = obj->base.write_domain;
3723 old_read_domains = obj->base.read_domains;
3725 /* It should now be out of any other write domains, and we can update
3726 * the domain values for our changes.
3728 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3729 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3731 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
3732 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
3737 intel_fb_obj_invalidate(obj, NULL);
3739 trace_i915_gem_object_change_domain(obj,
3743 /* And bump the LRU for this access */
3744 vma = i915_gem_obj_to_ggtt(obj);
3745 if (vma && drm_mm_node_allocated(&vma->node) && !obj->active)
3746 list_move_tail(&vma->mm_list,
3747 &to_i915(obj->base.dev)->gtt.base.inactive_list);
3752 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
3753 enum i915_cache_level cache_level)
3755 struct drm_device *dev = obj->base.dev;
3756 struct i915_vma *vma, *next;
3759 if (obj->cache_level == cache_level)
3762 if (i915_gem_obj_is_pinned(obj)) {
3763 DRM_DEBUG("can not change the cache level of pinned objects\n");
3767 list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) {
3768 if (!i915_gem_valid_gtt_space(vma, cache_level)) {
3769 ret = i915_vma_unbind(vma);
3775 if (i915_gem_obj_bound_any(obj)) {
3776 ret = i915_gem_object_finish_gpu(obj);
3780 i915_gem_object_finish_gtt(obj);
3782 /* Before SandyBridge, you could not use tiling or fence
3783 * registers with snooped memory, so relinquish any fences
3784 * currently pointing to our region in the aperture.
3786 if (INTEL_INFO(dev)->gen < 6) {
3787 ret = i915_gem_object_put_fence(obj);
3792 list_for_each_entry(vma, &obj->vma_list, vma_link)
3793 if (drm_mm_node_allocated(&vma->node)) {
3794 ret = i915_vma_bind(vma, cache_level,
3795 vma->bound & GLOBAL_BIND);
3801 list_for_each_entry(vma, &obj->vma_list, vma_link)
3802 vma->node.color = cache_level;
3803 obj->cache_level = cache_level;
3805 if (cpu_write_needs_clflush(obj)) {
3806 u32 old_read_domains, old_write_domain;
3808 /* If we're coming from LLC cached, then we haven't
3809 * actually been tracking whether the data is in the
3810 * CPU cache or not, since we only allow one bit set
3811 * in obj->write_domain and have been skipping the clflushes.
3812 * Just set it to the CPU cache for now.
3814 i915_gem_object_retire(obj);
3815 WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU);
3817 old_read_domains = obj->base.read_domains;
3818 old_write_domain = obj->base.write_domain;
3820 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3821 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3823 trace_i915_gem_object_change_domain(obj,
3831 int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data,
3832 struct drm_file *file)
3834 struct drm_i915_gem_caching *args = data;
3835 struct drm_i915_gem_object *obj;
3838 ret = i915_mutex_lock_interruptible(dev);
3842 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3843 if (&obj->base == NULL) {
3848 switch (obj->cache_level) {
3849 case I915_CACHE_LLC:
3850 case I915_CACHE_L3_LLC:
3851 args->caching = I915_CACHING_CACHED;
3855 args->caching = I915_CACHING_DISPLAY;
3859 args->caching = I915_CACHING_NONE;
3863 drm_gem_object_unreference(&obj->base);
3865 mutex_unlock(&dev->struct_mutex);
3869 int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
3870 struct drm_file *file)
3872 struct drm_i915_gem_caching *args = data;
3873 struct drm_i915_gem_object *obj;
3874 enum i915_cache_level level;
3877 switch (args->caching) {
3878 case I915_CACHING_NONE:
3879 level = I915_CACHE_NONE;
3881 case I915_CACHING_CACHED:
3882 level = I915_CACHE_LLC;
3884 case I915_CACHING_DISPLAY:
3885 level = HAS_WT(dev) ? I915_CACHE_WT : I915_CACHE_NONE;
3891 ret = i915_mutex_lock_interruptible(dev);
3895 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3896 if (&obj->base == NULL) {
3901 ret = i915_gem_object_set_cache_level(obj, level);
3903 drm_gem_object_unreference(&obj->base);
3905 mutex_unlock(&dev->struct_mutex);
3909 static bool is_pin_display(struct drm_i915_gem_object *obj)
3911 struct i915_vma *vma;
3913 vma = i915_gem_obj_to_ggtt(obj);
3917 /* There are 2 sources that pin objects:
3918 * 1. The display engine (scanouts, sprites, cursors);
3919 * 2. Reservations for execbuffer;
3921 * We can ignore reservations as we hold the struct_mutex and
3922 * are only called outside of the reservation path.
3924 return vma->pin_count;
3928 * Prepare buffer for display plane (scanout, cursors, etc).
3929 * Can be called from an uninterruptible phase (modesetting) and allows
3930 * any flushes to be pipelined (for pageflips).
3933 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
3935 struct intel_engine_cs *pipelined)
3937 u32 old_read_domains, old_write_domain;
3938 bool was_pin_display;
3941 if (pipelined != i915_gem_request_get_ring(obj->last_read_req)) {
3942 ret = i915_gem_object_sync(obj, pipelined);
3947 /* Mark the pin_display early so that we account for the
3948 * display coherency whilst setting up the cache domains.
3950 was_pin_display = obj->pin_display;
3951 obj->pin_display = true;
3953 /* The display engine is not coherent with the LLC cache on gen6. As
3954 * a result, we make sure that the pinning that is about to occur is
3955 * done with uncached PTEs. This is lowest common denominator for all
3958 * However for gen6+, we could do better by using the GFDT bit instead
3959 * of uncaching, which would allow us to flush all the LLC-cached data
3960 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3962 ret = i915_gem_object_set_cache_level(obj,
3963 HAS_WT(obj->base.dev) ? I915_CACHE_WT : I915_CACHE_NONE);
3965 goto err_unpin_display;
3967 /* As the user may map the buffer once pinned in the display plane
3968 * (e.g. libkms for the bootup splash), we have to ensure that we
3969 * always use map_and_fenceable for all scanout buffers.
3971 ret = i915_gem_obj_ggtt_pin(obj, alignment, PIN_MAPPABLE);
3973 goto err_unpin_display;
3975 i915_gem_object_flush_cpu_write_domain(obj, true);
3977 old_write_domain = obj->base.write_domain;
3978 old_read_domains = obj->base.read_domains;
3980 /* It should now be out of any other write domains, and we can update
3981 * the domain values for our changes.
3983 obj->base.write_domain = 0;
3984 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3986 trace_i915_gem_object_change_domain(obj,
3993 WARN_ON(was_pin_display != is_pin_display(obj));
3994 obj->pin_display = was_pin_display;
3999 i915_gem_object_unpin_from_display_plane(struct drm_i915_gem_object *obj)
4001 i915_gem_object_ggtt_unpin(obj);
4002 obj->pin_display = is_pin_display(obj);
4006 i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
4010 if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
4013 ret = i915_gem_object_wait_rendering(obj, false);
4017 /* Ensure that we invalidate the GPU's caches and TLBs. */
4018 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
4023 * Moves a single object to the CPU read, and possibly write domain.
4025 * This function returns when the move is complete, including waiting on
4029 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
4031 uint32_t old_write_domain, old_read_domains;
4034 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
4037 ret = i915_gem_object_wait_rendering(obj, !write);
4041 i915_gem_object_retire(obj);
4042 i915_gem_object_flush_gtt_write_domain(obj);
4044 old_write_domain = obj->base.write_domain;
4045 old_read_domains = obj->base.read_domains;
4047 /* Flush the CPU cache if it's still invalid. */
4048 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
4049 i915_gem_clflush_object(obj, false);
4051 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
4054 /* It should now be out of any other write domains, and we can update
4055 * the domain values for our changes.
4057 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
4059 /* If we're writing through the CPU, then the GPU read domains will
4060 * need to be invalidated at next use.
4063 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4064 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4068 intel_fb_obj_invalidate(obj, NULL);
4070 trace_i915_gem_object_change_domain(obj,
4077 /* Throttle our rendering by waiting until the ring has completed our requests
4078 * emitted over 20 msec ago.
4080 * Note that if we were to use the current jiffies each time around the loop,
4081 * we wouldn't escape the function with any frames outstanding if the time to
4082 * render a frame was over 20ms.
4084 * This should get us reasonable parallelism between CPU and GPU but also
4085 * relatively low latency when blocking on a particular request to finish.
4088 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
4090 struct drm_i915_private *dev_priv = dev->dev_private;
4091 struct drm_i915_file_private *file_priv = file->driver_priv;
4092 unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
4093 struct drm_i915_gem_request *request, *target = NULL;
4094 unsigned reset_counter;
4097 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
4101 ret = i915_gem_check_wedge(&dev_priv->gpu_error, false);
4105 spin_lock(&file_priv->mm.lock);
4106 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
4107 if (time_after_eq(request->emitted_jiffies, recent_enough))
4112 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
4114 i915_gem_request_reference(target);
4115 spin_unlock(&file_priv->mm.lock);
4120 ret = __i915_wait_request(target, reset_counter, true, NULL, NULL);
4122 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
4124 mutex_lock(&dev->struct_mutex);
4125 i915_gem_request_unreference(target);
4126 mutex_unlock(&dev->struct_mutex);
4132 i915_vma_misplaced(struct i915_vma *vma, uint32_t alignment, uint64_t flags)
4134 struct drm_i915_gem_object *obj = vma->obj;
4137 vma->node.start & (alignment - 1))
4140 if (flags & PIN_MAPPABLE && !obj->map_and_fenceable)
4143 if (flags & PIN_OFFSET_BIAS &&
4144 vma->node.start < (flags & PIN_OFFSET_MASK))
4151 i915_gem_object_pin_view(struct drm_i915_gem_object *obj,
4152 struct i915_address_space *vm,
4155 const struct i915_ggtt_view *view)
4157 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
4158 struct i915_vma *vma;
4162 if (WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base))
4165 if (WARN_ON(flags & (PIN_GLOBAL | PIN_MAPPABLE) && !i915_is_ggtt(vm)))
4168 if (WARN_ON((flags & (PIN_MAPPABLE | PIN_GLOBAL)) == PIN_MAPPABLE))
4171 vma = i915_gem_obj_to_vma_view(obj, vm, view);
4173 if (WARN_ON(vma->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT))
4176 if (i915_vma_misplaced(vma, alignment, flags)) {
4177 WARN(vma->pin_count,
4178 "bo is already pinned with incorrect alignment:"
4179 " offset=%lx, req.alignment=%x, req.map_and_fenceable=%d,"
4180 " obj->map_and_fenceable=%d\n",
4181 i915_gem_obj_offset_view(obj, vm, view->type),
4183 !!(flags & PIN_MAPPABLE),
4184 obj->map_and_fenceable);
4185 ret = i915_vma_unbind(vma);
4193 bound = vma ? vma->bound : 0;
4194 if (vma == NULL || !drm_mm_node_allocated(&vma->node)) {
4195 vma = i915_gem_object_bind_to_vm(obj, vm, alignment,
4198 return PTR_ERR(vma);
4201 if (flags & PIN_GLOBAL && !(vma->bound & GLOBAL_BIND)) {
4202 ret = i915_vma_bind(vma, obj->cache_level, GLOBAL_BIND);
4207 if ((bound ^ vma->bound) & GLOBAL_BIND) {
4208 bool mappable, fenceable;
4209 u32 fence_size, fence_alignment;
4211 fence_size = i915_gem_get_gtt_size(obj->base.dev,
4214 fence_alignment = i915_gem_get_gtt_alignment(obj->base.dev,
4219 fenceable = (vma->node.size == fence_size &&
4220 (vma->node.start & (fence_alignment - 1)) == 0);
4222 mappable = (vma->node.start + obj->base.size <=
4223 dev_priv->gtt.mappable_end);
4225 obj->map_and_fenceable = mappable && fenceable;
4228 WARN_ON(flags & PIN_MAPPABLE && !obj->map_and_fenceable);
4231 if (flags & PIN_MAPPABLE)
4232 obj->pin_mappable |= true;
4238 i915_gem_object_ggtt_unpin(struct drm_i915_gem_object *obj)
4240 struct i915_vma *vma = i915_gem_obj_to_ggtt(obj);
4243 BUG_ON(vma->pin_count == 0);
4244 BUG_ON(!i915_gem_obj_ggtt_bound(obj));
4246 if (--vma->pin_count == 0)
4247 obj->pin_mappable = false;
4251 i915_gem_object_pin_fence(struct drm_i915_gem_object *obj)
4253 if (obj->fence_reg != I915_FENCE_REG_NONE) {
4254 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
4255 struct i915_vma *ggtt_vma = i915_gem_obj_to_ggtt(obj);
4257 WARN_ON(!ggtt_vma ||
4258 dev_priv->fence_regs[obj->fence_reg].pin_count >
4259 ggtt_vma->pin_count);
4260 dev_priv->fence_regs[obj->fence_reg].pin_count++;
4267 i915_gem_object_unpin_fence(struct drm_i915_gem_object *obj)
4269 if (obj->fence_reg != I915_FENCE_REG_NONE) {
4270 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
4271 WARN_ON(dev_priv->fence_regs[obj->fence_reg].pin_count <= 0);
4272 dev_priv->fence_regs[obj->fence_reg].pin_count--;
4277 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
4278 struct drm_file *file)
4280 struct drm_i915_gem_busy *args = data;
4281 struct drm_i915_gem_object *obj;
4284 ret = i915_mutex_lock_interruptible(dev);
4288 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
4289 if (&obj->base == NULL) {
4294 /* Count all active objects as busy, even if they are currently not used
4295 * by the gpu. Users of this interface expect objects to eventually
4296 * become non-busy without any further actions, therefore emit any
4297 * necessary flushes here.
4299 ret = i915_gem_object_flush_active(obj);
4301 args->busy = obj->active;
4302 if (obj->last_read_req) {
4303 struct intel_engine_cs *ring;
4304 BUILD_BUG_ON(I915_NUM_RINGS > 16);
4305 ring = i915_gem_request_get_ring(obj->last_read_req);
4306 args->busy |= intel_ring_flag(ring) << 16;
4309 drm_gem_object_unreference(&obj->base);
4311 mutex_unlock(&dev->struct_mutex);
4316 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
4317 struct drm_file *file_priv)
4319 return i915_gem_ring_throttle(dev, file_priv);
4323 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
4324 struct drm_file *file_priv)
4326 struct drm_i915_private *dev_priv = dev->dev_private;
4327 struct drm_i915_gem_madvise *args = data;
4328 struct drm_i915_gem_object *obj;
4331 switch (args->madv) {
4332 case I915_MADV_DONTNEED:
4333 case I915_MADV_WILLNEED:
4339 ret = i915_mutex_lock_interruptible(dev);
4343 obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
4344 if (&obj->base == NULL) {
4349 if (i915_gem_obj_is_pinned(obj)) {
4355 obj->tiling_mode != I915_TILING_NONE &&
4356 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
4357 if (obj->madv == I915_MADV_WILLNEED)
4358 i915_gem_object_unpin_pages(obj);
4359 if (args->madv == I915_MADV_WILLNEED)
4360 i915_gem_object_pin_pages(obj);
4363 if (obj->madv != __I915_MADV_PURGED)
4364 obj->madv = args->madv;
4366 /* if the object is no longer attached, discard its backing storage */
4367 if (i915_gem_object_is_purgeable(obj) && obj->pages == NULL)
4368 i915_gem_object_truncate(obj);
4370 args->retained = obj->madv != __I915_MADV_PURGED;
4373 drm_gem_object_unreference(&obj->base);
4375 mutex_unlock(&dev->struct_mutex);
4379 void i915_gem_object_init(struct drm_i915_gem_object *obj,
4380 const struct drm_i915_gem_object_ops *ops)
4382 INIT_LIST_HEAD(&obj->global_list);
4383 INIT_LIST_HEAD(&obj->ring_list);
4384 INIT_LIST_HEAD(&obj->obj_exec_link);
4385 INIT_LIST_HEAD(&obj->vma_list);
4386 INIT_LIST_HEAD(&obj->batch_pool_list);
4390 obj->fence_reg = I915_FENCE_REG_NONE;
4391 obj->madv = I915_MADV_WILLNEED;
4393 i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size);
4396 static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
4397 .get_pages = i915_gem_object_get_pages_gtt,
4398 .put_pages = i915_gem_object_put_pages_gtt,
4401 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
4404 struct drm_i915_gem_object *obj;
4405 struct address_space *mapping;
4408 obj = i915_gem_object_alloc(dev);
4412 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
4413 i915_gem_object_free(obj);
4417 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
4418 if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) {
4419 /* 965gm cannot relocate objects above 4GiB. */
4420 mask &= ~__GFP_HIGHMEM;
4421 mask |= __GFP_DMA32;
4424 mapping = file_inode(obj->base.filp)->i_mapping;
4425 mapping_set_gfp_mask(mapping, mask);
4427 i915_gem_object_init(obj, &i915_gem_object_ops);
4429 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4430 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4433 /* On some devices, we can have the GPU use the LLC (the CPU
4434 * cache) for about a 10% performance improvement
4435 * compared to uncached. Graphics requests other than
4436 * display scanout are coherent with the CPU in
4437 * accessing this cache. This means in this mode we
4438 * don't need to clflush on the CPU side, and on the
4439 * GPU side we only need to flush internal caches to
4440 * get data visible to the CPU.
4442 * However, we maintain the display planes as UC, and so
4443 * need to rebind when first used as such.
4445 obj->cache_level = I915_CACHE_LLC;
4447 obj->cache_level = I915_CACHE_NONE;
4449 trace_i915_gem_object_create(obj);
4454 static bool discard_backing_storage(struct drm_i915_gem_object *obj)
4456 /* If we are the last user of the backing storage (be it shmemfs
4457 * pages or stolen etc), we know that the pages are going to be
4458 * immediately released. In this case, we can then skip copying
4459 * back the contents from the GPU.
4462 if (obj->madv != I915_MADV_WILLNEED)
4465 if (obj->base.filp == NULL)
4468 /* At first glance, this looks racy, but then again so would be
4469 * userspace racing mmap against close. However, the first external
4470 * reference to the filp can only be obtained through the
4471 * i915_gem_mmap_ioctl() which safeguards us against the user
4472 * acquiring such a reference whilst we are in the middle of
4473 * freeing the object.
4475 return atomic_long_read(&obj->base.filp->f_count) == 1;
4478 void i915_gem_free_object(struct drm_gem_object *gem_obj)
4480 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
4481 struct drm_device *dev = obj->base.dev;
4482 struct drm_i915_private *dev_priv = dev->dev_private;
4483 struct i915_vma *vma, *next;
4485 intel_runtime_pm_get(dev_priv);
4487 trace_i915_gem_object_destroy(obj);
4489 list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) {
4493 ret = i915_vma_unbind(vma);
4494 if (WARN_ON(ret == -ERESTARTSYS)) {
4495 bool was_interruptible;
4497 was_interruptible = dev_priv->mm.interruptible;
4498 dev_priv->mm.interruptible = false;
4500 WARN_ON(i915_vma_unbind(vma));
4502 dev_priv->mm.interruptible = was_interruptible;
4506 /* Stolen objects don't hold a ref, but do hold pin count. Fix that up
4507 * before progressing. */
4509 i915_gem_object_unpin_pages(obj);
4511 WARN_ON(obj->frontbuffer_bits);
4513 if (obj->pages && obj->madv == I915_MADV_WILLNEED &&
4514 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES &&
4515 obj->tiling_mode != I915_TILING_NONE)
4516 i915_gem_object_unpin_pages(obj);
4518 if (WARN_ON(obj->pages_pin_count))
4519 obj->pages_pin_count = 0;
4520 if (discard_backing_storage(obj))
4521 obj->madv = I915_MADV_DONTNEED;
4522 i915_gem_object_put_pages(obj);
4523 i915_gem_object_free_mmap_offset(obj);
4527 if (obj->base.import_attach)
4528 drm_prime_gem_destroy(&obj->base, NULL);
4530 if (obj->ops->release)
4531 obj->ops->release(obj);
4533 drm_gem_object_release(&obj->base);
4534 i915_gem_info_remove_obj(dev_priv, obj->base.size);
4537 i915_gem_object_free(obj);
4539 intel_runtime_pm_put(dev_priv);
4542 struct i915_vma *i915_gem_obj_to_vma_view(struct drm_i915_gem_object *obj,
4543 struct i915_address_space *vm,
4544 const struct i915_ggtt_view *view)
4546 struct i915_vma *vma;
4547 list_for_each_entry(vma, &obj->vma_list, vma_link)
4548 if (vma->vm == vm && vma->ggtt_view.type == view->type)
4554 void i915_gem_vma_destroy(struct i915_vma *vma)
4556 struct i915_address_space *vm = NULL;
4557 WARN_ON(vma->node.allocated);
4559 /* Keep the vma as a placeholder in the execbuffer reservation lists */
4560 if (!list_empty(&vma->exec_list))
4565 if (!i915_is_ggtt(vm))
4566 i915_ppgtt_put(i915_vm_to_ppgtt(vm));
4568 list_del(&vma->vma_link);
4574 i915_gem_stop_ringbuffers(struct drm_device *dev)
4576 struct drm_i915_private *dev_priv = dev->dev_private;
4577 struct intel_engine_cs *ring;
4580 for_each_ring(ring, dev_priv, i)
4581 dev_priv->gt.stop_ring(ring);
4585 i915_gem_suspend(struct drm_device *dev)
4587 struct drm_i915_private *dev_priv = dev->dev_private;
4590 mutex_lock(&dev->struct_mutex);
4591 ret = i915_gpu_idle(dev);
4595 i915_gem_retire_requests(dev);
4597 /* Under UMS, be paranoid and evict. */
4598 if (!drm_core_check_feature(dev, DRIVER_MODESET))
4599 i915_gem_evict_everything(dev);
4601 i915_gem_stop_ringbuffers(dev);
4602 mutex_unlock(&dev->struct_mutex);
4604 del_timer_sync(&dev_priv->gpu_error.hangcheck_timer);
4605 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
4606 flush_delayed_work(&dev_priv->mm.idle_work);
4608 /* Assert that we sucessfully flushed all the work and
4609 * reset the GPU back to its idle, low power state.
4611 WARN_ON(dev_priv->mm.busy);
4616 mutex_unlock(&dev->struct_mutex);
4620 int i915_gem_l3_remap(struct intel_engine_cs *ring, int slice)
4622 struct drm_device *dev = ring->dev;
4623 struct drm_i915_private *dev_priv = dev->dev_private;
4624 u32 reg_base = GEN7_L3LOG_BASE + (slice * 0x200);
4625 u32 *remap_info = dev_priv->l3_parity.remap_info[slice];
4628 if (!HAS_L3_DPF(dev) || !remap_info)
4631 ret = intel_ring_begin(ring, GEN7_L3LOG_SIZE / 4 * 3);
4636 * Note: We do not worry about the concurrent register cacheline hang
4637 * here because no other code should access these registers other than
4638 * at initialization time.
4640 for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) {
4641 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
4642 intel_ring_emit(ring, reg_base + i);
4643 intel_ring_emit(ring, remap_info[i/4]);
4646 intel_ring_advance(ring);
4651 void i915_gem_init_swizzling(struct drm_device *dev)
4653 struct drm_i915_private *dev_priv = dev->dev_private;
4655 if (INTEL_INFO(dev)->gen < 5 ||
4656 dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
4659 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
4660 DISP_TILE_SURFACE_SWIZZLING);
4665 I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
4667 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
4668 else if (IS_GEN7(dev))
4669 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
4670 else if (IS_GEN8(dev))
4671 I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW));
4677 intel_enable_blt(struct drm_device *dev)
4682 /* The blitter was dysfunctional on early prototypes */
4683 if (IS_GEN6(dev) && dev->pdev->revision < 8) {
4684 DRM_INFO("BLT not supported on this pre-production hardware;"
4685 " graphics performance will be degraded.\n");
4692 static void init_unused_ring(struct drm_device *dev, u32 base)
4694 struct drm_i915_private *dev_priv = dev->dev_private;
4696 I915_WRITE(RING_CTL(base), 0);
4697 I915_WRITE(RING_HEAD(base), 0);
4698 I915_WRITE(RING_TAIL(base), 0);
4699 I915_WRITE(RING_START(base), 0);
4702 static void init_unused_rings(struct drm_device *dev)
4705 init_unused_ring(dev, PRB1_BASE);
4706 init_unused_ring(dev, SRB0_BASE);
4707 init_unused_ring(dev, SRB1_BASE);
4708 init_unused_ring(dev, SRB2_BASE);
4709 init_unused_ring(dev, SRB3_BASE);
4710 } else if (IS_GEN2(dev)) {
4711 init_unused_ring(dev, SRB0_BASE);
4712 init_unused_ring(dev, SRB1_BASE);
4713 } else if (IS_GEN3(dev)) {
4714 init_unused_ring(dev, PRB1_BASE);
4715 init_unused_ring(dev, PRB2_BASE);
4719 int i915_gem_init_rings(struct drm_device *dev)
4721 struct drm_i915_private *dev_priv = dev->dev_private;
4724 ret = intel_init_render_ring_buffer(dev);
4729 ret = intel_init_bsd_ring_buffer(dev);
4731 goto cleanup_render_ring;
4734 if (intel_enable_blt(dev)) {
4735 ret = intel_init_blt_ring_buffer(dev);
4737 goto cleanup_bsd_ring;
4740 if (HAS_VEBOX(dev)) {
4741 ret = intel_init_vebox_ring_buffer(dev);
4743 goto cleanup_blt_ring;
4746 if (HAS_BSD2(dev)) {
4747 ret = intel_init_bsd2_ring_buffer(dev);
4749 goto cleanup_vebox_ring;
4752 ret = i915_gem_set_seqno(dev, ((u32)~0 - 0x1000));
4754 goto cleanup_bsd2_ring;
4759 intel_cleanup_ring_buffer(&dev_priv->ring[VCS2]);
4761 intel_cleanup_ring_buffer(&dev_priv->ring[VECS]);
4763 intel_cleanup_ring_buffer(&dev_priv->ring[BCS]);
4765 intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
4766 cleanup_render_ring:
4767 intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
4773 i915_gem_init_hw(struct drm_device *dev)
4775 struct drm_i915_private *dev_priv = dev->dev_private;
4776 struct intel_engine_cs *ring;
4779 if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt())
4782 if (dev_priv->ellc_size)
4783 I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf));
4785 if (IS_HASWELL(dev))
4786 I915_WRITE(MI_PREDICATE_RESULT_2, IS_HSW_GT3(dev) ?
4787 LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED);
4789 if (HAS_PCH_NOP(dev)) {
4790 if (IS_IVYBRIDGE(dev)) {
4791 u32 temp = I915_READ(GEN7_MSG_CTL);
4792 temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK);
4793 I915_WRITE(GEN7_MSG_CTL, temp);
4794 } else if (INTEL_INFO(dev)->gen >= 7) {
4795 u32 temp = I915_READ(HSW_NDE_RSTWRN_OPT);
4796 temp &= ~RESET_PCH_HANDSHAKE_ENABLE;
4797 I915_WRITE(HSW_NDE_RSTWRN_OPT, temp);
4801 i915_gem_init_swizzling(dev);
4804 * At least 830 can leave some of the unused rings
4805 * "active" (ie. head != tail) after resume which
4806 * will prevent c3 entry. Makes sure all unused rings
4809 init_unused_rings(dev);
4811 for_each_ring(ring, dev_priv, i) {
4812 ret = ring->init_hw(ring);
4817 for (i = 0; i < NUM_L3_SLICES(dev); i++)
4818 i915_gem_l3_remap(&dev_priv->ring[RCS], i);
4821 * XXX: Contexts should only be initialized once. Doing a switch to the
4822 * default context switch however is something we'd like to do after
4823 * reset or thaw (the latter may not actually be necessary for HW, but
4824 * goes with our code better). Context switching requires rings (for
4825 * the do_switch), but before enabling PPGTT. So don't move this.
4827 ret = i915_gem_context_enable(dev_priv);
4828 if (ret && ret != -EIO) {
4829 DRM_ERROR("Context enable failed %d\n", ret);
4830 i915_gem_cleanup_ringbuffer(dev);
4835 ret = i915_ppgtt_init_hw(dev);
4836 if (ret && ret != -EIO) {
4837 DRM_ERROR("PPGTT enable failed %d\n", ret);
4838 i915_gem_cleanup_ringbuffer(dev);
4844 int i915_gem_init(struct drm_device *dev)
4846 struct drm_i915_private *dev_priv = dev->dev_private;
4849 i915.enable_execlists = intel_sanitize_enable_execlists(dev,
4850 i915.enable_execlists);
4852 mutex_lock(&dev->struct_mutex);
4854 if (IS_VALLEYVIEW(dev)) {
4855 /* VLVA0 (potential hack), BIOS isn't actually waking us */
4856 I915_WRITE(VLV_GTLC_WAKE_CTRL, VLV_GTLC_ALLOWWAKEREQ);
4857 if (wait_for((I915_READ(VLV_GTLC_PW_STATUS) &
4858 VLV_GTLC_ALLOWWAKEACK), 10))
4859 DRM_DEBUG_DRIVER("allow wake ack timed out\n");
4862 if (!i915.enable_execlists) {
4863 dev_priv->gt.do_execbuf = i915_gem_ringbuffer_submission;
4864 dev_priv->gt.init_rings = i915_gem_init_rings;
4865 dev_priv->gt.cleanup_ring = intel_cleanup_ring_buffer;
4866 dev_priv->gt.stop_ring = intel_stop_ring_buffer;
4868 dev_priv->gt.do_execbuf = intel_execlists_submission;
4869 dev_priv->gt.init_rings = intel_logical_rings_init;
4870 dev_priv->gt.cleanup_ring = intel_logical_ring_cleanup;
4871 dev_priv->gt.stop_ring = intel_logical_ring_stop;
4874 ret = i915_gem_init_userptr(dev);
4878 i915_gem_init_global_gtt(dev);
4880 ret = i915_gem_context_init(dev);
4884 ret = dev_priv->gt.init_rings(dev);
4888 ret = i915_gem_init_hw(dev);
4890 /* Allow ring initialisation to fail by marking the GPU as
4891 * wedged. But we only want to do this where the GPU is angry,
4892 * for all other failure, such as an allocation failure, bail.
4894 DRM_ERROR("Failed to initialize GPU, declaring it wedged\n");
4895 atomic_set_mask(I915_WEDGED, &dev_priv->gpu_error.reset_counter);
4900 mutex_unlock(&dev->struct_mutex);
4906 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
4908 struct drm_i915_private *dev_priv = dev->dev_private;
4909 struct intel_engine_cs *ring;
4912 for_each_ring(ring, dev_priv, i)
4913 dev_priv->gt.cleanup_ring(ring);
4917 init_ring_lists(struct intel_engine_cs *ring)
4919 INIT_LIST_HEAD(&ring->active_list);
4920 INIT_LIST_HEAD(&ring->request_list);
4923 void i915_init_vm(struct drm_i915_private *dev_priv,
4924 struct i915_address_space *vm)
4926 if (!i915_is_ggtt(vm))
4927 drm_mm_init(&vm->mm, vm->start, vm->total);
4928 vm->dev = dev_priv->dev;
4929 INIT_LIST_HEAD(&vm->active_list);
4930 INIT_LIST_HEAD(&vm->inactive_list);
4931 INIT_LIST_HEAD(&vm->global_link);
4932 list_add_tail(&vm->global_link, &dev_priv->vm_list);
4936 i915_gem_load(struct drm_device *dev)
4938 struct drm_i915_private *dev_priv = dev->dev_private;
4942 kmem_cache_create("i915_gem_object",
4943 sizeof(struct drm_i915_gem_object), 0,
4947 INIT_LIST_HEAD(&dev_priv->vm_list);
4948 i915_init_vm(dev_priv, &dev_priv->gtt.base);
4950 INIT_LIST_HEAD(&dev_priv->context_list);
4951 INIT_LIST_HEAD(&dev_priv->mm.unbound_list);
4952 INIT_LIST_HEAD(&dev_priv->mm.bound_list);
4953 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4954 for (i = 0; i < I915_NUM_RINGS; i++)
4955 init_ring_lists(&dev_priv->ring[i]);
4956 for (i = 0; i < I915_MAX_NUM_FENCES; i++)
4957 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
4958 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
4959 i915_gem_retire_work_handler);
4960 INIT_DELAYED_WORK(&dev_priv->mm.idle_work,
4961 i915_gem_idle_work_handler);
4962 init_waitqueue_head(&dev_priv->gpu_error.reset_queue);
4964 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
4965 if (!drm_core_check_feature(dev, DRIVER_MODESET) && IS_GEN3(dev)) {
4966 I915_WRITE(MI_ARB_STATE,
4967 _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
4970 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
4972 /* Old X drivers will take 0-2 for front, back, depth buffers */
4973 if (!drm_core_check_feature(dev, DRIVER_MODESET))
4974 dev_priv->fence_reg_start = 3;
4976 if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev))
4977 dev_priv->num_fence_regs = 32;
4978 else if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
4979 dev_priv->num_fence_regs = 16;
4981 dev_priv->num_fence_regs = 8;
4983 /* Initialize fence registers to zero */
4984 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4985 i915_gem_restore_fences(dev);
4987 i915_gem_detect_bit_6_swizzle(dev);
4988 init_waitqueue_head(&dev_priv->pending_flip_queue);
4990 dev_priv->mm.interruptible = true;
4992 dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
4993 dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count;
4994 dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS;
4995 register_shrinker(&dev_priv->mm.shrinker);
4997 dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
4998 register_oom_notifier(&dev_priv->mm.oom_notifier);
5000 i915_gem_batch_pool_init(dev, &dev_priv->mm.batch_pool);
5002 mutex_init(&dev_priv->fb_tracking.lock);
5005 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
5007 struct drm_i915_file_private *file_priv = file->driver_priv;
5009 cancel_delayed_work_sync(&file_priv->mm.idle_work);
5011 /* Clean up our request list when the client is going away, so that
5012 * later retire_requests won't dereference our soon-to-be-gone
5015 spin_lock(&file_priv->mm.lock);
5016 while (!list_empty(&file_priv->mm.request_list)) {
5017 struct drm_i915_gem_request *request;
5019 request = list_first_entry(&file_priv->mm.request_list,
5020 struct drm_i915_gem_request,
5022 list_del(&request->client_list);
5023 request->file_priv = NULL;
5025 spin_unlock(&file_priv->mm.lock);
5029 i915_gem_file_idle_work_handler(struct work_struct *work)
5031 struct drm_i915_file_private *file_priv =
5032 container_of(work, typeof(*file_priv), mm.idle_work.work);
5034 atomic_set(&file_priv->rps_wait_boost, false);
5037 int i915_gem_open(struct drm_device *dev, struct drm_file *file)
5039 struct drm_i915_file_private *file_priv;
5042 DRM_DEBUG_DRIVER("\n");
5044 file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL);
5048 file->driver_priv = file_priv;
5049 file_priv->dev_priv = dev->dev_private;
5050 file_priv->file = file;
5052 spin_lock_init(&file_priv->mm.lock);
5053 INIT_LIST_HEAD(&file_priv->mm.request_list);
5054 INIT_DELAYED_WORK(&file_priv->mm.idle_work,
5055 i915_gem_file_idle_work_handler);
5057 ret = i915_gem_context_open(dev, file);
5065 * i915_gem_track_fb - update frontbuffer tracking
5066 * old: current GEM buffer for the frontbuffer slots
5067 * new: new GEM buffer for the frontbuffer slots
5068 * frontbuffer_bits: bitmask of frontbuffer slots
5070 * This updates the frontbuffer tracking bits @frontbuffer_bits by clearing them
5071 * from @old and setting them in @new. Both @old and @new can be NULL.
5073 void i915_gem_track_fb(struct drm_i915_gem_object *old,
5074 struct drm_i915_gem_object *new,
5075 unsigned frontbuffer_bits)
5078 WARN_ON(!mutex_is_locked(&old->base.dev->struct_mutex));
5079 WARN_ON(!(old->frontbuffer_bits & frontbuffer_bits));
5080 old->frontbuffer_bits &= ~frontbuffer_bits;
5084 WARN_ON(!mutex_is_locked(&new->base.dev->struct_mutex));
5085 WARN_ON(new->frontbuffer_bits & frontbuffer_bits);
5086 new->frontbuffer_bits |= frontbuffer_bits;
5090 static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task)
5092 if (!mutex_is_locked(mutex))
5095 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES)
5096 return mutex->owner == task;
5098 /* Since UP may be pre-empted, we cannot assume that we own the lock */
5103 static bool i915_gem_shrinker_lock(struct drm_device *dev, bool *unlock)
5105 if (!mutex_trylock(&dev->struct_mutex)) {
5106 if (!mutex_is_locked_by(&dev->struct_mutex, current))
5109 if (to_i915(dev)->mm.shrinker_no_lock_stealing)
5119 static int num_vma_bound(struct drm_i915_gem_object *obj)
5121 struct i915_vma *vma;
5124 list_for_each_entry(vma, &obj->vma_list, vma_link)
5125 if (drm_mm_node_allocated(&vma->node))
5131 static unsigned long
5132 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
5134 struct drm_i915_private *dev_priv =
5135 container_of(shrinker, struct drm_i915_private, mm.shrinker);
5136 struct drm_device *dev = dev_priv->dev;
5137 struct drm_i915_gem_object *obj;
5138 unsigned long count;
5141 if (!i915_gem_shrinker_lock(dev, &unlock))
5145 list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list)
5146 if (obj->pages_pin_count == 0)
5147 count += obj->base.size >> PAGE_SHIFT;
5149 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
5150 if (!i915_gem_obj_is_pinned(obj) &&
5151 obj->pages_pin_count == num_vma_bound(obj))
5152 count += obj->base.size >> PAGE_SHIFT;
5156 mutex_unlock(&dev->struct_mutex);
5161 /* All the new VM stuff */
5162 unsigned long i915_gem_obj_offset_view(struct drm_i915_gem_object *o,
5163 struct i915_address_space *vm,
5164 enum i915_ggtt_view_type view)
5166 struct drm_i915_private *dev_priv = o->base.dev->dev_private;
5167 struct i915_vma *vma;
5169 WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base);
5171 list_for_each_entry(vma, &o->vma_list, vma_link) {
5172 if (vma->vm == vm && vma->ggtt_view.type == view)
5173 return vma->node.start;
5176 WARN(1, "%s vma for this object not found.\n",
5177 i915_is_ggtt(vm) ? "global" : "ppgtt");
5181 bool i915_gem_obj_bound_view(struct drm_i915_gem_object *o,
5182 struct i915_address_space *vm,
5183 enum i915_ggtt_view_type view)
5185 struct i915_vma *vma;
5187 list_for_each_entry(vma, &o->vma_list, vma_link)
5188 if (vma->vm == vm &&
5189 vma->ggtt_view.type == view &&
5190 drm_mm_node_allocated(&vma->node))
5196 bool i915_gem_obj_bound_any(struct drm_i915_gem_object *o)
5198 struct i915_vma *vma;
5200 list_for_each_entry(vma, &o->vma_list, vma_link)
5201 if (drm_mm_node_allocated(&vma->node))
5207 unsigned long i915_gem_obj_size(struct drm_i915_gem_object *o,
5208 struct i915_address_space *vm)
5210 struct drm_i915_private *dev_priv = o->base.dev->dev_private;
5211 struct i915_vma *vma;
5213 WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base);
5215 BUG_ON(list_empty(&o->vma_list));
5217 list_for_each_entry(vma, &o->vma_list, vma_link)
5219 return vma->node.size;
5224 static unsigned long
5225 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
5227 struct drm_i915_private *dev_priv =
5228 container_of(shrinker, struct drm_i915_private, mm.shrinker);
5229 struct drm_device *dev = dev_priv->dev;
5230 unsigned long freed;
5233 if (!i915_gem_shrinker_lock(dev, &unlock))
5236 freed = i915_gem_shrink(dev_priv,
5239 I915_SHRINK_UNBOUND |
5240 I915_SHRINK_PURGEABLE);
5241 if (freed < sc->nr_to_scan)
5242 freed += i915_gem_shrink(dev_priv,
5243 sc->nr_to_scan - freed,
5245 I915_SHRINK_UNBOUND);
5247 mutex_unlock(&dev->struct_mutex);
5253 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
5255 struct drm_i915_private *dev_priv =
5256 container_of(nb, struct drm_i915_private, mm.oom_notifier);
5257 struct drm_device *dev = dev_priv->dev;
5258 struct drm_i915_gem_object *obj;
5259 unsigned long timeout = msecs_to_jiffies(5000) + 1;
5260 unsigned long pinned, bound, unbound, freed_pages;
5261 bool was_interruptible;
5264 while (!i915_gem_shrinker_lock(dev, &unlock) && --timeout) {
5265 schedule_timeout_killable(1);
5266 if (fatal_signal_pending(current))
5270 pr_err("Unable to purge GPU memory due lock contention.\n");
5274 was_interruptible = dev_priv->mm.interruptible;
5275 dev_priv->mm.interruptible = false;
5277 freed_pages = i915_gem_shrink_all(dev_priv);
5279 dev_priv->mm.interruptible = was_interruptible;
5281 /* Because we may be allocating inside our own driver, we cannot
5282 * assert that there are no objects with pinned pages that are not
5283 * being pointed to by hardware.
5285 unbound = bound = pinned = 0;
5286 list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) {
5287 if (!obj->base.filp) /* not backed by a freeable object */
5290 if (obj->pages_pin_count)
5291 pinned += obj->base.size;
5293 unbound += obj->base.size;
5295 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
5296 if (!obj->base.filp)
5299 if (obj->pages_pin_count)
5300 pinned += obj->base.size;
5302 bound += obj->base.size;
5306 mutex_unlock(&dev->struct_mutex);
5308 if (freed_pages || unbound || bound)
5309 pr_info("Purging GPU memory, %lu bytes freed, %lu bytes still pinned.\n",
5310 freed_pages << PAGE_SHIFT, pinned);
5311 if (unbound || bound)
5312 pr_err("%lu and %lu bytes still available in the "
5313 "bound and unbound GPU page lists.\n",
5316 *(unsigned long *)ptr += freed_pages;
5320 struct i915_vma *i915_gem_obj_to_ggtt(struct drm_i915_gem_object *obj)
5322 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj);
5323 struct i915_vma *vma;
5325 list_for_each_entry(vma, &obj->vma_list, vma_link)
5326 if (vma->vm == ggtt &&
5327 vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)