2 * Copyright © 2008-2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eric Anholt <eric@anholt.net>
29 #include <drm/drm_vma_manager.h>
30 #include <drm/i915_drm.h>
32 #include "i915_vgpu.h"
33 #include "i915_trace.h"
34 #include "intel_drv.h"
35 #include "intel_mocs.h"
36 #include <linux/shmem_fs.h>
37 #include <linux/slab.h>
38 #include <linux/swap.h>
39 #include <linux/pci.h>
40 #include <linux/dma-buf.h>
42 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);
43 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj);
45 i915_gem_object_retire__write(struct drm_i915_gem_object *obj);
47 i915_gem_object_retire__read(struct drm_i915_gem_object *obj, int ring);
49 static bool cpu_cache_is_coherent(struct drm_device *dev,
50 enum i915_cache_level level)
52 return HAS_LLC(dev) || level != I915_CACHE_NONE;
55 static bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj)
57 if (!cpu_cache_is_coherent(obj->base.dev, obj->cache_level))
60 return obj->pin_display;
63 /* some bookkeeping */
64 static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
67 spin_lock(&dev_priv->mm.object_stat_lock);
68 dev_priv->mm.object_count++;
69 dev_priv->mm.object_memory += size;
70 spin_unlock(&dev_priv->mm.object_stat_lock);
73 static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
76 spin_lock(&dev_priv->mm.object_stat_lock);
77 dev_priv->mm.object_count--;
78 dev_priv->mm.object_memory -= size;
79 spin_unlock(&dev_priv->mm.object_stat_lock);
83 i915_gem_wait_for_error(struct i915_gpu_error *error)
87 if (!i915_reset_in_progress(error))
91 * Only wait 10 seconds for the gpu reset to complete to avoid hanging
92 * userspace. If it takes that long something really bad is going on and
93 * we should simply try to bail out and fail as gracefully as possible.
95 ret = wait_event_interruptible_timeout(error->reset_queue,
96 !i915_reset_in_progress(error),
99 DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
101 } else if (ret < 0) {
108 int i915_mutex_lock_interruptible(struct drm_device *dev)
110 struct drm_i915_private *dev_priv = dev->dev_private;
113 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
117 ret = mutex_lock_interruptible(&dev->struct_mutex);
121 WARN_ON(i915_verify_lists(dev));
126 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
127 struct drm_file *file)
129 struct drm_i915_private *dev_priv = to_i915(dev);
130 struct i915_ggtt *ggtt = &dev_priv->ggtt;
131 struct drm_i915_gem_get_aperture *args = data;
132 struct i915_vma *vma;
136 mutex_lock(&dev->struct_mutex);
137 list_for_each_entry(vma, &ggtt->base.active_list, vm_link)
139 pinned += vma->node.size;
140 list_for_each_entry(vma, &ggtt->base.inactive_list, vm_link)
142 pinned += vma->node.size;
143 mutex_unlock(&dev->struct_mutex);
145 args->aper_size = ggtt->base.total;
146 args->aper_available_size = args->aper_size - pinned;
152 i915_gem_object_get_pages_phys(struct drm_i915_gem_object *obj)
154 struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
155 char *vaddr = obj->phys_handle->vaddr;
157 struct scatterlist *sg;
160 if (WARN_ON(i915_gem_object_needs_bit17_swizzle(obj)))
163 for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
167 page = shmem_read_mapping_page(mapping, i);
169 return PTR_ERR(page);
171 src = kmap_atomic(page);
172 memcpy(vaddr, src, PAGE_SIZE);
173 drm_clflush_virt_range(vaddr, PAGE_SIZE);
180 i915_gem_chipset_flush(obj->base.dev);
182 st = kmalloc(sizeof(*st), GFP_KERNEL);
186 if (sg_alloc_table(st, 1, GFP_KERNEL)) {
193 sg->length = obj->base.size;
195 sg_dma_address(sg) = obj->phys_handle->busaddr;
196 sg_dma_len(sg) = obj->base.size;
203 i915_gem_object_put_pages_phys(struct drm_i915_gem_object *obj)
207 BUG_ON(obj->madv == __I915_MADV_PURGED);
209 ret = i915_gem_object_set_to_cpu_domain(obj, true);
211 /* In the event of a disaster, abandon all caches and
214 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
217 if (obj->madv == I915_MADV_DONTNEED)
221 struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
222 char *vaddr = obj->phys_handle->vaddr;
225 for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
229 page = shmem_read_mapping_page(mapping, i);
233 dst = kmap_atomic(page);
234 drm_clflush_virt_range(vaddr, PAGE_SIZE);
235 memcpy(dst, vaddr, PAGE_SIZE);
238 set_page_dirty(page);
239 if (obj->madv == I915_MADV_WILLNEED)
240 mark_page_accessed(page);
247 sg_free_table(obj->pages);
252 i915_gem_object_release_phys(struct drm_i915_gem_object *obj)
254 drm_pci_free(obj->base.dev, obj->phys_handle);
257 static const struct drm_i915_gem_object_ops i915_gem_phys_ops = {
258 .get_pages = i915_gem_object_get_pages_phys,
259 .put_pages = i915_gem_object_put_pages_phys,
260 .release = i915_gem_object_release_phys,
264 drop_pages(struct drm_i915_gem_object *obj)
266 struct i915_vma *vma, *next;
269 drm_gem_object_reference(&obj->base);
270 list_for_each_entry_safe(vma, next, &obj->vma_list, obj_link)
271 if (i915_vma_unbind(vma))
274 ret = i915_gem_object_put_pages(obj);
275 drm_gem_object_unreference(&obj->base);
281 i915_gem_object_attach_phys(struct drm_i915_gem_object *obj,
284 drm_dma_handle_t *phys;
287 if (obj->phys_handle) {
288 if ((unsigned long)obj->phys_handle->vaddr & (align -1))
294 if (obj->madv != I915_MADV_WILLNEED)
297 if (obj->base.filp == NULL)
300 ret = drop_pages(obj);
304 /* create a new object */
305 phys = drm_pci_alloc(obj->base.dev, obj->base.size, align);
309 obj->phys_handle = phys;
310 obj->ops = &i915_gem_phys_ops;
312 return i915_gem_object_get_pages(obj);
316 i915_gem_phys_pwrite(struct drm_i915_gem_object *obj,
317 struct drm_i915_gem_pwrite *args,
318 struct drm_file *file_priv)
320 struct drm_device *dev = obj->base.dev;
321 void *vaddr = obj->phys_handle->vaddr + args->offset;
322 char __user *user_data = u64_to_user_ptr(args->data_ptr);
325 /* We manually control the domain here and pretend that it
326 * remains coherent i.e. in the GTT domain, like shmem_pwrite.
328 ret = i915_gem_object_wait_rendering(obj, false);
332 intel_fb_obj_invalidate(obj, ORIGIN_CPU);
333 if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
334 unsigned long unwritten;
336 /* The physical object once assigned is fixed for the lifetime
337 * of the obj, so we can safely drop the lock and continue
340 mutex_unlock(&dev->struct_mutex);
341 unwritten = copy_from_user(vaddr, user_data, args->size);
342 mutex_lock(&dev->struct_mutex);
349 drm_clflush_virt_range(vaddr, args->size);
350 i915_gem_chipset_flush(dev);
353 intel_fb_obj_flush(obj, false, ORIGIN_CPU);
357 void *i915_gem_object_alloc(struct drm_device *dev)
359 struct drm_i915_private *dev_priv = dev->dev_private;
360 return kmem_cache_zalloc(dev_priv->objects, GFP_KERNEL);
363 void i915_gem_object_free(struct drm_i915_gem_object *obj)
365 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
366 kmem_cache_free(dev_priv->objects, obj);
370 i915_gem_create(struct drm_file *file,
371 struct drm_device *dev,
375 struct drm_i915_gem_object *obj;
379 size = roundup(size, PAGE_SIZE);
383 /* Allocate the new object */
384 obj = i915_gem_alloc_object(dev, size);
388 ret = drm_gem_handle_create(file, &obj->base, &handle);
389 /* drop reference from allocate - handle holds it now */
390 drm_gem_object_unreference_unlocked(&obj->base);
399 i915_gem_dumb_create(struct drm_file *file,
400 struct drm_device *dev,
401 struct drm_mode_create_dumb *args)
403 /* have to work out size/pitch and return them */
404 args->pitch = ALIGN(args->width * DIV_ROUND_UP(args->bpp, 8), 64);
405 args->size = args->pitch * args->height;
406 return i915_gem_create(file, dev,
407 args->size, &args->handle);
411 * Creates a new mm object and returns a handle to it.
414 i915_gem_create_ioctl(struct drm_device *dev, void *data,
415 struct drm_file *file)
417 struct drm_i915_gem_create *args = data;
419 return i915_gem_create(file, dev,
420 args->size, &args->handle);
424 __copy_to_user_swizzled(char __user *cpu_vaddr,
425 const char *gpu_vaddr, int gpu_offset,
428 int ret, cpu_offset = 0;
431 int cacheline_end = ALIGN(gpu_offset + 1, 64);
432 int this_length = min(cacheline_end - gpu_offset, length);
433 int swizzled_gpu_offset = gpu_offset ^ 64;
435 ret = __copy_to_user(cpu_vaddr + cpu_offset,
436 gpu_vaddr + swizzled_gpu_offset,
441 cpu_offset += this_length;
442 gpu_offset += this_length;
443 length -= this_length;
450 __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset,
451 const char __user *cpu_vaddr,
454 int ret, cpu_offset = 0;
457 int cacheline_end = ALIGN(gpu_offset + 1, 64);
458 int this_length = min(cacheline_end - gpu_offset, length);
459 int swizzled_gpu_offset = gpu_offset ^ 64;
461 ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset,
462 cpu_vaddr + cpu_offset,
467 cpu_offset += this_length;
468 gpu_offset += this_length;
469 length -= this_length;
476 * Pins the specified object's pages and synchronizes the object with
477 * GPU accesses. Sets needs_clflush to non-zero if the caller should
478 * flush the object from the CPU cache.
480 int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj,
487 if (WARN_ON((obj->ops->flags & I915_GEM_OBJECT_HAS_STRUCT_PAGE) == 0))
490 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) {
491 /* If we're not in the cpu read domain, set ourself into the gtt
492 * read domain and manually flush cachelines (if required). This
493 * optimizes for the case when the gpu will dirty the data
494 * anyway again before the next pread happens. */
495 *needs_clflush = !cpu_cache_is_coherent(obj->base.dev,
497 ret = i915_gem_object_wait_rendering(obj, true);
502 ret = i915_gem_object_get_pages(obj);
506 i915_gem_object_pin_pages(obj);
511 /* Per-page copy function for the shmem pread fastpath.
512 * Flushes invalid cachelines before reading the target if
513 * needs_clflush is set. */
515 shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length,
516 char __user *user_data,
517 bool page_do_bit17_swizzling, bool needs_clflush)
522 if (unlikely(page_do_bit17_swizzling))
525 vaddr = kmap_atomic(page);
527 drm_clflush_virt_range(vaddr + shmem_page_offset,
529 ret = __copy_to_user_inatomic(user_data,
530 vaddr + shmem_page_offset,
532 kunmap_atomic(vaddr);
534 return ret ? -EFAULT : 0;
538 shmem_clflush_swizzled_range(char *addr, unsigned long length,
541 if (unlikely(swizzled)) {
542 unsigned long start = (unsigned long) addr;
543 unsigned long end = (unsigned long) addr + length;
545 /* For swizzling simply ensure that we always flush both
546 * channels. Lame, but simple and it works. Swizzled
547 * pwrite/pread is far from a hotpath - current userspace
548 * doesn't use it at all. */
549 start = round_down(start, 128);
550 end = round_up(end, 128);
552 drm_clflush_virt_range((void *)start, end - start);
554 drm_clflush_virt_range(addr, length);
559 /* Only difference to the fast-path function is that this can handle bit17
560 * and uses non-atomic copy and kmap functions. */
562 shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length,
563 char __user *user_data,
564 bool page_do_bit17_swizzling, bool needs_clflush)
571 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
573 page_do_bit17_swizzling);
575 if (page_do_bit17_swizzling)
576 ret = __copy_to_user_swizzled(user_data,
577 vaddr, shmem_page_offset,
580 ret = __copy_to_user(user_data,
581 vaddr + shmem_page_offset,
585 return ret ? - EFAULT : 0;
589 i915_gem_shmem_pread(struct drm_device *dev,
590 struct drm_i915_gem_object *obj,
591 struct drm_i915_gem_pread *args,
592 struct drm_file *file)
594 char __user *user_data;
597 int shmem_page_offset, page_length, ret = 0;
598 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
600 int needs_clflush = 0;
601 struct sg_page_iter sg_iter;
603 user_data = u64_to_user_ptr(args->data_ptr);
606 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
608 ret = i915_gem_obj_prepare_shmem_read(obj, &needs_clflush);
612 offset = args->offset;
614 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
615 offset >> PAGE_SHIFT) {
616 struct page *page = sg_page_iter_page(&sg_iter);
621 /* Operation in this page
623 * shmem_page_offset = offset within page in shmem file
624 * page_length = bytes to copy for this page
626 shmem_page_offset = offset_in_page(offset);
627 page_length = remain;
628 if ((shmem_page_offset + page_length) > PAGE_SIZE)
629 page_length = PAGE_SIZE - shmem_page_offset;
631 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
632 (page_to_phys(page) & (1 << 17)) != 0;
634 ret = shmem_pread_fast(page, shmem_page_offset, page_length,
635 user_data, page_do_bit17_swizzling,
640 mutex_unlock(&dev->struct_mutex);
642 if (likely(!i915.prefault_disable) && !prefaulted) {
643 ret = fault_in_multipages_writeable(user_data, remain);
644 /* Userspace is tricking us, but we've already clobbered
645 * its pages with the prefault and promised to write the
646 * data up to the first fault. Hence ignore any errors
647 * and just continue. */
652 ret = shmem_pread_slow(page, shmem_page_offset, page_length,
653 user_data, page_do_bit17_swizzling,
656 mutex_lock(&dev->struct_mutex);
662 remain -= page_length;
663 user_data += page_length;
664 offset += page_length;
668 i915_gem_object_unpin_pages(obj);
674 * Reads data from the object referenced by handle.
676 * On error, the contents of *data are undefined.
679 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
680 struct drm_file *file)
682 struct drm_i915_gem_pread *args = data;
683 struct drm_i915_gem_object *obj;
689 if (!access_ok(VERIFY_WRITE,
690 u64_to_user_ptr(args->data_ptr),
694 ret = i915_mutex_lock_interruptible(dev);
698 obj = to_intel_bo(drm_gem_object_lookup(file, args->handle));
699 if (&obj->base == NULL) {
704 /* Bounds check source. */
705 if (args->offset > obj->base.size ||
706 args->size > obj->base.size - args->offset) {
711 /* prime objects have no backing filp to GEM pread/pwrite
714 if (!obj->base.filp) {
719 trace_i915_gem_object_pread(obj, args->offset, args->size);
721 ret = i915_gem_shmem_pread(dev, obj, args, file);
724 drm_gem_object_unreference(&obj->base);
726 mutex_unlock(&dev->struct_mutex);
730 /* This is the fast write path which cannot handle
731 * page faults in the source data
735 fast_user_write(struct io_mapping *mapping,
736 loff_t page_base, int page_offset,
737 char __user *user_data,
740 void __iomem *vaddr_atomic;
742 unsigned long unwritten;
744 vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
745 /* We can use the cpu mem copy function because this is X86. */
746 vaddr = (void __force*)vaddr_atomic + page_offset;
747 unwritten = __copy_from_user_inatomic_nocache(vaddr,
749 io_mapping_unmap_atomic(vaddr_atomic);
754 * This is the fast pwrite path, where we copy the data directly from the
755 * user into the GTT, uncached.
758 i915_gem_gtt_pwrite_fast(struct drm_device *dev,
759 struct drm_i915_gem_object *obj,
760 struct drm_i915_gem_pwrite *args,
761 struct drm_file *file)
763 struct drm_i915_private *dev_priv = to_i915(dev);
764 struct i915_ggtt *ggtt = &dev_priv->ggtt;
766 loff_t offset, page_base;
767 char __user *user_data;
768 int page_offset, page_length, ret;
770 ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE | PIN_NONBLOCK);
774 ret = i915_gem_object_set_to_gtt_domain(obj, true);
778 ret = i915_gem_object_put_fence(obj);
782 user_data = u64_to_user_ptr(args->data_ptr);
785 offset = i915_gem_obj_ggtt_offset(obj) + args->offset;
787 intel_fb_obj_invalidate(obj, ORIGIN_GTT);
790 /* Operation in this page
792 * page_base = page offset within aperture
793 * page_offset = offset within page
794 * page_length = bytes to copy for this page
796 page_base = offset & PAGE_MASK;
797 page_offset = offset_in_page(offset);
798 page_length = remain;
799 if ((page_offset + remain) > PAGE_SIZE)
800 page_length = PAGE_SIZE - page_offset;
802 /* If we get a fault while copying data, then (presumably) our
803 * source page isn't available. Return the error and we'll
804 * retry in the slow path.
806 if (fast_user_write(ggtt->mappable, page_base,
807 page_offset, user_data, page_length)) {
812 remain -= page_length;
813 user_data += page_length;
814 offset += page_length;
818 intel_fb_obj_flush(obj, false, ORIGIN_GTT);
820 i915_gem_object_ggtt_unpin(obj);
825 /* Per-page copy function for the shmem pwrite fastpath.
826 * Flushes invalid cachelines before writing to the target if
827 * needs_clflush_before is set and flushes out any written cachelines after
828 * writing if needs_clflush is set. */
830 shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
831 char __user *user_data,
832 bool page_do_bit17_swizzling,
833 bool needs_clflush_before,
834 bool needs_clflush_after)
839 if (unlikely(page_do_bit17_swizzling))
842 vaddr = kmap_atomic(page);
843 if (needs_clflush_before)
844 drm_clflush_virt_range(vaddr + shmem_page_offset,
846 ret = __copy_from_user_inatomic(vaddr + shmem_page_offset,
847 user_data, page_length);
848 if (needs_clflush_after)
849 drm_clflush_virt_range(vaddr + shmem_page_offset,
851 kunmap_atomic(vaddr);
853 return ret ? -EFAULT : 0;
856 /* Only difference to the fast-path function is that this can handle bit17
857 * and uses non-atomic copy and kmap functions. */
859 shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
860 char __user *user_data,
861 bool page_do_bit17_swizzling,
862 bool needs_clflush_before,
863 bool needs_clflush_after)
869 if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
870 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
872 page_do_bit17_swizzling);
873 if (page_do_bit17_swizzling)
874 ret = __copy_from_user_swizzled(vaddr, shmem_page_offset,
878 ret = __copy_from_user(vaddr + shmem_page_offset,
881 if (needs_clflush_after)
882 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
884 page_do_bit17_swizzling);
887 return ret ? -EFAULT : 0;
891 i915_gem_shmem_pwrite(struct drm_device *dev,
892 struct drm_i915_gem_object *obj,
893 struct drm_i915_gem_pwrite *args,
894 struct drm_file *file)
898 char __user *user_data;
899 int shmem_page_offset, page_length, ret = 0;
900 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
901 int hit_slowpath = 0;
902 int needs_clflush_after = 0;
903 int needs_clflush_before = 0;
904 struct sg_page_iter sg_iter;
906 user_data = u64_to_user_ptr(args->data_ptr);
909 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
911 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
912 /* If we're not in the cpu write domain, set ourself into the gtt
913 * write domain and manually flush cachelines (if required). This
914 * optimizes for the case when the gpu will use the data
915 * right away and we therefore have to clflush anyway. */
916 needs_clflush_after = cpu_write_needs_clflush(obj);
917 ret = i915_gem_object_wait_rendering(obj, false);
921 /* Same trick applies to invalidate partially written cachelines read
923 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0)
924 needs_clflush_before =
925 !cpu_cache_is_coherent(dev, obj->cache_level);
927 ret = i915_gem_object_get_pages(obj);
931 intel_fb_obj_invalidate(obj, ORIGIN_CPU);
933 i915_gem_object_pin_pages(obj);
935 offset = args->offset;
938 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
939 offset >> PAGE_SHIFT) {
940 struct page *page = sg_page_iter_page(&sg_iter);
941 int partial_cacheline_write;
946 /* Operation in this page
948 * shmem_page_offset = offset within page in shmem file
949 * page_length = bytes to copy for this page
951 shmem_page_offset = offset_in_page(offset);
953 page_length = remain;
954 if ((shmem_page_offset + page_length) > PAGE_SIZE)
955 page_length = PAGE_SIZE - shmem_page_offset;
957 /* If we don't overwrite a cacheline completely we need to be
958 * careful to have up-to-date data by first clflushing. Don't
959 * overcomplicate things and flush the entire patch. */
960 partial_cacheline_write = needs_clflush_before &&
961 ((shmem_page_offset | page_length)
962 & (boot_cpu_data.x86_clflush_size - 1));
964 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
965 (page_to_phys(page) & (1 << 17)) != 0;
967 ret = shmem_pwrite_fast(page, shmem_page_offset, page_length,
968 user_data, page_do_bit17_swizzling,
969 partial_cacheline_write,
970 needs_clflush_after);
975 mutex_unlock(&dev->struct_mutex);
976 ret = shmem_pwrite_slow(page, shmem_page_offset, page_length,
977 user_data, page_do_bit17_swizzling,
978 partial_cacheline_write,
979 needs_clflush_after);
981 mutex_lock(&dev->struct_mutex);
987 remain -= page_length;
988 user_data += page_length;
989 offset += page_length;
993 i915_gem_object_unpin_pages(obj);
997 * Fixup: Flush cpu caches in case we didn't flush the dirty
998 * cachelines in-line while writing and the object moved
999 * out of the cpu write domain while we've dropped the lock.
1001 if (!needs_clflush_after &&
1002 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
1003 if (i915_gem_clflush_object(obj, obj->pin_display))
1004 needs_clflush_after = true;
1008 if (needs_clflush_after)
1009 i915_gem_chipset_flush(dev);
1011 obj->cache_dirty = true;
1013 intel_fb_obj_flush(obj, false, ORIGIN_CPU);
1018 * Writes data to the object referenced by handle.
1020 * On error, the contents of the buffer that were to be modified are undefined.
1023 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
1024 struct drm_file *file)
1026 struct drm_i915_private *dev_priv = dev->dev_private;
1027 struct drm_i915_gem_pwrite *args = data;
1028 struct drm_i915_gem_object *obj;
1031 if (args->size == 0)
1034 if (!access_ok(VERIFY_READ,
1035 u64_to_user_ptr(args->data_ptr),
1039 if (likely(!i915.prefault_disable)) {
1040 ret = fault_in_multipages_readable(u64_to_user_ptr(args->data_ptr),
1046 intel_runtime_pm_get(dev_priv);
1048 ret = i915_mutex_lock_interruptible(dev);
1052 obj = to_intel_bo(drm_gem_object_lookup(file, args->handle));
1053 if (&obj->base == NULL) {
1058 /* Bounds check destination. */
1059 if (args->offset > obj->base.size ||
1060 args->size > obj->base.size - args->offset) {
1065 /* prime objects have no backing filp to GEM pread/pwrite
1068 if (!obj->base.filp) {
1073 trace_i915_gem_object_pwrite(obj, args->offset, args->size);
1076 /* We can only do the GTT pwrite on untiled buffers, as otherwise
1077 * it would end up going through the fenced access, and we'll get
1078 * different detiling behavior between reading and writing.
1079 * pread/pwrite currently are reading and writing from the CPU
1080 * perspective, requiring manual detiling by the client.
1082 if (obj->tiling_mode == I915_TILING_NONE &&
1083 obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
1084 cpu_write_needs_clflush(obj)) {
1085 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
1086 /* Note that the gtt paths might fail with non-page-backed user
1087 * pointers (e.g. gtt mappings when moving data between
1088 * textures). Fallback to the shmem path in that case. */
1091 if (ret == -EFAULT || ret == -ENOSPC) {
1092 if (obj->phys_handle)
1093 ret = i915_gem_phys_pwrite(obj, args, file);
1095 ret = i915_gem_shmem_pwrite(dev, obj, args, file);
1099 drm_gem_object_unreference(&obj->base);
1101 mutex_unlock(&dev->struct_mutex);
1103 intel_runtime_pm_put(dev_priv);
1109 i915_gem_check_wedge(unsigned reset_counter, bool interruptible)
1111 if (__i915_terminally_wedged(reset_counter))
1114 if (__i915_reset_in_progress(reset_counter)) {
1115 /* Non-interruptible callers can't handle -EAGAIN, hence return
1116 * -EIO unconditionally for these. */
1126 static void fake_irq(unsigned long data)
1128 wake_up_process((struct task_struct *)data);
1131 static bool missed_irq(struct drm_i915_private *dev_priv,
1132 struct intel_engine_cs *engine)
1134 return test_bit(engine->id, &dev_priv->gpu_error.missed_irq_rings);
1137 static unsigned long local_clock_us(unsigned *cpu)
1141 /* Cheaply and approximately convert from nanoseconds to microseconds.
1142 * The result and subsequent calculations are also defined in the same
1143 * approximate microseconds units. The principal source of timing
1144 * error here is from the simple truncation.
1146 * Note that local_clock() is only defined wrt to the current CPU;
1147 * the comparisons are no longer valid if we switch CPUs. Instead of
1148 * blocking preemption for the entire busywait, we can detect the CPU
1149 * switch and use that as indicator of system load and a reason to
1150 * stop busywaiting, see busywait_stop().
1153 t = local_clock() >> 10;
1159 static bool busywait_stop(unsigned long timeout, unsigned cpu)
1163 if (time_after(local_clock_us(&this_cpu), timeout))
1166 return this_cpu != cpu;
1169 static int __i915_spin_request(struct drm_i915_gem_request *req, int state)
1171 unsigned long timeout;
1174 /* When waiting for high frequency requests, e.g. during synchronous
1175 * rendering split between the CPU and GPU, the finite amount of time
1176 * required to set up the irq and wait upon it limits the response
1177 * rate. By busywaiting on the request completion for a short while we
1178 * can service the high frequency waits as quick as possible. However,
1179 * if it is a slow request, we want to sleep as quickly as possible.
1180 * The tradeoff between waiting and sleeping is roughly the time it
1181 * takes to sleep on a request, on the order of a microsecond.
1184 if (req->engine->irq_refcount)
1187 /* Only spin if we know the GPU is processing this request */
1188 if (!i915_gem_request_started(req, true))
1191 timeout = local_clock_us(&cpu) + 5;
1192 while (!need_resched()) {
1193 if (i915_gem_request_completed(req, true))
1196 if (signal_pending_state(state, current))
1199 if (busywait_stop(timeout, cpu))
1202 cpu_relax_lowlatency();
1205 if (i915_gem_request_completed(req, false))
1212 * __i915_wait_request - wait until execution of request has finished
1214 * @interruptible: do an interruptible wait (normally yes)
1215 * @timeout: in - how long to wait (NULL forever); out - how much time remaining
1217 * Note: It is of utmost importance that the passed in seqno and reset_counter
1218 * values have been read by the caller in an smp safe manner. Where read-side
1219 * locks are involved, it is sufficient to read the reset_counter before
1220 * unlocking the lock that protects the seqno. For lockless tricks, the
1221 * reset_counter _must_ be read before, and an appropriate smp_rmb must be
1224 * Returns 0 if the request was found within the alloted time. Else returns the
1225 * errno with remaining time filled in timeout argument.
1227 int __i915_wait_request(struct drm_i915_gem_request *req,
1230 struct intel_rps_client *rps)
1232 struct intel_engine_cs *engine = i915_gem_request_get_engine(req);
1233 struct drm_device *dev = engine->dev;
1234 struct drm_i915_private *dev_priv = dev->dev_private;
1235 const bool irq_test_in_progress =
1236 ACCESS_ONCE(dev_priv->gpu_error.test_irq_rings) & intel_engine_flag(engine);
1237 int state = interruptible ? TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;
1239 unsigned long timeout_expire;
1240 s64 before = 0; /* Only to silence a compiler warning. */
1243 WARN(!intel_irqs_enabled(dev_priv), "IRQs disabled");
1245 if (list_empty(&req->list))
1248 if (i915_gem_request_completed(req, true))
1253 if (WARN_ON(*timeout < 0))
1259 timeout_expire = jiffies + nsecs_to_jiffies_timeout(*timeout);
1262 * Record current time in case interrupted by signal, or wedged.
1264 before = ktime_get_raw_ns();
1267 if (INTEL_INFO(dev_priv)->gen >= 6)
1268 gen6_rps_boost(dev_priv, rps, req->emitted_jiffies);
1270 trace_i915_gem_request_wait_begin(req);
1272 /* Optimistic spin for the next jiffie before touching IRQs */
1273 ret = __i915_spin_request(req, state);
1277 if (!irq_test_in_progress && WARN_ON(!engine->irq_get(engine))) {
1283 struct timer_list timer;
1285 prepare_to_wait(&engine->irq_queue, &wait, state);
1287 /* We need to check whether any gpu reset happened in between
1288 * the request being submitted and now. If a reset has occurred,
1289 * the request is effectively complete (we either are in the
1290 * process of or have discarded the rendering and completely
1291 * reset the GPU. The results of the request are lost and we
1292 * are free to continue on with the original operation.
1294 if (req->reset_counter != i915_reset_counter(&dev_priv->gpu_error)) {
1299 if (i915_gem_request_completed(req, false)) {
1304 if (signal_pending_state(state, current)) {
1309 if (timeout && time_after_eq(jiffies, timeout_expire)) {
1314 timer.function = NULL;
1315 if (timeout || missed_irq(dev_priv, engine)) {
1316 unsigned long expire;
1318 setup_timer_on_stack(&timer, fake_irq, (unsigned long)current);
1319 expire = missed_irq(dev_priv, engine) ? jiffies + 1 : timeout_expire;
1320 mod_timer(&timer, expire);
1325 if (timer.function) {
1326 del_singleshot_timer_sync(&timer);
1327 destroy_timer_on_stack(&timer);
1330 if (!irq_test_in_progress)
1331 engine->irq_put(engine);
1333 finish_wait(&engine->irq_queue, &wait);
1336 trace_i915_gem_request_wait_end(req);
1339 s64 tres = *timeout - (ktime_get_raw_ns() - before);
1341 *timeout = tres < 0 ? 0 : tres;
1344 * Apparently ktime isn't accurate enough and occasionally has a
1345 * bit of mismatch in the jiffies<->nsecs<->ktime loop. So patch
1346 * things up to make the test happy. We allow up to 1 jiffy.
1348 * This is a regrssion from the timespec->ktime conversion.
1350 if (ret == -ETIME && *timeout < jiffies_to_usecs(1)*1000)
1357 int i915_gem_request_add_to_client(struct drm_i915_gem_request *req,
1358 struct drm_file *file)
1360 struct drm_i915_file_private *file_priv;
1362 WARN_ON(!req || !file || req->file_priv);
1370 file_priv = file->driver_priv;
1372 spin_lock(&file_priv->mm.lock);
1373 req->file_priv = file_priv;
1374 list_add_tail(&req->client_list, &file_priv->mm.request_list);
1375 spin_unlock(&file_priv->mm.lock);
1377 req->pid = get_pid(task_pid(current));
1383 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
1385 struct drm_i915_file_private *file_priv = request->file_priv;
1390 spin_lock(&file_priv->mm.lock);
1391 list_del(&request->client_list);
1392 request->file_priv = NULL;
1393 spin_unlock(&file_priv->mm.lock);
1395 put_pid(request->pid);
1396 request->pid = NULL;
1399 static void i915_gem_request_retire(struct drm_i915_gem_request *request)
1401 trace_i915_gem_request_retire(request);
1403 /* We know the GPU must have read the request to have
1404 * sent us the seqno + interrupt, so use the position
1405 * of tail of the request to update the last known position
1408 * Note this requires that we are always called in request
1411 request->ringbuf->last_retired_head = request->postfix;
1413 list_del_init(&request->list);
1414 i915_gem_request_remove_from_client(request);
1416 i915_gem_request_unreference(request);
1420 __i915_gem_request_retire__upto(struct drm_i915_gem_request *req)
1422 struct intel_engine_cs *engine = req->engine;
1423 struct drm_i915_gem_request *tmp;
1425 lockdep_assert_held(&engine->dev->struct_mutex);
1427 if (list_empty(&req->list))
1431 tmp = list_first_entry(&engine->request_list,
1432 typeof(*tmp), list);
1434 i915_gem_request_retire(tmp);
1435 } while (tmp != req);
1437 WARN_ON(i915_verify_lists(engine->dev));
1441 * Waits for a request to be signaled, and cleans up the
1442 * request and object lists appropriately for that event.
1445 i915_wait_request(struct drm_i915_gem_request *req)
1447 struct drm_i915_private *dev_priv = req->i915;
1451 interruptible = dev_priv->mm.interruptible;
1453 BUG_ON(!mutex_is_locked(&dev_priv->dev->struct_mutex));
1455 ret = __i915_wait_request(req, interruptible, NULL, NULL);
1459 __i915_gem_request_retire__upto(req);
1464 * Ensures that all rendering to the object has completed and the object is
1465 * safe to unbind from the GTT or access from the CPU.
1468 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
1477 if (obj->last_write_req != NULL) {
1478 ret = i915_wait_request(obj->last_write_req);
1482 i = obj->last_write_req->engine->id;
1483 if (obj->last_read_req[i] == obj->last_write_req)
1484 i915_gem_object_retire__read(obj, i);
1486 i915_gem_object_retire__write(obj);
1489 for (i = 0; i < I915_NUM_ENGINES; i++) {
1490 if (obj->last_read_req[i] == NULL)
1493 ret = i915_wait_request(obj->last_read_req[i]);
1497 i915_gem_object_retire__read(obj, i);
1499 GEM_BUG_ON(obj->active);
1506 i915_gem_object_retire_request(struct drm_i915_gem_object *obj,
1507 struct drm_i915_gem_request *req)
1509 int ring = req->engine->id;
1511 if (obj->last_read_req[ring] == req)
1512 i915_gem_object_retire__read(obj, ring);
1513 else if (obj->last_write_req == req)
1514 i915_gem_object_retire__write(obj);
1516 __i915_gem_request_retire__upto(req);
1519 /* A nonblocking variant of the above wait. This is a highly dangerous routine
1520 * as the object state may change during this call.
1522 static __must_check int
1523 i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
1524 struct intel_rps_client *rps,
1527 struct drm_device *dev = obj->base.dev;
1528 struct drm_i915_private *dev_priv = dev->dev_private;
1529 struct drm_i915_gem_request *requests[I915_NUM_ENGINES];
1532 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1533 BUG_ON(!dev_priv->mm.interruptible);
1539 struct drm_i915_gem_request *req;
1541 req = obj->last_write_req;
1545 requests[n++] = i915_gem_request_reference(req);
1547 for (i = 0; i < I915_NUM_ENGINES; i++) {
1548 struct drm_i915_gem_request *req;
1550 req = obj->last_read_req[i];
1554 requests[n++] = i915_gem_request_reference(req);
1558 mutex_unlock(&dev->struct_mutex);
1560 for (i = 0; ret == 0 && i < n; i++)
1561 ret = __i915_wait_request(requests[i], true, NULL, rps);
1562 mutex_lock(&dev->struct_mutex);
1564 for (i = 0; i < n; i++) {
1566 i915_gem_object_retire_request(obj, requests[i]);
1567 i915_gem_request_unreference(requests[i]);
1573 static struct intel_rps_client *to_rps_client(struct drm_file *file)
1575 struct drm_i915_file_private *fpriv = file->driver_priv;
1580 * Called when user space prepares to use an object with the CPU, either
1581 * through the mmap ioctl's mapping or a GTT mapping.
1584 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1585 struct drm_file *file)
1587 struct drm_i915_gem_set_domain *args = data;
1588 struct drm_i915_gem_object *obj;
1589 uint32_t read_domains = args->read_domains;
1590 uint32_t write_domain = args->write_domain;
1593 /* Only handle setting domains to types used by the CPU. */
1594 if (write_domain & I915_GEM_GPU_DOMAINS)
1597 if (read_domains & I915_GEM_GPU_DOMAINS)
1600 /* Having something in the write domain implies it's in the read
1601 * domain, and only that read domain. Enforce that in the request.
1603 if (write_domain != 0 && read_domains != write_domain)
1606 ret = i915_mutex_lock_interruptible(dev);
1610 obj = to_intel_bo(drm_gem_object_lookup(file, args->handle));
1611 if (&obj->base == NULL) {
1616 /* Try to flush the object off the GPU without holding the lock.
1617 * We will repeat the flush holding the lock in the normal manner
1618 * to catch cases where we are gazumped.
1620 ret = i915_gem_object_wait_rendering__nonblocking(obj,
1621 to_rps_client(file),
1626 if (read_domains & I915_GEM_DOMAIN_GTT)
1627 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1629 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1631 if (write_domain != 0)
1632 intel_fb_obj_invalidate(obj,
1633 write_domain == I915_GEM_DOMAIN_GTT ?
1634 ORIGIN_GTT : ORIGIN_CPU);
1637 drm_gem_object_unreference(&obj->base);
1639 mutex_unlock(&dev->struct_mutex);
1644 * Called when user space has done writes to this buffer
1647 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1648 struct drm_file *file)
1650 struct drm_i915_gem_sw_finish *args = data;
1651 struct drm_i915_gem_object *obj;
1654 ret = i915_mutex_lock_interruptible(dev);
1658 obj = to_intel_bo(drm_gem_object_lookup(file, args->handle));
1659 if (&obj->base == NULL) {
1664 /* Pinned buffers may be scanout, so flush the cache */
1665 if (obj->pin_display)
1666 i915_gem_object_flush_cpu_write_domain(obj);
1668 drm_gem_object_unreference(&obj->base);
1670 mutex_unlock(&dev->struct_mutex);
1675 * Maps the contents of an object, returning the address it is mapped
1678 * While the mapping holds a reference on the contents of the object, it doesn't
1679 * imply a ref on the object itself.
1683 * DRM driver writers who look a this function as an example for how to do GEM
1684 * mmap support, please don't implement mmap support like here. The modern way
1685 * to implement DRM mmap support is with an mmap offset ioctl (like
1686 * i915_gem_mmap_gtt) and then using the mmap syscall on the DRM fd directly.
1687 * That way debug tooling like valgrind will understand what's going on, hiding
1688 * the mmap call in a driver private ioctl will break that. The i915 driver only
1689 * does cpu mmaps this way because we didn't know better.
1692 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1693 struct drm_file *file)
1695 struct drm_i915_gem_mmap *args = data;
1696 struct drm_gem_object *obj;
1699 if (args->flags & ~(I915_MMAP_WC))
1702 if (args->flags & I915_MMAP_WC && !boot_cpu_has(X86_FEATURE_PAT))
1705 obj = drm_gem_object_lookup(file, args->handle);
1709 /* prime objects have no backing filp to GEM mmap
1713 drm_gem_object_unreference_unlocked(obj);
1717 addr = vm_mmap(obj->filp, 0, args->size,
1718 PROT_READ | PROT_WRITE, MAP_SHARED,
1720 if (args->flags & I915_MMAP_WC) {
1721 struct mm_struct *mm = current->mm;
1722 struct vm_area_struct *vma;
1724 if (down_write_killable(&mm->mmap_sem)) {
1725 drm_gem_object_unreference_unlocked(obj);
1728 vma = find_vma(mm, addr);
1731 pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
1734 up_write(&mm->mmap_sem);
1736 drm_gem_object_unreference_unlocked(obj);
1737 if (IS_ERR((void *)addr))
1740 args->addr_ptr = (uint64_t) addr;
1746 * i915_gem_fault - fault a page into the GTT
1747 * @vma: VMA in question
1750 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1751 * from userspace. The fault handler takes care of binding the object to
1752 * the GTT (if needed), allocating and programming a fence register (again,
1753 * only if needed based on whether the old reg is still valid or the object
1754 * is tiled) and inserting a new PTE into the faulting process.
1756 * Note that the faulting process may involve evicting existing objects
1757 * from the GTT and/or fence registers to make room. So performance may
1758 * suffer if the GTT working set is large or there are few fence registers
1761 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1763 struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1764 struct drm_device *dev = obj->base.dev;
1765 struct drm_i915_private *dev_priv = to_i915(dev);
1766 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1767 struct i915_ggtt_view view = i915_ggtt_view_normal;
1768 pgoff_t page_offset;
1771 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1773 intel_runtime_pm_get(dev_priv);
1775 /* We don't use vmf->pgoff since that has the fake offset */
1776 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1779 ret = i915_mutex_lock_interruptible(dev);
1783 trace_i915_gem_object_fault(obj, page_offset, true, write);
1785 /* Try to flush the object off the GPU first without holding the lock.
1786 * Upon reacquiring the lock, we will perform our sanity checks and then
1787 * repeat the flush holding the lock in the normal manner to catch cases
1788 * where we are gazumped.
1790 ret = i915_gem_object_wait_rendering__nonblocking(obj, NULL, !write);
1794 /* Access to snoopable pages through the GTT is incoherent. */
1795 if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) {
1800 /* Use a partial view if the object is bigger than the aperture. */
1801 if (obj->base.size >= ggtt->mappable_end &&
1802 obj->tiling_mode == I915_TILING_NONE) {
1803 static const unsigned int chunk_size = 256; // 1 MiB
1805 memset(&view, 0, sizeof(view));
1806 view.type = I915_GGTT_VIEW_PARTIAL;
1807 view.params.partial.offset = rounddown(page_offset, chunk_size);
1808 view.params.partial.size =
1811 (vma->vm_end - vma->vm_start)/PAGE_SIZE -
1812 view.params.partial.offset);
1815 /* Now pin it into the GTT if needed */
1816 ret = i915_gem_object_ggtt_pin(obj, &view, 0, PIN_MAPPABLE);
1820 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1824 ret = i915_gem_object_get_fence(obj);
1828 /* Finally, remap it using the new GTT offset */
1829 pfn = ggtt->mappable_base +
1830 i915_gem_obj_ggtt_offset_view(obj, &view);
1833 if (unlikely(view.type == I915_GGTT_VIEW_PARTIAL)) {
1834 /* Overriding existing pages in partial view does not cause
1835 * us any trouble as TLBs are still valid because the fault
1836 * is due to userspace losing part of the mapping or never
1837 * having accessed it before (at this partials' range).
1839 unsigned long base = vma->vm_start +
1840 (view.params.partial.offset << PAGE_SHIFT);
1843 for (i = 0; i < view.params.partial.size; i++) {
1844 ret = vm_insert_pfn(vma, base + i * PAGE_SIZE, pfn + i);
1849 obj->fault_mappable = true;
1851 if (!obj->fault_mappable) {
1852 unsigned long size = min_t(unsigned long,
1853 vma->vm_end - vma->vm_start,
1857 for (i = 0; i < size >> PAGE_SHIFT; i++) {
1858 ret = vm_insert_pfn(vma,
1859 (unsigned long)vma->vm_start + i * PAGE_SIZE,
1865 obj->fault_mappable = true;
1867 ret = vm_insert_pfn(vma,
1868 (unsigned long)vmf->virtual_address,
1872 i915_gem_object_ggtt_unpin_view(obj, &view);
1874 mutex_unlock(&dev->struct_mutex);
1879 * We eat errors when the gpu is terminally wedged to avoid
1880 * userspace unduly crashing (gl has no provisions for mmaps to
1881 * fail). But any other -EIO isn't ours (e.g. swap in failure)
1882 * and so needs to be reported.
1884 if (!i915_terminally_wedged(&dev_priv->gpu_error)) {
1885 ret = VM_FAULT_SIGBUS;
1890 * EAGAIN means the gpu is hung and we'll wait for the error
1891 * handler to reset everything when re-faulting in
1892 * i915_mutex_lock_interruptible.
1899 * EBUSY is ok: this just means that another thread
1900 * already did the job.
1902 ret = VM_FAULT_NOPAGE;
1909 ret = VM_FAULT_SIGBUS;
1912 WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret);
1913 ret = VM_FAULT_SIGBUS;
1917 intel_runtime_pm_put(dev_priv);
1922 * i915_gem_release_mmap - remove physical page mappings
1923 * @obj: obj in question
1925 * Preserve the reservation of the mmapping with the DRM core code, but
1926 * relinquish ownership of the pages back to the system.
1928 * It is vital that we remove the page mapping if we have mapped a tiled
1929 * object through the GTT and then lose the fence register due to
1930 * resource pressure. Similarly if the object has been moved out of the
1931 * aperture, than pages mapped into userspace must be revoked. Removing the
1932 * mapping will then trigger a page fault on the next user access, allowing
1933 * fixup by i915_gem_fault().
1936 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1938 /* Serialisation between user GTT access and our code depends upon
1939 * revoking the CPU's PTE whilst the mutex is held. The next user
1940 * pagefault then has to wait until we release the mutex.
1942 lockdep_assert_held(&obj->base.dev->struct_mutex);
1944 if (!obj->fault_mappable)
1947 drm_vma_node_unmap(&obj->base.vma_node,
1948 obj->base.dev->anon_inode->i_mapping);
1950 /* Ensure that the CPU's PTE are revoked and there are not outstanding
1951 * memory transactions from userspace before we return. The TLB
1952 * flushing implied above by changing the PTE above *should* be
1953 * sufficient, an extra barrier here just provides us with a bit
1954 * of paranoid documentation about our requirement to serialise
1955 * memory writes before touching registers / GSM.
1959 obj->fault_mappable = false;
1963 i915_gem_release_all_mmaps(struct drm_i915_private *dev_priv)
1965 struct drm_i915_gem_object *obj;
1967 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list)
1968 i915_gem_release_mmap(obj);
1972 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1976 if (INTEL_INFO(dev)->gen >= 4 ||
1977 tiling_mode == I915_TILING_NONE)
1980 /* Previous chips need a power-of-two fence region when tiling */
1981 if (INTEL_INFO(dev)->gen == 3)
1982 gtt_size = 1024*1024;
1984 gtt_size = 512*1024;
1986 while (gtt_size < size)
1993 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1994 * @obj: object to check
1996 * Return the required GTT alignment for an object, taking into account
1997 * potential fence register mapping.
2000 i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size,
2001 int tiling_mode, bool fenced)
2004 * Minimum alignment is 4k (GTT page size), but might be greater
2005 * if a fence register is needed for the object.
2007 if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) ||
2008 tiling_mode == I915_TILING_NONE)
2012 * Previous chips need to be aligned to the size of the smallest
2013 * fence register that can contain the object.
2015 return i915_gem_get_gtt_size(dev, size, tiling_mode);
2018 static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj)
2020 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2023 dev_priv->mm.shrinker_no_lock_stealing = true;
2025 ret = drm_gem_create_mmap_offset(&obj->base);
2029 /* Badly fragmented mmap space? The only way we can recover
2030 * space is by destroying unwanted objects. We can't randomly release
2031 * mmap_offsets as userspace expects them to be persistent for the
2032 * lifetime of the objects. The closest we can is to release the
2033 * offsets on purgeable objects by truncating it and marking it purged,
2034 * which prevents userspace from ever using that object again.
2036 i915_gem_shrink(dev_priv,
2037 obj->base.size >> PAGE_SHIFT,
2039 I915_SHRINK_UNBOUND |
2040 I915_SHRINK_PURGEABLE);
2041 ret = drm_gem_create_mmap_offset(&obj->base);
2045 i915_gem_shrink_all(dev_priv);
2046 ret = drm_gem_create_mmap_offset(&obj->base);
2048 dev_priv->mm.shrinker_no_lock_stealing = false;
2053 static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)
2055 drm_gem_free_mmap_offset(&obj->base);
2059 i915_gem_mmap_gtt(struct drm_file *file,
2060 struct drm_device *dev,
2064 struct drm_i915_gem_object *obj;
2067 ret = i915_mutex_lock_interruptible(dev);
2071 obj = to_intel_bo(drm_gem_object_lookup(file, handle));
2072 if (&obj->base == NULL) {
2077 if (obj->madv != I915_MADV_WILLNEED) {
2078 DRM_DEBUG("Attempting to mmap a purgeable buffer\n");
2083 ret = i915_gem_object_create_mmap_offset(obj);
2087 *offset = drm_vma_node_offset_addr(&obj->base.vma_node);
2090 drm_gem_object_unreference(&obj->base);
2092 mutex_unlock(&dev->struct_mutex);
2097 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
2099 * @data: GTT mapping ioctl data
2100 * @file: GEM object info
2102 * Simply returns the fake offset to userspace so it can mmap it.
2103 * The mmap call will end up in drm_gem_mmap(), which will set things
2104 * up so we can get faults in the handler above.
2106 * The fault handler will take care of binding the object into the GTT
2107 * (since it may have been evicted to make room for something), allocating
2108 * a fence register, and mapping the appropriate aperture address into
2112 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
2113 struct drm_file *file)
2115 struct drm_i915_gem_mmap_gtt *args = data;
2117 return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
2120 /* Immediately discard the backing storage */
2122 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
2124 i915_gem_object_free_mmap_offset(obj);
2126 if (obj->base.filp == NULL)
2129 /* Our goal here is to return as much of the memory as
2130 * is possible back to the system as we are called from OOM.
2131 * To do this we must instruct the shmfs to drop all of its
2132 * backing pages, *now*.
2134 shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1);
2135 obj->madv = __I915_MADV_PURGED;
2138 /* Try to discard unwanted pages */
2140 i915_gem_object_invalidate(struct drm_i915_gem_object *obj)
2142 struct address_space *mapping;
2144 switch (obj->madv) {
2145 case I915_MADV_DONTNEED:
2146 i915_gem_object_truncate(obj);
2147 case __I915_MADV_PURGED:
2151 if (obj->base.filp == NULL)
2154 mapping = file_inode(obj->base.filp)->i_mapping,
2155 invalidate_mapping_pages(mapping, 0, (loff_t)-1);
2159 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
2161 struct sg_page_iter sg_iter;
2164 BUG_ON(obj->madv == __I915_MADV_PURGED);
2166 ret = i915_gem_object_set_to_cpu_domain(obj, true);
2168 /* In the event of a disaster, abandon all caches and
2169 * hope for the best.
2171 i915_gem_clflush_object(obj, true);
2172 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
2175 i915_gem_gtt_finish_object(obj);
2177 if (i915_gem_object_needs_bit17_swizzle(obj))
2178 i915_gem_object_save_bit_17_swizzle(obj);
2180 if (obj->madv == I915_MADV_DONTNEED)
2183 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
2184 struct page *page = sg_page_iter_page(&sg_iter);
2187 set_page_dirty(page);
2189 if (obj->madv == I915_MADV_WILLNEED)
2190 mark_page_accessed(page);
2196 sg_free_table(obj->pages);
2201 i915_gem_object_put_pages(struct drm_i915_gem_object *obj)
2203 const struct drm_i915_gem_object_ops *ops = obj->ops;
2205 if (obj->pages == NULL)
2208 if (obj->pages_pin_count)
2211 BUG_ON(i915_gem_obj_bound_any(obj));
2213 /* ->put_pages might need to allocate memory for the bit17 swizzle
2214 * array, hence protect them from being reaped by removing them from gtt
2216 list_del(&obj->global_list);
2219 if (is_vmalloc_addr(obj->mapping))
2220 vunmap(obj->mapping);
2222 kunmap(kmap_to_page(obj->mapping));
2223 obj->mapping = NULL;
2226 ops->put_pages(obj);
2229 i915_gem_object_invalidate(obj);
2235 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
2237 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2239 struct address_space *mapping;
2240 struct sg_table *st;
2241 struct scatterlist *sg;
2242 struct sg_page_iter sg_iter;
2244 unsigned long last_pfn = 0; /* suppress gcc warning */
2248 /* Assert that the object is not currently in any GPU domain. As it
2249 * wasn't in the GTT, there shouldn't be any way it could have been in
2252 BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
2253 BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
2255 st = kmalloc(sizeof(*st), GFP_KERNEL);
2259 page_count = obj->base.size / PAGE_SIZE;
2260 if (sg_alloc_table(st, page_count, GFP_KERNEL)) {
2265 /* Get the list of pages out of our struct file. They'll be pinned
2266 * at this point until we release them.
2268 * Fail silently without starting the shrinker
2270 mapping = file_inode(obj->base.filp)->i_mapping;
2271 gfp = mapping_gfp_constraint(mapping, ~(__GFP_IO | __GFP_RECLAIM));
2272 gfp |= __GFP_NORETRY | __GFP_NOWARN;
2275 for (i = 0; i < page_count; i++) {
2276 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2278 i915_gem_shrink(dev_priv,
2281 I915_SHRINK_UNBOUND |
2282 I915_SHRINK_PURGEABLE);
2283 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2286 /* We've tried hard to allocate the memory by reaping
2287 * our own buffer, now let the real VM do its job and
2288 * go down in flames if truly OOM.
2290 i915_gem_shrink_all(dev_priv);
2291 page = shmem_read_mapping_page(mapping, i);
2293 ret = PTR_ERR(page);
2297 #ifdef CONFIG_SWIOTLB
2298 if (swiotlb_nr_tbl()) {
2300 sg_set_page(sg, page, PAGE_SIZE, 0);
2305 if (!i || page_to_pfn(page) != last_pfn + 1) {
2309 sg_set_page(sg, page, PAGE_SIZE, 0);
2311 sg->length += PAGE_SIZE;
2313 last_pfn = page_to_pfn(page);
2315 /* Check that the i965g/gm workaround works. */
2316 WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
2318 #ifdef CONFIG_SWIOTLB
2319 if (!swiotlb_nr_tbl())
2324 ret = i915_gem_gtt_prepare_object(obj);
2328 if (i915_gem_object_needs_bit17_swizzle(obj))
2329 i915_gem_object_do_bit_17_swizzle(obj);
2331 if (obj->tiling_mode != I915_TILING_NONE &&
2332 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES)
2333 i915_gem_object_pin_pages(obj);
2339 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0)
2340 put_page(sg_page_iter_page(&sg_iter));
2344 /* shmemfs first checks if there is enough memory to allocate the page
2345 * and reports ENOSPC should there be insufficient, along with the usual
2346 * ENOMEM for a genuine allocation failure.
2348 * We use ENOSPC in our driver to mean that we have run out of aperture
2349 * space and so want to translate the error from shmemfs back to our
2350 * usual understanding of ENOMEM.
2358 /* Ensure that the associated pages are gathered from the backing storage
2359 * and pinned into our object. i915_gem_object_get_pages() may be called
2360 * multiple times before they are released by a single call to
2361 * i915_gem_object_put_pages() - once the pages are no longer referenced
2362 * either as a result of memory pressure (reaping pages under the shrinker)
2363 * or as the object is itself released.
2366 i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
2368 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2369 const struct drm_i915_gem_object_ops *ops = obj->ops;
2375 if (obj->madv != I915_MADV_WILLNEED) {
2376 DRM_DEBUG("Attempting to obtain a purgeable object\n");
2380 BUG_ON(obj->pages_pin_count);
2382 ret = ops->get_pages(obj);
2386 list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list);
2388 obj->get_page.sg = obj->pages->sgl;
2389 obj->get_page.last = 0;
2394 void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj)
2398 lockdep_assert_held(&obj->base.dev->struct_mutex);
2400 ret = i915_gem_object_get_pages(obj);
2402 return ERR_PTR(ret);
2404 i915_gem_object_pin_pages(obj);
2406 if (obj->mapping == NULL) {
2407 struct page **pages;
2410 if (obj->base.size == PAGE_SIZE)
2411 obj->mapping = kmap(sg_page(obj->pages->sgl));
2413 pages = drm_malloc_gfp(obj->base.size >> PAGE_SHIFT,
2416 if (pages != NULL) {
2417 struct sg_page_iter sg_iter;
2421 for_each_sg_page(obj->pages->sgl, &sg_iter,
2422 obj->pages->nents, 0)
2423 pages[n++] = sg_page_iter_page(&sg_iter);
2425 obj->mapping = vmap(pages, n, 0, PAGE_KERNEL);
2426 drm_free_large(pages);
2428 if (obj->mapping == NULL) {
2429 i915_gem_object_unpin_pages(obj);
2430 return ERR_PTR(-ENOMEM);
2434 return obj->mapping;
2437 void i915_vma_move_to_active(struct i915_vma *vma,
2438 struct drm_i915_gem_request *req)
2440 struct drm_i915_gem_object *obj = vma->obj;
2441 struct intel_engine_cs *engine;
2443 engine = i915_gem_request_get_engine(req);
2445 /* Add a reference if we're newly entering the active list. */
2446 if (obj->active == 0)
2447 drm_gem_object_reference(&obj->base);
2448 obj->active |= intel_engine_flag(engine);
2450 list_move_tail(&obj->engine_list[engine->id], &engine->active_list);
2451 i915_gem_request_assign(&obj->last_read_req[engine->id], req);
2453 list_move_tail(&vma->vm_link, &vma->vm->active_list);
2457 i915_gem_object_retire__write(struct drm_i915_gem_object *obj)
2459 GEM_BUG_ON(obj->last_write_req == NULL);
2460 GEM_BUG_ON(!(obj->active & intel_engine_flag(obj->last_write_req->engine)));
2462 i915_gem_request_assign(&obj->last_write_req, NULL);
2463 intel_fb_obj_flush(obj, true, ORIGIN_CS);
2467 i915_gem_object_retire__read(struct drm_i915_gem_object *obj, int ring)
2469 struct i915_vma *vma;
2471 GEM_BUG_ON(obj->last_read_req[ring] == NULL);
2472 GEM_BUG_ON(!(obj->active & (1 << ring)));
2474 list_del_init(&obj->engine_list[ring]);
2475 i915_gem_request_assign(&obj->last_read_req[ring], NULL);
2477 if (obj->last_write_req && obj->last_write_req->engine->id == ring)
2478 i915_gem_object_retire__write(obj);
2480 obj->active &= ~(1 << ring);
2484 /* Bump our place on the bound list to keep it roughly in LRU order
2485 * so that we don't steal from recently used but inactive objects
2486 * (unless we are forced to ofc!)
2488 list_move_tail(&obj->global_list,
2489 &to_i915(obj->base.dev)->mm.bound_list);
2491 list_for_each_entry(vma, &obj->vma_list, obj_link) {
2492 if (!list_empty(&vma->vm_link))
2493 list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
2496 i915_gem_request_assign(&obj->last_fenced_req, NULL);
2497 drm_gem_object_unreference(&obj->base);
2501 i915_gem_init_seqno(struct drm_device *dev, u32 seqno)
2503 struct drm_i915_private *dev_priv = dev->dev_private;
2504 struct intel_engine_cs *engine;
2507 /* Carefully retire all requests without writing to the rings */
2508 for_each_engine(engine, dev_priv) {
2509 ret = intel_engine_idle(engine);
2513 i915_gem_retire_requests(dev);
2515 /* Finally reset hw state */
2516 for_each_engine(engine, dev_priv)
2517 intel_ring_init_seqno(engine, seqno);
2522 int i915_gem_set_seqno(struct drm_device *dev, u32 seqno)
2524 struct drm_i915_private *dev_priv = dev->dev_private;
2530 /* HWS page needs to be set less than what we
2531 * will inject to ring
2533 ret = i915_gem_init_seqno(dev, seqno - 1);
2537 /* Carefully set the last_seqno value so that wrap
2538 * detection still works
2540 dev_priv->next_seqno = seqno;
2541 dev_priv->last_seqno = seqno - 1;
2542 if (dev_priv->last_seqno == 0)
2543 dev_priv->last_seqno--;
2549 i915_gem_get_seqno(struct drm_device *dev, u32 *seqno)
2551 struct drm_i915_private *dev_priv = dev->dev_private;
2553 /* reserve 0 for non-seqno */
2554 if (dev_priv->next_seqno == 0) {
2555 int ret = i915_gem_init_seqno(dev, 0);
2559 dev_priv->next_seqno = 1;
2562 *seqno = dev_priv->last_seqno = dev_priv->next_seqno++;
2567 * NB: This function is not allowed to fail. Doing so would mean the the
2568 * request is not being tracked for completion but the work itself is
2569 * going to happen on the hardware. This would be a Bad Thing(tm).
2571 void __i915_add_request(struct drm_i915_gem_request *request,
2572 struct drm_i915_gem_object *obj,
2575 struct intel_engine_cs *engine;
2576 struct drm_i915_private *dev_priv;
2577 struct intel_ringbuffer *ringbuf;
2581 if (WARN_ON(request == NULL))
2584 engine = request->engine;
2585 dev_priv = request->i915;
2586 ringbuf = request->ringbuf;
2589 * To ensure that this call will not fail, space for its emissions
2590 * should already have been reserved in the ring buffer. Let the ring
2591 * know that it is time to use that space up.
2593 intel_ring_reserved_space_use(ringbuf);
2595 request_start = intel_ring_get_tail(ringbuf);
2597 * Emit any outstanding flushes - execbuf can fail to emit the flush
2598 * after having emitted the batchbuffer command. Hence we need to fix
2599 * things up similar to emitting the lazy request. The difference here
2600 * is that the flush _must_ happen before the next request, no matter
2604 if (i915.enable_execlists)
2605 ret = logical_ring_flush_all_caches(request);
2607 ret = intel_ring_flush_all_caches(request);
2608 /* Not allowed to fail! */
2609 WARN(ret, "*_ring_flush_all_caches failed: %d!\n", ret);
2612 trace_i915_gem_request_add(request);
2614 request->head = request_start;
2616 /* Whilst this request exists, batch_obj will be on the
2617 * active_list, and so will hold the active reference. Only when this
2618 * request is retired will the the batch_obj be moved onto the
2619 * inactive_list and lose its active reference. Hence we do not need
2620 * to explicitly hold another reference here.
2622 request->batch_obj = obj;
2624 /* Seal the request and mark it as pending execution. Note that
2625 * we may inspect this state, without holding any locks, during
2626 * hangcheck. Hence we apply the barrier to ensure that we do not
2627 * see a more recent value in the hws than we are tracking.
2629 request->emitted_jiffies = jiffies;
2630 request->previous_seqno = engine->last_submitted_seqno;
2631 smp_store_mb(engine->last_submitted_seqno, request->seqno);
2632 list_add_tail(&request->list, &engine->request_list);
2634 /* Record the position of the start of the request so that
2635 * should we detect the updated seqno part-way through the
2636 * GPU processing the request, we never over-estimate the
2637 * position of the head.
2639 request->postfix = intel_ring_get_tail(ringbuf);
2641 if (i915.enable_execlists)
2642 ret = engine->emit_request(request);
2644 ret = engine->add_request(request);
2646 request->tail = intel_ring_get_tail(ringbuf);
2648 /* Not allowed to fail! */
2649 WARN(ret, "emit|add_request failed: %d!\n", ret);
2651 i915_queue_hangcheck(engine->dev);
2653 queue_delayed_work(dev_priv->wq,
2654 &dev_priv->mm.retire_work,
2655 round_jiffies_up_relative(HZ));
2656 intel_mark_busy(dev_priv->dev);
2658 /* Sanity check that the reserved size was large enough. */
2659 intel_ring_reserved_space_end(ringbuf);
2662 static bool i915_context_is_banned(struct drm_i915_private *dev_priv,
2663 const struct intel_context *ctx)
2665 unsigned long elapsed;
2667 elapsed = get_seconds() - ctx->hang_stats.guilty_ts;
2669 if (ctx->hang_stats.banned)
2672 if (ctx->hang_stats.ban_period_seconds &&
2673 elapsed <= ctx->hang_stats.ban_period_seconds) {
2674 if (!i915_gem_context_is_default(ctx)) {
2675 DRM_DEBUG("context hanging too fast, banning!\n");
2677 } else if (i915_stop_ring_allow_ban(dev_priv)) {
2678 if (i915_stop_ring_allow_warn(dev_priv))
2679 DRM_ERROR("gpu hanging too fast, banning!\n");
2687 static void i915_set_reset_status(struct drm_i915_private *dev_priv,
2688 struct intel_context *ctx,
2691 struct i915_ctx_hang_stats *hs;
2696 hs = &ctx->hang_stats;
2699 hs->banned = i915_context_is_banned(dev_priv, ctx);
2701 hs->guilty_ts = get_seconds();
2703 hs->batch_pending++;
2707 void i915_gem_request_free(struct kref *req_ref)
2709 struct drm_i915_gem_request *req = container_of(req_ref,
2711 struct intel_context *ctx = req->ctx;
2714 i915_gem_request_remove_from_client(req);
2717 if (i915.enable_execlists && ctx != req->i915->kernel_context)
2718 intel_lr_context_unpin(ctx, req->engine);
2720 i915_gem_context_unreference(ctx);
2723 kmem_cache_free(req->i915->requests, req);
2727 __i915_gem_request_alloc(struct intel_engine_cs *engine,
2728 struct intel_context *ctx,
2729 struct drm_i915_gem_request **req_out)
2731 struct drm_i915_private *dev_priv = to_i915(engine->dev);
2732 unsigned reset_counter = i915_reset_counter(&dev_priv->gpu_error);
2733 struct drm_i915_gem_request *req;
2741 /* ABI: Before userspace accesses the GPU (e.g. execbuffer), report
2742 * EIO if the GPU is already wedged, or EAGAIN to drop the struct_mutex
2745 ret = i915_gem_check_wedge(reset_counter, dev_priv->mm.interruptible);
2749 req = kmem_cache_zalloc(dev_priv->requests, GFP_KERNEL);
2753 ret = i915_gem_get_seqno(engine->dev, &req->seqno);
2757 kref_init(&req->ref);
2758 req->i915 = dev_priv;
2759 req->engine = engine;
2760 req->reset_counter = reset_counter;
2762 i915_gem_context_reference(req->ctx);
2764 if (i915.enable_execlists)
2765 ret = intel_logical_ring_alloc_request_extras(req);
2767 ret = intel_ring_alloc_request_extras(req);
2769 i915_gem_context_unreference(req->ctx);
2774 * Reserve space in the ring buffer for all the commands required to
2775 * eventually emit this request. This is to guarantee that the
2776 * i915_add_request() call can't fail. Note that the reserve may need
2777 * to be redone if the request is not actually submitted straight
2778 * away, e.g. because a GPU scheduler has deferred it.
2780 if (i915.enable_execlists)
2781 ret = intel_logical_ring_reserve_space(req);
2783 ret = intel_ring_reserve_space(req);
2786 * At this point, the request is fully allocated even if not
2787 * fully prepared. Thus it can be cleaned up using the proper
2790 intel_ring_reserved_space_cancel(req->ringbuf);
2791 i915_gem_request_unreference(req);
2799 kmem_cache_free(dev_priv->requests, req);
2804 * i915_gem_request_alloc - allocate a request structure
2806 * @engine: engine that we wish to issue the request on.
2807 * @ctx: context that the request will be associated with.
2808 * This can be NULL if the request is not directly related to
2809 * any specific user context, in which case this function will
2810 * choose an appropriate context to use.
2812 * Returns a pointer to the allocated request if successful,
2813 * or an error code if not.
2815 struct drm_i915_gem_request *
2816 i915_gem_request_alloc(struct intel_engine_cs *engine,
2817 struct intel_context *ctx)
2819 struct drm_i915_gem_request *req;
2823 ctx = to_i915(engine->dev)->kernel_context;
2824 err = __i915_gem_request_alloc(engine, ctx, &req);
2825 return err ? ERR_PTR(err) : req;
2828 struct drm_i915_gem_request *
2829 i915_gem_find_active_request(struct intel_engine_cs *engine)
2831 struct drm_i915_gem_request *request;
2833 list_for_each_entry(request, &engine->request_list, list) {
2834 if (i915_gem_request_completed(request, false))
2843 static void i915_gem_reset_engine_status(struct drm_i915_private *dev_priv,
2844 struct intel_engine_cs *engine)
2846 struct drm_i915_gem_request *request;
2849 request = i915_gem_find_active_request(engine);
2851 if (request == NULL)
2854 ring_hung = engine->hangcheck.score >= HANGCHECK_SCORE_RING_HUNG;
2856 i915_set_reset_status(dev_priv, request->ctx, ring_hung);
2858 list_for_each_entry_continue(request, &engine->request_list, list)
2859 i915_set_reset_status(dev_priv, request->ctx, false);
2862 static void i915_gem_reset_engine_cleanup(struct drm_i915_private *dev_priv,
2863 struct intel_engine_cs *engine)
2865 struct intel_ringbuffer *buffer;
2867 while (!list_empty(&engine->active_list)) {
2868 struct drm_i915_gem_object *obj;
2870 obj = list_first_entry(&engine->active_list,
2871 struct drm_i915_gem_object,
2872 engine_list[engine->id]);
2874 i915_gem_object_retire__read(obj, engine->id);
2878 * Clear the execlists queue up before freeing the requests, as those
2879 * are the ones that keep the context and ringbuffer backing objects
2883 if (i915.enable_execlists) {
2884 /* Ensure irq handler finishes or is cancelled. */
2885 tasklet_kill(&engine->irq_tasklet);
2887 spin_lock_bh(&engine->execlist_lock);
2888 /* list_splice_tail_init checks for empty lists */
2889 list_splice_tail_init(&engine->execlist_queue,
2890 &engine->execlist_retired_req_list);
2891 spin_unlock_bh(&engine->execlist_lock);
2893 intel_execlists_retire_requests(engine);
2897 * We must free the requests after all the corresponding objects have
2898 * been moved off active lists. Which is the same order as the normal
2899 * retire_requests function does. This is important if object hold
2900 * implicit references on things like e.g. ppgtt address spaces through
2903 while (!list_empty(&engine->request_list)) {
2904 struct drm_i915_gem_request *request;
2906 request = list_first_entry(&engine->request_list,
2907 struct drm_i915_gem_request,
2910 i915_gem_request_retire(request);
2913 /* Having flushed all requests from all queues, we know that all
2914 * ringbuffers must now be empty. However, since we do not reclaim
2915 * all space when retiring the request (to prevent HEADs colliding
2916 * with rapid ringbuffer wraparound) the amount of available space
2917 * upon reset is less than when we start. Do one more pass over
2918 * all the ringbuffers to reset last_retired_head.
2920 list_for_each_entry(buffer, &engine->buffers, link) {
2921 buffer->last_retired_head = buffer->tail;
2922 intel_ring_update_space(buffer);
2925 intel_ring_init_seqno(engine, engine->last_submitted_seqno);
2928 void i915_gem_reset(struct drm_device *dev)
2930 struct drm_i915_private *dev_priv = dev->dev_private;
2931 struct intel_engine_cs *engine;
2934 * Before we free the objects from the requests, we need to inspect
2935 * them for finding the guilty party. As the requests only borrow
2936 * their reference to the objects, the inspection must be done first.
2938 for_each_engine(engine, dev_priv)
2939 i915_gem_reset_engine_status(dev_priv, engine);
2941 for_each_engine(engine, dev_priv)
2942 i915_gem_reset_engine_cleanup(dev_priv, engine);
2944 i915_gem_context_reset(dev);
2946 i915_gem_restore_fences(dev);
2948 WARN_ON(i915_verify_lists(dev));
2952 * This function clears the request list as sequence numbers are passed.
2955 i915_gem_retire_requests_ring(struct intel_engine_cs *engine)
2957 WARN_ON(i915_verify_lists(engine->dev));
2959 /* Retire requests first as we use it above for the early return.
2960 * If we retire requests last, we may use a later seqno and so clear
2961 * the requests lists without clearing the active list, leading to
2964 while (!list_empty(&engine->request_list)) {
2965 struct drm_i915_gem_request *request;
2967 request = list_first_entry(&engine->request_list,
2968 struct drm_i915_gem_request,
2971 if (!i915_gem_request_completed(request, true))
2974 i915_gem_request_retire(request);
2977 /* Move any buffers on the active list that are no longer referenced
2978 * by the ringbuffer to the flushing/inactive lists as appropriate,
2979 * before we free the context associated with the requests.
2981 while (!list_empty(&engine->active_list)) {
2982 struct drm_i915_gem_object *obj;
2984 obj = list_first_entry(&engine->active_list,
2985 struct drm_i915_gem_object,
2986 engine_list[engine->id]);
2988 if (!list_empty(&obj->last_read_req[engine->id]->list))
2991 i915_gem_object_retire__read(obj, engine->id);
2994 if (unlikely(engine->trace_irq_req &&
2995 i915_gem_request_completed(engine->trace_irq_req, true))) {
2996 engine->irq_put(engine);
2997 i915_gem_request_assign(&engine->trace_irq_req, NULL);
3000 WARN_ON(i915_verify_lists(engine->dev));
3004 i915_gem_retire_requests(struct drm_device *dev)
3006 struct drm_i915_private *dev_priv = dev->dev_private;
3007 struct intel_engine_cs *engine;
3010 for_each_engine(engine, dev_priv) {
3011 i915_gem_retire_requests_ring(engine);
3012 idle &= list_empty(&engine->request_list);
3013 if (i915.enable_execlists) {
3014 spin_lock_bh(&engine->execlist_lock);
3015 idle &= list_empty(&engine->execlist_queue);
3016 spin_unlock_bh(&engine->execlist_lock);
3018 intel_execlists_retire_requests(engine);
3023 mod_delayed_work(dev_priv->wq,
3024 &dev_priv->mm.idle_work,
3025 msecs_to_jiffies(100));
3031 i915_gem_retire_work_handler(struct work_struct *work)
3033 struct drm_i915_private *dev_priv =
3034 container_of(work, typeof(*dev_priv), mm.retire_work.work);
3035 struct drm_device *dev = dev_priv->dev;
3038 /* Come back later if the device is busy... */
3040 if (mutex_trylock(&dev->struct_mutex)) {
3041 idle = i915_gem_retire_requests(dev);
3042 mutex_unlock(&dev->struct_mutex);
3045 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
3046 round_jiffies_up_relative(HZ));
3050 i915_gem_idle_work_handler(struct work_struct *work)
3052 struct drm_i915_private *dev_priv =
3053 container_of(work, typeof(*dev_priv), mm.idle_work.work);
3054 struct drm_device *dev = dev_priv->dev;
3055 struct intel_engine_cs *engine;
3057 for_each_engine(engine, dev_priv)
3058 if (!list_empty(&engine->request_list))
3061 /* we probably should sync with hangcheck here, using cancel_work_sync.
3062 * Also locking seems to be fubar here, engine->request_list is protected
3063 * by dev->struct_mutex. */
3065 intel_mark_idle(dev);
3067 if (mutex_trylock(&dev->struct_mutex)) {
3068 for_each_engine(engine, dev_priv)
3069 i915_gem_batch_pool_fini(&engine->batch_pool);
3071 mutex_unlock(&dev->struct_mutex);
3076 * Ensures that an object will eventually get non-busy by flushing any required
3077 * write domains, emitting any outstanding lazy request and retiring and
3078 * completed requests.
3081 i915_gem_object_flush_active(struct drm_i915_gem_object *obj)
3088 for (i = 0; i < I915_NUM_ENGINES; i++) {
3089 struct drm_i915_gem_request *req;
3091 req = obj->last_read_req[i];
3095 if (list_empty(&req->list))
3098 if (i915_gem_request_completed(req, true)) {
3099 __i915_gem_request_retire__upto(req);
3101 i915_gem_object_retire__read(obj, i);
3109 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
3110 * @DRM_IOCTL_ARGS: standard ioctl arguments
3112 * Returns 0 if successful, else an error is returned with the remaining time in
3113 * the timeout parameter.
3114 * -ETIME: object is still busy after timeout
3115 * -ERESTARTSYS: signal interrupted the wait
3116 * -ENONENT: object doesn't exist
3117 * Also possible, but rare:
3118 * -EAGAIN: GPU wedged
3120 * -ENODEV: Internal IRQ fail
3121 * -E?: The add request failed
3123 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
3124 * non-zero timeout parameter the wait ioctl will wait for the given number of
3125 * nanoseconds on an object becoming unbusy. Since the wait itself does so
3126 * without holding struct_mutex the object may become re-busied before this
3127 * function completes. A similar but shorter * race condition exists in the busy
3131 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
3133 struct drm_i915_gem_wait *args = data;
3134 struct drm_i915_gem_object *obj;
3135 struct drm_i915_gem_request *req[I915_NUM_ENGINES];
3139 if (args->flags != 0)
3142 ret = i915_mutex_lock_interruptible(dev);
3146 obj = to_intel_bo(drm_gem_object_lookup(file, args->bo_handle));
3147 if (&obj->base == NULL) {
3148 mutex_unlock(&dev->struct_mutex);
3152 /* Need to make sure the object gets inactive eventually. */
3153 ret = i915_gem_object_flush_active(obj);
3160 /* Do this after OLR check to make sure we make forward progress polling
3161 * on this IOCTL with a timeout == 0 (like busy ioctl)
3163 if (args->timeout_ns == 0) {
3168 drm_gem_object_unreference(&obj->base);
3170 for (i = 0; i < I915_NUM_ENGINES; i++) {
3171 if (obj->last_read_req[i] == NULL)
3174 req[n++] = i915_gem_request_reference(obj->last_read_req[i]);
3177 mutex_unlock(&dev->struct_mutex);
3179 for (i = 0; i < n; i++) {
3181 ret = __i915_wait_request(req[i], true,
3182 args->timeout_ns > 0 ? &args->timeout_ns : NULL,
3183 to_rps_client(file));
3184 i915_gem_request_unreference__unlocked(req[i]);
3189 drm_gem_object_unreference(&obj->base);
3190 mutex_unlock(&dev->struct_mutex);
3195 __i915_gem_object_sync(struct drm_i915_gem_object *obj,
3196 struct intel_engine_cs *to,
3197 struct drm_i915_gem_request *from_req,
3198 struct drm_i915_gem_request **to_req)
3200 struct intel_engine_cs *from;
3203 from = i915_gem_request_get_engine(from_req);
3207 if (i915_gem_request_completed(from_req, true))
3210 if (!i915_semaphore_is_enabled(obj->base.dev)) {
3211 struct drm_i915_private *i915 = to_i915(obj->base.dev);
3212 ret = __i915_wait_request(from_req,
3213 i915->mm.interruptible,
3215 &i915->rps.semaphores);
3219 i915_gem_object_retire_request(obj, from_req);
3221 int idx = intel_ring_sync_index(from, to);
3222 u32 seqno = i915_gem_request_get_seqno(from_req);
3226 if (seqno <= from->semaphore.sync_seqno[idx])
3229 if (*to_req == NULL) {
3230 struct drm_i915_gem_request *req;
3232 req = i915_gem_request_alloc(to, NULL);
3234 return PTR_ERR(req);
3239 trace_i915_gem_ring_sync_to(*to_req, from, from_req);
3240 ret = to->semaphore.sync_to(*to_req, from, seqno);
3244 /* We use last_read_req because sync_to()
3245 * might have just caused seqno wrap under
3248 from->semaphore.sync_seqno[idx] =
3249 i915_gem_request_get_seqno(obj->last_read_req[from->id]);
3256 * i915_gem_object_sync - sync an object to a ring.
3258 * @obj: object which may be in use on another ring.
3259 * @to: ring we wish to use the object on. May be NULL.
3260 * @to_req: request we wish to use the object for. See below.
3261 * This will be allocated and returned if a request is
3262 * required but not passed in.
3264 * This code is meant to abstract object synchronization with the GPU.
3265 * Calling with NULL implies synchronizing the object with the CPU
3266 * rather than a particular GPU ring. Conceptually we serialise writes
3267 * between engines inside the GPU. We only allow one engine to write
3268 * into a buffer at any time, but multiple readers. To ensure each has
3269 * a coherent view of memory, we must:
3271 * - If there is an outstanding write request to the object, the new
3272 * request must wait for it to complete (either CPU or in hw, requests
3273 * on the same ring will be naturally ordered).
3275 * - If we are a write request (pending_write_domain is set), the new
3276 * request must wait for outstanding read requests to complete.
3278 * For CPU synchronisation (NULL to) no request is required. For syncing with
3279 * rings to_req must be non-NULL. However, a request does not have to be
3280 * pre-allocated. If *to_req is NULL and sync commands will be emitted then a
3281 * request will be allocated automatically and returned through *to_req. Note
3282 * that it is not guaranteed that commands will be emitted (because the system
3283 * might already be idle). Hence there is no need to create a request that
3284 * might never have any work submitted. Note further that if a request is
3285 * returned in *to_req, it is the responsibility of the caller to submit
3286 * that request (after potentially adding more work to it).
3288 * Returns 0 if successful, else propagates up the lower layer error.
3291 i915_gem_object_sync(struct drm_i915_gem_object *obj,
3292 struct intel_engine_cs *to,
3293 struct drm_i915_gem_request **to_req)
3295 const bool readonly = obj->base.pending_write_domain == 0;
3296 struct drm_i915_gem_request *req[I915_NUM_ENGINES];
3303 return i915_gem_object_wait_rendering(obj, readonly);
3307 if (obj->last_write_req)
3308 req[n++] = obj->last_write_req;
3310 for (i = 0; i < I915_NUM_ENGINES; i++)
3311 if (obj->last_read_req[i])
3312 req[n++] = obj->last_read_req[i];
3314 for (i = 0; i < n; i++) {
3315 ret = __i915_gem_object_sync(obj, to, req[i], to_req);
3323 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
3325 u32 old_write_domain, old_read_domains;
3327 /* Force a pagefault for domain tracking on next user access */
3328 i915_gem_release_mmap(obj);
3330 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
3333 old_read_domains = obj->base.read_domains;
3334 old_write_domain = obj->base.write_domain;
3336 obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
3337 obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
3339 trace_i915_gem_object_change_domain(obj,
3344 static int __i915_vma_unbind(struct i915_vma *vma, bool wait)
3346 struct drm_i915_gem_object *obj = vma->obj;
3347 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3350 if (list_empty(&vma->obj_link))
3353 if (!drm_mm_node_allocated(&vma->node)) {
3354 i915_gem_vma_destroy(vma);
3361 BUG_ON(obj->pages == NULL);
3364 ret = i915_gem_object_wait_rendering(obj, false);
3369 if (vma->is_ggtt && vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL) {
3370 i915_gem_object_finish_gtt(obj);
3372 /* release the fence reg _after_ flushing */
3373 ret = i915_gem_object_put_fence(obj);
3378 trace_i915_vma_unbind(vma);
3380 vma->vm->unbind_vma(vma);
3383 list_del_init(&vma->vm_link);
3385 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL) {
3386 obj->map_and_fenceable = false;
3387 } else if (vma->ggtt_view.pages) {
3388 sg_free_table(vma->ggtt_view.pages);
3389 kfree(vma->ggtt_view.pages);
3391 vma->ggtt_view.pages = NULL;
3394 drm_mm_remove_node(&vma->node);
3395 i915_gem_vma_destroy(vma);
3397 /* Since the unbound list is global, only move to that list if
3398 * no more VMAs exist. */
3399 if (list_empty(&obj->vma_list))
3400 list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list);
3402 /* And finally now the object is completely decoupled from this vma,
3403 * we can drop its hold on the backing storage and allow it to be
3404 * reaped by the shrinker.
3406 i915_gem_object_unpin_pages(obj);
3411 int i915_vma_unbind(struct i915_vma *vma)
3413 return __i915_vma_unbind(vma, true);
3416 int __i915_vma_unbind_no_wait(struct i915_vma *vma)
3418 return __i915_vma_unbind(vma, false);
3421 int i915_gpu_idle(struct drm_device *dev)
3423 struct drm_i915_private *dev_priv = dev->dev_private;
3424 struct intel_engine_cs *engine;
3427 /* Flush everything onto the inactive list. */
3428 for_each_engine(engine, dev_priv) {
3429 if (!i915.enable_execlists) {
3430 struct drm_i915_gem_request *req;
3432 req = i915_gem_request_alloc(engine, NULL);
3434 return PTR_ERR(req);
3436 ret = i915_switch_context(req);
3437 i915_add_request_no_flush(req);
3442 ret = intel_engine_idle(engine);
3447 WARN_ON(i915_verify_lists(dev));
3451 static bool i915_gem_valid_gtt_space(struct i915_vma *vma,
3452 unsigned long cache_level)
3454 struct drm_mm_node *gtt_space = &vma->node;
3455 struct drm_mm_node *other;
3458 * On some machines we have to be careful when putting differing types
3459 * of snoopable memory together to avoid the prefetcher crossing memory
3460 * domains and dying. During vm initialisation, we decide whether or not
3461 * these constraints apply and set the drm_mm.color_adjust
3464 if (vma->vm->mm.color_adjust == NULL)
3467 if (!drm_mm_node_allocated(gtt_space))
3470 if (list_empty(>t_space->node_list))
3473 other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list);
3474 if (other->allocated && !other->hole_follows && other->color != cache_level)
3477 other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list);
3478 if (other->allocated && !gtt_space->hole_follows && other->color != cache_level)
3485 * Finds free space in the GTT aperture and binds the object or a view of it
3488 static struct i915_vma *
3489 i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj,
3490 struct i915_address_space *vm,
3491 const struct i915_ggtt_view *ggtt_view,
3495 struct drm_device *dev = obj->base.dev;
3496 struct drm_i915_private *dev_priv = to_i915(dev);
3497 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3498 u32 fence_alignment, unfenced_alignment;
3499 u32 search_flag, alloc_flag;
3501 u64 size, fence_size;
3502 struct i915_vma *vma;
3505 if (i915_is_ggtt(vm)) {
3508 if (WARN_ON(!ggtt_view))
3509 return ERR_PTR(-EINVAL);
3511 view_size = i915_ggtt_view_size(obj, ggtt_view);
3513 fence_size = i915_gem_get_gtt_size(dev,
3516 fence_alignment = i915_gem_get_gtt_alignment(dev,
3520 unfenced_alignment = i915_gem_get_gtt_alignment(dev,
3524 size = flags & PIN_MAPPABLE ? fence_size : view_size;
3526 fence_size = i915_gem_get_gtt_size(dev,
3529 fence_alignment = i915_gem_get_gtt_alignment(dev,
3533 unfenced_alignment =
3534 i915_gem_get_gtt_alignment(dev,
3538 size = flags & PIN_MAPPABLE ? fence_size : obj->base.size;
3541 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
3543 if (flags & PIN_MAPPABLE)
3544 end = min_t(u64, end, ggtt->mappable_end);
3545 if (flags & PIN_ZONE_4G)
3546 end = min_t(u64, end, (1ULL << 32) - PAGE_SIZE);
3549 alignment = flags & PIN_MAPPABLE ? fence_alignment :
3551 if (flags & PIN_MAPPABLE && alignment & (fence_alignment - 1)) {
3552 DRM_DEBUG("Invalid object (view type=%u) alignment requested %u\n",
3553 ggtt_view ? ggtt_view->type : 0,
3555 return ERR_PTR(-EINVAL);
3558 /* If binding the object/GGTT view requires more space than the entire
3559 * aperture has, reject it early before evicting everything in a vain
3560 * attempt to find space.
3563 DRM_DEBUG("Attempting to bind an object (view type=%u) larger than the aperture: size=%llu > %s aperture=%llu\n",
3564 ggtt_view ? ggtt_view->type : 0,
3566 flags & PIN_MAPPABLE ? "mappable" : "total",
3568 return ERR_PTR(-E2BIG);
3571 ret = i915_gem_object_get_pages(obj);
3573 return ERR_PTR(ret);
3575 i915_gem_object_pin_pages(obj);
3577 vma = ggtt_view ? i915_gem_obj_lookup_or_create_ggtt_vma(obj, ggtt_view) :
3578 i915_gem_obj_lookup_or_create_vma(obj, vm);
3583 if (flags & PIN_OFFSET_FIXED) {
3584 uint64_t offset = flags & PIN_OFFSET_MASK;
3586 if (offset & (alignment - 1) || offset + size > end) {
3590 vma->node.start = offset;
3591 vma->node.size = size;
3592 vma->node.color = obj->cache_level;
3593 ret = drm_mm_reserve_node(&vm->mm, &vma->node);
3595 ret = i915_gem_evict_for_vma(vma);
3597 ret = drm_mm_reserve_node(&vm->mm, &vma->node);
3602 if (flags & PIN_HIGH) {
3603 search_flag = DRM_MM_SEARCH_BELOW;
3604 alloc_flag = DRM_MM_CREATE_TOP;
3606 search_flag = DRM_MM_SEARCH_DEFAULT;
3607 alloc_flag = DRM_MM_CREATE_DEFAULT;
3611 ret = drm_mm_insert_node_in_range_generic(&vm->mm, &vma->node,
3618 ret = i915_gem_evict_something(dev, vm, size, alignment,
3628 if (WARN_ON(!i915_gem_valid_gtt_space(vma, obj->cache_level))) {
3630 goto err_remove_node;
3633 trace_i915_vma_bind(vma, flags);
3634 ret = i915_vma_bind(vma, obj->cache_level, flags);
3636 goto err_remove_node;
3638 list_move_tail(&obj->global_list, &dev_priv->mm.bound_list);
3639 list_add_tail(&vma->vm_link, &vm->inactive_list);
3644 drm_mm_remove_node(&vma->node);
3646 i915_gem_vma_destroy(vma);
3649 i915_gem_object_unpin_pages(obj);
3654 i915_gem_clflush_object(struct drm_i915_gem_object *obj,
3657 /* If we don't have a page list set up, then we're not pinned
3658 * to GPU, and we can ignore the cache flush because it'll happen
3659 * again at bind time.
3661 if (obj->pages == NULL)
3665 * Stolen memory is always coherent with the GPU as it is explicitly
3666 * marked as wc by the system, or the system is cache-coherent.
3668 if (obj->stolen || obj->phys_handle)
3671 /* If the GPU is snooping the contents of the CPU cache,
3672 * we do not need to manually clear the CPU cache lines. However,
3673 * the caches are only snooped when the render cache is
3674 * flushed/invalidated. As we always have to emit invalidations
3675 * and flushes when moving into and out of the RENDER domain, correct
3676 * snooping behaviour occurs naturally as the result of our domain
3679 if (!force && cpu_cache_is_coherent(obj->base.dev, obj->cache_level)) {
3680 obj->cache_dirty = true;
3684 trace_i915_gem_object_clflush(obj);
3685 drm_clflush_sg(obj->pages);
3686 obj->cache_dirty = false;
3691 /** Flushes the GTT write domain for the object if it's dirty. */
3693 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
3695 uint32_t old_write_domain;
3697 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
3700 /* No actual flushing is required for the GTT write domain. Writes
3701 * to it immediately go to main memory as far as we know, so there's
3702 * no chipset flush. It also doesn't land in render cache.
3704 * However, we do have to enforce the order so that all writes through
3705 * the GTT land before any writes to the device, such as updates to
3710 old_write_domain = obj->base.write_domain;
3711 obj->base.write_domain = 0;
3713 intel_fb_obj_flush(obj, false, ORIGIN_GTT);
3715 trace_i915_gem_object_change_domain(obj,
3716 obj->base.read_domains,
3720 /** Flushes the CPU write domain for the object if it's dirty. */
3722 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj)
3724 uint32_t old_write_domain;
3726 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
3729 if (i915_gem_clflush_object(obj, obj->pin_display))
3730 i915_gem_chipset_flush(obj->base.dev);
3732 old_write_domain = obj->base.write_domain;
3733 obj->base.write_domain = 0;
3735 intel_fb_obj_flush(obj, false, ORIGIN_CPU);
3737 trace_i915_gem_object_change_domain(obj,
3738 obj->base.read_domains,
3743 * Moves a single object to the GTT read, and possibly write domain.
3745 * This function returns when the move is complete, including waiting on
3749 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
3751 struct drm_device *dev = obj->base.dev;
3752 struct drm_i915_private *dev_priv = to_i915(dev);
3753 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3754 uint32_t old_write_domain, old_read_domains;
3755 struct i915_vma *vma;
3758 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
3761 ret = i915_gem_object_wait_rendering(obj, !write);
3765 /* Flush and acquire obj->pages so that we are coherent through
3766 * direct access in memory with previous cached writes through
3767 * shmemfs and that our cache domain tracking remains valid.
3768 * For example, if the obj->filp was moved to swap without us
3769 * being notified and releasing the pages, we would mistakenly
3770 * continue to assume that the obj remained out of the CPU cached
3773 ret = i915_gem_object_get_pages(obj);
3777 i915_gem_object_flush_cpu_write_domain(obj);
3779 /* Serialise direct access to this object with the barriers for
3780 * coherent writes from the GPU, by effectively invalidating the
3781 * GTT domain upon first access.
3783 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
3786 old_write_domain = obj->base.write_domain;
3787 old_read_domains = obj->base.read_domains;
3789 /* It should now be out of any other write domains, and we can update
3790 * the domain values for our changes.
3792 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3793 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3795 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
3796 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
3800 trace_i915_gem_object_change_domain(obj,
3804 /* And bump the LRU for this access */
3805 vma = i915_gem_obj_to_ggtt(obj);
3806 if (vma && drm_mm_node_allocated(&vma->node) && !obj->active)
3807 list_move_tail(&vma->vm_link,
3808 &ggtt->base.inactive_list);
3814 * Changes the cache-level of an object across all VMA.
3816 * After this function returns, the object will be in the new cache-level
3817 * across all GTT and the contents of the backing storage will be coherent,
3818 * with respect to the new cache-level. In order to keep the backing storage
3819 * coherent for all users, we only allow a single cache level to be set
3820 * globally on the object and prevent it from being changed whilst the
3821 * hardware is reading from the object. That is if the object is currently
3822 * on the scanout it will be set to uncached (or equivalent display
3823 * cache coherency) and all non-MOCS GPU access will also be uncached so
3824 * that all direct access to the scanout remains coherent.
3826 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
3827 enum i915_cache_level cache_level)
3829 struct drm_device *dev = obj->base.dev;
3830 struct i915_vma *vma, *next;
3834 if (obj->cache_level == cache_level)
3837 /* Inspect the list of currently bound VMA and unbind any that would
3838 * be invalid given the new cache-level. This is principally to
3839 * catch the issue of the CS prefetch crossing page boundaries and
3840 * reading an invalid PTE on older architectures.
3842 list_for_each_entry_safe(vma, next, &obj->vma_list, obj_link) {
3843 if (!drm_mm_node_allocated(&vma->node))
3846 if (vma->pin_count) {
3847 DRM_DEBUG("can not change the cache level of pinned objects\n");
3851 if (!i915_gem_valid_gtt_space(vma, cache_level)) {
3852 ret = i915_vma_unbind(vma);
3859 /* We can reuse the existing drm_mm nodes but need to change the
3860 * cache-level on the PTE. We could simply unbind them all and
3861 * rebind with the correct cache-level on next use. However since
3862 * we already have a valid slot, dma mapping, pages etc, we may as
3863 * rewrite the PTE in the belief that doing so tramples upon less
3864 * state and so involves less work.
3867 /* Before we change the PTE, the GPU must not be accessing it.
3868 * If we wait upon the object, we know that all the bound
3869 * VMA are no longer active.
3871 ret = i915_gem_object_wait_rendering(obj, false);
3875 if (!HAS_LLC(dev) && cache_level != I915_CACHE_NONE) {
3876 /* Access to snoopable pages through the GTT is
3877 * incoherent and on some machines causes a hard
3878 * lockup. Relinquish the CPU mmaping to force
3879 * userspace to refault in the pages and we can
3880 * then double check if the GTT mapping is still
3881 * valid for that pointer access.
3883 i915_gem_release_mmap(obj);
3885 /* As we no longer need a fence for GTT access,
3886 * we can relinquish it now (and so prevent having
3887 * to steal a fence from someone else on the next
3888 * fence request). Note GPU activity would have
3889 * dropped the fence as all snoopable access is
3890 * supposed to be linear.
3892 ret = i915_gem_object_put_fence(obj);
3896 /* We either have incoherent backing store and
3897 * so no GTT access or the architecture is fully
3898 * coherent. In such cases, existing GTT mmaps
3899 * ignore the cache bit in the PTE and we can
3900 * rewrite it without confusing the GPU or having
3901 * to force userspace to fault back in its mmaps.
3905 list_for_each_entry(vma, &obj->vma_list, obj_link) {
3906 if (!drm_mm_node_allocated(&vma->node))
3909 ret = i915_vma_bind(vma, cache_level, PIN_UPDATE);
3915 list_for_each_entry(vma, &obj->vma_list, obj_link)
3916 vma->node.color = cache_level;
3917 obj->cache_level = cache_level;
3920 /* Flush the dirty CPU caches to the backing storage so that the
3921 * object is now coherent at its new cache level (with respect
3922 * to the access domain).
3924 if (obj->cache_dirty &&
3925 obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
3926 cpu_write_needs_clflush(obj)) {
3927 if (i915_gem_clflush_object(obj, true))
3928 i915_gem_chipset_flush(obj->base.dev);
3934 int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data,
3935 struct drm_file *file)
3937 struct drm_i915_gem_caching *args = data;
3938 struct drm_i915_gem_object *obj;
3940 obj = to_intel_bo(drm_gem_object_lookup(file, args->handle));
3941 if (&obj->base == NULL)
3944 switch (obj->cache_level) {
3945 case I915_CACHE_LLC:
3946 case I915_CACHE_L3_LLC:
3947 args->caching = I915_CACHING_CACHED;
3951 args->caching = I915_CACHING_DISPLAY;
3955 args->caching = I915_CACHING_NONE;
3959 drm_gem_object_unreference_unlocked(&obj->base);
3963 int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
3964 struct drm_file *file)
3966 struct drm_i915_private *dev_priv = dev->dev_private;
3967 struct drm_i915_gem_caching *args = data;
3968 struct drm_i915_gem_object *obj;
3969 enum i915_cache_level level;
3972 switch (args->caching) {
3973 case I915_CACHING_NONE:
3974 level = I915_CACHE_NONE;
3976 case I915_CACHING_CACHED:
3978 * Due to a HW issue on BXT A stepping, GPU stores via a
3979 * snooped mapping may leave stale data in a corresponding CPU
3980 * cacheline, whereas normally such cachelines would get
3983 if (!HAS_LLC(dev) && !HAS_SNOOP(dev))
3986 level = I915_CACHE_LLC;
3988 case I915_CACHING_DISPLAY:
3989 level = HAS_WT(dev) ? I915_CACHE_WT : I915_CACHE_NONE;
3995 intel_runtime_pm_get(dev_priv);
3997 ret = i915_mutex_lock_interruptible(dev);
4001 obj = to_intel_bo(drm_gem_object_lookup(file, args->handle));
4002 if (&obj->base == NULL) {
4007 ret = i915_gem_object_set_cache_level(obj, level);
4009 drm_gem_object_unreference(&obj->base);
4011 mutex_unlock(&dev->struct_mutex);
4013 intel_runtime_pm_put(dev_priv);
4019 * Prepare buffer for display plane (scanout, cursors, etc).
4020 * Can be called from an uninterruptible phase (modesetting) and allows
4021 * any flushes to be pipelined (for pageflips).
4024 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
4026 const struct i915_ggtt_view *view)
4028 u32 old_read_domains, old_write_domain;
4031 /* Mark the pin_display early so that we account for the
4032 * display coherency whilst setting up the cache domains.
4036 /* The display engine is not coherent with the LLC cache on gen6. As
4037 * a result, we make sure that the pinning that is about to occur is
4038 * done with uncached PTEs. This is lowest common denominator for all
4041 * However for gen6+, we could do better by using the GFDT bit instead
4042 * of uncaching, which would allow us to flush all the LLC-cached data
4043 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
4045 ret = i915_gem_object_set_cache_level(obj,
4046 HAS_WT(obj->base.dev) ? I915_CACHE_WT : I915_CACHE_NONE);
4048 goto err_unpin_display;
4050 /* As the user may map the buffer once pinned in the display plane
4051 * (e.g. libkms for the bootup splash), we have to ensure that we
4052 * always use map_and_fenceable for all scanout buffers.
4054 ret = i915_gem_object_ggtt_pin(obj, view, alignment,
4055 view->type == I915_GGTT_VIEW_NORMAL ?
4058 goto err_unpin_display;
4060 i915_gem_object_flush_cpu_write_domain(obj);
4062 old_write_domain = obj->base.write_domain;
4063 old_read_domains = obj->base.read_domains;
4065 /* It should now be out of any other write domains, and we can update
4066 * the domain values for our changes.
4068 obj->base.write_domain = 0;
4069 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
4071 trace_i915_gem_object_change_domain(obj,
4083 i915_gem_object_unpin_from_display_plane(struct drm_i915_gem_object *obj,
4084 const struct i915_ggtt_view *view)
4086 if (WARN_ON(obj->pin_display == 0))
4089 i915_gem_object_ggtt_unpin_view(obj, view);
4095 * Moves a single object to the CPU read, and possibly write domain.
4097 * This function returns when the move is complete, including waiting on
4101 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
4103 uint32_t old_write_domain, old_read_domains;
4106 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
4109 ret = i915_gem_object_wait_rendering(obj, !write);
4113 i915_gem_object_flush_gtt_write_domain(obj);
4115 old_write_domain = obj->base.write_domain;
4116 old_read_domains = obj->base.read_domains;
4118 /* Flush the CPU cache if it's still invalid. */
4119 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
4120 i915_gem_clflush_object(obj, false);
4122 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
4125 /* It should now be out of any other write domains, and we can update
4126 * the domain values for our changes.
4128 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
4130 /* If we're writing through the CPU, then the GPU read domains will
4131 * need to be invalidated at next use.
4134 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4135 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4138 trace_i915_gem_object_change_domain(obj,
4145 /* Throttle our rendering by waiting until the ring has completed our requests
4146 * emitted over 20 msec ago.
4148 * Note that if we were to use the current jiffies each time around the loop,
4149 * we wouldn't escape the function with any frames outstanding if the time to
4150 * render a frame was over 20ms.
4152 * This should get us reasonable parallelism between CPU and GPU but also
4153 * relatively low latency when blocking on a particular request to finish.
4156 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
4158 struct drm_i915_private *dev_priv = dev->dev_private;
4159 struct drm_i915_file_private *file_priv = file->driver_priv;
4160 unsigned long recent_enough = jiffies - DRM_I915_THROTTLE_JIFFIES;
4161 struct drm_i915_gem_request *request, *target = NULL;
4164 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
4168 /* ABI: return -EIO if already wedged */
4169 if (i915_terminally_wedged(&dev_priv->gpu_error))
4172 spin_lock(&file_priv->mm.lock);
4173 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
4174 if (time_after_eq(request->emitted_jiffies, recent_enough))
4178 * Note that the request might not have been submitted yet.
4179 * In which case emitted_jiffies will be zero.
4181 if (!request->emitted_jiffies)
4187 i915_gem_request_reference(target);
4188 spin_unlock(&file_priv->mm.lock);
4193 ret = __i915_wait_request(target, true, NULL, NULL);
4195 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
4197 i915_gem_request_unreference__unlocked(target);
4203 i915_vma_misplaced(struct i915_vma *vma, uint32_t alignment, uint64_t flags)
4205 struct drm_i915_gem_object *obj = vma->obj;
4208 vma->node.start & (alignment - 1))
4211 if (flags & PIN_MAPPABLE && !obj->map_and_fenceable)
4214 if (flags & PIN_OFFSET_BIAS &&
4215 vma->node.start < (flags & PIN_OFFSET_MASK))
4218 if (flags & PIN_OFFSET_FIXED &&
4219 vma->node.start != (flags & PIN_OFFSET_MASK))
4225 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
4227 struct drm_i915_gem_object *obj = vma->obj;
4228 bool mappable, fenceable;
4229 u32 fence_size, fence_alignment;
4231 fence_size = i915_gem_get_gtt_size(obj->base.dev,
4234 fence_alignment = i915_gem_get_gtt_alignment(obj->base.dev,
4239 fenceable = (vma->node.size == fence_size &&
4240 (vma->node.start & (fence_alignment - 1)) == 0);
4242 mappable = (vma->node.start + fence_size <=
4243 to_i915(obj->base.dev)->ggtt.mappable_end);
4245 obj->map_and_fenceable = mappable && fenceable;
4249 i915_gem_object_do_pin(struct drm_i915_gem_object *obj,
4250 struct i915_address_space *vm,
4251 const struct i915_ggtt_view *ggtt_view,
4255 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
4256 struct i915_vma *vma;
4260 if (WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base))
4263 if (WARN_ON(flags & (PIN_GLOBAL | PIN_MAPPABLE) && !i915_is_ggtt(vm)))
4266 if (WARN_ON((flags & (PIN_MAPPABLE | PIN_GLOBAL)) == PIN_MAPPABLE))
4269 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
4272 vma = ggtt_view ? i915_gem_obj_to_ggtt_view(obj, ggtt_view) :
4273 i915_gem_obj_to_vma(obj, vm);
4276 if (WARN_ON(vma->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT))
4279 if (i915_vma_misplaced(vma, alignment, flags)) {
4280 WARN(vma->pin_count,
4281 "bo is already pinned in %s with incorrect alignment:"
4282 " offset=%08x %08x, req.alignment=%x, req.map_and_fenceable=%d,"
4283 " obj->map_and_fenceable=%d\n",
4284 ggtt_view ? "ggtt" : "ppgtt",
4285 upper_32_bits(vma->node.start),
4286 lower_32_bits(vma->node.start),
4288 !!(flags & PIN_MAPPABLE),
4289 obj->map_and_fenceable);
4290 ret = i915_vma_unbind(vma);
4298 bound = vma ? vma->bound : 0;
4299 if (vma == NULL || !drm_mm_node_allocated(&vma->node)) {
4300 vma = i915_gem_object_bind_to_vm(obj, vm, ggtt_view, alignment,
4303 return PTR_ERR(vma);
4305 ret = i915_vma_bind(vma, obj->cache_level, flags);
4310 if (ggtt_view && ggtt_view->type == I915_GGTT_VIEW_NORMAL &&
4311 (bound ^ vma->bound) & GLOBAL_BIND) {
4312 __i915_vma_set_map_and_fenceable(vma);
4313 WARN_ON(flags & PIN_MAPPABLE && !obj->map_and_fenceable);
4321 i915_gem_object_pin(struct drm_i915_gem_object *obj,
4322 struct i915_address_space *vm,
4326 return i915_gem_object_do_pin(obj, vm,
4327 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL,
4332 i915_gem_object_ggtt_pin(struct drm_i915_gem_object *obj,
4333 const struct i915_ggtt_view *view,
4337 struct drm_device *dev = obj->base.dev;
4338 struct drm_i915_private *dev_priv = to_i915(dev);
4339 struct i915_ggtt *ggtt = &dev_priv->ggtt;
4343 return i915_gem_object_do_pin(obj, &ggtt->base, view,
4344 alignment, flags | PIN_GLOBAL);
4348 i915_gem_object_ggtt_unpin_view(struct drm_i915_gem_object *obj,
4349 const struct i915_ggtt_view *view)
4351 struct i915_vma *vma = i915_gem_obj_to_ggtt_view(obj, view);
4353 WARN_ON(vma->pin_count == 0);
4354 WARN_ON(!i915_gem_obj_ggtt_bound_view(obj, view));
4360 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
4361 struct drm_file *file)
4363 struct drm_i915_gem_busy *args = data;
4364 struct drm_i915_gem_object *obj;
4367 ret = i915_mutex_lock_interruptible(dev);
4371 obj = to_intel_bo(drm_gem_object_lookup(file, args->handle));
4372 if (&obj->base == NULL) {
4377 /* Count all active objects as busy, even if they are currently not used
4378 * by the gpu. Users of this interface expect objects to eventually
4379 * become non-busy without any further actions, therefore emit any
4380 * necessary flushes here.
4382 ret = i915_gem_object_flush_active(obj);
4390 for (i = 0; i < I915_NUM_ENGINES; i++) {
4391 struct drm_i915_gem_request *req;
4393 req = obj->last_read_req[i];
4395 args->busy |= 1 << (16 + req->engine->exec_id);
4397 if (obj->last_write_req)
4398 args->busy |= obj->last_write_req->engine->exec_id;
4402 drm_gem_object_unreference(&obj->base);
4404 mutex_unlock(&dev->struct_mutex);
4409 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
4410 struct drm_file *file_priv)
4412 return i915_gem_ring_throttle(dev, file_priv);
4416 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
4417 struct drm_file *file_priv)
4419 struct drm_i915_private *dev_priv = dev->dev_private;
4420 struct drm_i915_gem_madvise *args = data;
4421 struct drm_i915_gem_object *obj;
4424 switch (args->madv) {
4425 case I915_MADV_DONTNEED:
4426 case I915_MADV_WILLNEED:
4432 ret = i915_mutex_lock_interruptible(dev);
4436 obj = to_intel_bo(drm_gem_object_lookup(file_priv, args->handle));
4437 if (&obj->base == NULL) {
4442 if (i915_gem_obj_is_pinned(obj)) {
4448 obj->tiling_mode != I915_TILING_NONE &&
4449 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
4450 if (obj->madv == I915_MADV_WILLNEED)
4451 i915_gem_object_unpin_pages(obj);
4452 if (args->madv == I915_MADV_WILLNEED)
4453 i915_gem_object_pin_pages(obj);
4456 if (obj->madv != __I915_MADV_PURGED)
4457 obj->madv = args->madv;
4459 /* if the object is no longer attached, discard its backing storage */
4460 if (obj->madv == I915_MADV_DONTNEED && obj->pages == NULL)
4461 i915_gem_object_truncate(obj);
4463 args->retained = obj->madv != __I915_MADV_PURGED;
4466 drm_gem_object_unreference(&obj->base);
4468 mutex_unlock(&dev->struct_mutex);
4472 void i915_gem_object_init(struct drm_i915_gem_object *obj,
4473 const struct drm_i915_gem_object_ops *ops)
4477 INIT_LIST_HEAD(&obj->global_list);
4478 for (i = 0; i < I915_NUM_ENGINES; i++)
4479 INIT_LIST_HEAD(&obj->engine_list[i]);
4480 INIT_LIST_HEAD(&obj->obj_exec_link);
4481 INIT_LIST_HEAD(&obj->vma_list);
4482 INIT_LIST_HEAD(&obj->batch_pool_link);
4486 obj->fence_reg = I915_FENCE_REG_NONE;
4487 obj->madv = I915_MADV_WILLNEED;
4489 i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size);
4492 static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
4493 .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE,
4494 .get_pages = i915_gem_object_get_pages_gtt,
4495 .put_pages = i915_gem_object_put_pages_gtt,
4498 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
4501 struct drm_i915_gem_object *obj;
4502 struct address_space *mapping;
4505 obj = i915_gem_object_alloc(dev);
4509 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
4510 i915_gem_object_free(obj);
4514 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
4515 if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) {
4516 /* 965gm cannot relocate objects above 4GiB. */
4517 mask &= ~__GFP_HIGHMEM;
4518 mask |= __GFP_DMA32;
4521 mapping = file_inode(obj->base.filp)->i_mapping;
4522 mapping_set_gfp_mask(mapping, mask);
4524 i915_gem_object_init(obj, &i915_gem_object_ops);
4526 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4527 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4530 /* On some devices, we can have the GPU use the LLC (the CPU
4531 * cache) for about a 10% performance improvement
4532 * compared to uncached. Graphics requests other than
4533 * display scanout are coherent with the CPU in
4534 * accessing this cache. This means in this mode we
4535 * don't need to clflush on the CPU side, and on the
4536 * GPU side we only need to flush internal caches to
4537 * get data visible to the CPU.
4539 * However, we maintain the display planes as UC, and so
4540 * need to rebind when first used as such.
4542 obj->cache_level = I915_CACHE_LLC;
4544 obj->cache_level = I915_CACHE_NONE;
4546 trace_i915_gem_object_create(obj);
4551 static bool discard_backing_storage(struct drm_i915_gem_object *obj)
4553 /* If we are the last user of the backing storage (be it shmemfs
4554 * pages or stolen etc), we know that the pages are going to be
4555 * immediately released. In this case, we can then skip copying
4556 * back the contents from the GPU.
4559 if (obj->madv != I915_MADV_WILLNEED)
4562 if (obj->base.filp == NULL)
4565 /* At first glance, this looks racy, but then again so would be
4566 * userspace racing mmap against close. However, the first external
4567 * reference to the filp can only be obtained through the
4568 * i915_gem_mmap_ioctl() which safeguards us against the user
4569 * acquiring such a reference whilst we are in the middle of
4570 * freeing the object.
4572 return atomic_long_read(&obj->base.filp->f_count) == 1;
4575 void i915_gem_free_object(struct drm_gem_object *gem_obj)
4577 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
4578 struct drm_device *dev = obj->base.dev;
4579 struct drm_i915_private *dev_priv = dev->dev_private;
4580 struct i915_vma *vma, *next;
4582 intel_runtime_pm_get(dev_priv);
4584 trace_i915_gem_object_destroy(obj);
4586 list_for_each_entry_safe(vma, next, &obj->vma_list, obj_link) {
4590 ret = i915_vma_unbind(vma);
4591 if (WARN_ON(ret == -ERESTARTSYS)) {
4592 bool was_interruptible;
4594 was_interruptible = dev_priv->mm.interruptible;
4595 dev_priv->mm.interruptible = false;
4597 WARN_ON(i915_vma_unbind(vma));
4599 dev_priv->mm.interruptible = was_interruptible;
4603 /* Stolen objects don't hold a ref, but do hold pin count. Fix that up
4604 * before progressing. */
4606 i915_gem_object_unpin_pages(obj);
4608 WARN_ON(obj->frontbuffer_bits);
4610 if (obj->pages && obj->madv == I915_MADV_WILLNEED &&
4611 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES &&
4612 obj->tiling_mode != I915_TILING_NONE)
4613 i915_gem_object_unpin_pages(obj);
4615 if (WARN_ON(obj->pages_pin_count))
4616 obj->pages_pin_count = 0;
4617 if (discard_backing_storage(obj))
4618 obj->madv = I915_MADV_DONTNEED;
4619 i915_gem_object_put_pages(obj);
4620 i915_gem_object_free_mmap_offset(obj);
4624 if (obj->base.import_attach)
4625 drm_prime_gem_destroy(&obj->base, NULL);
4627 if (obj->ops->release)
4628 obj->ops->release(obj);
4630 drm_gem_object_release(&obj->base);
4631 i915_gem_info_remove_obj(dev_priv, obj->base.size);
4634 i915_gem_object_free(obj);
4636 intel_runtime_pm_put(dev_priv);
4639 struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj,
4640 struct i915_address_space *vm)
4642 struct i915_vma *vma;
4643 list_for_each_entry(vma, &obj->vma_list, obj_link) {
4644 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL &&
4651 struct i915_vma *i915_gem_obj_to_ggtt_view(struct drm_i915_gem_object *obj,
4652 const struct i915_ggtt_view *view)
4654 struct drm_device *dev = obj->base.dev;
4655 struct drm_i915_private *dev_priv = to_i915(dev);
4656 struct i915_ggtt *ggtt = &dev_priv->ggtt;
4657 struct i915_vma *vma;
4661 list_for_each_entry(vma, &obj->vma_list, obj_link)
4662 if (vma->vm == &ggtt->base &&
4663 i915_ggtt_view_equal(&vma->ggtt_view, view))
4668 void i915_gem_vma_destroy(struct i915_vma *vma)
4670 WARN_ON(vma->node.allocated);
4672 /* Keep the vma as a placeholder in the execbuffer reservation lists */
4673 if (!list_empty(&vma->exec_list))
4677 i915_ppgtt_put(i915_vm_to_ppgtt(vma->vm));
4679 list_del(&vma->obj_link);
4681 kmem_cache_free(to_i915(vma->obj->base.dev)->vmas, vma);
4685 i915_gem_stop_engines(struct drm_device *dev)
4687 struct drm_i915_private *dev_priv = dev->dev_private;
4688 struct intel_engine_cs *engine;
4690 for_each_engine(engine, dev_priv)
4691 dev_priv->gt.stop_engine(engine);
4695 i915_gem_suspend(struct drm_device *dev)
4697 struct drm_i915_private *dev_priv = dev->dev_private;
4700 mutex_lock(&dev->struct_mutex);
4701 ret = i915_gpu_idle(dev);
4705 i915_gem_retire_requests(dev);
4707 i915_gem_stop_engines(dev);
4708 mutex_unlock(&dev->struct_mutex);
4710 cancel_delayed_work_sync(&dev_priv->gpu_error.hangcheck_work);
4711 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
4712 flush_delayed_work(&dev_priv->mm.idle_work);
4714 /* Assert that we sucessfully flushed all the work and
4715 * reset the GPU back to its idle, low power state.
4717 WARN_ON(dev_priv->mm.busy);
4722 mutex_unlock(&dev->struct_mutex);
4726 int i915_gem_l3_remap(struct drm_i915_gem_request *req, int slice)
4728 struct intel_engine_cs *engine = req->engine;
4729 struct drm_device *dev = engine->dev;
4730 struct drm_i915_private *dev_priv = dev->dev_private;
4731 u32 *remap_info = dev_priv->l3_parity.remap_info[slice];
4734 if (!HAS_L3_DPF(dev) || !remap_info)
4737 ret = intel_ring_begin(req, GEN7_L3LOG_SIZE / 4 * 3);
4742 * Note: We do not worry about the concurrent register cacheline hang
4743 * here because no other code should access these registers other than
4744 * at initialization time.
4746 for (i = 0; i < GEN7_L3LOG_SIZE / 4; i++) {
4747 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(1));
4748 intel_ring_emit_reg(engine, GEN7_L3LOG(slice, i));
4749 intel_ring_emit(engine, remap_info[i]);
4752 intel_ring_advance(engine);
4757 void i915_gem_init_swizzling(struct drm_device *dev)
4759 struct drm_i915_private *dev_priv = dev->dev_private;
4761 if (INTEL_INFO(dev)->gen < 5 ||
4762 dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
4765 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
4766 DISP_TILE_SURFACE_SWIZZLING);
4771 I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
4773 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
4774 else if (IS_GEN7(dev))
4775 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
4776 else if (IS_GEN8(dev))
4777 I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW));
4782 static void init_unused_ring(struct drm_device *dev, u32 base)
4784 struct drm_i915_private *dev_priv = dev->dev_private;
4786 I915_WRITE(RING_CTL(base), 0);
4787 I915_WRITE(RING_HEAD(base), 0);
4788 I915_WRITE(RING_TAIL(base), 0);
4789 I915_WRITE(RING_START(base), 0);
4792 static void init_unused_rings(struct drm_device *dev)
4795 init_unused_ring(dev, PRB1_BASE);
4796 init_unused_ring(dev, SRB0_BASE);
4797 init_unused_ring(dev, SRB1_BASE);
4798 init_unused_ring(dev, SRB2_BASE);
4799 init_unused_ring(dev, SRB3_BASE);
4800 } else if (IS_GEN2(dev)) {
4801 init_unused_ring(dev, SRB0_BASE);
4802 init_unused_ring(dev, SRB1_BASE);
4803 } else if (IS_GEN3(dev)) {
4804 init_unused_ring(dev, PRB1_BASE);
4805 init_unused_ring(dev, PRB2_BASE);
4809 int i915_gem_init_engines(struct drm_device *dev)
4811 struct drm_i915_private *dev_priv = dev->dev_private;
4814 ret = intel_init_render_ring_buffer(dev);
4819 ret = intel_init_bsd_ring_buffer(dev);
4821 goto cleanup_render_ring;
4825 ret = intel_init_blt_ring_buffer(dev);
4827 goto cleanup_bsd_ring;
4830 if (HAS_VEBOX(dev)) {
4831 ret = intel_init_vebox_ring_buffer(dev);
4833 goto cleanup_blt_ring;
4836 if (HAS_BSD2(dev)) {
4837 ret = intel_init_bsd2_ring_buffer(dev);
4839 goto cleanup_vebox_ring;
4845 intel_cleanup_engine(&dev_priv->engine[VECS]);
4847 intel_cleanup_engine(&dev_priv->engine[BCS]);
4849 intel_cleanup_engine(&dev_priv->engine[VCS]);
4850 cleanup_render_ring:
4851 intel_cleanup_engine(&dev_priv->engine[RCS]);
4857 i915_gem_init_hw(struct drm_device *dev)
4859 struct drm_i915_private *dev_priv = dev->dev_private;
4860 struct intel_engine_cs *engine;
4863 if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt())
4866 /* Double layer security blanket, see i915_gem_init() */
4867 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4869 if (HAS_EDRAM(dev) && INTEL_GEN(dev_priv) < 9)
4870 I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf));
4872 if (IS_HASWELL(dev))
4873 I915_WRITE(MI_PREDICATE_RESULT_2, IS_HSW_GT3(dev) ?
4874 LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED);
4876 if (HAS_PCH_NOP(dev)) {
4877 if (IS_IVYBRIDGE(dev)) {
4878 u32 temp = I915_READ(GEN7_MSG_CTL);
4879 temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK);
4880 I915_WRITE(GEN7_MSG_CTL, temp);
4881 } else if (INTEL_INFO(dev)->gen >= 7) {
4882 u32 temp = I915_READ(HSW_NDE_RSTWRN_OPT);
4883 temp &= ~RESET_PCH_HANDSHAKE_ENABLE;
4884 I915_WRITE(HSW_NDE_RSTWRN_OPT, temp);
4888 i915_gem_init_swizzling(dev);
4891 * At least 830 can leave some of the unused rings
4892 * "active" (ie. head != tail) after resume which
4893 * will prevent c3 entry. Makes sure all unused rings
4896 init_unused_rings(dev);
4898 BUG_ON(!dev_priv->kernel_context);
4900 ret = i915_ppgtt_init_hw(dev);
4902 DRM_ERROR("PPGTT enable HW failed %d\n", ret);
4906 /* Need to do basic initialisation of all rings first: */
4907 for_each_engine(engine, dev_priv) {
4908 ret = engine->init_hw(engine);
4913 intel_mocs_init_l3cc_table(dev);
4915 /* We can't enable contexts until all firmware is loaded */
4916 if (HAS_GUC_UCODE(dev)) {
4917 ret = intel_guc_ucode_load(dev);
4919 DRM_ERROR("Failed to initialize GuC, error %d\n", ret);
4926 * Increment the next seqno by 0x100 so we have a visible break
4927 * on re-initialisation
4929 ret = i915_gem_set_seqno(dev, dev_priv->next_seqno+0x100);
4933 /* Now it is safe to go back round and do everything else: */
4934 for_each_engine(engine, dev_priv) {
4935 struct drm_i915_gem_request *req;
4937 req = i915_gem_request_alloc(engine, NULL);
4943 if (engine->id == RCS) {
4944 for (j = 0; j < NUM_L3_SLICES(dev); j++) {
4945 ret = i915_gem_l3_remap(req, j);
4951 ret = i915_ppgtt_init_ring(req);
4955 ret = i915_gem_context_enable(req);
4960 i915_add_request_no_flush(req);
4962 DRM_ERROR("Failed to enable %s, error=%d\n",
4964 i915_gem_cleanup_engines(dev);
4970 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4974 int i915_gem_init(struct drm_device *dev)
4976 struct drm_i915_private *dev_priv = dev->dev_private;
4979 i915.enable_execlists = intel_sanitize_enable_execlists(dev,
4980 i915.enable_execlists);
4982 mutex_lock(&dev->struct_mutex);
4984 if (!i915.enable_execlists) {
4985 dev_priv->gt.execbuf_submit = i915_gem_ringbuffer_submission;
4986 dev_priv->gt.init_engines = i915_gem_init_engines;
4987 dev_priv->gt.cleanup_engine = intel_cleanup_engine;
4988 dev_priv->gt.stop_engine = intel_stop_engine;
4990 dev_priv->gt.execbuf_submit = intel_execlists_submission;
4991 dev_priv->gt.init_engines = intel_logical_rings_init;
4992 dev_priv->gt.cleanup_engine = intel_logical_ring_cleanup;
4993 dev_priv->gt.stop_engine = intel_logical_ring_stop;
4996 /* This is just a security blanket to placate dragons.
4997 * On some systems, we very sporadically observe that the first TLBs
4998 * used by the CS may be stale, despite us poking the TLB reset. If
4999 * we hold the forcewake during initialisation these problems
5000 * just magically go away.
5002 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5004 ret = i915_gem_init_userptr(dev);
5008 i915_gem_init_ggtt(dev);
5010 ret = i915_gem_context_init(dev);
5014 ret = dev_priv->gt.init_engines(dev);
5018 ret = i915_gem_init_hw(dev);
5020 /* Allow ring initialisation to fail by marking the GPU as
5021 * wedged. But we only want to do this where the GPU is angry,
5022 * for all other failure, such as an allocation failure, bail.
5024 DRM_ERROR("Failed to initialize GPU, declaring it wedged\n");
5025 atomic_or(I915_WEDGED, &dev_priv->gpu_error.reset_counter);
5030 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5031 mutex_unlock(&dev->struct_mutex);
5037 i915_gem_cleanup_engines(struct drm_device *dev)
5039 struct drm_i915_private *dev_priv = dev->dev_private;
5040 struct intel_engine_cs *engine;
5042 for_each_engine(engine, dev_priv)
5043 dev_priv->gt.cleanup_engine(engine);
5045 if (i915.enable_execlists)
5047 * Neither the BIOS, ourselves or any other kernel
5048 * expects the system to be in execlists mode on startup,
5049 * so we need to reset the GPU back to legacy mode.
5051 intel_gpu_reset(dev, ALL_ENGINES);
5055 init_engine_lists(struct intel_engine_cs *engine)
5057 INIT_LIST_HEAD(&engine->active_list);
5058 INIT_LIST_HEAD(&engine->request_list);
5062 i915_gem_load_init_fences(struct drm_i915_private *dev_priv)
5064 struct drm_device *dev = dev_priv->dev;
5066 if (INTEL_INFO(dev_priv)->gen >= 7 && !IS_VALLEYVIEW(dev_priv) &&
5067 !IS_CHERRYVIEW(dev_priv))
5068 dev_priv->num_fence_regs = 32;
5069 else if (INTEL_INFO(dev_priv)->gen >= 4 || IS_I945G(dev_priv) ||
5070 IS_I945GM(dev_priv) || IS_G33(dev_priv))
5071 dev_priv->num_fence_regs = 16;
5073 dev_priv->num_fence_regs = 8;
5075 if (intel_vgpu_active(dev))
5076 dev_priv->num_fence_regs =
5077 I915_READ(vgtif_reg(avail_rs.fence_num));
5079 /* Initialize fence registers to zero */
5080 i915_gem_restore_fences(dev);
5082 i915_gem_detect_bit_6_swizzle(dev);
5086 i915_gem_load_init(struct drm_device *dev)
5088 struct drm_i915_private *dev_priv = dev->dev_private;
5092 kmem_cache_create("i915_gem_object",
5093 sizeof(struct drm_i915_gem_object), 0,
5097 kmem_cache_create("i915_gem_vma",
5098 sizeof(struct i915_vma), 0,
5101 dev_priv->requests =
5102 kmem_cache_create("i915_gem_request",
5103 sizeof(struct drm_i915_gem_request), 0,
5107 INIT_LIST_HEAD(&dev_priv->vm_list);
5108 INIT_LIST_HEAD(&dev_priv->context_list);
5109 INIT_LIST_HEAD(&dev_priv->mm.unbound_list);
5110 INIT_LIST_HEAD(&dev_priv->mm.bound_list);
5111 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
5112 for (i = 0; i < I915_NUM_ENGINES; i++)
5113 init_engine_lists(&dev_priv->engine[i]);
5114 for (i = 0; i < I915_MAX_NUM_FENCES; i++)
5115 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
5116 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
5117 i915_gem_retire_work_handler);
5118 INIT_DELAYED_WORK(&dev_priv->mm.idle_work,
5119 i915_gem_idle_work_handler);
5120 init_waitqueue_head(&dev_priv->gpu_error.reset_queue);
5122 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
5125 * Set initial sequence number for requests.
5126 * Using this number allows the wraparound to happen early,
5127 * catching any obvious problems.
5129 dev_priv->next_seqno = ((u32)~0 - 0x1100);
5130 dev_priv->last_seqno = ((u32)~0 - 0x1101);
5132 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
5134 init_waitqueue_head(&dev_priv->pending_flip_queue);
5136 dev_priv->mm.interruptible = true;
5138 mutex_init(&dev_priv->fb_tracking.lock);
5141 void i915_gem_load_cleanup(struct drm_device *dev)
5143 struct drm_i915_private *dev_priv = to_i915(dev);
5145 kmem_cache_destroy(dev_priv->requests);
5146 kmem_cache_destroy(dev_priv->vmas);
5147 kmem_cache_destroy(dev_priv->objects);
5150 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
5152 struct drm_i915_file_private *file_priv = file->driver_priv;
5154 /* Clean up our request list when the client is going away, so that
5155 * later retire_requests won't dereference our soon-to-be-gone
5158 spin_lock(&file_priv->mm.lock);
5159 while (!list_empty(&file_priv->mm.request_list)) {
5160 struct drm_i915_gem_request *request;
5162 request = list_first_entry(&file_priv->mm.request_list,
5163 struct drm_i915_gem_request,
5165 list_del(&request->client_list);
5166 request->file_priv = NULL;
5168 spin_unlock(&file_priv->mm.lock);
5170 if (!list_empty(&file_priv->rps.link)) {
5171 spin_lock(&to_i915(dev)->rps.client_lock);
5172 list_del(&file_priv->rps.link);
5173 spin_unlock(&to_i915(dev)->rps.client_lock);
5177 int i915_gem_open(struct drm_device *dev, struct drm_file *file)
5179 struct drm_i915_file_private *file_priv;
5182 DRM_DEBUG_DRIVER("\n");
5184 file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL);
5188 file->driver_priv = file_priv;
5189 file_priv->dev_priv = dev->dev_private;
5190 file_priv->file = file;
5191 INIT_LIST_HEAD(&file_priv->rps.link);
5193 spin_lock_init(&file_priv->mm.lock);
5194 INIT_LIST_HEAD(&file_priv->mm.request_list);
5196 file_priv->bsd_ring = -1;
5198 ret = i915_gem_context_open(dev, file);
5206 * i915_gem_track_fb - update frontbuffer tracking
5207 * @old: current GEM buffer for the frontbuffer slots
5208 * @new: new GEM buffer for the frontbuffer slots
5209 * @frontbuffer_bits: bitmask of frontbuffer slots
5211 * This updates the frontbuffer tracking bits @frontbuffer_bits by clearing them
5212 * from @old and setting them in @new. Both @old and @new can be NULL.
5214 void i915_gem_track_fb(struct drm_i915_gem_object *old,
5215 struct drm_i915_gem_object *new,
5216 unsigned frontbuffer_bits)
5219 WARN_ON(!mutex_is_locked(&old->base.dev->struct_mutex));
5220 WARN_ON(!(old->frontbuffer_bits & frontbuffer_bits));
5221 old->frontbuffer_bits &= ~frontbuffer_bits;
5225 WARN_ON(!mutex_is_locked(&new->base.dev->struct_mutex));
5226 WARN_ON(new->frontbuffer_bits & frontbuffer_bits);
5227 new->frontbuffer_bits |= frontbuffer_bits;
5231 /* All the new VM stuff */
5232 u64 i915_gem_obj_offset(struct drm_i915_gem_object *o,
5233 struct i915_address_space *vm)
5235 struct drm_i915_private *dev_priv = o->base.dev->dev_private;
5236 struct i915_vma *vma;
5238 WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base);
5240 list_for_each_entry(vma, &o->vma_list, obj_link) {
5242 vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL)
5245 return vma->node.start;
5248 WARN(1, "%s vma for this object not found.\n",
5249 i915_is_ggtt(vm) ? "global" : "ppgtt");
5253 u64 i915_gem_obj_ggtt_offset_view(struct drm_i915_gem_object *o,
5254 const struct i915_ggtt_view *view)
5256 struct drm_i915_private *dev_priv = to_i915(o->base.dev);
5257 struct i915_ggtt *ggtt = &dev_priv->ggtt;
5258 struct i915_vma *vma;
5260 list_for_each_entry(vma, &o->vma_list, obj_link)
5261 if (vma->vm == &ggtt->base &&
5262 i915_ggtt_view_equal(&vma->ggtt_view, view))
5263 return vma->node.start;
5265 WARN(1, "global vma for this object not found. (view=%u)\n", view->type);
5269 bool i915_gem_obj_bound(struct drm_i915_gem_object *o,
5270 struct i915_address_space *vm)
5272 struct i915_vma *vma;
5274 list_for_each_entry(vma, &o->vma_list, obj_link) {
5276 vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL)
5278 if (vma->vm == vm && drm_mm_node_allocated(&vma->node))
5285 bool i915_gem_obj_ggtt_bound_view(struct drm_i915_gem_object *o,
5286 const struct i915_ggtt_view *view)
5288 struct drm_i915_private *dev_priv = to_i915(o->base.dev);
5289 struct i915_ggtt *ggtt = &dev_priv->ggtt;
5290 struct i915_vma *vma;
5292 list_for_each_entry(vma, &o->vma_list, obj_link)
5293 if (vma->vm == &ggtt->base &&
5294 i915_ggtt_view_equal(&vma->ggtt_view, view) &&
5295 drm_mm_node_allocated(&vma->node))
5301 bool i915_gem_obj_bound_any(struct drm_i915_gem_object *o)
5303 struct i915_vma *vma;
5305 list_for_each_entry(vma, &o->vma_list, obj_link)
5306 if (drm_mm_node_allocated(&vma->node))
5312 unsigned long i915_gem_obj_size(struct drm_i915_gem_object *o,
5313 struct i915_address_space *vm)
5315 struct drm_i915_private *dev_priv = o->base.dev->dev_private;
5316 struct i915_vma *vma;
5318 WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base);
5320 BUG_ON(list_empty(&o->vma_list));
5322 list_for_each_entry(vma, &o->vma_list, obj_link) {
5324 vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL)
5327 return vma->node.size;
5332 bool i915_gem_obj_is_pinned(struct drm_i915_gem_object *obj)
5334 struct i915_vma *vma;
5335 list_for_each_entry(vma, &obj->vma_list, obj_link)
5336 if (vma->pin_count > 0)
5342 /* Like i915_gem_object_get_page(), but mark the returned page dirty */
5344 i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj, int n)
5348 /* Only default objects have per-page dirty tracking */
5349 if (WARN_ON((obj->ops->flags & I915_GEM_OBJECT_HAS_STRUCT_PAGE) == 0))
5352 page = i915_gem_object_get_page(obj, n);
5353 set_page_dirty(page);
5357 /* Allocate a new GEM object and fill it with the supplied data */
5358 struct drm_i915_gem_object *
5359 i915_gem_object_create_from_data(struct drm_device *dev,
5360 const void *data, size_t size)
5362 struct drm_i915_gem_object *obj;
5363 struct sg_table *sg;
5367 obj = i915_gem_alloc_object(dev, round_up(size, PAGE_SIZE));
5368 if (IS_ERR_OR_NULL(obj))
5371 ret = i915_gem_object_set_to_cpu_domain(obj, true);
5375 ret = i915_gem_object_get_pages(obj);
5379 i915_gem_object_pin_pages(obj);
5381 bytes = sg_copy_from_buffer(sg->sgl, sg->nents, (void *)data, size);
5382 obj->dirty = 1; /* Backing store is now out of date */
5383 i915_gem_object_unpin_pages(obj);
5385 if (WARN_ON(bytes != size)) {
5386 DRM_ERROR("Incomplete copy, wrote %zu of %zu", bytes, size);
5394 drm_gem_object_unreference(&obj->base);
5395 return ERR_PTR(ret);