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 <linux/shmem_fs.h>
36 #include <linux/slab.h>
37 #include <linux/swap.h>
38 #include <linux/pci.h>
39 #include <linux/dma-buf.h>
41 #define RQ_BUG_ON(expr)
43 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);
44 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj);
46 i915_gem_object_retire__write(struct drm_i915_gem_object *obj);
48 i915_gem_object_retire__read(struct drm_i915_gem_object *obj, int ring);
50 static bool cpu_cache_is_coherent(struct drm_device *dev,
51 enum i915_cache_level level)
53 return HAS_LLC(dev) || level != I915_CACHE_NONE;
56 static bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj)
58 if (!cpu_cache_is_coherent(obj->base.dev, obj->cache_level))
61 return obj->pin_display;
64 /* some bookkeeping */
65 static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
68 spin_lock(&dev_priv->mm.object_stat_lock);
69 dev_priv->mm.object_count++;
70 dev_priv->mm.object_memory += size;
71 spin_unlock(&dev_priv->mm.object_stat_lock);
74 static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
77 spin_lock(&dev_priv->mm.object_stat_lock);
78 dev_priv->mm.object_count--;
79 dev_priv->mm.object_memory -= size;
80 spin_unlock(&dev_priv->mm.object_stat_lock);
84 i915_gem_wait_for_error(struct i915_gpu_error *error)
88 #define EXIT_COND (!i915_reset_in_progress(error) || \
89 i915_terminally_wedged(error))
94 * Only wait 10 seconds for the gpu reset to complete to avoid hanging
95 * userspace. If it takes that long something really bad is going on and
96 * we should simply try to bail out and fail as gracefully as possible.
98 ret = wait_event_interruptible_timeout(error->reset_queue,
102 DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
104 } else if (ret < 0) {
112 int i915_mutex_lock_interruptible(struct drm_device *dev)
114 struct drm_i915_private *dev_priv = dev->dev_private;
117 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
121 ret = mutex_lock_interruptible(&dev->struct_mutex);
125 WARN_ON(i915_verify_lists(dev));
130 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
131 struct drm_file *file)
133 struct drm_i915_private *dev_priv = dev->dev_private;
134 struct drm_i915_gem_get_aperture *args = data;
135 struct i915_gtt *ggtt = &dev_priv->gtt;
136 struct i915_vma *vma;
140 mutex_lock(&dev->struct_mutex);
141 list_for_each_entry(vma, &ggtt->base.active_list, mm_list)
143 pinned += vma->node.size;
144 list_for_each_entry(vma, &ggtt->base.inactive_list, mm_list)
146 pinned += vma->node.size;
147 mutex_unlock(&dev->struct_mutex);
149 args->aper_size = dev_priv->gtt.base.total;
150 args->aper_available_size = args->aper_size - pinned;
156 i915_gem_object_get_pages_phys(struct drm_i915_gem_object *obj)
158 struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
159 char *vaddr = obj->phys_handle->vaddr;
161 struct scatterlist *sg;
164 if (WARN_ON(i915_gem_object_needs_bit17_swizzle(obj)))
167 for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
171 page = shmem_read_mapping_page(mapping, i);
173 return PTR_ERR(page);
175 src = kmap_atomic(page);
176 memcpy(vaddr, src, PAGE_SIZE);
177 drm_clflush_virt_range(vaddr, PAGE_SIZE);
180 page_cache_release(page);
184 i915_gem_chipset_flush(obj->base.dev);
186 st = kmalloc(sizeof(*st), GFP_KERNEL);
190 if (sg_alloc_table(st, 1, GFP_KERNEL)) {
197 sg->length = obj->base.size;
199 sg_dma_address(sg) = obj->phys_handle->busaddr;
200 sg_dma_len(sg) = obj->base.size;
207 i915_gem_object_put_pages_phys(struct drm_i915_gem_object *obj)
211 BUG_ON(obj->madv == __I915_MADV_PURGED);
213 ret = i915_gem_object_set_to_cpu_domain(obj, true);
215 /* In the event of a disaster, abandon all caches and
218 WARN_ON(ret != -EIO);
219 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
222 if (obj->madv == I915_MADV_DONTNEED)
226 struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
227 char *vaddr = obj->phys_handle->vaddr;
230 for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
234 page = shmem_read_mapping_page(mapping, i);
238 dst = kmap_atomic(page);
239 drm_clflush_virt_range(vaddr, PAGE_SIZE);
240 memcpy(dst, vaddr, PAGE_SIZE);
243 set_page_dirty(page);
244 if (obj->madv == I915_MADV_WILLNEED)
245 mark_page_accessed(page);
246 page_cache_release(page);
252 sg_free_table(obj->pages);
257 i915_gem_object_release_phys(struct drm_i915_gem_object *obj)
259 drm_pci_free(obj->base.dev, obj->phys_handle);
262 static const struct drm_i915_gem_object_ops i915_gem_phys_ops = {
263 .get_pages = i915_gem_object_get_pages_phys,
264 .put_pages = i915_gem_object_put_pages_phys,
265 .release = i915_gem_object_release_phys,
269 drop_pages(struct drm_i915_gem_object *obj)
271 struct i915_vma *vma, *next;
274 drm_gem_object_reference(&obj->base);
275 list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link)
276 if (i915_vma_unbind(vma))
279 ret = i915_gem_object_put_pages(obj);
280 drm_gem_object_unreference(&obj->base);
286 i915_gem_object_attach_phys(struct drm_i915_gem_object *obj,
289 drm_dma_handle_t *phys;
292 if (obj->phys_handle) {
293 if ((unsigned long)obj->phys_handle->vaddr & (align -1))
299 if (obj->madv != I915_MADV_WILLNEED)
302 if (obj->base.filp == NULL)
305 ret = drop_pages(obj);
309 /* create a new object */
310 phys = drm_pci_alloc(obj->base.dev, obj->base.size, align);
314 obj->phys_handle = phys;
315 obj->ops = &i915_gem_phys_ops;
317 return i915_gem_object_get_pages(obj);
321 i915_gem_phys_pwrite(struct drm_i915_gem_object *obj,
322 struct drm_i915_gem_pwrite *args,
323 struct drm_file *file_priv)
325 struct drm_device *dev = obj->base.dev;
326 void *vaddr = obj->phys_handle->vaddr + args->offset;
327 char __user *user_data = to_user_ptr(args->data_ptr);
330 /* We manually control the domain here and pretend that it
331 * remains coherent i.e. in the GTT domain, like shmem_pwrite.
333 ret = i915_gem_object_wait_rendering(obj, false);
337 intel_fb_obj_invalidate(obj, ORIGIN_CPU);
338 if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
339 unsigned long unwritten;
341 /* The physical object once assigned is fixed for the lifetime
342 * of the obj, so we can safely drop the lock and continue
345 mutex_unlock(&dev->struct_mutex);
346 unwritten = copy_from_user(vaddr, user_data, args->size);
347 mutex_lock(&dev->struct_mutex);
354 drm_clflush_virt_range(vaddr, args->size);
355 i915_gem_chipset_flush(dev);
358 intel_fb_obj_flush(obj, false, ORIGIN_CPU);
362 void *i915_gem_object_alloc(struct drm_device *dev)
364 struct drm_i915_private *dev_priv = dev->dev_private;
365 return kmem_cache_zalloc(dev_priv->objects, GFP_KERNEL);
368 void i915_gem_object_free(struct drm_i915_gem_object *obj)
370 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
371 kmem_cache_free(dev_priv->objects, obj);
375 i915_gem_create(struct drm_file *file,
376 struct drm_device *dev,
380 struct drm_i915_gem_object *obj;
384 size = roundup(size, PAGE_SIZE);
388 /* Allocate the new object */
389 obj = i915_gem_alloc_object(dev, size);
393 ret = drm_gem_handle_create(file, &obj->base, &handle);
394 /* drop reference from allocate - handle holds it now */
395 drm_gem_object_unreference_unlocked(&obj->base);
404 i915_gem_dumb_create(struct drm_file *file,
405 struct drm_device *dev,
406 struct drm_mode_create_dumb *args)
408 /* have to work out size/pitch and return them */
409 args->pitch = ALIGN(args->width * DIV_ROUND_UP(args->bpp, 8), 64);
410 args->size = args->pitch * args->height;
411 return i915_gem_create(file, dev,
412 args->size, &args->handle);
416 * Creates a new mm object and returns a handle to it.
419 i915_gem_create_ioctl(struct drm_device *dev, void *data,
420 struct drm_file *file)
422 struct drm_i915_gem_create *args = data;
424 return i915_gem_create(file, dev,
425 args->size, &args->handle);
429 __copy_to_user_swizzled(char __user *cpu_vaddr,
430 const char *gpu_vaddr, int gpu_offset,
433 int ret, cpu_offset = 0;
436 int cacheline_end = ALIGN(gpu_offset + 1, 64);
437 int this_length = min(cacheline_end - gpu_offset, length);
438 int swizzled_gpu_offset = gpu_offset ^ 64;
440 ret = __copy_to_user(cpu_vaddr + cpu_offset,
441 gpu_vaddr + swizzled_gpu_offset,
446 cpu_offset += this_length;
447 gpu_offset += this_length;
448 length -= this_length;
455 __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset,
456 const char __user *cpu_vaddr,
459 int ret, cpu_offset = 0;
462 int cacheline_end = ALIGN(gpu_offset + 1, 64);
463 int this_length = min(cacheline_end - gpu_offset, length);
464 int swizzled_gpu_offset = gpu_offset ^ 64;
466 ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset,
467 cpu_vaddr + cpu_offset,
472 cpu_offset += this_length;
473 gpu_offset += this_length;
474 length -= this_length;
481 * Pins the specified object's pages and synchronizes the object with
482 * GPU accesses. Sets needs_clflush to non-zero if the caller should
483 * flush the object from the CPU cache.
485 int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj,
495 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) {
496 /* If we're not in the cpu read domain, set ourself into the gtt
497 * read domain and manually flush cachelines (if required). This
498 * optimizes for the case when the gpu will dirty the data
499 * anyway again before the next pread happens. */
500 *needs_clflush = !cpu_cache_is_coherent(obj->base.dev,
502 ret = i915_gem_object_wait_rendering(obj, true);
507 ret = i915_gem_object_get_pages(obj);
511 i915_gem_object_pin_pages(obj);
516 /* Per-page copy function for the shmem pread fastpath.
517 * Flushes invalid cachelines before reading the target if
518 * needs_clflush is set. */
520 shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length,
521 char __user *user_data,
522 bool page_do_bit17_swizzling, bool needs_clflush)
527 if (unlikely(page_do_bit17_swizzling))
530 vaddr = kmap_atomic(page);
532 drm_clflush_virt_range(vaddr + shmem_page_offset,
534 ret = __copy_to_user_inatomic(user_data,
535 vaddr + shmem_page_offset,
537 kunmap_atomic(vaddr);
539 return ret ? -EFAULT : 0;
543 shmem_clflush_swizzled_range(char *addr, unsigned long length,
546 if (unlikely(swizzled)) {
547 unsigned long start = (unsigned long) addr;
548 unsigned long end = (unsigned long) addr + length;
550 /* For swizzling simply ensure that we always flush both
551 * channels. Lame, but simple and it works. Swizzled
552 * pwrite/pread is far from a hotpath - current userspace
553 * doesn't use it at all. */
554 start = round_down(start, 128);
555 end = round_up(end, 128);
557 drm_clflush_virt_range((void *)start, end - start);
559 drm_clflush_virt_range(addr, length);
564 /* Only difference to the fast-path function is that this can handle bit17
565 * and uses non-atomic copy and kmap functions. */
567 shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length,
568 char __user *user_data,
569 bool page_do_bit17_swizzling, bool needs_clflush)
576 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
578 page_do_bit17_swizzling);
580 if (page_do_bit17_swizzling)
581 ret = __copy_to_user_swizzled(user_data,
582 vaddr, shmem_page_offset,
585 ret = __copy_to_user(user_data,
586 vaddr + shmem_page_offset,
590 return ret ? - EFAULT : 0;
594 i915_gem_shmem_pread(struct drm_device *dev,
595 struct drm_i915_gem_object *obj,
596 struct drm_i915_gem_pread *args,
597 struct drm_file *file)
599 char __user *user_data;
602 int shmem_page_offset, page_length, ret = 0;
603 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
605 int needs_clflush = 0;
606 struct sg_page_iter sg_iter;
608 user_data = to_user_ptr(args->data_ptr);
611 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
613 ret = i915_gem_obj_prepare_shmem_read(obj, &needs_clflush);
617 offset = args->offset;
619 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
620 offset >> PAGE_SHIFT) {
621 struct page *page = sg_page_iter_page(&sg_iter);
626 /* Operation in this page
628 * shmem_page_offset = offset within page in shmem file
629 * page_length = bytes to copy for this page
631 shmem_page_offset = offset_in_page(offset);
632 page_length = remain;
633 if ((shmem_page_offset + page_length) > PAGE_SIZE)
634 page_length = PAGE_SIZE - shmem_page_offset;
636 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
637 (page_to_phys(page) & (1 << 17)) != 0;
639 ret = shmem_pread_fast(page, shmem_page_offset, page_length,
640 user_data, page_do_bit17_swizzling,
645 mutex_unlock(&dev->struct_mutex);
647 if (likely(!i915.prefault_disable) && !prefaulted) {
648 ret = fault_in_multipages_writeable(user_data, remain);
649 /* Userspace is tricking us, but we've already clobbered
650 * its pages with the prefault and promised to write the
651 * data up to the first fault. Hence ignore any errors
652 * and just continue. */
657 ret = shmem_pread_slow(page, shmem_page_offset, page_length,
658 user_data, page_do_bit17_swizzling,
661 mutex_lock(&dev->struct_mutex);
667 remain -= page_length;
668 user_data += page_length;
669 offset += page_length;
673 i915_gem_object_unpin_pages(obj);
679 * Reads data from the object referenced by handle.
681 * On error, the contents of *data are undefined.
684 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
685 struct drm_file *file)
687 struct drm_i915_gem_pread *args = data;
688 struct drm_i915_gem_object *obj;
694 if (!access_ok(VERIFY_WRITE,
695 to_user_ptr(args->data_ptr),
699 ret = i915_mutex_lock_interruptible(dev);
703 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
704 if (&obj->base == NULL) {
709 /* Bounds check source. */
710 if (args->offset > obj->base.size ||
711 args->size > obj->base.size - args->offset) {
716 /* prime objects have no backing filp to GEM pread/pwrite
719 if (!obj->base.filp) {
724 trace_i915_gem_object_pread(obj, args->offset, args->size);
726 ret = i915_gem_shmem_pread(dev, obj, args, file);
729 drm_gem_object_unreference(&obj->base);
731 mutex_unlock(&dev->struct_mutex);
735 /* This is the fast write path which cannot handle
736 * page faults in the source data
740 fast_user_write(struct io_mapping *mapping,
741 loff_t page_base, int page_offset,
742 char __user *user_data,
745 void __iomem *vaddr_atomic;
747 unsigned long unwritten;
749 vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
750 /* We can use the cpu mem copy function because this is X86. */
751 vaddr = (void __force*)vaddr_atomic + page_offset;
752 unwritten = __copy_from_user_inatomic_nocache(vaddr,
754 io_mapping_unmap_atomic(vaddr_atomic);
759 * This is the fast pwrite path, where we copy the data directly from the
760 * user into the GTT, uncached.
763 i915_gem_gtt_pwrite_fast(struct drm_device *dev,
764 struct drm_i915_gem_object *obj,
765 struct drm_i915_gem_pwrite *args,
766 struct drm_file *file)
768 struct drm_i915_private *dev_priv = dev->dev_private;
770 loff_t offset, page_base;
771 char __user *user_data;
772 int page_offset, page_length, ret;
774 ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE | PIN_NONBLOCK);
778 ret = i915_gem_object_set_to_gtt_domain(obj, true);
782 ret = i915_gem_object_put_fence(obj);
786 user_data = to_user_ptr(args->data_ptr);
789 offset = i915_gem_obj_ggtt_offset(obj) + args->offset;
791 intel_fb_obj_invalidate(obj, ORIGIN_GTT);
794 /* Operation in this page
796 * page_base = page offset within aperture
797 * page_offset = offset within page
798 * page_length = bytes to copy for this page
800 page_base = offset & PAGE_MASK;
801 page_offset = offset_in_page(offset);
802 page_length = remain;
803 if ((page_offset + remain) > PAGE_SIZE)
804 page_length = PAGE_SIZE - page_offset;
806 /* If we get a fault while copying data, then (presumably) our
807 * source page isn't available. Return the error and we'll
808 * retry in the slow path.
810 if (fast_user_write(dev_priv->gtt.mappable, page_base,
811 page_offset, user_data, page_length)) {
816 remain -= page_length;
817 user_data += page_length;
818 offset += page_length;
822 intel_fb_obj_flush(obj, false, ORIGIN_GTT);
824 i915_gem_object_ggtt_unpin(obj);
829 /* Per-page copy function for the shmem pwrite fastpath.
830 * Flushes invalid cachelines before writing to the target if
831 * needs_clflush_before is set and flushes out any written cachelines after
832 * writing if needs_clflush is set. */
834 shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
835 char __user *user_data,
836 bool page_do_bit17_swizzling,
837 bool needs_clflush_before,
838 bool needs_clflush_after)
843 if (unlikely(page_do_bit17_swizzling))
846 vaddr = kmap_atomic(page);
847 if (needs_clflush_before)
848 drm_clflush_virt_range(vaddr + shmem_page_offset,
850 ret = __copy_from_user_inatomic(vaddr + shmem_page_offset,
851 user_data, page_length);
852 if (needs_clflush_after)
853 drm_clflush_virt_range(vaddr + shmem_page_offset,
855 kunmap_atomic(vaddr);
857 return ret ? -EFAULT : 0;
860 /* Only difference to the fast-path function is that this can handle bit17
861 * and uses non-atomic copy and kmap functions. */
863 shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
864 char __user *user_data,
865 bool page_do_bit17_swizzling,
866 bool needs_clflush_before,
867 bool needs_clflush_after)
873 if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
874 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
876 page_do_bit17_swizzling);
877 if (page_do_bit17_swizzling)
878 ret = __copy_from_user_swizzled(vaddr, shmem_page_offset,
882 ret = __copy_from_user(vaddr + shmem_page_offset,
885 if (needs_clflush_after)
886 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
888 page_do_bit17_swizzling);
891 return ret ? -EFAULT : 0;
895 i915_gem_shmem_pwrite(struct drm_device *dev,
896 struct drm_i915_gem_object *obj,
897 struct drm_i915_gem_pwrite *args,
898 struct drm_file *file)
902 char __user *user_data;
903 int shmem_page_offset, page_length, ret = 0;
904 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
905 int hit_slowpath = 0;
906 int needs_clflush_after = 0;
907 int needs_clflush_before = 0;
908 struct sg_page_iter sg_iter;
910 user_data = to_user_ptr(args->data_ptr);
913 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
915 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
916 /* If we're not in the cpu write domain, set ourself into the gtt
917 * write domain and manually flush cachelines (if required). This
918 * optimizes for the case when the gpu will use the data
919 * right away and we therefore have to clflush anyway. */
920 needs_clflush_after = cpu_write_needs_clflush(obj);
921 ret = i915_gem_object_wait_rendering(obj, false);
925 /* Same trick applies to invalidate partially written cachelines read
927 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0)
928 needs_clflush_before =
929 !cpu_cache_is_coherent(dev, obj->cache_level);
931 ret = i915_gem_object_get_pages(obj);
935 intel_fb_obj_invalidate(obj, ORIGIN_CPU);
937 i915_gem_object_pin_pages(obj);
939 offset = args->offset;
942 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
943 offset >> PAGE_SHIFT) {
944 struct page *page = sg_page_iter_page(&sg_iter);
945 int partial_cacheline_write;
950 /* Operation in this page
952 * shmem_page_offset = offset within page in shmem file
953 * page_length = bytes to copy for this page
955 shmem_page_offset = offset_in_page(offset);
957 page_length = remain;
958 if ((shmem_page_offset + page_length) > PAGE_SIZE)
959 page_length = PAGE_SIZE - shmem_page_offset;
961 /* If we don't overwrite a cacheline completely we need to be
962 * careful to have up-to-date data by first clflushing. Don't
963 * overcomplicate things and flush the entire patch. */
964 partial_cacheline_write = needs_clflush_before &&
965 ((shmem_page_offset | page_length)
966 & (boot_cpu_data.x86_clflush_size - 1));
968 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
969 (page_to_phys(page) & (1 << 17)) != 0;
971 ret = shmem_pwrite_fast(page, shmem_page_offset, page_length,
972 user_data, page_do_bit17_swizzling,
973 partial_cacheline_write,
974 needs_clflush_after);
979 mutex_unlock(&dev->struct_mutex);
980 ret = shmem_pwrite_slow(page, shmem_page_offset, page_length,
981 user_data, page_do_bit17_swizzling,
982 partial_cacheline_write,
983 needs_clflush_after);
985 mutex_lock(&dev->struct_mutex);
991 remain -= page_length;
992 user_data += page_length;
993 offset += page_length;
997 i915_gem_object_unpin_pages(obj);
1001 * Fixup: Flush cpu caches in case we didn't flush the dirty
1002 * cachelines in-line while writing and the object moved
1003 * out of the cpu write domain while we've dropped the lock.
1005 if (!needs_clflush_after &&
1006 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
1007 if (i915_gem_clflush_object(obj, obj->pin_display))
1008 needs_clflush_after = true;
1012 if (needs_clflush_after)
1013 i915_gem_chipset_flush(dev);
1015 obj->cache_dirty = true;
1017 intel_fb_obj_flush(obj, false, ORIGIN_CPU);
1022 * Writes data to the object referenced by handle.
1024 * On error, the contents of the buffer that were to be modified are undefined.
1027 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
1028 struct drm_file *file)
1030 struct drm_i915_private *dev_priv = dev->dev_private;
1031 struct drm_i915_gem_pwrite *args = data;
1032 struct drm_i915_gem_object *obj;
1035 if (args->size == 0)
1038 if (!access_ok(VERIFY_READ,
1039 to_user_ptr(args->data_ptr),
1043 if (likely(!i915.prefault_disable)) {
1044 ret = fault_in_multipages_readable(to_user_ptr(args->data_ptr),
1050 intel_runtime_pm_get(dev_priv);
1052 ret = i915_mutex_lock_interruptible(dev);
1056 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1057 if (&obj->base == NULL) {
1062 /* Bounds check destination. */
1063 if (args->offset > obj->base.size ||
1064 args->size > obj->base.size - args->offset) {
1069 /* prime objects have no backing filp to GEM pread/pwrite
1072 if (!obj->base.filp) {
1077 trace_i915_gem_object_pwrite(obj, args->offset, args->size);
1080 /* We can only do the GTT pwrite on untiled buffers, as otherwise
1081 * it would end up going through the fenced access, and we'll get
1082 * different detiling behavior between reading and writing.
1083 * pread/pwrite currently are reading and writing from the CPU
1084 * perspective, requiring manual detiling by the client.
1086 if (obj->tiling_mode == I915_TILING_NONE &&
1087 obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
1088 cpu_write_needs_clflush(obj)) {
1089 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
1090 /* Note that the gtt paths might fail with non-page-backed user
1091 * pointers (e.g. gtt mappings when moving data between
1092 * textures). Fallback to the shmem path in that case. */
1095 if (ret == -EFAULT || ret == -ENOSPC) {
1096 if (obj->phys_handle)
1097 ret = i915_gem_phys_pwrite(obj, args, file);
1099 ret = i915_gem_shmem_pwrite(dev, obj, args, file);
1103 drm_gem_object_unreference(&obj->base);
1105 mutex_unlock(&dev->struct_mutex);
1107 intel_runtime_pm_put(dev_priv);
1113 i915_gem_check_wedge(struct i915_gpu_error *error,
1116 if (i915_reset_in_progress(error)) {
1117 /* Non-interruptible callers can't handle -EAGAIN, hence return
1118 * -EIO unconditionally for these. */
1122 /* Recovery complete, but the reset failed ... */
1123 if (i915_terminally_wedged(error))
1127 * Check if GPU Reset is in progress - we need intel_ring_begin
1128 * to work properly to reinit the hw state while the gpu is
1129 * still marked as reset-in-progress. Handle this with a flag.
1131 if (!error->reload_in_reset)
1138 static void fake_irq(unsigned long data)
1140 wake_up_process((struct task_struct *)data);
1143 static bool missed_irq(struct drm_i915_private *dev_priv,
1144 struct intel_engine_cs *ring)
1146 return test_bit(ring->id, &dev_priv->gpu_error.missed_irq_rings);
1149 static int __i915_spin_request(struct drm_i915_gem_request *req)
1151 unsigned long timeout;
1153 if (i915_gem_request_get_ring(req)->irq_refcount)
1156 timeout = jiffies + 1;
1157 while (!need_resched()) {
1158 if (i915_gem_request_completed(req, true))
1161 if (time_after_eq(jiffies, timeout))
1164 cpu_relax_lowlatency();
1166 if (i915_gem_request_completed(req, false))
1173 * __i915_wait_request - wait until execution of request has finished
1175 * @reset_counter: reset sequence associated with the given request
1176 * @interruptible: do an interruptible wait (normally yes)
1177 * @timeout: in - how long to wait (NULL forever); out - how much time remaining
1179 * Note: It is of utmost importance that the passed in seqno and reset_counter
1180 * values have been read by the caller in an smp safe manner. Where read-side
1181 * locks are involved, it is sufficient to read the reset_counter before
1182 * unlocking the lock that protects the seqno. For lockless tricks, the
1183 * reset_counter _must_ be read before, and an appropriate smp_rmb must be
1186 * Returns 0 if the request was found within the alloted time. Else returns the
1187 * errno with remaining time filled in timeout argument.
1189 int __i915_wait_request(struct drm_i915_gem_request *req,
1190 unsigned reset_counter,
1193 struct intel_rps_client *rps)
1195 struct intel_engine_cs *ring = i915_gem_request_get_ring(req);
1196 struct drm_device *dev = ring->dev;
1197 struct drm_i915_private *dev_priv = dev->dev_private;
1198 const bool irq_test_in_progress =
1199 ACCESS_ONCE(dev_priv->gpu_error.test_irq_rings) & intel_ring_flag(ring);
1201 unsigned long timeout_expire;
1205 WARN(!intel_irqs_enabled(dev_priv), "IRQs disabled");
1207 if (list_empty(&req->list))
1210 if (i915_gem_request_completed(req, true))
1213 timeout_expire = timeout ?
1214 jiffies + nsecs_to_jiffies_timeout((u64)*timeout) : 0;
1216 if (INTEL_INFO(dev_priv)->gen >= 6)
1217 gen6_rps_boost(dev_priv, rps, req->emitted_jiffies);
1219 /* Record current time in case interrupted by signal, or wedged */
1220 trace_i915_gem_request_wait_begin(req);
1221 before = ktime_get_raw_ns();
1223 /* Optimistic spin for the next jiffie before touching IRQs */
1224 ret = __i915_spin_request(req);
1228 if (!irq_test_in_progress && WARN_ON(!ring->irq_get(ring))) {
1234 struct timer_list timer;
1236 prepare_to_wait(&ring->irq_queue, &wait,
1237 interruptible ? TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE);
1239 /* We need to check whether any gpu reset happened in between
1240 * the caller grabbing the seqno and now ... */
1241 if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) {
1242 /* ... but upgrade the -EAGAIN to an -EIO if the gpu
1243 * is truely gone. */
1244 ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1250 if (i915_gem_request_completed(req, false)) {
1255 if (interruptible && signal_pending(current)) {
1260 if (timeout && time_after_eq(jiffies, timeout_expire)) {
1265 timer.function = NULL;
1266 if (timeout || missed_irq(dev_priv, ring)) {
1267 unsigned long expire;
1269 setup_timer_on_stack(&timer, fake_irq, (unsigned long)current);
1270 expire = missed_irq(dev_priv, ring) ? jiffies + 1 : timeout_expire;
1271 mod_timer(&timer, expire);
1276 if (timer.function) {
1277 del_singleshot_timer_sync(&timer);
1278 destroy_timer_on_stack(&timer);
1281 if (!irq_test_in_progress)
1282 ring->irq_put(ring);
1284 finish_wait(&ring->irq_queue, &wait);
1287 now = ktime_get_raw_ns();
1288 trace_i915_gem_request_wait_end(req);
1291 s64 tres = *timeout - (now - before);
1293 *timeout = tres < 0 ? 0 : tres;
1296 * Apparently ktime isn't accurate enough and occasionally has a
1297 * bit of mismatch in the jiffies<->nsecs<->ktime loop. So patch
1298 * things up to make the test happy. We allow up to 1 jiffy.
1300 * This is a regrssion from the timespec->ktime conversion.
1302 if (ret == -ETIME && *timeout < jiffies_to_usecs(1)*1000)
1309 int i915_gem_request_add_to_client(struct drm_i915_gem_request *req,
1310 struct drm_file *file)
1312 struct drm_i915_private *dev_private;
1313 struct drm_i915_file_private *file_priv;
1315 WARN_ON(!req || !file || req->file_priv);
1323 dev_private = req->ring->dev->dev_private;
1324 file_priv = file->driver_priv;
1326 spin_lock(&file_priv->mm.lock);
1327 req->file_priv = file_priv;
1328 list_add_tail(&req->client_list, &file_priv->mm.request_list);
1329 spin_unlock(&file_priv->mm.lock);
1331 req->pid = get_pid(task_pid(current));
1337 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
1339 struct drm_i915_file_private *file_priv = request->file_priv;
1344 spin_lock(&file_priv->mm.lock);
1345 list_del(&request->client_list);
1346 request->file_priv = NULL;
1347 spin_unlock(&file_priv->mm.lock);
1349 put_pid(request->pid);
1350 request->pid = NULL;
1353 static void i915_gem_request_retire(struct drm_i915_gem_request *request)
1355 trace_i915_gem_request_retire(request);
1357 /* We know the GPU must have read the request to have
1358 * sent us the seqno + interrupt, so use the position
1359 * of tail of the request to update the last known position
1362 * Note this requires that we are always called in request
1365 request->ringbuf->last_retired_head = request->postfix;
1367 list_del_init(&request->list);
1368 i915_gem_request_remove_from_client(request);
1370 i915_gem_request_unreference(request);
1374 __i915_gem_request_retire__upto(struct drm_i915_gem_request *req)
1376 struct intel_engine_cs *engine = req->ring;
1377 struct drm_i915_gem_request *tmp;
1379 lockdep_assert_held(&engine->dev->struct_mutex);
1381 if (list_empty(&req->list))
1385 tmp = list_first_entry(&engine->request_list,
1386 typeof(*tmp), list);
1388 i915_gem_request_retire(tmp);
1389 } while (tmp != req);
1391 WARN_ON(i915_verify_lists(engine->dev));
1395 * Waits for a request to be signaled, and cleans up the
1396 * request and object lists appropriately for that event.
1399 i915_wait_request(struct drm_i915_gem_request *req)
1401 struct drm_device *dev;
1402 struct drm_i915_private *dev_priv;
1406 BUG_ON(req == NULL);
1408 dev = req->ring->dev;
1409 dev_priv = dev->dev_private;
1410 interruptible = dev_priv->mm.interruptible;
1412 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1414 ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1418 ret = __i915_wait_request(req,
1419 atomic_read(&dev_priv->gpu_error.reset_counter),
1420 interruptible, NULL, NULL);
1424 __i915_gem_request_retire__upto(req);
1429 * Ensures that all rendering to the object has completed and the object is
1430 * safe to unbind from the GTT or access from the CPU.
1433 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
1442 if (obj->last_write_req != NULL) {
1443 ret = i915_wait_request(obj->last_write_req);
1447 i = obj->last_write_req->ring->id;
1448 if (obj->last_read_req[i] == obj->last_write_req)
1449 i915_gem_object_retire__read(obj, i);
1451 i915_gem_object_retire__write(obj);
1454 for (i = 0; i < I915_NUM_RINGS; i++) {
1455 if (obj->last_read_req[i] == NULL)
1458 ret = i915_wait_request(obj->last_read_req[i]);
1462 i915_gem_object_retire__read(obj, i);
1464 RQ_BUG_ON(obj->active);
1471 i915_gem_object_retire_request(struct drm_i915_gem_object *obj,
1472 struct drm_i915_gem_request *req)
1474 int ring = req->ring->id;
1476 if (obj->last_read_req[ring] == req)
1477 i915_gem_object_retire__read(obj, ring);
1478 else if (obj->last_write_req == req)
1479 i915_gem_object_retire__write(obj);
1481 __i915_gem_request_retire__upto(req);
1484 /* A nonblocking variant of the above wait. This is a highly dangerous routine
1485 * as the object state may change during this call.
1487 static __must_check int
1488 i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
1489 struct intel_rps_client *rps,
1492 struct drm_device *dev = obj->base.dev;
1493 struct drm_i915_private *dev_priv = dev->dev_private;
1494 struct drm_i915_gem_request *requests[I915_NUM_RINGS];
1495 unsigned reset_counter;
1498 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1499 BUG_ON(!dev_priv->mm.interruptible);
1504 ret = i915_gem_check_wedge(&dev_priv->gpu_error, true);
1508 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
1511 struct drm_i915_gem_request *req;
1513 req = obj->last_write_req;
1517 requests[n++] = i915_gem_request_reference(req);
1519 for (i = 0; i < I915_NUM_RINGS; i++) {
1520 struct drm_i915_gem_request *req;
1522 req = obj->last_read_req[i];
1526 requests[n++] = i915_gem_request_reference(req);
1530 mutex_unlock(&dev->struct_mutex);
1531 for (i = 0; ret == 0 && i < n; i++)
1532 ret = __i915_wait_request(requests[i], reset_counter, true,
1534 mutex_lock(&dev->struct_mutex);
1536 for (i = 0; i < n; i++) {
1538 i915_gem_object_retire_request(obj, requests[i]);
1539 i915_gem_request_unreference(requests[i]);
1545 static struct intel_rps_client *to_rps_client(struct drm_file *file)
1547 struct drm_i915_file_private *fpriv = file->driver_priv;
1552 * Called when user space prepares to use an object with the CPU, either
1553 * through the mmap ioctl's mapping or a GTT mapping.
1556 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1557 struct drm_file *file)
1559 struct drm_i915_gem_set_domain *args = data;
1560 struct drm_i915_gem_object *obj;
1561 uint32_t read_domains = args->read_domains;
1562 uint32_t write_domain = args->write_domain;
1565 /* Only handle setting domains to types used by the CPU. */
1566 if (write_domain & I915_GEM_GPU_DOMAINS)
1569 if (read_domains & I915_GEM_GPU_DOMAINS)
1572 /* Having something in the write domain implies it's in the read
1573 * domain, and only that read domain. Enforce that in the request.
1575 if (write_domain != 0 && read_domains != write_domain)
1578 ret = i915_mutex_lock_interruptible(dev);
1582 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1583 if (&obj->base == NULL) {
1588 /* Try to flush the object off the GPU without holding the lock.
1589 * We will repeat the flush holding the lock in the normal manner
1590 * to catch cases where we are gazumped.
1592 ret = i915_gem_object_wait_rendering__nonblocking(obj,
1593 to_rps_client(file),
1598 if (read_domains & I915_GEM_DOMAIN_GTT)
1599 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1601 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1603 if (write_domain != 0)
1604 intel_fb_obj_invalidate(obj,
1605 write_domain == I915_GEM_DOMAIN_GTT ?
1606 ORIGIN_GTT : ORIGIN_CPU);
1609 drm_gem_object_unreference(&obj->base);
1611 mutex_unlock(&dev->struct_mutex);
1616 * Called when user space has done writes to this buffer
1619 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1620 struct drm_file *file)
1622 struct drm_i915_gem_sw_finish *args = data;
1623 struct drm_i915_gem_object *obj;
1626 ret = i915_mutex_lock_interruptible(dev);
1630 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1631 if (&obj->base == NULL) {
1636 /* Pinned buffers may be scanout, so flush the cache */
1637 if (obj->pin_display)
1638 i915_gem_object_flush_cpu_write_domain(obj);
1640 drm_gem_object_unreference(&obj->base);
1642 mutex_unlock(&dev->struct_mutex);
1647 * Maps the contents of an object, returning the address it is mapped
1650 * While the mapping holds a reference on the contents of the object, it doesn't
1651 * imply a ref on the object itself.
1655 * DRM driver writers who look a this function as an example for how to do GEM
1656 * mmap support, please don't implement mmap support like here. The modern way
1657 * to implement DRM mmap support is with an mmap offset ioctl (like
1658 * i915_gem_mmap_gtt) and then using the mmap syscall on the DRM fd directly.
1659 * That way debug tooling like valgrind will understand what's going on, hiding
1660 * the mmap call in a driver private ioctl will break that. The i915 driver only
1661 * does cpu mmaps this way because we didn't know better.
1664 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1665 struct drm_file *file)
1667 struct drm_i915_gem_mmap *args = data;
1668 struct drm_gem_object *obj;
1671 if (args->flags & ~(I915_MMAP_WC))
1674 if (args->flags & I915_MMAP_WC && !cpu_has_pat)
1677 obj = drm_gem_object_lookup(dev, file, args->handle);
1681 /* prime objects have no backing filp to GEM mmap
1685 drm_gem_object_unreference_unlocked(obj);
1689 addr = vm_mmap(obj->filp, 0, args->size,
1690 PROT_READ | PROT_WRITE, MAP_SHARED,
1692 if (args->flags & I915_MMAP_WC) {
1693 struct mm_struct *mm = current->mm;
1694 struct vm_area_struct *vma;
1696 down_write(&mm->mmap_sem);
1697 vma = find_vma(mm, addr);
1700 pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
1703 up_write(&mm->mmap_sem);
1705 drm_gem_object_unreference_unlocked(obj);
1706 if (IS_ERR((void *)addr))
1709 args->addr_ptr = (uint64_t) addr;
1715 * i915_gem_fault - fault a page into the GTT
1716 * @vma: VMA in question
1719 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1720 * from userspace. The fault handler takes care of binding the object to
1721 * the GTT (if needed), allocating and programming a fence register (again,
1722 * only if needed based on whether the old reg is still valid or the object
1723 * is tiled) and inserting a new PTE into the faulting process.
1725 * Note that the faulting process may involve evicting existing objects
1726 * from the GTT and/or fence registers to make room. So performance may
1727 * suffer if the GTT working set is large or there are few fence registers
1730 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1732 struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1733 struct drm_device *dev = obj->base.dev;
1734 struct drm_i915_private *dev_priv = dev->dev_private;
1735 struct i915_ggtt_view view = i915_ggtt_view_normal;
1736 pgoff_t page_offset;
1739 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1741 intel_runtime_pm_get(dev_priv);
1743 /* We don't use vmf->pgoff since that has the fake offset */
1744 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1747 ret = i915_mutex_lock_interruptible(dev);
1751 trace_i915_gem_object_fault(obj, page_offset, true, write);
1753 /* Try to flush the object off the GPU first without holding the lock.
1754 * Upon reacquiring the lock, we will perform our sanity checks and then
1755 * repeat the flush holding the lock in the normal manner to catch cases
1756 * where we are gazumped.
1758 ret = i915_gem_object_wait_rendering__nonblocking(obj, NULL, !write);
1762 /* Access to snoopable pages through the GTT is incoherent. */
1763 if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) {
1768 /* Use a partial view if the object is bigger than the aperture. */
1769 if (obj->base.size >= dev_priv->gtt.mappable_end &&
1770 obj->tiling_mode == I915_TILING_NONE) {
1771 static const unsigned int chunk_size = 256; // 1 MiB
1773 memset(&view, 0, sizeof(view));
1774 view.type = I915_GGTT_VIEW_PARTIAL;
1775 view.params.partial.offset = rounddown(page_offset, chunk_size);
1776 view.params.partial.size =
1779 (vma->vm_end - vma->vm_start)/PAGE_SIZE -
1780 view.params.partial.offset);
1783 /* Now pin it into the GTT if needed */
1784 ret = i915_gem_object_ggtt_pin(obj, &view, 0, PIN_MAPPABLE);
1788 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1792 ret = i915_gem_object_get_fence(obj);
1796 /* Finally, remap it using the new GTT offset */
1797 pfn = dev_priv->gtt.mappable_base +
1798 i915_gem_obj_ggtt_offset_view(obj, &view);
1801 if (unlikely(view.type == I915_GGTT_VIEW_PARTIAL)) {
1802 /* Overriding existing pages in partial view does not cause
1803 * us any trouble as TLBs are still valid because the fault
1804 * is due to userspace losing part of the mapping or never
1805 * having accessed it before (at this partials' range).
1807 unsigned long base = vma->vm_start +
1808 (view.params.partial.offset << PAGE_SHIFT);
1811 for (i = 0; i < view.params.partial.size; i++) {
1812 ret = vm_insert_pfn(vma, base + i * PAGE_SIZE, pfn + i);
1817 obj->fault_mappable = true;
1819 if (!obj->fault_mappable) {
1820 unsigned long size = min_t(unsigned long,
1821 vma->vm_end - vma->vm_start,
1825 for (i = 0; i < size >> PAGE_SHIFT; i++) {
1826 ret = vm_insert_pfn(vma,
1827 (unsigned long)vma->vm_start + i * PAGE_SIZE,
1833 obj->fault_mappable = true;
1835 ret = vm_insert_pfn(vma,
1836 (unsigned long)vmf->virtual_address,
1840 i915_gem_object_ggtt_unpin_view(obj, &view);
1842 mutex_unlock(&dev->struct_mutex);
1847 * We eat errors when the gpu is terminally wedged to avoid
1848 * userspace unduly crashing (gl has no provisions for mmaps to
1849 * fail). But any other -EIO isn't ours (e.g. swap in failure)
1850 * and so needs to be reported.
1852 if (!i915_terminally_wedged(&dev_priv->gpu_error)) {
1853 ret = VM_FAULT_SIGBUS;
1858 * EAGAIN means the gpu is hung and we'll wait for the error
1859 * handler to reset everything when re-faulting in
1860 * i915_mutex_lock_interruptible.
1867 * EBUSY is ok: this just means that another thread
1868 * already did the job.
1870 ret = VM_FAULT_NOPAGE;
1877 ret = VM_FAULT_SIGBUS;
1880 WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret);
1881 ret = VM_FAULT_SIGBUS;
1885 intel_runtime_pm_put(dev_priv);
1890 * i915_gem_release_mmap - remove physical page mappings
1891 * @obj: obj in question
1893 * Preserve the reservation of the mmapping with the DRM core code, but
1894 * relinquish ownership of the pages back to the system.
1896 * It is vital that we remove the page mapping if we have mapped a tiled
1897 * object through the GTT and then lose the fence register due to
1898 * resource pressure. Similarly if the object has been moved out of the
1899 * aperture, than pages mapped into userspace must be revoked. Removing the
1900 * mapping will then trigger a page fault on the next user access, allowing
1901 * fixup by i915_gem_fault().
1904 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1906 if (!obj->fault_mappable)
1909 drm_vma_node_unmap(&obj->base.vma_node,
1910 obj->base.dev->anon_inode->i_mapping);
1911 obj->fault_mappable = false;
1915 i915_gem_release_all_mmaps(struct drm_i915_private *dev_priv)
1917 struct drm_i915_gem_object *obj;
1919 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list)
1920 i915_gem_release_mmap(obj);
1924 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1928 if (INTEL_INFO(dev)->gen >= 4 ||
1929 tiling_mode == I915_TILING_NONE)
1932 /* Previous chips need a power-of-two fence region when tiling */
1933 if (INTEL_INFO(dev)->gen == 3)
1934 gtt_size = 1024*1024;
1936 gtt_size = 512*1024;
1938 while (gtt_size < size)
1945 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1946 * @obj: object to check
1948 * Return the required GTT alignment for an object, taking into account
1949 * potential fence register mapping.
1952 i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size,
1953 int tiling_mode, bool fenced)
1956 * Minimum alignment is 4k (GTT page size), but might be greater
1957 * if a fence register is needed for the object.
1959 if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) ||
1960 tiling_mode == I915_TILING_NONE)
1964 * Previous chips need to be aligned to the size of the smallest
1965 * fence register that can contain the object.
1967 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1970 static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj)
1972 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1975 if (drm_vma_node_has_offset(&obj->base.vma_node))
1978 dev_priv->mm.shrinker_no_lock_stealing = true;
1980 ret = drm_gem_create_mmap_offset(&obj->base);
1984 /* Badly fragmented mmap space? The only way we can recover
1985 * space is by destroying unwanted objects. We can't randomly release
1986 * mmap_offsets as userspace expects them to be persistent for the
1987 * lifetime of the objects. The closest we can is to release the
1988 * offsets on purgeable objects by truncating it and marking it purged,
1989 * which prevents userspace from ever using that object again.
1991 i915_gem_shrink(dev_priv,
1992 obj->base.size >> PAGE_SHIFT,
1994 I915_SHRINK_UNBOUND |
1995 I915_SHRINK_PURGEABLE);
1996 ret = drm_gem_create_mmap_offset(&obj->base);
2000 i915_gem_shrink_all(dev_priv);
2001 ret = drm_gem_create_mmap_offset(&obj->base);
2003 dev_priv->mm.shrinker_no_lock_stealing = false;
2008 static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)
2010 drm_gem_free_mmap_offset(&obj->base);
2014 i915_gem_mmap_gtt(struct drm_file *file,
2015 struct drm_device *dev,
2019 struct drm_i915_gem_object *obj;
2022 ret = i915_mutex_lock_interruptible(dev);
2026 obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
2027 if (&obj->base == NULL) {
2032 if (obj->madv != I915_MADV_WILLNEED) {
2033 DRM_DEBUG("Attempting to mmap a purgeable buffer\n");
2038 ret = i915_gem_object_create_mmap_offset(obj);
2042 *offset = drm_vma_node_offset_addr(&obj->base.vma_node);
2045 drm_gem_object_unreference(&obj->base);
2047 mutex_unlock(&dev->struct_mutex);
2052 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
2054 * @data: GTT mapping ioctl data
2055 * @file: GEM object info
2057 * Simply returns the fake offset to userspace so it can mmap it.
2058 * The mmap call will end up in drm_gem_mmap(), which will set things
2059 * up so we can get faults in the handler above.
2061 * The fault handler will take care of binding the object into the GTT
2062 * (since it may have been evicted to make room for something), allocating
2063 * a fence register, and mapping the appropriate aperture address into
2067 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
2068 struct drm_file *file)
2070 struct drm_i915_gem_mmap_gtt *args = data;
2072 return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
2075 /* Immediately discard the backing storage */
2077 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
2079 i915_gem_object_free_mmap_offset(obj);
2081 if (obj->base.filp == NULL)
2084 /* Our goal here is to return as much of the memory as
2085 * is possible back to the system as we are called from OOM.
2086 * To do this we must instruct the shmfs to drop all of its
2087 * backing pages, *now*.
2089 shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1);
2090 obj->madv = __I915_MADV_PURGED;
2093 /* Try to discard unwanted pages */
2095 i915_gem_object_invalidate(struct drm_i915_gem_object *obj)
2097 struct address_space *mapping;
2099 switch (obj->madv) {
2100 case I915_MADV_DONTNEED:
2101 i915_gem_object_truncate(obj);
2102 case __I915_MADV_PURGED:
2106 if (obj->base.filp == NULL)
2109 mapping = file_inode(obj->base.filp)->i_mapping,
2110 invalidate_mapping_pages(mapping, 0, (loff_t)-1);
2114 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
2116 struct sg_page_iter sg_iter;
2119 BUG_ON(obj->madv == __I915_MADV_PURGED);
2121 ret = i915_gem_object_set_to_cpu_domain(obj, true);
2123 /* In the event of a disaster, abandon all caches and
2124 * hope for the best.
2126 WARN_ON(ret != -EIO);
2127 i915_gem_clflush_object(obj, true);
2128 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
2131 i915_gem_gtt_finish_object(obj);
2133 if (i915_gem_object_needs_bit17_swizzle(obj))
2134 i915_gem_object_save_bit_17_swizzle(obj);
2136 if (obj->madv == I915_MADV_DONTNEED)
2139 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
2140 struct page *page = sg_page_iter_page(&sg_iter);
2143 set_page_dirty(page);
2145 if (obj->madv == I915_MADV_WILLNEED)
2146 mark_page_accessed(page);
2148 page_cache_release(page);
2152 sg_free_table(obj->pages);
2157 i915_gem_object_put_pages(struct drm_i915_gem_object *obj)
2159 const struct drm_i915_gem_object_ops *ops = obj->ops;
2161 if (obj->pages == NULL)
2164 if (obj->pages_pin_count)
2167 BUG_ON(i915_gem_obj_bound_any(obj));
2169 /* ->put_pages might need to allocate memory for the bit17 swizzle
2170 * array, hence protect them from being reaped by removing them from gtt
2172 list_del(&obj->global_list);
2174 ops->put_pages(obj);
2177 i915_gem_object_invalidate(obj);
2183 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
2185 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2187 struct address_space *mapping;
2188 struct sg_table *st;
2189 struct scatterlist *sg;
2190 struct sg_page_iter sg_iter;
2192 unsigned long last_pfn = 0; /* suppress gcc warning */
2196 /* Assert that the object is not currently in any GPU domain. As it
2197 * wasn't in the GTT, there shouldn't be any way it could have been in
2200 BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
2201 BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
2203 st = kmalloc(sizeof(*st), GFP_KERNEL);
2207 page_count = obj->base.size / PAGE_SIZE;
2208 if (sg_alloc_table(st, page_count, GFP_KERNEL)) {
2213 /* Get the list of pages out of our struct file. They'll be pinned
2214 * at this point until we release them.
2216 * Fail silently without starting the shrinker
2218 mapping = file_inode(obj->base.filp)->i_mapping;
2219 gfp = mapping_gfp_mask(mapping);
2220 gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD;
2221 gfp &= ~(__GFP_IO | __GFP_WAIT);
2224 for (i = 0; i < page_count; i++) {
2225 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2227 i915_gem_shrink(dev_priv,
2230 I915_SHRINK_UNBOUND |
2231 I915_SHRINK_PURGEABLE);
2232 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2235 /* We've tried hard to allocate the memory by reaping
2236 * our own buffer, now let the real VM do its job and
2237 * go down in flames if truly OOM.
2239 i915_gem_shrink_all(dev_priv);
2240 page = shmem_read_mapping_page(mapping, i);
2242 ret = PTR_ERR(page);
2246 #ifdef CONFIG_SWIOTLB
2247 if (swiotlb_nr_tbl()) {
2249 sg_set_page(sg, page, PAGE_SIZE, 0);
2254 if (!i || page_to_pfn(page) != last_pfn + 1) {
2258 sg_set_page(sg, page, PAGE_SIZE, 0);
2260 sg->length += PAGE_SIZE;
2262 last_pfn = page_to_pfn(page);
2264 /* Check that the i965g/gm workaround works. */
2265 WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
2267 #ifdef CONFIG_SWIOTLB
2268 if (!swiotlb_nr_tbl())
2273 ret = i915_gem_gtt_prepare_object(obj);
2277 if (i915_gem_object_needs_bit17_swizzle(obj))
2278 i915_gem_object_do_bit_17_swizzle(obj);
2280 if (obj->tiling_mode != I915_TILING_NONE &&
2281 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES)
2282 i915_gem_object_pin_pages(obj);
2288 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0)
2289 page_cache_release(sg_page_iter_page(&sg_iter));
2293 /* shmemfs first checks if there is enough memory to allocate the page
2294 * and reports ENOSPC should there be insufficient, along with the usual
2295 * ENOMEM for a genuine allocation failure.
2297 * We use ENOSPC in our driver to mean that we have run out of aperture
2298 * space and so want to translate the error from shmemfs back to our
2299 * usual understanding of ENOMEM.
2307 /* Ensure that the associated pages are gathered from the backing storage
2308 * and pinned into our object. i915_gem_object_get_pages() may be called
2309 * multiple times before they are released by a single call to
2310 * i915_gem_object_put_pages() - once the pages are no longer referenced
2311 * either as a result of memory pressure (reaping pages under the shrinker)
2312 * or as the object is itself released.
2315 i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
2317 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2318 const struct drm_i915_gem_object_ops *ops = obj->ops;
2324 if (obj->madv != I915_MADV_WILLNEED) {
2325 DRM_DEBUG("Attempting to obtain a purgeable object\n");
2329 BUG_ON(obj->pages_pin_count);
2331 ret = ops->get_pages(obj);
2335 list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list);
2337 obj->get_page.sg = obj->pages->sgl;
2338 obj->get_page.last = 0;
2343 void i915_vma_move_to_active(struct i915_vma *vma,
2344 struct drm_i915_gem_request *req)
2346 struct drm_i915_gem_object *obj = vma->obj;
2347 struct intel_engine_cs *ring;
2349 ring = i915_gem_request_get_ring(req);
2351 /* Add a reference if we're newly entering the active list. */
2352 if (obj->active == 0)
2353 drm_gem_object_reference(&obj->base);
2354 obj->active |= intel_ring_flag(ring);
2356 list_move_tail(&obj->ring_list[ring->id], &ring->active_list);
2357 i915_gem_request_assign(&obj->last_read_req[ring->id], req);
2359 list_move_tail(&vma->mm_list, &vma->vm->active_list);
2363 i915_gem_object_retire__write(struct drm_i915_gem_object *obj)
2365 RQ_BUG_ON(obj->last_write_req == NULL);
2366 RQ_BUG_ON(!(obj->active & intel_ring_flag(obj->last_write_req->ring)));
2368 i915_gem_request_assign(&obj->last_write_req, NULL);
2369 intel_fb_obj_flush(obj, true, ORIGIN_CS);
2373 i915_gem_object_retire__read(struct drm_i915_gem_object *obj, int ring)
2375 struct i915_vma *vma;
2377 RQ_BUG_ON(obj->last_read_req[ring] == NULL);
2378 RQ_BUG_ON(!(obj->active & (1 << ring)));
2380 list_del_init(&obj->ring_list[ring]);
2381 i915_gem_request_assign(&obj->last_read_req[ring], NULL);
2383 if (obj->last_write_req && obj->last_write_req->ring->id == ring)
2384 i915_gem_object_retire__write(obj);
2386 obj->active &= ~(1 << ring);
2390 /* Bump our place on the bound list to keep it roughly in LRU order
2391 * so that we don't steal from recently used but inactive objects
2392 * (unless we are forced to ofc!)
2394 list_move_tail(&obj->global_list,
2395 &to_i915(obj->base.dev)->mm.bound_list);
2397 list_for_each_entry(vma, &obj->vma_list, vma_link) {
2398 if (!list_empty(&vma->mm_list))
2399 list_move_tail(&vma->mm_list, &vma->vm->inactive_list);
2402 i915_gem_request_assign(&obj->last_fenced_req, NULL);
2403 drm_gem_object_unreference(&obj->base);
2407 i915_gem_init_seqno(struct drm_device *dev, u32 seqno)
2409 struct drm_i915_private *dev_priv = dev->dev_private;
2410 struct intel_engine_cs *ring;
2413 /* Carefully retire all requests without writing to the rings */
2414 for_each_ring(ring, dev_priv, i) {
2415 ret = intel_ring_idle(ring);
2419 i915_gem_retire_requests(dev);
2421 /* Finally reset hw state */
2422 for_each_ring(ring, dev_priv, i) {
2423 intel_ring_init_seqno(ring, seqno);
2425 for (j = 0; j < ARRAY_SIZE(ring->semaphore.sync_seqno); j++)
2426 ring->semaphore.sync_seqno[j] = 0;
2432 int i915_gem_set_seqno(struct drm_device *dev, u32 seqno)
2434 struct drm_i915_private *dev_priv = dev->dev_private;
2440 /* HWS page needs to be set less than what we
2441 * will inject to ring
2443 ret = i915_gem_init_seqno(dev, seqno - 1);
2447 /* Carefully set the last_seqno value so that wrap
2448 * detection still works
2450 dev_priv->next_seqno = seqno;
2451 dev_priv->last_seqno = seqno - 1;
2452 if (dev_priv->last_seqno == 0)
2453 dev_priv->last_seqno--;
2459 i915_gem_get_seqno(struct drm_device *dev, u32 *seqno)
2461 struct drm_i915_private *dev_priv = dev->dev_private;
2463 /* reserve 0 for non-seqno */
2464 if (dev_priv->next_seqno == 0) {
2465 int ret = i915_gem_init_seqno(dev, 0);
2469 dev_priv->next_seqno = 1;
2472 *seqno = dev_priv->last_seqno = dev_priv->next_seqno++;
2477 * NB: This function is not allowed to fail. Doing so would mean the the
2478 * request is not being tracked for completion but the work itself is
2479 * going to happen on the hardware. This would be a Bad Thing(tm).
2481 void __i915_add_request(struct drm_i915_gem_request *request,
2482 struct drm_i915_gem_object *obj,
2485 struct intel_engine_cs *ring;
2486 struct drm_i915_private *dev_priv;
2487 struct intel_ringbuffer *ringbuf;
2491 if (WARN_ON(request == NULL))
2494 ring = request->ring;
2495 dev_priv = ring->dev->dev_private;
2496 ringbuf = request->ringbuf;
2499 * To ensure that this call will not fail, space for its emissions
2500 * should already have been reserved in the ring buffer. Let the ring
2501 * know that it is time to use that space up.
2503 intel_ring_reserved_space_use(ringbuf);
2505 request_start = intel_ring_get_tail(ringbuf);
2507 * Emit any outstanding flushes - execbuf can fail to emit the flush
2508 * after having emitted the batchbuffer command. Hence we need to fix
2509 * things up similar to emitting the lazy request. The difference here
2510 * is that the flush _must_ happen before the next request, no matter
2514 if (i915.enable_execlists)
2515 ret = logical_ring_flush_all_caches(request);
2517 ret = intel_ring_flush_all_caches(request);
2518 /* Not allowed to fail! */
2519 WARN(ret, "*_ring_flush_all_caches failed: %d!\n", ret);
2522 /* Record the position of the start of the request so that
2523 * should we detect the updated seqno part-way through the
2524 * GPU processing the request, we never over-estimate the
2525 * position of the head.
2527 request->postfix = intel_ring_get_tail(ringbuf);
2529 if (i915.enable_execlists)
2530 ret = ring->emit_request(request);
2532 ret = ring->add_request(request);
2534 request->tail = intel_ring_get_tail(ringbuf);
2536 /* Not allowed to fail! */
2537 WARN(ret, "emit|add_request failed: %d!\n", ret);
2539 request->head = request_start;
2541 /* Whilst this request exists, batch_obj will be on the
2542 * active_list, and so will hold the active reference. Only when this
2543 * request is retired will the the batch_obj be moved onto the
2544 * inactive_list and lose its active reference. Hence we do not need
2545 * to explicitly hold another reference here.
2547 request->batch_obj = obj;
2549 request->emitted_jiffies = jiffies;
2550 ring->last_submitted_seqno = request->seqno;
2551 list_add_tail(&request->list, &ring->request_list);
2553 trace_i915_gem_request_add(request);
2555 i915_queue_hangcheck(ring->dev);
2557 queue_delayed_work(dev_priv->wq,
2558 &dev_priv->mm.retire_work,
2559 round_jiffies_up_relative(HZ));
2560 intel_mark_busy(dev_priv->dev);
2562 /* Sanity check that the reserved size was large enough. */
2563 intel_ring_reserved_space_end(ringbuf);
2566 static bool i915_context_is_banned(struct drm_i915_private *dev_priv,
2567 const struct intel_context *ctx)
2569 unsigned long elapsed;
2571 elapsed = get_seconds() - ctx->hang_stats.guilty_ts;
2573 if (ctx->hang_stats.banned)
2576 if (ctx->hang_stats.ban_period_seconds &&
2577 elapsed <= ctx->hang_stats.ban_period_seconds) {
2578 if (!i915_gem_context_is_default(ctx)) {
2579 DRM_DEBUG("context hanging too fast, banning!\n");
2581 } else if (i915_stop_ring_allow_ban(dev_priv)) {
2582 if (i915_stop_ring_allow_warn(dev_priv))
2583 DRM_ERROR("gpu hanging too fast, banning!\n");
2591 static void i915_set_reset_status(struct drm_i915_private *dev_priv,
2592 struct intel_context *ctx,
2595 struct i915_ctx_hang_stats *hs;
2600 hs = &ctx->hang_stats;
2603 hs->banned = i915_context_is_banned(dev_priv, ctx);
2605 hs->guilty_ts = get_seconds();
2607 hs->batch_pending++;
2611 void i915_gem_request_free(struct kref *req_ref)
2613 struct drm_i915_gem_request *req = container_of(req_ref,
2615 struct intel_context *ctx = req->ctx;
2618 i915_gem_request_remove_from_client(req);
2621 if (i915.enable_execlists) {
2622 if (ctx != req->ring->default_context)
2623 intel_lr_context_unpin(req);
2626 i915_gem_context_unreference(ctx);
2629 kmem_cache_free(req->i915->requests, req);
2632 int i915_gem_request_alloc(struct intel_engine_cs *ring,
2633 struct intel_context *ctx,
2634 struct drm_i915_gem_request **req_out)
2636 struct drm_i915_private *dev_priv = to_i915(ring->dev);
2637 struct drm_i915_gem_request *req;
2645 req = kmem_cache_zalloc(dev_priv->requests, GFP_KERNEL);
2649 ret = i915_gem_get_seqno(ring->dev, &req->seqno);
2653 kref_init(&req->ref);
2654 req->i915 = dev_priv;
2657 i915_gem_context_reference(req->ctx);
2659 if (i915.enable_execlists)
2660 ret = intel_logical_ring_alloc_request_extras(req);
2662 ret = intel_ring_alloc_request_extras(req);
2664 i915_gem_context_unreference(req->ctx);
2669 * Reserve space in the ring buffer for all the commands required to
2670 * eventually emit this request. This is to guarantee that the
2671 * i915_add_request() call can't fail. Note that the reserve may need
2672 * to be redone if the request is not actually submitted straight
2673 * away, e.g. because a GPU scheduler has deferred it.
2675 if (i915.enable_execlists)
2676 ret = intel_logical_ring_reserve_space(req);
2678 ret = intel_ring_reserve_space(req);
2681 * At this point, the request is fully allocated even if not
2682 * fully prepared. Thus it can be cleaned up using the proper
2685 i915_gem_request_cancel(req);
2693 kmem_cache_free(dev_priv->requests, req);
2697 void i915_gem_request_cancel(struct drm_i915_gem_request *req)
2699 intel_ring_reserved_space_cancel(req->ringbuf);
2701 i915_gem_request_unreference(req);
2704 struct drm_i915_gem_request *
2705 i915_gem_find_active_request(struct intel_engine_cs *ring)
2707 struct drm_i915_gem_request *request;
2709 list_for_each_entry(request, &ring->request_list, list) {
2710 if (i915_gem_request_completed(request, false))
2719 static void i915_gem_reset_ring_status(struct drm_i915_private *dev_priv,
2720 struct intel_engine_cs *ring)
2722 struct drm_i915_gem_request *request;
2725 request = i915_gem_find_active_request(ring);
2727 if (request == NULL)
2730 ring_hung = ring->hangcheck.score >= HANGCHECK_SCORE_RING_HUNG;
2732 i915_set_reset_status(dev_priv, request->ctx, ring_hung);
2734 list_for_each_entry_continue(request, &ring->request_list, list)
2735 i915_set_reset_status(dev_priv, request->ctx, false);
2738 static void i915_gem_reset_ring_cleanup(struct drm_i915_private *dev_priv,
2739 struct intel_engine_cs *ring)
2741 while (!list_empty(&ring->active_list)) {
2742 struct drm_i915_gem_object *obj;
2744 obj = list_first_entry(&ring->active_list,
2745 struct drm_i915_gem_object,
2746 ring_list[ring->id]);
2748 i915_gem_object_retire__read(obj, ring->id);
2752 * Clear the execlists queue up before freeing the requests, as those
2753 * are the ones that keep the context and ringbuffer backing objects
2756 while (!list_empty(&ring->execlist_queue)) {
2757 struct drm_i915_gem_request *submit_req;
2759 submit_req = list_first_entry(&ring->execlist_queue,
2760 struct drm_i915_gem_request,
2762 list_del(&submit_req->execlist_link);
2764 if (submit_req->ctx != ring->default_context)
2765 intel_lr_context_unpin(submit_req);
2767 i915_gem_request_unreference(submit_req);
2771 * We must free the requests after all the corresponding objects have
2772 * been moved off active lists. Which is the same order as the normal
2773 * retire_requests function does. This is important if object hold
2774 * implicit references on things like e.g. ppgtt address spaces through
2777 while (!list_empty(&ring->request_list)) {
2778 struct drm_i915_gem_request *request;
2780 request = list_first_entry(&ring->request_list,
2781 struct drm_i915_gem_request,
2784 i915_gem_request_retire(request);
2788 void i915_gem_reset(struct drm_device *dev)
2790 struct drm_i915_private *dev_priv = dev->dev_private;
2791 struct intel_engine_cs *ring;
2795 * Before we free the objects from the requests, we need to inspect
2796 * them for finding the guilty party. As the requests only borrow
2797 * their reference to the objects, the inspection must be done first.
2799 for_each_ring(ring, dev_priv, i)
2800 i915_gem_reset_ring_status(dev_priv, ring);
2802 for_each_ring(ring, dev_priv, i)
2803 i915_gem_reset_ring_cleanup(dev_priv, ring);
2805 i915_gem_context_reset(dev);
2807 i915_gem_restore_fences(dev);
2809 WARN_ON(i915_verify_lists(dev));
2813 * This function clears the request list as sequence numbers are passed.
2816 i915_gem_retire_requests_ring(struct intel_engine_cs *ring)
2818 WARN_ON(i915_verify_lists(ring->dev));
2820 /* Retire requests first as we use it above for the early return.
2821 * If we retire requests last, we may use a later seqno and so clear
2822 * the requests lists without clearing the active list, leading to
2825 while (!list_empty(&ring->request_list)) {
2826 struct drm_i915_gem_request *request;
2828 request = list_first_entry(&ring->request_list,
2829 struct drm_i915_gem_request,
2832 if (!i915_gem_request_completed(request, true))
2835 i915_gem_request_retire(request);
2838 /* Move any buffers on the active list that are no longer referenced
2839 * by the ringbuffer to the flushing/inactive lists as appropriate,
2840 * before we free the context associated with the requests.
2842 while (!list_empty(&ring->active_list)) {
2843 struct drm_i915_gem_object *obj;
2845 obj = list_first_entry(&ring->active_list,
2846 struct drm_i915_gem_object,
2847 ring_list[ring->id]);
2849 if (!list_empty(&obj->last_read_req[ring->id]->list))
2852 i915_gem_object_retire__read(obj, ring->id);
2855 if (unlikely(ring->trace_irq_req &&
2856 i915_gem_request_completed(ring->trace_irq_req, true))) {
2857 ring->irq_put(ring);
2858 i915_gem_request_assign(&ring->trace_irq_req, NULL);
2861 WARN_ON(i915_verify_lists(ring->dev));
2865 i915_gem_retire_requests(struct drm_device *dev)
2867 struct drm_i915_private *dev_priv = dev->dev_private;
2868 struct intel_engine_cs *ring;
2872 for_each_ring(ring, dev_priv, i) {
2873 i915_gem_retire_requests_ring(ring);
2874 idle &= list_empty(&ring->request_list);
2875 if (i915.enable_execlists) {
2876 unsigned long flags;
2878 spin_lock_irqsave(&ring->execlist_lock, flags);
2879 idle &= list_empty(&ring->execlist_queue);
2880 spin_unlock_irqrestore(&ring->execlist_lock, flags);
2882 intel_execlists_retire_requests(ring);
2887 mod_delayed_work(dev_priv->wq,
2888 &dev_priv->mm.idle_work,
2889 msecs_to_jiffies(100));
2895 i915_gem_retire_work_handler(struct work_struct *work)
2897 struct drm_i915_private *dev_priv =
2898 container_of(work, typeof(*dev_priv), mm.retire_work.work);
2899 struct drm_device *dev = dev_priv->dev;
2902 /* Come back later if the device is busy... */
2904 if (mutex_trylock(&dev->struct_mutex)) {
2905 idle = i915_gem_retire_requests(dev);
2906 mutex_unlock(&dev->struct_mutex);
2909 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
2910 round_jiffies_up_relative(HZ));
2914 i915_gem_idle_work_handler(struct work_struct *work)
2916 struct drm_i915_private *dev_priv =
2917 container_of(work, typeof(*dev_priv), mm.idle_work.work);
2918 struct drm_device *dev = dev_priv->dev;
2919 struct intel_engine_cs *ring;
2922 for_each_ring(ring, dev_priv, i)
2923 if (!list_empty(&ring->request_list))
2926 intel_mark_idle(dev);
2928 if (mutex_trylock(&dev->struct_mutex)) {
2929 struct intel_engine_cs *ring;
2932 for_each_ring(ring, dev_priv, i)
2933 i915_gem_batch_pool_fini(&ring->batch_pool);
2935 mutex_unlock(&dev->struct_mutex);
2940 * Ensures that an object will eventually get non-busy by flushing any required
2941 * write domains, emitting any outstanding lazy request and retiring and
2942 * completed requests.
2945 i915_gem_object_flush_active(struct drm_i915_gem_object *obj)
2952 for (i = 0; i < I915_NUM_RINGS; i++) {
2953 struct drm_i915_gem_request *req;
2955 req = obj->last_read_req[i];
2959 if (list_empty(&req->list))
2962 if (i915_gem_request_completed(req, true)) {
2963 __i915_gem_request_retire__upto(req);
2965 i915_gem_object_retire__read(obj, i);
2973 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
2974 * @DRM_IOCTL_ARGS: standard ioctl arguments
2976 * Returns 0 if successful, else an error is returned with the remaining time in
2977 * the timeout parameter.
2978 * -ETIME: object is still busy after timeout
2979 * -ERESTARTSYS: signal interrupted the wait
2980 * -ENONENT: object doesn't exist
2981 * Also possible, but rare:
2982 * -EAGAIN: GPU wedged
2984 * -ENODEV: Internal IRQ fail
2985 * -E?: The add request failed
2987 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
2988 * non-zero timeout parameter the wait ioctl will wait for the given number of
2989 * nanoseconds on an object becoming unbusy. Since the wait itself does so
2990 * without holding struct_mutex the object may become re-busied before this
2991 * function completes. A similar but shorter * race condition exists in the busy
2995 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
2997 struct drm_i915_private *dev_priv = dev->dev_private;
2998 struct drm_i915_gem_wait *args = data;
2999 struct drm_i915_gem_object *obj;
3000 struct drm_i915_gem_request *req[I915_NUM_RINGS];
3001 unsigned reset_counter;
3005 if (args->flags != 0)
3008 ret = i915_mutex_lock_interruptible(dev);
3012 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle));
3013 if (&obj->base == NULL) {
3014 mutex_unlock(&dev->struct_mutex);
3018 /* Need to make sure the object gets inactive eventually. */
3019 ret = i915_gem_object_flush_active(obj);
3026 /* Do this after OLR check to make sure we make forward progress polling
3027 * on this IOCTL with a timeout == 0 (like busy ioctl)
3029 if (args->timeout_ns == 0) {
3034 drm_gem_object_unreference(&obj->base);
3035 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
3037 for (i = 0; i < I915_NUM_RINGS; i++) {
3038 if (obj->last_read_req[i] == NULL)
3041 req[n++] = i915_gem_request_reference(obj->last_read_req[i]);
3044 mutex_unlock(&dev->struct_mutex);
3046 for (i = 0; i < n; i++) {
3048 ret = __i915_wait_request(req[i], reset_counter, true,
3049 args->timeout_ns > 0 ? &args->timeout_ns : NULL,
3051 i915_gem_request_unreference__unlocked(req[i]);
3056 drm_gem_object_unreference(&obj->base);
3057 mutex_unlock(&dev->struct_mutex);
3062 __i915_gem_object_sync(struct drm_i915_gem_object *obj,
3063 struct intel_engine_cs *to,
3064 struct drm_i915_gem_request *from_req,
3065 struct drm_i915_gem_request **to_req)
3067 struct intel_engine_cs *from;
3070 from = i915_gem_request_get_ring(from_req);
3074 if (i915_gem_request_completed(from_req, true))
3077 if (!i915_semaphore_is_enabled(obj->base.dev)) {
3078 struct drm_i915_private *i915 = to_i915(obj->base.dev);
3079 ret = __i915_wait_request(from_req,
3080 atomic_read(&i915->gpu_error.reset_counter),
3081 i915->mm.interruptible,
3083 &i915->rps.semaphores);
3087 i915_gem_object_retire_request(obj, from_req);
3089 int idx = intel_ring_sync_index(from, to);
3090 u32 seqno = i915_gem_request_get_seqno(from_req);
3094 if (seqno <= from->semaphore.sync_seqno[idx])
3097 if (*to_req == NULL) {
3098 ret = i915_gem_request_alloc(to, to->default_context, to_req);
3103 trace_i915_gem_ring_sync_to(*to_req, from, from_req);
3104 ret = to->semaphore.sync_to(*to_req, from, seqno);
3108 /* We use last_read_req because sync_to()
3109 * might have just caused seqno wrap under
3112 from->semaphore.sync_seqno[idx] =
3113 i915_gem_request_get_seqno(obj->last_read_req[from->id]);
3120 * i915_gem_object_sync - sync an object to a ring.
3122 * @obj: object which may be in use on another ring.
3123 * @to: ring we wish to use the object on. May be NULL.
3124 * @to_req: request we wish to use the object for. See below.
3125 * This will be allocated and returned if a request is
3126 * required but not passed in.
3128 * This code is meant to abstract object synchronization with the GPU.
3129 * Calling with NULL implies synchronizing the object with the CPU
3130 * rather than a particular GPU ring. Conceptually we serialise writes
3131 * between engines inside the GPU. We only allow one engine to write
3132 * into a buffer at any time, but multiple readers. To ensure each has
3133 * a coherent view of memory, we must:
3135 * - If there is an outstanding write request to the object, the new
3136 * request must wait for it to complete (either CPU or in hw, requests
3137 * on the same ring will be naturally ordered).
3139 * - If we are a write request (pending_write_domain is set), the new
3140 * request must wait for outstanding read requests to complete.
3142 * For CPU synchronisation (NULL to) no request is required. For syncing with
3143 * rings to_req must be non-NULL. However, a request does not have to be
3144 * pre-allocated. If *to_req is NULL and sync commands will be emitted then a
3145 * request will be allocated automatically and returned through *to_req. Note
3146 * that it is not guaranteed that commands will be emitted (because the system
3147 * might already be idle). Hence there is no need to create a request that
3148 * might never have any work submitted. Note further that if a request is
3149 * returned in *to_req, it is the responsibility of the caller to submit
3150 * that request (after potentially adding more work to it).
3152 * Returns 0 if successful, else propagates up the lower layer error.
3155 i915_gem_object_sync(struct drm_i915_gem_object *obj,
3156 struct intel_engine_cs *to,
3157 struct drm_i915_gem_request **to_req)
3159 const bool readonly = obj->base.pending_write_domain == 0;
3160 struct drm_i915_gem_request *req[I915_NUM_RINGS];
3167 return i915_gem_object_wait_rendering(obj, readonly);
3171 if (obj->last_write_req)
3172 req[n++] = obj->last_write_req;
3174 for (i = 0; i < I915_NUM_RINGS; i++)
3175 if (obj->last_read_req[i])
3176 req[n++] = obj->last_read_req[i];
3178 for (i = 0; i < n; i++) {
3179 ret = __i915_gem_object_sync(obj, to, req[i], to_req);
3187 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
3189 u32 old_write_domain, old_read_domains;
3191 /* Force a pagefault for domain tracking on next user access */
3192 i915_gem_release_mmap(obj);
3194 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
3197 /* Wait for any direct GTT access to complete */
3200 old_read_domains = obj->base.read_domains;
3201 old_write_domain = obj->base.write_domain;
3203 obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
3204 obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
3206 trace_i915_gem_object_change_domain(obj,
3211 int i915_vma_unbind(struct i915_vma *vma)
3213 struct drm_i915_gem_object *obj = vma->obj;
3214 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3217 if (list_empty(&vma->vma_link))
3220 if (!drm_mm_node_allocated(&vma->node)) {
3221 i915_gem_vma_destroy(vma);
3228 BUG_ON(obj->pages == NULL);
3230 ret = i915_gem_object_wait_rendering(obj, false);
3234 if (i915_is_ggtt(vma->vm) &&
3235 vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL) {
3236 i915_gem_object_finish_gtt(obj);
3238 /* release the fence reg _after_ flushing */
3239 ret = i915_gem_object_put_fence(obj);
3244 trace_i915_vma_unbind(vma);
3246 vma->vm->unbind_vma(vma);
3249 list_del_init(&vma->mm_list);
3250 if (i915_is_ggtt(vma->vm)) {
3251 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL) {
3252 obj->map_and_fenceable = false;
3253 } else if (vma->ggtt_view.pages) {
3254 sg_free_table(vma->ggtt_view.pages);
3255 kfree(vma->ggtt_view.pages);
3257 vma->ggtt_view.pages = NULL;
3260 drm_mm_remove_node(&vma->node);
3261 i915_gem_vma_destroy(vma);
3263 /* Since the unbound list is global, only move to that list if
3264 * no more VMAs exist. */
3265 if (list_empty(&obj->vma_list))
3266 list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list);
3268 /* And finally now the object is completely decoupled from this vma,
3269 * we can drop its hold on the backing storage and allow it to be
3270 * reaped by the shrinker.
3272 i915_gem_object_unpin_pages(obj);
3277 int i915_gpu_idle(struct drm_device *dev)
3279 struct drm_i915_private *dev_priv = dev->dev_private;
3280 struct intel_engine_cs *ring;
3283 /* Flush everything onto the inactive list. */
3284 for_each_ring(ring, dev_priv, i) {
3285 if (!i915.enable_execlists) {
3286 struct drm_i915_gem_request *req;
3288 ret = i915_gem_request_alloc(ring, ring->default_context, &req);
3292 ret = i915_switch_context(req);
3294 i915_gem_request_cancel(req);
3298 i915_add_request_no_flush(req);
3301 ret = intel_ring_idle(ring);
3306 WARN_ON(i915_verify_lists(dev));
3310 static bool i915_gem_valid_gtt_space(struct i915_vma *vma,
3311 unsigned long cache_level)
3313 struct drm_mm_node *gtt_space = &vma->node;
3314 struct drm_mm_node *other;
3317 * On some machines we have to be careful when putting differing types
3318 * of snoopable memory together to avoid the prefetcher crossing memory
3319 * domains and dying. During vm initialisation, we decide whether or not
3320 * these constraints apply and set the drm_mm.color_adjust
3323 if (vma->vm->mm.color_adjust == NULL)
3326 if (!drm_mm_node_allocated(gtt_space))
3329 if (list_empty(>t_space->node_list))
3332 other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list);
3333 if (other->allocated && !other->hole_follows && other->color != cache_level)
3336 other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list);
3337 if (other->allocated && !gtt_space->hole_follows && other->color != cache_level)
3344 * Finds free space in the GTT aperture and binds the object or a view of it
3347 static struct i915_vma *
3348 i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj,
3349 struct i915_address_space *vm,
3350 const struct i915_ggtt_view *ggtt_view,
3354 struct drm_device *dev = obj->base.dev;
3355 struct drm_i915_private *dev_priv = dev->dev_private;
3356 u32 fence_alignment, unfenced_alignment;
3357 u64 size, fence_size;
3359 flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
3361 flags & PIN_MAPPABLE ? dev_priv->gtt.mappable_end : vm->total;
3362 struct i915_vma *vma;
3365 if (i915_is_ggtt(vm)) {
3368 if (WARN_ON(!ggtt_view))
3369 return ERR_PTR(-EINVAL);
3371 view_size = i915_ggtt_view_size(obj, ggtt_view);
3373 fence_size = i915_gem_get_gtt_size(dev,
3376 fence_alignment = i915_gem_get_gtt_alignment(dev,
3380 unfenced_alignment = i915_gem_get_gtt_alignment(dev,
3384 size = flags & PIN_MAPPABLE ? fence_size : view_size;
3386 fence_size = i915_gem_get_gtt_size(dev,
3389 fence_alignment = i915_gem_get_gtt_alignment(dev,
3393 unfenced_alignment =
3394 i915_gem_get_gtt_alignment(dev,
3398 size = flags & PIN_MAPPABLE ? fence_size : obj->base.size;
3402 alignment = flags & PIN_MAPPABLE ? fence_alignment :
3404 if (flags & PIN_MAPPABLE && alignment & (fence_alignment - 1)) {
3405 DRM_DEBUG("Invalid object (view type=%u) alignment requested %u\n",
3406 ggtt_view ? ggtt_view->type : 0,
3408 return ERR_PTR(-EINVAL);
3411 /* If binding the object/GGTT view requires more space than the entire
3412 * aperture has, reject it early before evicting everything in a vain
3413 * attempt to find space.
3416 DRM_DEBUG("Attempting to bind an object (view type=%u) larger than the aperture: size=%llu > %s aperture=%llu\n",
3417 ggtt_view ? ggtt_view->type : 0,
3419 flags & PIN_MAPPABLE ? "mappable" : "total",
3421 return ERR_PTR(-E2BIG);
3424 ret = i915_gem_object_get_pages(obj);
3426 return ERR_PTR(ret);
3428 i915_gem_object_pin_pages(obj);
3430 vma = ggtt_view ? i915_gem_obj_lookup_or_create_ggtt_vma(obj, ggtt_view) :
3431 i915_gem_obj_lookup_or_create_vma(obj, vm);
3437 ret = drm_mm_insert_node_in_range_generic(&vm->mm, &vma->node,
3441 DRM_MM_SEARCH_DEFAULT,
3442 DRM_MM_CREATE_DEFAULT);
3444 ret = i915_gem_evict_something(dev, vm, size, alignment,
3453 if (WARN_ON(!i915_gem_valid_gtt_space(vma, obj->cache_level))) {
3455 goto err_remove_node;
3458 trace_i915_vma_bind(vma, flags);
3459 ret = i915_vma_bind(vma, obj->cache_level, flags);
3461 goto err_remove_node;
3463 list_move_tail(&obj->global_list, &dev_priv->mm.bound_list);
3464 list_add_tail(&vma->mm_list, &vm->inactive_list);
3469 drm_mm_remove_node(&vma->node);
3471 i915_gem_vma_destroy(vma);
3474 i915_gem_object_unpin_pages(obj);
3479 i915_gem_clflush_object(struct drm_i915_gem_object *obj,
3482 /* If we don't have a page list set up, then we're not pinned
3483 * to GPU, and we can ignore the cache flush because it'll happen
3484 * again at bind time.
3486 if (obj->pages == NULL)
3490 * Stolen memory is always coherent with the GPU as it is explicitly
3491 * marked as wc by the system, or the system is cache-coherent.
3493 if (obj->stolen || obj->phys_handle)
3496 /* If the GPU is snooping the contents of the CPU cache,
3497 * we do not need to manually clear the CPU cache lines. However,
3498 * the caches are only snooped when the render cache is
3499 * flushed/invalidated. As we always have to emit invalidations
3500 * and flushes when moving into and out of the RENDER domain, correct
3501 * snooping behaviour occurs naturally as the result of our domain
3504 if (!force && cpu_cache_is_coherent(obj->base.dev, obj->cache_level)) {
3505 obj->cache_dirty = true;
3509 trace_i915_gem_object_clflush(obj);
3510 drm_clflush_sg(obj->pages);
3511 obj->cache_dirty = false;
3516 /** Flushes the GTT write domain for the object if it's dirty. */
3518 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
3520 uint32_t old_write_domain;
3522 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
3525 /* No actual flushing is required for the GTT write domain. Writes
3526 * to it immediately go to main memory as far as we know, so there's
3527 * no chipset flush. It also doesn't land in render cache.
3529 * However, we do have to enforce the order so that all writes through
3530 * the GTT land before any writes to the device, such as updates to
3535 old_write_domain = obj->base.write_domain;
3536 obj->base.write_domain = 0;
3538 intel_fb_obj_flush(obj, false, ORIGIN_GTT);
3540 trace_i915_gem_object_change_domain(obj,
3541 obj->base.read_domains,
3545 /** Flushes the CPU write domain for the object if it's dirty. */
3547 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj)
3549 uint32_t old_write_domain;
3551 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
3554 if (i915_gem_clflush_object(obj, obj->pin_display))
3555 i915_gem_chipset_flush(obj->base.dev);
3557 old_write_domain = obj->base.write_domain;
3558 obj->base.write_domain = 0;
3560 intel_fb_obj_flush(obj, false, ORIGIN_CPU);
3562 trace_i915_gem_object_change_domain(obj,
3563 obj->base.read_domains,
3568 * Moves a single object to the GTT read, and possibly write domain.
3570 * This function returns when the move is complete, including waiting on
3574 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
3576 uint32_t old_write_domain, old_read_domains;
3577 struct i915_vma *vma;
3580 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
3583 ret = i915_gem_object_wait_rendering(obj, !write);
3587 /* Flush and acquire obj->pages so that we are coherent through
3588 * direct access in memory with previous cached writes through
3589 * shmemfs and that our cache domain tracking remains valid.
3590 * For example, if the obj->filp was moved to swap without us
3591 * being notified and releasing the pages, we would mistakenly
3592 * continue to assume that the obj remained out of the CPU cached
3595 ret = i915_gem_object_get_pages(obj);
3599 i915_gem_object_flush_cpu_write_domain(obj);
3601 /* Serialise direct access to this object with the barriers for
3602 * coherent writes from the GPU, by effectively invalidating the
3603 * GTT domain upon first access.
3605 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
3608 old_write_domain = obj->base.write_domain;
3609 old_read_domains = obj->base.read_domains;
3611 /* It should now be out of any other write domains, and we can update
3612 * the domain values for our changes.
3614 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3615 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3617 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
3618 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
3622 trace_i915_gem_object_change_domain(obj,
3626 /* And bump the LRU for this access */
3627 vma = i915_gem_obj_to_ggtt(obj);
3628 if (vma && drm_mm_node_allocated(&vma->node) && !obj->active)
3629 list_move_tail(&vma->mm_list,
3630 &to_i915(obj->base.dev)->gtt.base.inactive_list);
3635 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
3636 enum i915_cache_level cache_level)
3638 struct drm_device *dev = obj->base.dev;
3639 struct i915_vma *vma, *next;
3642 if (obj->cache_level == cache_level)
3645 if (i915_gem_obj_is_pinned(obj)) {
3646 DRM_DEBUG("can not change the cache level of pinned objects\n");
3650 list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) {
3651 if (!i915_gem_valid_gtt_space(vma, cache_level)) {
3652 ret = i915_vma_unbind(vma);
3658 if (i915_gem_obj_bound_any(obj)) {
3659 ret = i915_gem_object_wait_rendering(obj, false);
3663 i915_gem_object_finish_gtt(obj);
3665 /* Before SandyBridge, you could not use tiling or fence
3666 * registers with snooped memory, so relinquish any fences
3667 * currently pointing to our region in the aperture.
3669 if (INTEL_INFO(dev)->gen < 6) {
3670 ret = i915_gem_object_put_fence(obj);
3675 list_for_each_entry(vma, &obj->vma_list, vma_link)
3676 if (drm_mm_node_allocated(&vma->node)) {
3677 ret = i915_vma_bind(vma, cache_level,
3684 list_for_each_entry(vma, &obj->vma_list, vma_link)
3685 vma->node.color = cache_level;
3686 obj->cache_level = cache_level;
3689 if (obj->cache_dirty &&
3690 obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
3691 cpu_write_needs_clflush(obj)) {
3692 if (i915_gem_clflush_object(obj, true))
3693 i915_gem_chipset_flush(obj->base.dev);
3699 int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data,
3700 struct drm_file *file)
3702 struct drm_i915_gem_caching *args = data;
3703 struct drm_i915_gem_object *obj;
3705 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3706 if (&obj->base == NULL)
3709 switch (obj->cache_level) {
3710 case I915_CACHE_LLC:
3711 case I915_CACHE_L3_LLC:
3712 args->caching = I915_CACHING_CACHED;
3716 args->caching = I915_CACHING_DISPLAY;
3720 args->caching = I915_CACHING_NONE;
3724 drm_gem_object_unreference_unlocked(&obj->base);
3728 int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
3729 struct drm_file *file)
3731 struct drm_i915_gem_caching *args = data;
3732 struct drm_i915_gem_object *obj;
3733 enum i915_cache_level level;
3736 switch (args->caching) {
3737 case I915_CACHING_NONE:
3738 level = I915_CACHE_NONE;
3740 case I915_CACHING_CACHED:
3742 * Due to a HW issue on BXT A stepping, GPU stores via a
3743 * snooped mapping may leave stale data in a corresponding CPU
3744 * cacheline, whereas normally such cachelines would get
3747 if (IS_BROXTON(dev) && INTEL_REVID(dev) < BXT_REVID_B0)
3750 level = I915_CACHE_LLC;
3752 case I915_CACHING_DISPLAY:
3753 level = HAS_WT(dev) ? I915_CACHE_WT : I915_CACHE_NONE;
3759 ret = i915_mutex_lock_interruptible(dev);
3763 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3764 if (&obj->base == NULL) {
3769 ret = i915_gem_object_set_cache_level(obj, level);
3771 drm_gem_object_unreference(&obj->base);
3773 mutex_unlock(&dev->struct_mutex);
3778 * Prepare buffer for display plane (scanout, cursors, etc).
3779 * Can be called from an uninterruptible phase (modesetting) and allows
3780 * any flushes to be pipelined (for pageflips).
3783 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
3785 struct intel_engine_cs *pipelined,
3786 struct drm_i915_gem_request **pipelined_request,
3787 const struct i915_ggtt_view *view)
3789 u32 old_read_domains, old_write_domain;
3792 ret = i915_gem_object_sync(obj, pipelined, pipelined_request);
3796 /* Mark the pin_display early so that we account for the
3797 * display coherency whilst setting up the cache domains.
3801 /* The display engine is not coherent with the LLC cache on gen6. As
3802 * a result, we make sure that the pinning that is about to occur is
3803 * done with uncached PTEs. This is lowest common denominator for all
3806 * However for gen6+, we could do better by using the GFDT bit instead
3807 * of uncaching, which would allow us to flush all the LLC-cached data
3808 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3810 ret = i915_gem_object_set_cache_level(obj,
3811 HAS_WT(obj->base.dev) ? I915_CACHE_WT : I915_CACHE_NONE);
3813 goto err_unpin_display;
3815 /* As the user may map the buffer once pinned in the display plane
3816 * (e.g. libkms for the bootup splash), we have to ensure that we
3817 * always use map_and_fenceable for all scanout buffers.
3819 ret = i915_gem_object_ggtt_pin(obj, view, alignment,
3820 view->type == I915_GGTT_VIEW_NORMAL ?
3823 goto err_unpin_display;
3825 i915_gem_object_flush_cpu_write_domain(obj);
3827 old_write_domain = obj->base.write_domain;
3828 old_read_domains = obj->base.read_domains;
3830 /* It should now be out of any other write domains, and we can update
3831 * the domain values for our changes.
3833 obj->base.write_domain = 0;
3834 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3836 trace_i915_gem_object_change_domain(obj,
3848 i915_gem_object_unpin_from_display_plane(struct drm_i915_gem_object *obj,
3849 const struct i915_ggtt_view *view)
3851 if (WARN_ON(obj->pin_display == 0))
3854 i915_gem_object_ggtt_unpin_view(obj, view);
3860 * Moves a single object to the CPU read, and possibly write domain.
3862 * This function returns when the move is complete, including waiting on
3866 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
3868 uint32_t old_write_domain, old_read_domains;
3871 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
3874 ret = i915_gem_object_wait_rendering(obj, !write);
3878 i915_gem_object_flush_gtt_write_domain(obj);
3880 old_write_domain = obj->base.write_domain;
3881 old_read_domains = obj->base.read_domains;
3883 /* Flush the CPU cache if it's still invalid. */
3884 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
3885 i915_gem_clflush_object(obj, false);
3887 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3890 /* It should now be out of any other write domains, and we can update
3891 * the domain values for our changes.
3893 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3895 /* If we're writing through the CPU, then the GPU read domains will
3896 * need to be invalidated at next use.
3899 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3900 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3903 trace_i915_gem_object_change_domain(obj,
3910 /* Throttle our rendering by waiting until the ring has completed our requests
3911 * emitted over 20 msec ago.
3913 * Note that if we were to use the current jiffies each time around the loop,
3914 * we wouldn't escape the function with any frames outstanding if the time to
3915 * render a frame was over 20ms.
3917 * This should get us reasonable parallelism between CPU and GPU but also
3918 * relatively low latency when blocking on a particular request to finish.
3921 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
3923 struct drm_i915_private *dev_priv = dev->dev_private;
3924 struct drm_i915_file_private *file_priv = file->driver_priv;
3925 unsigned long recent_enough = jiffies - DRM_I915_THROTTLE_JIFFIES;
3926 struct drm_i915_gem_request *request, *target = NULL;
3927 unsigned reset_counter;
3930 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
3934 ret = i915_gem_check_wedge(&dev_priv->gpu_error, false);
3938 spin_lock(&file_priv->mm.lock);
3939 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
3940 if (time_after_eq(request->emitted_jiffies, recent_enough))
3944 * Note that the request might not have been submitted yet.
3945 * In which case emitted_jiffies will be zero.
3947 if (!request->emitted_jiffies)
3952 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
3954 i915_gem_request_reference(target);
3955 spin_unlock(&file_priv->mm.lock);
3960 ret = __i915_wait_request(target, reset_counter, true, NULL, NULL);
3962 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
3964 i915_gem_request_unreference__unlocked(target);
3970 i915_vma_misplaced(struct i915_vma *vma, uint32_t alignment, uint64_t flags)
3972 struct drm_i915_gem_object *obj = vma->obj;
3975 vma->node.start & (alignment - 1))
3978 if (flags & PIN_MAPPABLE && !obj->map_and_fenceable)
3981 if (flags & PIN_OFFSET_BIAS &&
3982 vma->node.start < (flags & PIN_OFFSET_MASK))
3989 i915_gem_object_do_pin(struct drm_i915_gem_object *obj,
3990 struct i915_address_space *vm,
3991 const struct i915_ggtt_view *ggtt_view,
3995 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3996 struct i915_vma *vma;
4000 if (WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base))
4003 if (WARN_ON(flags & (PIN_GLOBAL | PIN_MAPPABLE) && !i915_is_ggtt(vm)))
4006 if (WARN_ON((flags & (PIN_MAPPABLE | PIN_GLOBAL)) == PIN_MAPPABLE))
4009 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
4012 vma = ggtt_view ? i915_gem_obj_to_ggtt_view(obj, ggtt_view) :
4013 i915_gem_obj_to_vma(obj, vm);
4016 return PTR_ERR(vma);
4019 if (WARN_ON(vma->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT))
4022 if (i915_vma_misplaced(vma, alignment, flags)) {
4023 WARN(vma->pin_count,
4024 "bo is already pinned in %s with incorrect alignment:"
4025 " offset=%08x %08x, req.alignment=%x, req.map_and_fenceable=%d,"
4026 " obj->map_and_fenceable=%d\n",
4027 ggtt_view ? "ggtt" : "ppgtt",
4028 upper_32_bits(vma->node.start),
4029 lower_32_bits(vma->node.start),
4031 !!(flags & PIN_MAPPABLE),
4032 obj->map_and_fenceable);
4033 ret = i915_vma_unbind(vma);
4041 bound = vma ? vma->bound : 0;
4042 if (vma == NULL || !drm_mm_node_allocated(&vma->node)) {
4043 vma = i915_gem_object_bind_to_vm(obj, vm, ggtt_view, alignment,
4046 return PTR_ERR(vma);
4048 ret = i915_vma_bind(vma, obj->cache_level, flags);
4053 if (ggtt_view && ggtt_view->type == I915_GGTT_VIEW_NORMAL &&
4054 (bound ^ vma->bound) & GLOBAL_BIND) {
4055 bool mappable, fenceable;
4056 u32 fence_size, fence_alignment;
4058 fence_size = i915_gem_get_gtt_size(obj->base.dev,
4061 fence_alignment = i915_gem_get_gtt_alignment(obj->base.dev,
4066 fenceable = (vma->node.size == fence_size &&
4067 (vma->node.start & (fence_alignment - 1)) == 0);
4069 mappable = (vma->node.start + fence_size <=
4070 dev_priv->gtt.mappable_end);
4072 obj->map_and_fenceable = mappable && fenceable;
4074 WARN_ON(flags & PIN_MAPPABLE && !obj->map_and_fenceable);
4082 i915_gem_object_pin(struct drm_i915_gem_object *obj,
4083 struct i915_address_space *vm,
4087 return i915_gem_object_do_pin(obj, vm,
4088 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL,
4093 i915_gem_object_ggtt_pin(struct drm_i915_gem_object *obj,
4094 const struct i915_ggtt_view *view,
4098 if (WARN_ONCE(!view, "no view specified"))
4101 return i915_gem_object_do_pin(obj, i915_obj_to_ggtt(obj), view,
4102 alignment, flags | PIN_GLOBAL);
4106 i915_gem_object_ggtt_unpin_view(struct drm_i915_gem_object *obj,
4107 const struct i915_ggtt_view *view)
4109 struct i915_vma *vma = i915_gem_obj_to_ggtt_view(obj, view);
4112 WARN_ON(vma->pin_count == 0);
4113 WARN_ON(!i915_gem_obj_ggtt_bound_view(obj, view));
4119 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
4120 struct drm_file *file)
4122 struct drm_i915_gem_busy *args = data;
4123 struct drm_i915_gem_object *obj;
4126 ret = i915_mutex_lock_interruptible(dev);
4130 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
4131 if (&obj->base == NULL) {
4136 /* Count all active objects as busy, even if they are currently not used
4137 * by the gpu. Users of this interface expect objects to eventually
4138 * become non-busy without any further actions, therefore emit any
4139 * necessary flushes here.
4141 ret = i915_gem_object_flush_active(obj);
4145 BUILD_BUG_ON(I915_NUM_RINGS > 16);
4146 args->busy = obj->active << 16;
4147 if (obj->last_write_req)
4148 args->busy |= obj->last_write_req->ring->id;
4151 drm_gem_object_unreference(&obj->base);
4153 mutex_unlock(&dev->struct_mutex);
4158 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
4159 struct drm_file *file_priv)
4161 return i915_gem_ring_throttle(dev, file_priv);
4165 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
4166 struct drm_file *file_priv)
4168 struct drm_i915_private *dev_priv = dev->dev_private;
4169 struct drm_i915_gem_madvise *args = data;
4170 struct drm_i915_gem_object *obj;
4173 switch (args->madv) {
4174 case I915_MADV_DONTNEED:
4175 case I915_MADV_WILLNEED:
4181 ret = i915_mutex_lock_interruptible(dev);
4185 obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
4186 if (&obj->base == NULL) {
4191 if (i915_gem_obj_is_pinned(obj)) {
4197 obj->tiling_mode != I915_TILING_NONE &&
4198 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
4199 if (obj->madv == I915_MADV_WILLNEED)
4200 i915_gem_object_unpin_pages(obj);
4201 if (args->madv == I915_MADV_WILLNEED)
4202 i915_gem_object_pin_pages(obj);
4205 if (obj->madv != __I915_MADV_PURGED)
4206 obj->madv = args->madv;
4208 /* if the object is no longer attached, discard its backing storage */
4209 if (obj->madv == I915_MADV_DONTNEED && obj->pages == NULL)
4210 i915_gem_object_truncate(obj);
4212 args->retained = obj->madv != __I915_MADV_PURGED;
4215 drm_gem_object_unreference(&obj->base);
4217 mutex_unlock(&dev->struct_mutex);
4221 void i915_gem_object_init(struct drm_i915_gem_object *obj,
4222 const struct drm_i915_gem_object_ops *ops)
4226 INIT_LIST_HEAD(&obj->global_list);
4227 for (i = 0; i < I915_NUM_RINGS; i++)
4228 INIT_LIST_HEAD(&obj->ring_list[i]);
4229 INIT_LIST_HEAD(&obj->obj_exec_link);
4230 INIT_LIST_HEAD(&obj->vma_list);
4231 INIT_LIST_HEAD(&obj->batch_pool_link);
4235 obj->fence_reg = I915_FENCE_REG_NONE;
4236 obj->madv = I915_MADV_WILLNEED;
4238 i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size);
4241 static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
4242 .get_pages = i915_gem_object_get_pages_gtt,
4243 .put_pages = i915_gem_object_put_pages_gtt,
4246 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
4249 struct drm_i915_gem_object *obj;
4250 struct address_space *mapping;
4253 obj = i915_gem_object_alloc(dev);
4257 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
4258 i915_gem_object_free(obj);
4262 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
4263 if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) {
4264 /* 965gm cannot relocate objects above 4GiB. */
4265 mask &= ~__GFP_HIGHMEM;
4266 mask |= __GFP_DMA32;
4269 mapping = file_inode(obj->base.filp)->i_mapping;
4270 mapping_set_gfp_mask(mapping, mask);
4272 i915_gem_object_init(obj, &i915_gem_object_ops);
4274 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4275 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4278 /* On some devices, we can have the GPU use the LLC (the CPU
4279 * cache) for about a 10% performance improvement
4280 * compared to uncached. Graphics requests other than
4281 * display scanout are coherent with the CPU in
4282 * accessing this cache. This means in this mode we
4283 * don't need to clflush on the CPU side, and on the
4284 * GPU side we only need to flush internal caches to
4285 * get data visible to the CPU.
4287 * However, we maintain the display planes as UC, and so
4288 * need to rebind when first used as such.
4290 obj->cache_level = I915_CACHE_LLC;
4292 obj->cache_level = I915_CACHE_NONE;
4294 trace_i915_gem_object_create(obj);
4299 static bool discard_backing_storage(struct drm_i915_gem_object *obj)
4301 /* If we are the last user of the backing storage (be it shmemfs
4302 * pages or stolen etc), we know that the pages are going to be
4303 * immediately released. In this case, we can then skip copying
4304 * back the contents from the GPU.
4307 if (obj->madv != I915_MADV_WILLNEED)
4310 if (obj->base.filp == NULL)
4313 /* At first glance, this looks racy, but then again so would be
4314 * userspace racing mmap against close. However, the first external
4315 * reference to the filp can only be obtained through the
4316 * i915_gem_mmap_ioctl() which safeguards us against the user
4317 * acquiring such a reference whilst we are in the middle of
4318 * freeing the object.
4320 return atomic_long_read(&obj->base.filp->f_count) == 1;
4323 void i915_gem_free_object(struct drm_gem_object *gem_obj)
4325 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
4326 struct drm_device *dev = obj->base.dev;
4327 struct drm_i915_private *dev_priv = dev->dev_private;
4328 struct i915_vma *vma, *next;
4330 intel_runtime_pm_get(dev_priv);
4332 trace_i915_gem_object_destroy(obj);
4334 list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) {
4338 ret = i915_vma_unbind(vma);
4339 if (WARN_ON(ret == -ERESTARTSYS)) {
4340 bool was_interruptible;
4342 was_interruptible = dev_priv->mm.interruptible;
4343 dev_priv->mm.interruptible = false;
4345 WARN_ON(i915_vma_unbind(vma));
4347 dev_priv->mm.interruptible = was_interruptible;
4351 /* Stolen objects don't hold a ref, but do hold pin count. Fix that up
4352 * before progressing. */
4354 i915_gem_object_unpin_pages(obj);
4356 WARN_ON(obj->frontbuffer_bits);
4358 if (obj->pages && obj->madv == I915_MADV_WILLNEED &&
4359 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES &&
4360 obj->tiling_mode != I915_TILING_NONE)
4361 i915_gem_object_unpin_pages(obj);
4363 if (WARN_ON(obj->pages_pin_count))
4364 obj->pages_pin_count = 0;
4365 if (discard_backing_storage(obj))
4366 obj->madv = I915_MADV_DONTNEED;
4367 i915_gem_object_put_pages(obj);
4368 i915_gem_object_free_mmap_offset(obj);
4372 if (obj->base.import_attach)
4373 drm_prime_gem_destroy(&obj->base, NULL);
4375 if (obj->ops->release)
4376 obj->ops->release(obj);
4378 drm_gem_object_release(&obj->base);
4379 i915_gem_info_remove_obj(dev_priv, obj->base.size);
4382 i915_gem_object_free(obj);
4384 intel_runtime_pm_put(dev_priv);
4387 struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj,
4388 struct i915_address_space *vm)
4390 struct i915_vma *vma;
4391 list_for_each_entry(vma, &obj->vma_list, vma_link) {
4392 if (i915_is_ggtt(vma->vm) &&
4393 vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL)
4401 struct i915_vma *i915_gem_obj_to_ggtt_view(struct drm_i915_gem_object *obj,
4402 const struct i915_ggtt_view *view)
4404 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj);
4405 struct i915_vma *vma;
4407 if (WARN_ONCE(!view, "no view specified"))
4408 return ERR_PTR(-EINVAL);
4410 list_for_each_entry(vma, &obj->vma_list, vma_link)
4411 if (vma->vm == ggtt &&
4412 i915_ggtt_view_equal(&vma->ggtt_view, view))
4417 void i915_gem_vma_destroy(struct i915_vma *vma)
4419 struct i915_address_space *vm = NULL;
4420 WARN_ON(vma->node.allocated);
4422 /* Keep the vma as a placeholder in the execbuffer reservation lists */
4423 if (!list_empty(&vma->exec_list))
4428 if (!i915_is_ggtt(vm))
4429 i915_ppgtt_put(i915_vm_to_ppgtt(vm));
4431 list_del(&vma->vma_link);
4433 kmem_cache_free(to_i915(vma->obj->base.dev)->vmas, vma);
4437 i915_gem_stop_ringbuffers(struct drm_device *dev)
4439 struct drm_i915_private *dev_priv = dev->dev_private;
4440 struct intel_engine_cs *ring;
4443 for_each_ring(ring, dev_priv, i)
4444 dev_priv->gt.stop_ring(ring);
4448 i915_gem_suspend(struct drm_device *dev)
4450 struct drm_i915_private *dev_priv = dev->dev_private;
4453 mutex_lock(&dev->struct_mutex);
4454 ret = i915_gpu_idle(dev);
4458 i915_gem_retire_requests(dev);
4460 i915_gem_stop_ringbuffers(dev);
4461 mutex_unlock(&dev->struct_mutex);
4463 cancel_delayed_work_sync(&dev_priv->gpu_error.hangcheck_work);
4464 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
4465 flush_delayed_work(&dev_priv->mm.idle_work);
4467 /* Assert that we sucessfully flushed all the work and
4468 * reset the GPU back to its idle, low power state.
4470 WARN_ON(dev_priv->mm.busy);
4475 mutex_unlock(&dev->struct_mutex);
4479 int i915_gem_l3_remap(struct drm_i915_gem_request *req, int slice)
4481 struct intel_engine_cs *ring = req->ring;
4482 struct drm_device *dev = ring->dev;
4483 struct drm_i915_private *dev_priv = dev->dev_private;
4484 u32 reg_base = GEN7_L3LOG_BASE + (slice * 0x200);
4485 u32 *remap_info = dev_priv->l3_parity.remap_info[slice];
4488 if (!HAS_L3_DPF(dev) || !remap_info)
4491 ret = intel_ring_begin(req, GEN7_L3LOG_SIZE / 4 * 3);
4496 * Note: We do not worry about the concurrent register cacheline hang
4497 * here because no other code should access these registers other than
4498 * at initialization time.
4500 for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) {
4501 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
4502 intel_ring_emit(ring, reg_base + i);
4503 intel_ring_emit(ring, remap_info[i/4]);
4506 intel_ring_advance(ring);
4511 void i915_gem_init_swizzling(struct drm_device *dev)
4513 struct drm_i915_private *dev_priv = dev->dev_private;
4515 if (INTEL_INFO(dev)->gen < 5 ||
4516 dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
4519 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
4520 DISP_TILE_SURFACE_SWIZZLING);
4525 I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
4527 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
4528 else if (IS_GEN7(dev))
4529 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
4530 else if (IS_GEN8(dev))
4531 I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW));
4537 intel_enable_blt(struct drm_device *dev)
4542 /* The blitter was dysfunctional on early prototypes */
4543 if (IS_GEN6(dev) && dev->pdev->revision < 8) {
4544 DRM_INFO("BLT not supported on this pre-production hardware;"
4545 " graphics performance will be degraded.\n");
4552 static void init_unused_ring(struct drm_device *dev, u32 base)
4554 struct drm_i915_private *dev_priv = dev->dev_private;
4556 I915_WRITE(RING_CTL(base), 0);
4557 I915_WRITE(RING_HEAD(base), 0);
4558 I915_WRITE(RING_TAIL(base), 0);
4559 I915_WRITE(RING_START(base), 0);
4562 static void init_unused_rings(struct drm_device *dev)
4565 init_unused_ring(dev, PRB1_BASE);
4566 init_unused_ring(dev, SRB0_BASE);
4567 init_unused_ring(dev, SRB1_BASE);
4568 init_unused_ring(dev, SRB2_BASE);
4569 init_unused_ring(dev, SRB3_BASE);
4570 } else if (IS_GEN2(dev)) {
4571 init_unused_ring(dev, SRB0_BASE);
4572 init_unused_ring(dev, SRB1_BASE);
4573 } else if (IS_GEN3(dev)) {
4574 init_unused_ring(dev, PRB1_BASE);
4575 init_unused_ring(dev, PRB2_BASE);
4579 int i915_gem_init_rings(struct drm_device *dev)
4581 struct drm_i915_private *dev_priv = dev->dev_private;
4584 ret = intel_init_render_ring_buffer(dev);
4589 ret = intel_init_bsd_ring_buffer(dev);
4591 goto cleanup_render_ring;
4594 if (intel_enable_blt(dev)) {
4595 ret = intel_init_blt_ring_buffer(dev);
4597 goto cleanup_bsd_ring;
4600 if (HAS_VEBOX(dev)) {
4601 ret = intel_init_vebox_ring_buffer(dev);
4603 goto cleanup_blt_ring;
4606 if (HAS_BSD2(dev)) {
4607 ret = intel_init_bsd2_ring_buffer(dev);
4609 goto cleanup_vebox_ring;
4615 intel_cleanup_ring_buffer(&dev_priv->ring[VECS]);
4617 intel_cleanup_ring_buffer(&dev_priv->ring[BCS]);
4619 intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
4620 cleanup_render_ring:
4621 intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
4627 i915_gem_init_hw(struct drm_device *dev)
4629 struct drm_i915_private *dev_priv = dev->dev_private;
4630 struct intel_engine_cs *ring;
4633 if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt())
4636 /* Double layer security blanket, see i915_gem_init() */
4637 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4639 if (dev_priv->ellc_size)
4640 I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf));
4642 if (IS_HASWELL(dev))
4643 I915_WRITE(MI_PREDICATE_RESULT_2, IS_HSW_GT3(dev) ?
4644 LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED);
4646 if (HAS_PCH_NOP(dev)) {
4647 if (IS_IVYBRIDGE(dev)) {
4648 u32 temp = I915_READ(GEN7_MSG_CTL);
4649 temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK);
4650 I915_WRITE(GEN7_MSG_CTL, temp);
4651 } else if (INTEL_INFO(dev)->gen >= 7) {
4652 u32 temp = I915_READ(HSW_NDE_RSTWRN_OPT);
4653 temp &= ~RESET_PCH_HANDSHAKE_ENABLE;
4654 I915_WRITE(HSW_NDE_RSTWRN_OPT, temp);
4658 i915_gem_init_swizzling(dev);
4661 * At least 830 can leave some of the unused rings
4662 * "active" (ie. head != tail) after resume which
4663 * will prevent c3 entry. Makes sure all unused rings
4666 init_unused_rings(dev);
4668 BUG_ON(!dev_priv->ring[RCS].default_context);
4670 ret = i915_ppgtt_init_hw(dev);
4672 DRM_ERROR("PPGTT enable HW failed %d\n", ret);
4676 /* Need to do basic initialisation of all rings first: */
4677 for_each_ring(ring, dev_priv, i) {
4678 ret = ring->init_hw(ring);
4683 /* We can't enable contexts until all firmware is loaded */
4684 if (HAS_GUC_UCODE(dev)) {
4685 ret = intel_guc_ucode_load(dev);
4688 * If we got an error and GuC submission is enabled, map
4689 * the error to -EIO so the GPU will be declared wedged.
4690 * OTOH, if we didn't intend to use the GuC anyway, just
4691 * discard the error and carry on.
4693 DRM_ERROR("Failed to initialize GuC, error %d%s\n", ret,
4694 i915.enable_guc_submission ? "" :
4696 ret = i915.enable_guc_submission ? -EIO : 0;
4703 * Increment the next seqno by 0x100 so we have a visible break
4704 * on re-initialisation
4706 ret = i915_gem_set_seqno(dev, dev_priv->next_seqno+0x100);
4710 /* Now it is safe to go back round and do everything else: */
4711 for_each_ring(ring, dev_priv, i) {
4712 struct drm_i915_gem_request *req;
4714 WARN_ON(!ring->default_context);
4716 ret = i915_gem_request_alloc(ring, ring->default_context, &req);
4718 i915_gem_cleanup_ringbuffer(dev);
4722 if (ring->id == RCS) {
4723 for (j = 0; j < NUM_L3_SLICES(dev); j++)
4724 i915_gem_l3_remap(req, j);
4727 ret = i915_ppgtt_init_ring(req);
4728 if (ret && ret != -EIO) {
4729 DRM_ERROR("PPGTT enable ring #%d failed %d\n", i, ret);
4730 i915_gem_request_cancel(req);
4731 i915_gem_cleanup_ringbuffer(dev);
4735 ret = i915_gem_context_enable(req);
4736 if (ret && ret != -EIO) {
4737 DRM_ERROR("Context enable ring #%d failed %d\n", i, ret);
4738 i915_gem_request_cancel(req);
4739 i915_gem_cleanup_ringbuffer(dev);
4743 i915_add_request_no_flush(req);
4747 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4751 int i915_gem_init(struct drm_device *dev)
4753 struct drm_i915_private *dev_priv = dev->dev_private;
4756 i915.enable_execlists = intel_sanitize_enable_execlists(dev,
4757 i915.enable_execlists);
4759 mutex_lock(&dev->struct_mutex);
4761 if (IS_VALLEYVIEW(dev)) {
4762 /* VLVA0 (potential hack), BIOS isn't actually waking us */
4763 I915_WRITE(VLV_GTLC_WAKE_CTRL, VLV_GTLC_ALLOWWAKEREQ);
4764 if (wait_for((I915_READ(VLV_GTLC_PW_STATUS) &
4765 VLV_GTLC_ALLOWWAKEACK), 10))
4766 DRM_DEBUG_DRIVER("allow wake ack timed out\n");
4769 if (!i915.enable_execlists) {
4770 dev_priv->gt.execbuf_submit = i915_gem_ringbuffer_submission;
4771 dev_priv->gt.init_rings = i915_gem_init_rings;
4772 dev_priv->gt.cleanup_ring = intel_cleanup_ring_buffer;
4773 dev_priv->gt.stop_ring = intel_stop_ring_buffer;
4775 dev_priv->gt.execbuf_submit = intel_execlists_submission;
4776 dev_priv->gt.init_rings = intel_logical_rings_init;
4777 dev_priv->gt.cleanup_ring = intel_logical_ring_cleanup;
4778 dev_priv->gt.stop_ring = intel_logical_ring_stop;
4781 /* This is just a security blanket to placate dragons.
4782 * On some systems, we very sporadically observe that the first TLBs
4783 * used by the CS may be stale, despite us poking the TLB reset. If
4784 * we hold the forcewake during initialisation these problems
4785 * just magically go away.
4787 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4789 ret = i915_gem_init_userptr(dev);
4793 i915_gem_init_global_gtt(dev);
4795 ret = i915_gem_context_init(dev);
4799 ret = dev_priv->gt.init_rings(dev);
4803 ret = i915_gem_init_hw(dev);
4805 /* Allow ring initialisation to fail by marking the GPU as
4806 * wedged. But we only want to do this where the GPU is angry,
4807 * for all other failure, such as an allocation failure, bail.
4809 DRM_ERROR("Failed to initialize GPU, declaring it wedged\n");
4810 atomic_or(I915_WEDGED, &dev_priv->gpu_error.reset_counter);
4815 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4816 mutex_unlock(&dev->struct_mutex);
4822 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
4824 struct drm_i915_private *dev_priv = dev->dev_private;
4825 struct intel_engine_cs *ring;
4828 for_each_ring(ring, dev_priv, i)
4829 dev_priv->gt.cleanup_ring(ring);
4831 if (i915.enable_execlists)
4833 * Neither the BIOS, ourselves or any other kernel
4834 * expects the system to be in execlists mode on startup,
4835 * so we need to reset the GPU back to legacy mode.
4837 intel_gpu_reset(dev);
4841 init_ring_lists(struct intel_engine_cs *ring)
4843 INIT_LIST_HEAD(&ring->active_list);
4844 INIT_LIST_HEAD(&ring->request_list);
4848 i915_gem_load(struct drm_device *dev)
4850 struct drm_i915_private *dev_priv = dev->dev_private;
4854 kmem_cache_create("i915_gem_object",
4855 sizeof(struct drm_i915_gem_object), 0,
4859 kmem_cache_create("i915_gem_vma",
4860 sizeof(struct i915_vma), 0,
4863 dev_priv->requests =
4864 kmem_cache_create("i915_gem_request",
4865 sizeof(struct drm_i915_gem_request), 0,
4869 INIT_LIST_HEAD(&dev_priv->vm_list);
4870 INIT_LIST_HEAD(&dev_priv->context_list);
4871 INIT_LIST_HEAD(&dev_priv->mm.unbound_list);
4872 INIT_LIST_HEAD(&dev_priv->mm.bound_list);
4873 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4874 for (i = 0; i < I915_NUM_RINGS; i++)
4875 init_ring_lists(&dev_priv->ring[i]);
4876 for (i = 0; i < I915_MAX_NUM_FENCES; i++)
4877 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
4878 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
4879 i915_gem_retire_work_handler);
4880 INIT_DELAYED_WORK(&dev_priv->mm.idle_work,
4881 i915_gem_idle_work_handler);
4882 init_waitqueue_head(&dev_priv->gpu_error.reset_queue);
4884 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
4886 if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev))
4887 dev_priv->num_fence_regs = 32;
4888 else if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
4889 dev_priv->num_fence_regs = 16;
4891 dev_priv->num_fence_regs = 8;
4893 if (intel_vgpu_active(dev))
4894 dev_priv->num_fence_regs =
4895 I915_READ(vgtif_reg(avail_rs.fence_num));
4898 * Set initial sequence number for requests.
4899 * Using this number allows the wraparound to happen early,
4900 * catching any obvious problems.
4902 dev_priv->next_seqno = ((u32)~0 - 0x1100);
4903 dev_priv->last_seqno = ((u32)~0 - 0x1101);
4905 /* Initialize fence registers to zero */
4906 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4907 i915_gem_restore_fences(dev);
4909 i915_gem_detect_bit_6_swizzle(dev);
4910 init_waitqueue_head(&dev_priv->pending_flip_queue);
4912 dev_priv->mm.interruptible = true;
4914 i915_gem_shrinker_init(dev_priv);
4916 mutex_init(&dev_priv->fb_tracking.lock);
4919 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
4921 struct drm_i915_file_private *file_priv = file->driver_priv;
4923 /* Clean up our request list when the client is going away, so that
4924 * later retire_requests won't dereference our soon-to-be-gone
4927 spin_lock(&file_priv->mm.lock);
4928 while (!list_empty(&file_priv->mm.request_list)) {
4929 struct drm_i915_gem_request *request;
4931 request = list_first_entry(&file_priv->mm.request_list,
4932 struct drm_i915_gem_request,
4934 list_del(&request->client_list);
4935 request->file_priv = NULL;
4937 spin_unlock(&file_priv->mm.lock);
4939 if (!list_empty(&file_priv->rps.link)) {
4940 spin_lock(&to_i915(dev)->rps.client_lock);
4941 list_del(&file_priv->rps.link);
4942 spin_unlock(&to_i915(dev)->rps.client_lock);
4946 int i915_gem_open(struct drm_device *dev, struct drm_file *file)
4948 struct drm_i915_file_private *file_priv;
4951 DRM_DEBUG_DRIVER("\n");
4953 file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL);
4957 file->driver_priv = file_priv;
4958 file_priv->dev_priv = dev->dev_private;
4959 file_priv->file = file;
4960 INIT_LIST_HEAD(&file_priv->rps.link);
4962 spin_lock_init(&file_priv->mm.lock);
4963 INIT_LIST_HEAD(&file_priv->mm.request_list);
4965 ret = i915_gem_context_open(dev, file);
4973 * i915_gem_track_fb - update frontbuffer tracking
4974 * @old: current GEM buffer for the frontbuffer slots
4975 * @new: new GEM buffer for the frontbuffer slots
4976 * @frontbuffer_bits: bitmask of frontbuffer slots
4978 * This updates the frontbuffer tracking bits @frontbuffer_bits by clearing them
4979 * from @old and setting them in @new. Both @old and @new can be NULL.
4981 void i915_gem_track_fb(struct drm_i915_gem_object *old,
4982 struct drm_i915_gem_object *new,
4983 unsigned frontbuffer_bits)
4986 WARN_ON(!mutex_is_locked(&old->base.dev->struct_mutex));
4987 WARN_ON(!(old->frontbuffer_bits & frontbuffer_bits));
4988 old->frontbuffer_bits &= ~frontbuffer_bits;
4992 WARN_ON(!mutex_is_locked(&new->base.dev->struct_mutex));
4993 WARN_ON(new->frontbuffer_bits & frontbuffer_bits);
4994 new->frontbuffer_bits |= frontbuffer_bits;
4998 /* All the new VM stuff */
4999 u64 i915_gem_obj_offset(struct drm_i915_gem_object *o,
5000 struct i915_address_space *vm)
5002 struct drm_i915_private *dev_priv = o->base.dev->dev_private;
5003 struct i915_vma *vma;
5005 WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base);
5007 list_for_each_entry(vma, &o->vma_list, vma_link) {
5008 if (i915_is_ggtt(vma->vm) &&
5009 vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL)
5012 return vma->node.start;
5015 WARN(1, "%s vma for this object not found.\n",
5016 i915_is_ggtt(vm) ? "global" : "ppgtt");
5020 u64 i915_gem_obj_ggtt_offset_view(struct drm_i915_gem_object *o,
5021 const struct i915_ggtt_view *view)
5023 struct i915_address_space *ggtt = i915_obj_to_ggtt(o);
5024 struct i915_vma *vma;
5026 list_for_each_entry(vma, &o->vma_list, vma_link)
5027 if (vma->vm == ggtt &&
5028 i915_ggtt_view_equal(&vma->ggtt_view, view))
5029 return vma->node.start;
5031 WARN(1, "global vma for this object not found. (view=%u)\n", view->type);
5035 bool i915_gem_obj_bound(struct drm_i915_gem_object *o,
5036 struct i915_address_space *vm)
5038 struct i915_vma *vma;
5040 list_for_each_entry(vma, &o->vma_list, vma_link) {
5041 if (i915_is_ggtt(vma->vm) &&
5042 vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL)
5044 if (vma->vm == vm && drm_mm_node_allocated(&vma->node))
5051 bool i915_gem_obj_ggtt_bound_view(struct drm_i915_gem_object *o,
5052 const struct i915_ggtt_view *view)
5054 struct i915_address_space *ggtt = i915_obj_to_ggtt(o);
5055 struct i915_vma *vma;
5057 list_for_each_entry(vma, &o->vma_list, vma_link)
5058 if (vma->vm == ggtt &&
5059 i915_ggtt_view_equal(&vma->ggtt_view, view) &&
5060 drm_mm_node_allocated(&vma->node))
5066 bool i915_gem_obj_bound_any(struct drm_i915_gem_object *o)
5068 struct i915_vma *vma;
5070 list_for_each_entry(vma, &o->vma_list, vma_link)
5071 if (drm_mm_node_allocated(&vma->node))
5077 unsigned long i915_gem_obj_size(struct drm_i915_gem_object *o,
5078 struct i915_address_space *vm)
5080 struct drm_i915_private *dev_priv = o->base.dev->dev_private;
5081 struct i915_vma *vma;
5083 WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base);
5085 BUG_ON(list_empty(&o->vma_list));
5087 list_for_each_entry(vma, &o->vma_list, vma_link) {
5088 if (i915_is_ggtt(vma->vm) &&
5089 vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL)
5092 return vma->node.size;
5097 bool i915_gem_obj_is_pinned(struct drm_i915_gem_object *obj)
5099 struct i915_vma *vma;
5100 list_for_each_entry(vma, &obj->vma_list, vma_link)
5101 if (vma->pin_count > 0)
5107 /* Allocate a new GEM object and fill it with the supplied data */
5108 struct drm_i915_gem_object *
5109 i915_gem_object_create_from_data(struct drm_device *dev,
5110 const void *data, size_t size)
5112 struct drm_i915_gem_object *obj;
5113 struct sg_table *sg;
5117 obj = i915_gem_alloc_object(dev, round_up(size, PAGE_SIZE));
5118 if (IS_ERR_OR_NULL(obj))
5121 ret = i915_gem_object_set_to_cpu_domain(obj, true);
5125 ret = i915_gem_object_get_pages(obj);
5129 i915_gem_object_pin_pages(obj);
5131 bytes = sg_copy_from_buffer(sg->sgl, sg->nents, (void *)data, size);
5132 i915_gem_object_unpin_pages(obj);
5134 if (WARN_ON(bytes != size)) {
5135 DRM_ERROR("Incomplete copy, wrote %zu of %zu", bytes, size);
5143 drm_gem_object_unreference(&obj->base);
5144 return ERR_PTR(ret);