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 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);
42 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj);
43 static __must_check int
44 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
47 i915_gem_object_retire(struct drm_i915_gem_object *obj);
49 static void i915_gem_write_fence(struct drm_device *dev, int reg,
50 struct drm_i915_gem_object *obj);
51 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
52 struct drm_i915_fence_reg *fence,
55 static bool cpu_cache_is_coherent(struct drm_device *dev,
56 enum i915_cache_level level)
58 return HAS_LLC(dev) || level != I915_CACHE_NONE;
61 static bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj)
63 if (!cpu_cache_is_coherent(obj->base.dev, obj->cache_level))
66 return obj->pin_display;
69 static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object *obj)
72 i915_gem_release_mmap(obj);
74 /* As we do not have an associated fence register, we will force
75 * a tiling change if we ever need to acquire one.
77 obj->fence_dirty = false;
78 obj->fence_reg = I915_FENCE_REG_NONE;
81 /* some bookkeeping */
82 static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
85 spin_lock(&dev_priv->mm.object_stat_lock);
86 dev_priv->mm.object_count++;
87 dev_priv->mm.object_memory += size;
88 spin_unlock(&dev_priv->mm.object_stat_lock);
91 static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
94 spin_lock(&dev_priv->mm.object_stat_lock);
95 dev_priv->mm.object_count--;
96 dev_priv->mm.object_memory -= size;
97 spin_unlock(&dev_priv->mm.object_stat_lock);
101 i915_gem_wait_for_error(struct i915_gpu_error *error)
105 #define EXIT_COND (!i915_reset_in_progress(error) || \
106 i915_terminally_wedged(error))
111 * Only wait 10 seconds for the gpu reset to complete to avoid hanging
112 * userspace. If it takes that long something really bad is going on and
113 * we should simply try to bail out and fail as gracefully as possible.
115 ret = wait_event_interruptible_timeout(error->reset_queue,
119 DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
121 } else if (ret < 0) {
129 int i915_mutex_lock_interruptible(struct drm_device *dev)
131 struct drm_i915_private *dev_priv = dev->dev_private;
134 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
138 ret = mutex_lock_interruptible(&dev->struct_mutex);
142 WARN_ON(i915_verify_lists(dev));
147 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
148 struct drm_file *file)
150 struct drm_i915_private *dev_priv = dev->dev_private;
151 struct drm_i915_gem_get_aperture *args = data;
152 struct drm_i915_gem_object *obj;
156 mutex_lock(&dev->struct_mutex);
157 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list)
158 if (i915_gem_obj_is_pinned(obj))
159 pinned += i915_gem_obj_ggtt_size(obj);
160 mutex_unlock(&dev->struct_mutex);
162 args->aper_size = dev_priv->gtt.base.total;
163 args->aper_available_size = args->aper_size - pinned;
169 i915_gem_object_get_pages_phys(struct drm_i915_gem_object *obj)
171 struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
172 char *vaddr = obj->phys_handle->vaddr;
174 struct scatterlist *sg;
177 if (WARN_ON(i915_gem_object_needs_bit17_swizzle(obj)))
180 for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
184 page = shmem_read_mapping_page(mapping, i);
186 return PTR_ERR(page);
188 src = kmap_atomic(page);
189 memcpy(vaddr, src, PAGE_SIZE);
190 drm_clflush_virt_range(vaddr, PAGE_SIZE);
193 page_cache_release(page);
197 i915_gem_chipset_flush(obj->base.dev);
199 st = kmalloc(sizeof(*st), GFP_KERNEL);
203 if (sg_alloc_table(st, 1, GFP_KERNEL)) {
210 sg->length = obj->base.size;
212 sg_dma_address(sg) = obj->phys_handle->busaddr;
213 sg_dma_len(sg) = obj->base.size;
216 obj->has_dma_mapping = true;
221 i915_gem_object_put_pages_phys(struct drm_i915_gem_object *obj)
225 BUG_ON(obj->madv == __I915_MADV_PURGED);
227 ret = i915_gem_object_set_to_cpu_domain(obj, true);
229 /* In the event of a disaster, abandon all caches and
232 WARN_ON(ret != -EIO);
233 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
236 if (obj->madv == I915_MADV_DONTNEED)
240 struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
241 char *vaddr = obj->phys_handle->vaddr;
244 for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
248 page = shmem_read_mapping_page(mapping, i);
252 dst = kmap_atomic(page);
253 drm_clflush_virt_range(vaddr, PAGE_SIZE);
254 memcpy(dst, vaddr, PAGE_SIZE);
257 set_page_dirty(page);
258 if (obj->madv == I915_MADV_WILLNEED)
259 mark_page_accessed(page);
260 page_cache_release(page);
266 sg_free_table(obj->pages);
269 obj->has_dma_mapping = false;
273 i915_gem_object_release_phys(struct drm_i915_gem_object *obj)
275 drm_pci_free(obj->base.dev, obj->phys_handle);
278 static const struct drm_i915_gem_object_ops i915_gem_phys_ops = {
279 .get_pages = i915_gem_object_get_pages_phys,
280 .put_pages = i915_gem_object_put_pages_phys,
281 .release = i915_gem_object_release_phys,
285 drop_pages(struct drm_i915_gem_object *obj)
287 struct i915_vma *vma, *next;
290 drm_gem_object_reference(&obj->base);
291 list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link)
292 if (i915_vma_unbind(vma))
295 ret = i915_gem_object_put_pages(obj);
296 drm_gem_object_unreference(&obj->base);
302 i915_gem_object_attach_phys(struct drm_i915_gem_object *obj,
305 drm_dma_handle_t *phys;
308 if (obj->phys_handle) {
309 if ((unsigned long)obj->phys_handle->vaddr & (align -1))
315 if (obj->madv != I915_MADV_WILLNEED)
318 if (obj->base.filp == NULL)
321 ret = drop_pages(obj);
325 /* create a new object */
326 phys = drm_pci_alloc(obj->base.dev, obj->base.size, align);
330 obj->phys_handle = phys;
331 obj->ops = &i915_gem_phys_ops;
333 return i915_gem_object_get_pages(obj);
337 i915_gem_phys_pwrite(struct drm_i915_gem_object *obj,
338 struct drm_i915_gem_pwrite *args,
339 struct drm_file *file_priv)
341 struct drm_device *dev = obj->base.dev;
342 void *vaddr = obj->phys_handle->vaddr + args->offset;
343 char __user *user_data = to_user_ptr(args->data_ptr);
346 /* We manually control the domain here and pretend that it
347 * remains coherent i.e. in the GTT domain, like shmem_pwrite.
349 ret = i915_gem_object_wait_rendering(obj, false);
353 intel_fb_obj_invalidate(obj, NULL, ORIGIN_CPU);
354 if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
355 unsigned long unwritten;
357 /* The physical object once assigned is fixed for the lifetime
358 * of the obj, so we can safely drop the lock and continue
361 mutex_unlock(&dev->struct_mutex);
362 unwritten = copy_from_user(vaddr, user_data, args->size);
363 mutex_lock(&dev->struct_mutex);
370 drm_clflush_virt_range(vaddr, args->size);
371 i915_gem_chipset_flush(dev);
374 intel_fb_obj_flush(obj, false);
378 void *i915_gem_object_alloc(struct drm_device *dev)
380 struct drm_i915_private *dev_priv = dev->dev_private;
381 return kmem_cache_zalloc(dev_priv->slab, GFP_KERNEL);
384 void i915_gem_object_free(struct drm_i915_gem_object *obj)
386 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
387 kmem_cache_free(dev_priv->slab, obj);
391 i915_gem_create(struct drm_file *file,
392 struct drm_device *dev,
396 struct drm_i915_gem_object *obj;
400 size = roundup(size, PAGE_SIZE);
404 /* Allocate the new object */
405 obj = i915_gem_alloc_object(dev, size);
409 ret = drm_gem_handle_create(file, &obj->base, &handle);
410 /* drop reference from allocate - handle holds it now */
411 drm_gem_object_unreference_unlocked(&obj->base);
420 i915_gem_dumb_create(struct drm_file *file,
421 struct drm_device *dev,
422 struct drm_mode_create_dumb *args)
424 /* have to work out size/pitch and return them */
425 args->pitch = ALIGN(args->width * DIV_ROUND_UP(args->bpp, 8), 64);
426 args->size = args->pitch * args->height;
427 return i915_gem_create(file, dev,
428 args->size, &args->handle);
432 * Creates a new mm object and returns a handle to it.
435 i915_gem_create_ioctl(struct drm_device *dev, void *data,
436 struct drm_file *file)
438 struct drm_i915_gem_create *args = data;
440 return i915_gem_create(file, dev,
441 args->size, &args->handle);
445 __copy_to_user_swizzled(char __user *cpu_vaddr,
446 const char *gpu_vaddr, int gpu_offset,
449 int ret, cpu_offset = 0;
452 int cacheline_end = ALIGN(gpu_offset + 1, 64);
453 int this_length = min(cacheline_end - gpu_offset, length);
454 int swizzled_gpu_offset = gpu_offset ^ 64;
456 ret = __copy_to_user(cpu_vaddr + cpu_offset,
457 gpu_vaddr + swizzled_gpu_offset,
462 cpu_offset += this_length;
463 gpu_offset += this_length;
464 length -= this_length;
471 __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset,
472 const char __user *cpu_vaddr,
475 int ret, cpu_offset = 0;
478 int cacheline_end = ALIGN(gpu_offset + 1, 64);
479 int this_length = min(cacheline_end - gpu_offset, length);
480 int swizzled_gpu_offset = gpu_offset ^ 64;
482 ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset,
483 cpu_vaddr + cpu_offset,
488 cpu_offset += this_length;
489 gpu_offset += this_length;
490 length -= this_length;
497 * Pins the specified object's pages and synchronizes the object with
498 * GPU accesses. Sets needs_clflush to non-zero if the caller should
499 * flush the object from the CPU cache.
501 int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj,
511 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) {
512 /* If we're not in the cpu read domain, set ourself into the gtt
513 * read domain and manually flush cachelines (if required). This
514 * optimizes for the case when the gpu will dirty the data
515 * anyway again before the next pread happens. */
516 *needs_clflush = !cpu_cache_is_coherent(obj->base.dev,
518 ret = i915_gem_object_wait_rendering(obj, true);
522 i915_gem_object_retire(obj);
525 ret = i915_gem_object_get_pages(obj);
529 i915_gem_object_pin_pages(obj);
534 /* Per-page copy function for the shmem pread fastpath.
535 * Flushes invalid cachelines before reading the target if
536 * needs_clflush is set. */
538 shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length,
539 char __user *user_data,
540 bool page_do_bit17_swizzling, bool needs_clflush)
545 if (unlikely(page_do_bit17_swizzling))
548 vaddr = kmap_atomic(page);
550 drm_clflush_virt_range(vaddr + shmem_page_offset,
552 ret = __copy_to_user_inatomic(user_data,
553 vaddr + shmem_page_offset,
555 kunmap_atomic(vaddr);
557 return ret ? -EFAULT : 0;
561 shmem_clflush_swizzled_range(char *addr, unsigned long length,
564 if (unlikely(swizzled)) {
565 unsigned long start = (unsigned long) addr;
566 unsigned long end = (unsigned long) addr + length;
568 /* For swizzling simply ensure that we always flush both
569 * channels. Lame, but simple and it works. Swizzled
570 * pwrite/pread is far from a hotpath - current userspace
571 * doesn't use it at all. */
572 start = round_down(start, 128);
573 end = round_up(end, 128);
575 drm_clflush_virt_range((void *)start, end - start);
577 drm_clflush_virt_range(addr, length);
582 /* Only difference to the fast-path function is that this can handle bit17
583 * and uses non-atomic copy and kmap functions. */
585 shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length,
586 char __user *user_data,
587 bool page_do_bit17_swizzling, bool needs_clflush)
594 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
596 page_do_bit17_swizzling);
598 if (page_do_bit17_swizzling)
599 ret = __copy_to_user_swizzled(user_data,
600 vaddr, shmem_page_offset,
603 ret = __copy_to_user(user_data,
604 vaddr + shmem_page_offset,
608 return ret ? - EFAULT : 0;
612 i915_gem_shmem_pread(struct drm_device *dev,
613 struct drm_i915_gem_object *obj,
614 struct drm_i915_gem_pread *args,
615 struct drm_file *file)
617 char __user *user_data;
620 int shmem_page_offset, page_length, ret = 0;
621 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
623 int needs_clflush = 0;
624 struct sg_page_iter sg_iter;
626 user_data = to_user_ptr(args->data_ptr);
629 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
631 ret = i915_gem_obj_prepare_shmem_read(obj, &needs_clflush);
635 offset = args->offset;
637 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
638 offset >> PAGE_SHIFT) {
639 struct page *page = sg_page_iter_page(&sg_iter);
644 /* Operation in this page
646 * shmem_page_offset = offset within page in shmem file
647 * page_length = bytes to copy for this page
649 shmem_page_offset = offset_in_page(offset);
650 page_length = remain;
651 if ((shmem_page_offset + page_length) > PAGE_SIZE)
652 page_length = PAGE_SIZE - shmem_page_offset;
654 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
655 (page_to_phys(page) & (1 << 17)) != 0;
657 ret = shmem_pread_fast(page, shmem_page_offset, page_length,
658 user_data, page_do_bit17_swizzling,
663 mutex_unlock(&dev->struct_mutex);
665 if (likely(!i915.prefault_disable) && !prefaulted) {
666 ret = fault_in_multipages_writeable(user_data, remain);
667 /* Userspace is tricking us, but we've already clobbered
668 * its pages with the prefault and promised to write the
669 * data up to the first fault. Hence ignore any errors
670 * and just continue. */
675 ret = shmem_pread_slow(page, shmem_page_offset, page_length,
676 user_data, page_do_bit17_swizzling,
679 mutex_lock(&dev->struct_mutex);
685 remain -= page_length;
686 user_data += page_length;
687 offset += page_length;
691 i915_gem_object_unpin_pages(obj);
697 * Reads data from the object referenced by handle.
699 * On error, the contents of *data are undefined.
702 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
703 struct drm_file *file)
705 struct drm_i915_gem_pread *args = data;
706 struct drm_i915_gem_object *obj;
712 if (!access_ok(VERIFY_WRITE,
713 to_user_ptr(args->data_ptr),
717 ret = i915_mutex_lock_interruptible(dev);
721 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
722 if (&obj->base == NULL) {
727 /* Bounds check source. */
728 if (args->offset > obj->base.size ||
729 args->size > obj->base.size - args->offset) {
734 /* prime objects have no backing filp to GEM pread/pwrite
737 if (!obj->base.filp) {
742 trace_i915_gem_object_pread(obj, args->offset, args->size);
744 ret = i915_gem_shmem_pread(dev, obj, args, file);
747 drm_gem_object_unreference(&obj->base);
749 mutex_unlock(&dev->struct_mutex);
753 /* This is the fast write path which cannot handle
754 * page faults in the source data
758 fast_user_write(struct io_mapping *mapping,
759 loff_t page_base, int page_offset,
760 char __user *user_data,
763 void __iomem *vaddr_atomic;
765 unsigned long unwritten;
767 vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
768 /* We can use the cpu mem copy function because this is X86. */
769 vaddr = (void __force*)vaddr_atomic + page_offset;
770 unwritten = __copy_from_user_inatomic_nocache(vaddr,
772 io_mapping_unmap_atomic(vaddr_atomic);
777 * This is the fast pwrite path, where we copy the data directly from the
778 * user into the GTT, uncached.
781 i915_gem_gtt_pwrite_fast(struct drm_device *dev,
782 struct drm_i915_gem_object *obj,
783 struct drm_i915_gem_pwrite *args,
784 struct drm_file *file)
786 struct drm_i915_private *dev_priv = dev->dev_private;
788 loff_t offset, page_base;
789 char __user *user_data;
790 int page_offset, page_length, ret;
792 ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE | PIN_NONBLOCK);
796 ret = i915_gem_object_set_to_gtt_domain(obj, true);
800 ret = i915_gem_object_put_fence(obj);
804 user_data = to_user_ptr(args->data_ptr);
807 offset = i915_gem_obj_ggtt_offset(obj) + args->offset;
809 intel_fb_obj_invalidate(obj, NULL, ORIGIN_GTT);
812 /* Operation in this page
814 * page_base = page offset within aperture
815 * page_offset = offset within page
816 * page_length = bytes to copy for this page
818 page_base = offset & PAGE_MASK;
819 page_offset = offset_in_page(offset);
820 page_length = remain;
821 if ((page_offset + remain) > PAGE_SIZE)
822 page_length = PAGE_SIZE - page_offset;
824 /* If we get a fault while copying data, then (presumably) our
825 * source page isn't available. Return the error and we'll
826 * retry in the slow path.
828 if (fast_user_write(dev_priv->gtt.mappable, page_base,
829 page_offset, user_data, page_length)) {
834 remain -= page_length;
835 user_data += page_length;
836 offset += page_length;
840 intel_fb_obj_flush(obj, false);
842 i915_gem_object_ggtt_unpin(obj);
847 /* Per-page copy function for the shmem pwrite fastpath.
848 * Flushes invalid cachelines before writing to the target if
849 * needs_clflush_before is set and flushes out any written cachelines after
850 * writing if needs_clflush is set. */
852 shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
853 char __user *user_data,
854 bool page_do_bit17_swizzling,
855 bool needs_clflush_before,
856 bool needs_clflush_after)
861 if (unlikely(page_do_bit17_swizzling))
864 vaddr = kmap_atomic(page);
865 if (needs_clflush_before)
866 drm_clflush_virt_range(vaddr + shmem_page_offset,
868 ret = __copy_from_user_inatomic(vaddr + shmem_page_offset,
869 user_data, page_length);
870 if (needs_clflush_after)
871 drm_clflush_virt_range(vaddr + shmem_page_offset,
873 kunmap_atomic(vaddr);
875 return ret ? -EFAULT : 0;
878 /* Only difference to the fast-path function is that this can handle bit17
879 * and uses non-atomic copy and kmap functions. */
881 shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
882 char __user *user_data,
883 bool page_do_bit17_swizzling,
884 bool needs_clflush_before,
885 bool needs_clflush_after)
891 if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
892 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
894 page_do_bit17_swizzling);
895 if (page_do_bit17_swizzling)
896 ret = __copy_from_user_swizzled(vaddr, shmem_page_offset,
900 ret = __copy_from_user(vaddr + shmem_page_offset,
903 if (needs_clflush_after)
904 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
906 page_do_bit17_swizzling);
909 return ret ? -EFAULT : 0;
913 i915_gem_shmem_pwrite(struct drm_device *dev,
914 struct drm_i915_gem_object *obj,
915 struct drm_i915_gem_pwrite *args,
916 struct drm_file *file)
920 char __user *user_data;
921 int shmem_page_offset, page_length, ret = 0;
922 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
923 int hit_slowpath = 0;
924 int needs_clflush_after = 0;
925 int needs_clflush_before = 0;
926 struct sg_page_iter sg_iter;
928 user_data = to_user_ptr(args->data_ptr);
931 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
933 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
934 /* If we're not in the cpu write domain, set ourself into the gtt
935 * write domain and manually flush cachelines (if required). This
936 * optimizes for the case when the gpu will use the data
937 * right away and we therefore have to clflush anyway. */
938 needs_clflush_after = cpu_write_needs_clflush(obj);
939 ret = i915_gem_object_wait_rendering(obj, false);
943 i915_gem_object_retire(obj);
945 /* Same trick applies to invalidate partially written cachelines read
947 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0)
948 needs_clflush_before =
949 !cpu_cache_is_coherent(dev, obj->cache_level);
951 ret = i915_gem_object_get_pages(obj);
955 intel_fb_obj_invalidate(obj, NULL, ORIGIN_CPU);
957 i915_gem_object_pin_pages(obj);
959 offset = args->offset;
962 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
963 offset >> PAGE_SHIFT) {
964 struct page *page = sg_page_iter_page(&sg_iter);
965 int partial_cacheline_write;
970 /* Operation in this page
972 * shmem_page_offset = offset within page in shmem file
973 * page_length = bytes to copy for this page
975 shmem_page_offset = offset_in_page(offset);
977 page_length = remain;
978 if ((shmem_page_offset + page_length) > PAGE_SIZE)
979 page_length = PAGE_SIZE - shmem_page_offset;
981 /* If we don't overwrite a cacheline completely we need to be
982 * careful to have up-to-date data by first clflushing. Don't
983 * overcomplicate things and flush the entire patch. */
984 partial_cacheline_write = needs_clflush_before &&
985 ((shmem_page_offset | page_length)
986 & (boot_cpu_data.x86_clflush_size - 1));
988 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
989 (page_to_phys(page) & (1 << 17)) != 0;
991 ret = shmem_pwrite_fast(page, shmem_page_offset, page_length,
992 user_data, page_do_bit17_swizzling,
993 partial_cacheline_write,
994 needs_clflush_after);
999 mutex_unlock(&dev->struct_mutex);
1000 ret = shmem_pwrite_slow(page, shmem_page_offset, page_length,
1001 user_data, page_do_bit17_swizzling,
1002 partial_cacheline_write,
1003 needs_clflush_after);
1005 mutex_lock(&dev->struct_mutex);
1011 remain -= page_length;
1012 user_data += page_length;
1013 offset += page_length;
1017 i915_gem_object_unpin_pages(obj);
1021 * Fixup: Flush cpu caches in case we didn't flush the dirty
1022 * cachelines in-line while writing and the object moved
1023 * out of the cpu write domain while we've dropped the lock.
1025 if (!needs_clflush_after &&
1026 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
1027 if (i915_gem_clflush_object(obj, obj->pin_display))
1028 i915_gem_chipset_flush(dev);
1032 if (needs_clflush_after)
1033 i915_gem_chipset_flush(dev);
1035 intel_fb_obj_flush(obj, false);
1040 * Writes data to the object referenced by handle.
1042 * On error, the contents of the buffer that were to be modified are undefined.
1045 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
1046 struct drm_file *file)
1048 struct drm_i915_private *dev_priv = dev->dev_private;
1049 struct drm_i915_gem_pwrite *args = data;
1050 struct drm_i915_gem_object *obj;
1053 if (args->size == 0)
1056 if (!access_ok(VERIFY_READ,
1057 to_user_ptr(args->data_ptr),
1061 if (likely(!i915.prefault_disable)) {
1062 ret = fault_in_multipages_readable(to_user_ptr(args->data_ptr),
1068 intel_runtime_pm_get(dev_priv);
1070 ret = i915_mutex_lock_interruptible(dev);
1074 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1075 if (&obj->base == NULL) {
1080 /* Bounds check destination. */
1081 if (args->offset > obj->base.size ||
1082 args->size > obj->base.size - args->offset) {
1087 /* prime objects have no backing filp to GEM pread/pwrite
1090 if (!obj->base.filp) {
1095 trace_i915_gem_object_pwrite(obj, args->offset, args->size);
1098 /* We can only do the GTT pwrite on untiled buffers, as otherwise
1099 * it would end up going through the fenced access, and we'll get
1100 * different detiling behavior between reading and writing.
1101 * pread/pwrite currently are reading and writing from the CPU
1102 * perspective, requiring manual detiling by the client.
1104 if (obj->tiling_mode == I915_TILING_NONE &&
1105 obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
1106 cpu_write_needs_clflush(obj)) {
1107 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
1108 /* Note that the gtt paths might fail with non-page-backed user
1109 * pointers (e.g. gtt mappings when moving data between
1110 * textures). Fallback to the shmem path in that case. */
1113 if (ret == -EFAULT || ret == -ENOSPC) {
1114 if (obj->phys_handle)
1115 ret = i915_gem_phys_pwrite(obj, args, file);
1117 ret = i915_gem_shmem_pwrite(dev, obj, args, file);
1121 drm_gem_object_unreference(&obj->base);
1123 mutex_unlock(&dev->struct_mutex);
1125 intel_runtime_pm_put(dev_priv);
1131 i915_gem_check_wedge(struct i915_gpu_error *error,
1134 if (i915_reset_in_progress(error)) {
1135 /* Non-interruptible callers can't handle -EAGAIN, hence return
1136 * -EIO unconditionally for these. */
1140 /* Recovery complete, but the reset failed ... */
1141 if (i915_terminally_wedged(error))
1145 * Check if GPU Reset is in progress - we need intel_ring_begin
1146 * to work properly to reinit the hw state while the gpu is
1147 * still marked as reset-in-progress. Handle this with a flag.
1149 if (!error->reload_in_reset)
1157 * Compare arbitrary request against outstanding lazy request. Emit on match.
1160 i915_gem_check_olr(struct drm_i915_gem_request *req)
1164 WARN_ON(!mutex_is_locked(&req->ring->dev->struct_mutex));
1167 if (req == req->ring->outstanding_lazy_request)
1168 ret = i915_add_request(req->ring);
1173 static void fake_irq(unsigned long data)
1175 wake_up_process((struct task_struct *)data);
1178 static bool missed_irq(struct drm_i915_private *dev_priv,
1179 struct intel_engine_cs *ring)
1181 return test_bit(ring->id, &dev_priv->gpu_error.missed_irq_rings);
1184 static bool can_wait_boost(struct drm_i915_file_private *file_priv)
1186 if (file_priv == NULL)
1189 return !atomic_xchg(&file_priv->rps_wait_boost, true);
1193 * __i915_wait_request - wait until execution of request has finished
1195 * @reset_counter: reset sequence associated with the given request
1196 * @interruptible: do an interruptible wait (normally yes)
1197 * @timeout: in - how long to wait (NULL forever); out - how much time remaining
1199 * Note: It is of utmost importance that the passed in seqno and reset_counter
1200 * values have been read by the caller in an smp safe manner. Where read-side
1201 * locks are involved, it is sufficient to read the reset_counter before
1202 * unlocking the lock that protects the seqno. For lockless tricks, the
1203 * reset_counter _must_ be read before, and an appropriate smp_rmb must be
1206 * Returns 0 if the request was found within the alloted time. Else returns the
1207 * errno with remaining time filled in timeout argument.
1209 int __i915_wait_request(struct drm_i915_gem_request *req,
1210 unsigned reset_counter,
1213 struct drm_i915_file_private *file_priv)
1215 struct intel_engine_cs *ring = i915_gem_request_get_ring(req);
1216 struct drm_device *dev = ring->dev;
1217 struct drm_i915_private *dev_priv = dev->dev_private;
1218 const bool irq_test_in_progress =
1219 ACCESS_ONCE(dev_priv->gpu_error.test_irq_rings) & intel_ring_flag(ring);
1221 unsigned long timeout_expire;
1225 WARN(!intel_irqs_enabled(dev_priv), "IRQs disabled");
1227 if (i915_gem_request_completed(req, true))
1230 timeout_expire = timeout ?
1231 jiffies + nsecs_to_jiffies_timeout((u64)*timeout) : 0;
1233 if (INTEL_INFO(dev)->gen >= 6 && ring->id == RCS && can_wait_boost(file_priv)) {
1234 gen6_rps_boost(dev_priv);
1236 mod_delayed_work(dev_priv->wq,
1237 &file_priv->mm.idle_work,
1238 msecs_to_jiffies(100));
1241 if (!irq_test_in_progress && WARN_ON(!ring->irq_get(ring)))
1244 /* Record current time in case interrupted by signal, or wedged */
1245 trace_i915_gem_request_wait_begin(req);
1246 before = ktime_get_raw_ns();
1248 struct timer_list timer;
1250 prepare_to_wait(&ring->irq_queue, &wait,
1251 interruptible ? TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE);
1253 /* We need to check whether any gpu reset happened in between
1254 * the caller grabbing the seqno and now ... */
1255 if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) {
1256 /* ... but upgrade the -EAGAIN to an -EIO if the gpu
1257 * is truely gone. */
1258 ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1264 if (i915_gem_request_completed(req, false)) {
1269 if (interruptible && signal_pending(current)) {
1274 if (timeout && time_after_eq(jiffies, timeout_expire)) {
1279 timer.function = NULL;
1280 if (timeout || missed_irq(dev_priv, ring)) {
1281 unsigned long expire;
1283 setup_timer_on_stack(&timer, fake_irq, (unsigned long)current);
1284 expire = missed_irq(dev_priv, ring) ? jiffies + 1 : timeout_expire;
1285 mod_timer(&timer, expire);
1290 if (timer.function) {
1291 del_singleshot_timer_sync(&timer);
1292 destroy_timer_on_stack(&timer);
1295 now = ktime_get_raw_ns();
1296 trace_i915_gem_request_wait_end(req);
1298 if (!irq_test_in_progress)
1299 ring->irq_put(ring);
1301 finish_wait(&ring->irq_queue, &wait);
1304 s64 tres = *timeout - (now - before);
1306 *timeout = tres < 0 ? 0 : tres;
1309 * Apparently ktime isn't accurate enough and occasionally has a
1310 * bit of mismatch in the jiffies<->nsecs<->ktime loop. So patch
1311 * things up to make the test happy. We allow up to 1 jiffy.
1313 * This is a regrssion from the timespec->ktime conversion.
1315 if (ret == -ETIME && *timeout < jiffies_to_usecs(1)*1000)
1323 * Waits for a request to be signaled, and cleans up the
1324 * request and object lists appropriately for that event.
1327 i915_wait_request(struct drm_i915_gem_request *req)
1329 struct drm_device *dev;
1330 struct drm_i915_private *dev_priv;
1332 unsigned reset_counter;
1335 BUG_ON(req == NULL);
1337 dev = req->ring->dev;
1338 dev_priv = dev->dev_private;
1339 interruptible = dev_priv->mm.interruptible;
1341 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1343 ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1347 ret = i915_gem_check_olr(req);
1351 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
1352 i915_gem_request_reference(req);
1353 ret = __i915_wait_request(req, reset_counter,
1354 interruptible, NULL, NULL);
1355 i915_gem_request_unreference(req);
1360 i915_gem_object_wait_rendering__tail(struct drm_i915_gem_object *obj)
1365 /* Manually manage the write flush as we may have not yet
1366 * retired the buffer.
1368 * Note that the last_write_req is always the earlier of
1369 * the two (read/write) requests, so if we haved successfully waited,
1370 * we know we have passed the last write.
1372 i915_gem_request_assign(&obj->last_write_req, NULL);
1378 * Ensures that all rendering to the object has completed and the object is
1379 * safe to unbind from the GTT or access from the CPU.
1381 static __must_check int
1382 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
1385 struct drm_i915_gem_request *req;
1388 req = readonly ? obj->last_write_req : obj->last_read_req;
1392 ret = i915_wait_request(req);
1396 return i915_gem_object_wait_rendering__tail(obj);
1399 /* A nonblocking variant of the above wait. This is a highly dangerous routine
1400 * as the object state may change during this call.
1402 static __must_check int
1403 i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
1404 struct drm_i915_file_private *file_priv,
1407 struct drm_i915_gem_request *req;
1408 struct drm_device *dev = obj->base.dev;
1409 struct drm_i915_private *dev_priv = dev->dev_private;
1410 unsigned reset_counter;
1413 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1414 BUG_ON(!dev_priv->mm.interruptible);
1416 req = readonly ? obj->last_write_req : obj->last_read_req;
1420 ret = i915_gem_check_wedge(&dev_priv->gpu_error, true);
1424 ret = i915_gem_check_olr(req);
1428 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
1429 i915_gem_request_reference(req);
1430 mutex_unlock(&dev->struct_mutex);
1431 ret = __i915_wait_request(req, reset_counter, true, NULL, file_priv);
1432 mutex_lock(&dev->struct_mutex);
1433 i915_gem_request_unreference(req);
1437 return i915_gem_object_wait_rendering__tail(obj);
1441 * Called when user space prepares to use an object with the CPU, either
1442 * through the mmap ioctl's mapping or a GTT mapping.
1445 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1446 struct drm_file *file)
1448 struct drm_i915_gem_set_domain *args = data;
1449 struct drm_i915_gem_object *obj;
1450 uint32_t read_domains = args->read_domains;
1451 uint32_t write_domain = args->write_domain;
1454 /* Only handle setting domains to types used by the CPU. */
1455 if (write_domain & I915_GEM_GPU_DOMAINS)
1458 if (read_domains & I915_GEM_GPU_DOMAINS)
1461 /* Having something in the write domain implies it's in the read
1462 * domain, and only that read domain. Enforce that in the request.
1464 if (write_domain != 0 && read_domains != write_domain)
1467 ret = i915_mutex_lock_interruptible(dev);
1471 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1472 if (&obj->base == NULL) {
1477 /* Try to flush the object off the GPU without holding the lock.
1478 * We will repeat the flush holding the lock in the normal manner
1479 * to catch cases where we are gazumped.
1481 ret = i915_gem_object_wait_rendering__nonblocking(obj,
1487 if (read_domains & I915_GEM_DOMAIN_GTT)
1488 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1490 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1493 drm_gem_object_unreference(&obj->base);
1495 mutex_unlock(&dev->struct_mutex);
1500 * Called when user space has done writes to this buffer
1503 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1504 struct drm_file *file)
1506 struct drm_i915_gem_sw_finish *args = data;
1507 struct drm_i915_gem_object *obj;
1510 ret = i915_mutex_lock_interruptible(dev);
1514 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1515 if (&obj->base == NULL) {
1520 /* Pinned buffers may be scanout, so flush the cache */
1521 if (obj->pin_display)
1522 i915_gem_object_flush_cpu_write_domain(obj);
1524 drm_gem_object_unreference(&obj->base);
1526 mutex_unlock(&dev->struct_mutex);
1531 * Maps the contents of an object, returning the address it is mapped
1534 * While the mapping holds a reference on the contents of the object, it doesn't
1535 * imply a ref on the object itself.
1539 * DRM driver writers who look a this function as an example for how to do GEM
1540 * mmap support, please don't implement mmap support like here. The modern way
1541 * to implement DRM mmap support is with an mmap offset ioctl (like
1542 * i915_gem_mmap_gtt) and then using the mmap syscall on the DRM fd directly.
1543 * That way debug tooling like valgrind will understand what's going on, hiding
1544 * the mmap call in a driver private ioctl will break that. The i915 driver only
1545 * does cpu mmaps this way because we didn't know better.
1548 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1549 struct drm_file *file)
1551 struct drm_i915_gem_mmap *args = data;
1552 struct drm_gem_object *obj;
1555 if (args->flags & ~(I915_MMAP_WC))
1558 if (args->flags & I915_MMAP_WC && !cpu_has_pat)
1561 obj = drm_gem_object_lookup(dev, file, args->handle);
1565 /* prime objects have no backing filp to GEM mmap
1569 drm_gem_object_unreference_unlocked(obj);
1573 addr = vm_mmap(obj->filp, 0, args->size,
1574 PROT_READ | PROT_WRITE, MAP_SHARED,
1576 if (args->flags & I915_MMAP_WC) {
1577 struct mm_struct *mm = current->mm;
1578 struct vm_area_struct *vma;
1580 down_write(&mm->mmap_sem);
1581 vma = find_vma(mm, addr);
1584 pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
1587 up_write(&mm->mmap_sem);
1589 drm_gem_object_unreference_unlocked(obj);
1590 if (IS_ERR((void *)addr))
1593 args->addr_ptr = (uint64_t) addr;
1599 * i915_gem_fault - fault a page into the GTT
1600 * vma: VMA in question
1603 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1604 * from userspace. The fault handler takes care of binding the object to
1605 * the GTT (if needed), allocating and programming a fence register (again,
1606 * only if needed based on whether the old reg is still valid or the object
1607 * is tiled) and inserting a new PTE into the faulting process.
1609 * Note that the faulting process may involve evicting existing objects
1610 * from the GTT and/or fence registers to make room. So performance may
1611 * suffer if the GTT working set is large or there are few fence registers
1614 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1616 struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1617 struct drm_device *dev = obj->base.dev;
1618 struct drm_i915_private *dev_priv = dev->dev_private;
1619 pgoff_t page_offset;
1622 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1624 intel_runtime_pm_get(dev_priv);
1626 /* We don't use vmf->pgoff since that has the fake offset */
1627 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1630 ret = i915_mutex_lock_interruptible(dev);
1634 trace_i915_gem_object_fault(obj, page_offset, true, write);
1636 /* Try to flush the object off the GPU first without holding the lock.
1637 * Upon reacquiring the lock, we will perform our sanity checks and then
1638 * repeat the flush holding the lock in the normal manner to catch cases
1639 * where we are gazumped.
1641 ret = i915_gem_object_wait_rendering__nonblocking(obj, NULL, !write);
1645 /* Access to snoopable pages through the GTT is incoherent. */
1646 if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) {
1651 /* Now bind it into the GTT if needed */
1652 ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE);
1656 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1660 ret = i915_gem_object_get_fence(obj);
1664 /* Finally, remap it using the new GTT offset */
1665 pfn = dev_priv->gtt.mappable_base + i915_gem_obj_ggtt_offset(obj);
1668 if (!obj->fault_mappable) {
1669 unsigned long size = min_t(unsigned long,
1670 vma->vm_end - vma->vm_start,
1674 for (i = 0; i < size >> PAGE_SHIFT; i++) {
1675 ret = vm_insert_pfn(vma,
1676 (unsigned long)vma->vm_start + i * PAGE_SIZE,
1682 obj->fault_mappable = true;
1684 ret = vm_insert_pfn(vma,
1685 (unsigned long)vmf->virtual_address,
1688 i915_gem_object_ggtt_unpin(obj);
1690 mutex_unlock(&dev->struct_mutex);
1695 * We eat errors when the gpu is terminally wedged to avoid
1696 * userspace unduly crashing (gl has no provisions for mmaps to
1697 * fail). But any other -EIO isn't ours (e.g. swap in failure)
1698 * and so needs to be reported.
1700 if (!i915_terminally_wedged(&dev_priv->gpu_error)) {
1701 ret = VM_FAULT_SIGBUS;
1706 * EAGAIN means the gpu is hung and we'll wait for the error
1707 * handler to reset everything when re-faulting in
1708 * i915_mutex_lock_interruptible.
1715 * EBUSY is ok: this just means that another thread
1716 * already did the job.
1718 ret = VM_FAULT_NOPAGE;
1725 ret = VM_FAULT_SIGBUS;
1728 WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret);
1729 ret = VM_FAULT_SIGBUS;
1733 intel_runtime_pm_put(dev_priv);
1738 * i915_gem_release_mmap - remove physical page mappings
1739 * @obj: obj in question
1741 * Preserve the reservation of the mmapping with the DRM core code, but
1742 * relinquish ownership of the pages back to the system.
1744 * It is vital that we remove the page mapping if we have mapped a tiled
1745 * object through the GTT and then lose the fence register due to
1746 * resource pressure. Similarly if the object has been moved out of the
1747 * aperture, than pages mapped into userspace must be revoked. Removing the
1748 * mapping will then trigger a page fault on the next user access, allowing
1749 * fixup by i915_gem_fault().
1752 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1754 if (!obj->fault_mappable)
1757 drm_vma_node_unmap(&obj->base.vma_node,
1758 obj->base.dev->anon_inode->i_mapping);
1759 obj->fault_mappable = false;
1763 i915_gem_release_all_mmaps(struct drm_i915_private *dev_priv)
1765 struct drm_i915_gem_object *obj;
1767 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list)
1768 i915_gem_release_mmap(obj);
1772 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1776 if (INTEL_INFO(dev)->gen >= 4 ||
1777 tiling_mode == I915_TILING_NONE)
1780 /* Previous chips need a power-of-two fence region when tiling */
1781 if (INTEL_INFO(dev)->gen == 3)
1782 gtt_size = 1024*1024;
1784 gtt_size = 512*1024;
1786 while (gtt_size < size)
1793 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1794 * @obj: object to check
1796 * Return the required GTT alignment for an object, taking into account
1797 * potential fence register mapping.
1800 i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size,
1801 int tiling_mode, bool fenced)
1804 * Minimum alignment is 4k (GTT page size), but might be greater
1805 * if a fence register is needed for the object.
1807 if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) ||
1808 tiling_mode == I915_TILING_NONE)
1812 * Previous chips need to be aligned to the size of the smallest
1813 * fence register that can contain the object.
1815 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1818 static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj)
1820 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1823 if (drm_vma_node_has_offset(&obj->base.vma_node))
1826 dev_priv->mm.shrinker_no_lock_stealing = true;
1828 ret = drm_gem_create_mmap_offset(&obj->base);
1832 /* Badly fragmented mmap space? The only way we can recover
1833 * space is by destroying unwanted objects. We can't randomly release
1834 * mmap_offsets as userspace expects them to be persistent for the
1835 * lifetime of the objects. The closest we can is to release the
1836 * offsets on purgeable objects by truncating it and marking it purged,
1837 * which prevents userspace from ever using that object again.
1839 i915_gem_shrink(dev_priv,
1840 obj->base.size >> PAGE_SHIFT,
1842 I915_SHRINK_UNBOUND |
1843 I915_SHRINK_PURGEABLE);
1844 ret = drm_gem_create_mmap_offset(&obj->base);
1848 i915_gem_shrink_all(dev_priv);
1849 ret = drm_gem_create_mmap_offset(&obj->base);
1851 dev_priv->mm.shrinker_no_lock_stealing = false;
1856 static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)
1858 drm_gem_free_mmap_offset(&obj->base);
1862 i915_gem_mmap_gtt(struct drm_file *file,
1863 struct drm_device *dev,
1867 struct drm_i915_private *dev_priv = dev->dev_private;
1868 struct drm_i915_gem_object *obj;
1871 ret = i915_mutex_lock_interruptible(dev);
1875 obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
1876 if (&obj->base == NULL) {
1881 if (obj->base.size > dev_priv->gtt.mappable_end) {
1886 if (obj->madv != I915_MADV_WILLNEED) {
1887 DRM_DEBUG("Attempting to mmap a purgeable buffer\n");
1892 ret = i915_gem_object_create_mmap_offset(obj);
1896 *offset = drm_vma_node_offset_addr(&obj->base.vma_node);
1899 drm_gem_object_unreference(&obj->base);
1901 mutex_unlock(&dev->struct_mutex);
1906 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1908 * @data: GTT mapping ioctl data
1909 * @file: GEM object info
1911 * Simply returns the fake offset to userspace so it can mmap it.
1912 * The mmap call will end up in drm_gem_mmap(), which will set things
1913 * up so we can get faults in the handler above.
1915 * The fault handler will take care of binding the object into the GTT
1916 * (since it may have been evicted to make room for something), allocating
1917 * a fence register, and mapping the appropriate aperture address into
1921 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1922 struct drm_file *file)
1924 struct drm_i915_gem_mmap_gtt *args = data;
1926 return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
1929 /* Immediately discard the backing storage */
1931 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
1933 i915_gem_object_free_mmap_offset(obj);
1935 if (obj->base.filp == NULL)
1938 /* Our goal here is to return as much of the memory as
1939 * is possible back to the system as we are called from OOM.
1940 * To do this we must instruct the shmfs to drop all of its
1941 * backing pages, *now*.
1943 shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1);
1944 obj->madv = __I915_MADV_PURGED;
1947 /* Try to discard unwanted pages */
1949 i915_gem_object_invalidate(struct drm_i915_gem_object *obj)
1951 struct address_space *mapping;
1953 switch (obj->madv) {
1954 case I915_MADV_DONTNEED:
1955 i915_gem_object_truncate(obj);
1956 case __I915_MADV_PURGED:
1960 if (obj->base.filp == NULL)
1963 mapping = file_inode(obj->base.filp)->i_mapping,
1964 invalidate_mapping_pages(mapping, 0, (loff_t)-1);
1968 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
1970 struct sg_page_iter sg_iter;
1973 BUG_ON(obj->madv == __I915_MADV_PURGED);
1975 ret = i915_gem_object_set_to_cpu_domain(obj, true);
1977 /* In the event of a disaster, abandon all caches and
1978 * hope for the best.
1980 WARN_ON(ret != -EIO);
1981 i915_gem_clflush_object(obj, true);
1982 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
1985 if (i915_gem_object_needs_bit17_swizzle(obj))
1986 i915_gem_object_save_bit_17_swizzle(obj);
1988 if (obj->madv == I915_MADV_DONTNEED)
1991 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
1992 struct page *page = sg_page_iter_page(&sg_iter);
1995 set_page_dirty(page);
1997 if (obj->madv == I915_MADV_WILLNEED)
1998 mark_page_accessed(page);
2000 page_cache_release(page);
2004 sg_free_table(obj->pages);
2009 i915_gem_object_put_pages(struct drm_i915_gem_object *obj)
2011 const struct drm_i915_gem_object_ops *ops = obj->ops;
2013 if (obj->pages == NULL)
2016 if (obj->pages_pin_count)
2019 BUG_ON(i915_gem_obj_bound_any(obj));
2021 /* ->put_pages might need to allocate memory for the bit17 swizzle
2022 * array, hence protect them from being reaped by removing them from gtt
2024 list_del(&obj->global_list);
2026 ops->put_pages(obj);
2029 i915_gem_object_invalidate(obj);
2035 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
2037 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2039 struct address_space *mapping;
2040 struct sg_table *st;
2041 struct scatterlist *sg;
2042 struct sg_page_iter sg_iter;
2044 unsigned long last_pfn = 0; /* suppress gcc warning */
2047 /* Assert that the object is not currently in any GPU domain. As it
2048 * wasn't in the GTT, there shouldn't be any way it could have been in
2051 BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
2052 BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
2054 st = kmalloc(sizeof(*st), GFP_KERNEL);
2058 page_count = obj->base.size / PAGE_SIZE;
2059 if (sg_alloc_table(st, page_count, GFP_KERNEL)) {
2064 /* Get the list of pages out of our struct file. They'll be pinned
2065 * at this point until we release them.
2067 * Fail silently without starting the shrinker
2069 mapping = file_inode(obj->base.filp)->i_mapping;
2070 gfp = mapping_gfp_mask(mapping);
2071 gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD;
2072 gfp &= ~(__GFP_IO | __GFP_WAIT);
2075 for (i = 0; i < page_count; i++) {
2076 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2078 i915_gem_shrink(dev_priv,
2081 I915_SHRINK_UNBOUND |
2082 I915_SHRINK_PURGEABLE);
2083 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2086 /* We've tried hard to allocate the memory by reaping
2087 * our own buffer, now let the real VM do its job and
2088 * go down in flames if truly OOM.
2090 i915_gem_shrink_all(dev_priv);
2091 page = shmem_read_mapping_page(mapping, i);
2095 #ifdef CONFIG_SWIOTLB
2096 if (swiotlb_nr_tbl()) {
2098 sg_set_page(sg, page, PAGE_SIZE, 0);
2103 if (!i || page_to_pfn(page) != last_pfn + 1) {
2107 sg_set_page(sg, page, PAGE_SIZE, 0);
2109 sg->length += PAGE_SIZE;
2111 last_pfn = page_to_pfn(page);
2113 /* Check that the i965g/gm workaround works. */
2114 WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
2116 #ifdef CONFIG_SWIOTLB
2117 if (!swiotlb_nr_tbl())
2122 if (i915_gem_object_needs_bit17_swizzle(obj))
2123 i915_gem_object_do_bit_17_swizzle(obj);
2125 if (obj->tiling_mode != I915_TILING_NONE &&
2126 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES)
2127 i915_gem_object_pin_pages(obj);
2133 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0)
2134 page_cache_release(sg_page_iter_page(&sg_iter));
2138 /* shmemfs first checks if there is enough memory to allocate the page
2139 * and reports ENOSPC should there be insufficient, along with the usual
2140 * ENOMEM for a genuine allocation failure.
2142 * We use ENOSPC in our driver to mean that we have run out of aperture
2143 * space and so want to translate the error from shmemfs back to our
2144 * usual understanding of ENOMEM.
2146 if (PTR_ERR(page) == -ENOSPC)
2149 return PTR_ERR(page);
2152 /* Ensure that the associated pages are gathered from the backing storage
2153 * and pinned into our object. i915_gem_object_get_pages() may be called
2154 * multiple times before they are released by a single call to
2155 * i915_gem_object_put_pages() - once the pages are no longer referenced
2156 * either as a result of memory pressure (reaping pages under the shrinker)
2157 * or as the object is itself released.
2160 i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
2162 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2163 const struct drm_i915_gem_object_ops *ops = obj->ops;
2169 if (obj->madv != I915_MADV_WILLNEED) {
2170 DRM_DEBUG("Attempting to obtain a purgeable object\n");
2174 BUG_ON(obj->pages_pin_count);
2176 ret = ops->get_pages(obj);
2180 list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list);
2185 i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
2186 struct intel_engine_cs *ring)
2188 struct drm_i915_gem_request *req;
2189 struct intel_engine_cs *old_ring;
2191 BUG_ON(ring == NULL);
2193 req = intel_ring_get_request(ring);
2194 old_ring = i915_gem_request_get_ring(obj->last_read_req);
2196 if (old_ring != ring && obj->last_write_req) {
2197 /* Keep the request relative to the current ring */
2198 i915_gem_request_assign(&obj->last_write_req, req);
2201 /* Add a reference if we're newly entering the active list. */
2203 drm_gem_object_reference(&obj->base);
2207 list_move_tail(&obj->ring_list, &ring->active_list);
2209 i915_gem_request_assign(&obj->last_read_req, req);
2212 void i915_vma_move_to_active(struct i915_vma *vma,
2213 struct intel_engine_cs *ring)
2215 list_move_tail(&vma->mm_list, &vma->vm->active_list);
2216 return i915_gem_object_move_to_active(vma->obj, ring);
2220 i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
2222 struct i915_vma *vma;
2224 BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS);
2225 BUG_ON(!obj->active);
2227 list_for_each_entry(vma, &obj->vma_list, vma_link) {
2228 if (!list_empty(&vma->mm_list))
2229 list_move_tail(&vma->mm_list, &vma->vm->inactive_list);
2232 intel_fb_obj_flush(obj, true);
2234 list_del_init(&obj->ring_list);
2236 i915_gem_request_assign(&obj->last_read_req, NULL);
2237 i915_gem_request_assign(&obj->last_write_req, NULL);
2238 obj->base.write_domain = 0;
2240 i915_gem_request_assign(&obj->last_fenced_req, NULL);
2243 drm_gem_object_unreference(&obj->base);
2245 WARN_ON(i915_verify_lists(dev));
2249 i915_gem_object_retire(struct drm_i915_gem_object *obj)
2251 if (obj->last_read_req == NULL)
2254 if (i915_gem_request_completed(obj->last_read_req, true))
2255 i915_gem_object_move_to_inactive(obj);
2259 i915_gem_init_seqno(struct drm_device *dev, u32 seqno)
2261 struct drm_i915_private *dev_priv = dev->dev_private;
2262 struct intel_engine_cs *ring;
2265 /* Carefully retire all requests without writing to the rings */
2266 for_each_ring(ring, dev_priv, i) {
2267 ret = intel_ring_idle(ring);
2271 i915_gem_retire_requests(dev);
2273 /* Finally reset hw state */
2274 for_each_ring(ring, dev_priv, i) {
2275 intel_ring_init_seqno(ring, seqno);
2277 for (j = 0; j < ARRAY_SIZE(ring->semaphore.sync_seqno); j++)
2278 ring->semaphore.sync_seqno[j] = 0;
2284 int i915_gem_set_seqno(struct drm_device *dev, u32 seqno)
2286 struct drm_i915_private *dev_priv = dev->dev_private;
2292 /* HWS page needs to be set less than what we
2293 * will inject to ring
2295 ret = i915_gem_init_seqno(dev, seqno - 1);
2299 /* Carefully set the last_seqno value so that wrap
2300 * detection still works
2302 dev_priv->next_seqno = seqno;
2303 dev_priv->last_seqno = seqno - 1;
2304 if (dev_priv->last_seqno == 0)
2305 dev_priv->last_seqno--;
2311 i915_gem_get_seqno(struct drm_device *dev, u32 *seqno)
2313 struct drm_i915_private *dev_priv = dev->dev_private;
2315 /* reserve 0 for non-seqno */
2316 if (dev_priv->next_seqno == 0) {
2317 int ret = i915_gem_init_seqno(dev, 0);
2321 dev_priv->next_seqno = 1;
2324 *seqno = dev_priv->last_seqno = dev_priv->next_seqno++;
2328 int __i915_add_request(struct intel_engine_cs *ring,
2329 struct drm_file *file,
2330 struct drm_i915_gem_object *obj)
2332 struct drm_i915_private *dev_priv = ring->dev->dev_private;
2333 struct drm_i915_gem_request *request;
2334 struct intel_ringbuffer *ringbuf;
2338 request = ring->outstanding_lazy_request;
2339 if (WARN_ON(request == NULL))
2342 if (i915.enable_execlists) {
2343 ringbuf = request->ctx->engine[ring->id].ringbuf;
2345 ringbuf = ring->buffer;
2347 request_start = intel_ring_get_tail(ringbuf);
2349 * Emit any outstanding flushes - execbuf can fail to emit the flush
2350 * after having emitted the batchbuffer command. Hence we need to fix
2351 * things up similar to emitting the lazy request. The difference here
2352 * is that the flush _must_ happen before the next request, no matter
2355 if (i915.enable_execlists) {
2356 ret = logical_ring_flush_all_caches(ringbuf, request->ctx);
2360 ret = intel_ring_flush_all_caches(ring);
2365 /* Record the position of the start of the request so that
2366 * should we detect the updated seqno part-way through the
2367 * GPU processing the request, we never over-estimate the
2368 * position of the head.
2370 request->postfix = intel_ring_get_tail(ringbuf);
2372 if (i915.enable_execlists) {
2373 ret = ring->emit_request(ringbuf, request);
2377 ret = ring->add_request(ring);
2381 request->tail = intel_ring_get_tail(ringbuf);
2384 request->head = request_start;
2386 /* Whilst this request exists, batch_obj will be on the
2387 * active_list, and so will hold the active reference. Only when this
2388 * request is retired will the the batch_obj be moved onto the
2389 * inactive_list and lose its active reference. Hence we do not need
2390 * to explicitly hold another reference here.
2392 request->batch_obj = obj;
2394 if (!i915.enable_execlists) {
2395 /* Hold a reference to the current context so that we can inspect
2396 * it later in case a hangcheck error event fires.
2398 request->ctx = ring->last_context;
2400 i915_gem_context_reference(request->ctx);
2403 request->emitted_jiffies = jiffies;
2404 ring->last_submitted_seqno = request->seqno;
2405 list_add_tail(&request->list, &ring->request_list);
2406 request->file_priv = NULL;
2409 struct drm_i915_file_private *file_priv = file->driver_priv;
2411 spin_lock(&file_priv->mm.lock);
2412 request->file_priv = file_priv;
2413 list_add_tail(&request->client_list,
2414 &file_priv->mm.request_list);
2415 spin_unlock(&file_priv->mm.lock);
2417 request->pid = get_pid(task_pid(current));
2420 trace_i915_gem_request_add(request);
2421 ring->outstanding_lazy_request = NULL;
2423 i915_queue_hangcheck(ring->dev);
2425 cancel_delayed_work_sync(&dev_priv->mm.idle_work);
2426 queue_delayed_work(dev_priv->wq,
2427 &dev_priv->mm.retire_work,
2428 round_jiffies_up_relative(HZ));
2429 intel_mark_busy(dev_priv->dev);
2435 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
2437 struct drm_i915_file_private *file_priv = request->file_priv;
2442 spin_lock(&file_priv->mm.lock);
2443 list_del(&request->client_list);
2444 request->file_priv = NULL;
2445 spin_unlock(&file_priv->mm.lock);
2448 static bool i915_context_is_banned(struct drm_i915_private *dev_priv,
2449 const struct intel_context *ctx)
2451 unsigned long elapsed;
2453 elapsed = get_seconds() - ctx->hang_stats.guilty_ts;
2455 if (ctx->hang_stats.banned)
2458 if (ctx->hang_stats.ban_period_seconds &&
2459 elapsed <= ctx->hang_stats.ban_period_seconds) {
2460 if (!i915_gem_context_is_default(ctx)) {
2461 DRM_DEBUG("context hanging too fast, banning!\n");
2463 } else if (i915_stop_ring_allow_ban(dev_priv)) {
2464 if (i915_stop_ring_allow_warn(dev_priv))
2465 DRM_ERROR("gpu hanging too fast, banning!\n");
2473 static void i915_set_reset_status(struct drm_i915_private *dev_priv,
2474 struct intel_context *ctx,
2477 struct i915_ctx_hang_stats *hs;
2482 hs = &ctx->hang_stats;
2485 hs->banned = i915_context_is_banned(dev_priv, ctx);
2487 hs->guilty_ts = get_seconds();
2489 hs->batch_pending++;
2493 static void i915_gem_free_request(struct drm_i915_gem_request *request)
2495 list_del(&request->list);
2496 i915_gem_request_remove_from_client(request);
2498 put_pid(request->pid);
2500 i915_gem_request_unreference(request);
2503 void i915_gem_request_free(struct kref *req_ref)
2505 struct drm_i915_gem_request *req = container_of(req_ref,
2507 struct intel_context *ctx = req->ctx;
2510 if (i915.enable_execlists) {
2511 struct intel_engine_cs *ring = req->ring;
2513 if (ctx != ring->default_context)
2514 intel_lr_context_unpin(ring, ctx);
2517 i915_gem_context_unreference(ctx);
2523 struct drm_i915_gem_request *
2524 i915_gem_find_active_request(struct intel_engine_cs *ring)
2526 struct drm_i915_gem_request *request;
2528 list_for_each_entry(request, &ring->request_list, list) {
2529 if (i915_gem_request_completed(request, false))
2538 static void i915_gem_reset_ring_status(struct drm_i915_private *dev_priv,
2539 struct intel_engine_cs *ring)
2541 struct drm_i915_gem_request *request;
2544 request = i915_gem_find_active_request(ring);
2546 if (request == NULL)
2549 ring_hung = ring->hangcheck.score >= HANGCHECK_SCORE_RING_HUNG;
2551 i915_set_reset_status(dev_priv, request->ctx, ring_hung);
2553 list_for_each_entry_continue(request, &ring->request_list, list)
2554 i915_set_reset_status(dev_priv, request->ctx, false);
2557 static void i915_gem_reset_ring_cleanup(struct drm_i915_private *dev_priv,
2558 struct intel_engine_cs *ring)
2560 while (!list_empty(&ring->active_list)) {
2561 struct drm_i915_gem_object *obj;
2563 obj = list_first_entry(&ring->active_list,
2564 struct drm_i915_gem_object,
2567 i915_gem_object_move_to_inactive(obj);
2571 * Clear the execlists queue up before freeing the requests, as those
2572 * are the ones that keep the context and ringbuffer backing objects
2575 while (!list_empty(&ring->execlist_queue)) {
2576 struct drm_i915_gem_request *submit_req;
2578 submit_req = list_first_entry(&ring->execlist_queue,
2579 struct drm_i915_gem_request,
2581 list_del(&submit_req->execlist_link);
2582 intel_runtime_pm_put(dev_priv);
2584 if (submit_req->ctx != ring->default_context)
2585 intel_lr_context_unpin(ring, submit_req->ctx);
2587 i915_gem_request_unreference(submit_req);
2591 * We must free the requests after all the corresponding objects have
2592 * been moved off active lists. Which is the same order as the normal
2593 * retire_requests function does. This is important if object hold
2594 * implicit references on things like e.g. ppgtt address spaces through
2597 while (!list_empty(&ring->request_list)) {
2598 struct drm_i915_gem_request *request;
2600 request = list_first_entry(&ring->request_list,
2601 struct drm_i915_gem_request,
2604 i915_gem_free_request(request);
2607 /* This may not have been flushed before the reset, so clean it now */
2608 i915_gem_request_assign(&ring->outstanding_lazy_request, NULL);
2611 void i915_gem_restore_fences(struct drm_device *dev)
2613 struct drm_i915_private *dev_priv = dev->dev_private;
2616 for (i = 0; i < dev_priv->num_fence_regs; i++) {
2617 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
2620 * Commit delayed tiling changes if we have an object still
2621 * attached to the fence, otherwise just clear the fence.
2624 i915_gem_object_update_fence(reg->obj, reg,
2625 reg->obj->tiling_mode);
2627 i915_gem_write_fence(dev, i, NULL);
2632 void i915_gem_reset(struct drm_device *dev)
2634 struct drm_i915_private *dev_priv = dev->dev_private;
2635 struct intel_engine_cs *ring;
2639 * Before we free the objects from the requests, we need to inspect
2640 * them for finding the guilty party. As the requests only borrow
2641 * their reference to the objects, the inspection must be done first.
2643 for_each_ring(ring, dev_priv, i)
2644 i915_gem_reset_ring_status(dev_priv, ring);
2646 for_each_ring(ring, dev_priv, i)
2647 i915_gem_reset_ring_cleanup(dev_priv, ring);
2649 i915_gem_context_reset(dev);
2651 i915_gem_restore_fences(dev);
2655 * This function clears the request list as sequence numbers are passed.
2658 i915_gem_retire_requests_ring(struct intel_engine_cs *ring)
2660 if (list_empty(&ring->request_list))
2663 WARN_ON(i915_verify_lists(ring->dev));
2665 /* Retire requests first as we use it above for the early return.
2666 * If we retire requests last, we may use a later seqno and so clear
2667 * the requests lists without clearing the active list, leading to
2670 while (!list_empty(&ring->request_list)) {
2671 struct drm_i915_gem_request *request;
2673 request = list_first_entry(&ring->request_list,
2674 struct drm_i915_gem_request,
2677 if (!i915_gem_request_completed(request, true))
2680 trace_i915_gem_request_retire(request);
2682 /* We know the GPU must have read the request to have
2683 * sent us the seqno + interrupt, so use the position
2684 * of tail of the request to update the last known position
2687 request->ringbuf->last_retired_head = request->postfix;
2689 i915_gem_free_request(request);
2692 /* Move any buffers on the active list that are no longer referenced
2693 * by the ringbuffer to the flushing/inactive lists as appropriate,
2694 * before we free the context associated with the requests.
2696 while (!list_empty(&ring->active_list)) {
2697 struct drm_i915_gem_object *obj;
2699 obj = list_first_entry(&ring->active_list,
2700 struct drm_i915_gem_object,
2703 if (!i915_gem_request_completed(obj->last_read_req, true))
2706 i915_gem_object_move_to_inactive(obj);
2709 if (unlikely(ring->trace_irq_req &&
2710 i915_gem_request_completed(ring->trace_irq_req, true))) {
2711 ring->irq_put(ring);
2712 i915_gem_request_assign(&ring->trace_irq_req, NULL);
2715 WARN_ON(i915_verify_lists(ring->dev));
2719 i915_gem_retire_requests(struct drm_device *dev)
2721 struct drm_i915_private *dev_priv = dev->dev_private;
2722 struct intel_engine_cs *ring;
2726 for_each_ring(ring, dev_priv, i) {
2727 i915_gem_retire_requests_ring(ring);
2728 idle &= list_empty(&ring->request_list);
2729 if (i915.enable_execlists) {
2730 unsigned long flags;
2732 spin_lock_irqsave(&ring->execlist_lock, flags);
2733 idle &= list_empty(&ring->execlist_queue);
2734 spin_unlock_irqrestore(&ring->execlist_lock, flags);
2736 intel_execlists_retire_requests(ring);
2741 mod_delayed_work(dev_priv->wq,
2742 &dev_priv->mm.idle_work,
2743 msecs_to_jiffies(100));
2749 i915_gem_retire_work_handler(struct work_struct *work)
2751 struct drm_i915_private *dev_priv =
2752 container_of(work, typeof(*dev_priv), mm.retire_work.work);
2753 struct drm_device *dev = dev_priv->dev;
2756 /* Come back later if the device is busy... */
2758 if (mutex_trylock(&dev->struct_mutex)) {
2759 idle = i915_gem_retire_requests(dev);
2760 mutex_unlock(&dev->struct_mutex);
2763 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
2764 round_jiffies_up_relative(HZ));
2768 i915_gem_idle_work_handler(struct work_struct *work)
2770 struct drm_i915_private *dev_priv =
2771 container_of(work, typeof(*dev_priv), mm.idle_work.work);
2773 intel_mark_idle(dev_priv->dev);
2777 * Ensures that an object will eventually get non-busy by flushing any required
2778 * write domains, emitting any outstanding lazy request and retiring and
2779 * completed requests.
2782 i915_gem_object_flush_active(struct drm_i915_gem_object *obj)
2784 struct intel_engine_cs *ring;
2788 ring = i915_gem_request_get_ring(obj->last_read_req);
2790 ret = i915_gem_check_olr(obj->last_read_req);
2794 i915_gem_retire_requests_ring(ring);
2801 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
2802 * @DRM_IOCTL_ARGS: standard ioctl arguments
2804 * Returns 0 if successful, else an error is returned with the remaining time in
2805 * the timeout parameter.
2806 * -ETIME: object is still busy after timeout
2807 * -ERESTARTSYS: signal interrupted the wait
2808 * -ENONENT: object doesn't exist
2809 * Also possible, but rare:
2810 * -EAGAIN: GPU wedged
2812 * -ENODEV: Internal IRQ fail
2813 * -E?: The add request failed
2815 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
2816 * non-zero timeout parameter the wait ioctl will wait for the given number of
2817 * nanoseconds on an object becoming unbusy. Since the wait itself does so
2818 * without holding struct_mutex the object may become re-busied before this
2819 * function completes. A similar but shorter * race condition exists in the busy
2823 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
2825 struct drm_i915_private *dev_priv = dev->dev_private;
2826 struct drm_i915_gem_wait *args = data;
2827 struct drm_i915_gem_object *obj;
2828 struct drm_i915_gem_request *req;
2829 unsigned reset_counter;
2832 if (args->flags != 0)
2835 ret = i915_mutex_lock_interruptible(dev);
2839 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle));
2840 if (&obj->base == NULL) {
2841 mutex_unlock(&dev->struct_mutex);
2845 /* Need to make sure the object gets inactive eventually. */
2846 ret = i915_gem_object_flush_active(obj);
2850 if (!obj->active || !obj->last_read_req)
2853 req = obj->last_read_req;
2855 /* Do this after OLR check to make sure we make forward progress polling
2856 * on this IOCTL with a timeout == 0 (like busy ioctl)
2858 if (args->timeout_ns == 0) {
2863 drm_gem_object_unreference(&obj->base);
2864 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
2865 i915_gem_request_reference(req);
2866 mutex_unlock(&dev->struct_mutex);
2868 ret = __i915_wait_request(req, reset_counter, true,
2869 args->timeout_ns > 0 ? &args->timeout_ns : NULL,
2871 mutex_lock(&dev->struct_mutex);
2872 i915_gem_request_unreference(req);
2873 mutex_unlock(&dev->struct_mutex);
2877 drm_gem_object_unreference(&obj->base);
2878 mutex_unlock(&dev->struct_mutex);
2883 * i915_gem_object_sync - sync an object to a ring.
2885 * @obj: object which may be in use on another ring.
2886 * @to: ring we wish to use the object on. May be NULL.
2888 * This code is meant to abstract object synchronization with the GPU.
2889 * Calling with NULL implies synchronizing the object with the CPU
2890 * rather than a particular GPU ring.
2892 * Returns 0 if successful, else propagates up the lower layer error.
2895 i915_gem_object_sync(struct drm_i915_gem_object *obj,
2896 struct intel_engine_cs *to)
2898 struct intel_engine_cs *from;
2902 from = i915_gem_request_get_ring(obj->last_read_req);
2904 if (from == NULL || to == from)
2907 if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev))
2908 return i915_gem_object_wait_rendering(obj, false);
2910 idx = intel_ring_sync_index(from, to);
2912 seqno = i915_gem_request_get_seqno(obj->last_read_req);
2913 /* Optimization: Avoid semaphore sync when we are sure we already
2914 * waited for an object with higher seqno */
2915 if (seqno <= from->semaphore.sync_seqno[idx])
2918 ret = i915_gem_check_olr(obj->last_read_req);
2922 trace_i915_gem_ring_sync_to(from, to, obj->last_read_req);
2923 ret = to->semaphore.sync_to(to, from, seqno);
2925 /* We use last_read_req because sync_to()
2926 * might have just caused seqno wrap under
2929 from->semaphore.sync_seqno[idx] =
2930 i915_gem_request_get_seqno(obj->last_read_req);
2935 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
2937 u32 old_write_domain, old_read_domains;
2939 /* Force a pagefault for domain tracking on next user access */
2940 i915_gem_release_mmap(obj);
2942 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
2945 /* Wait for any direct GTT access to complete */
2948 old_read_domains = obj->base.read_domains;
2949 old_write_domain = obj->base.write_domain;
2951 obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
2952 obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
2954 trace_i915_gem_object_change_domain(obj,
2959 int i915_vma_unbind(struct i915_vma *vma)
2961 struct drm_i915_gem_object *obj = vma->obj;
2962 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2965 if (list_empty(&vma->vma_link))
2968 if (!drm_mm_node_allocated(&vma->node)) {
2969 i915_gem_vma_destroy(vma);
2976 BUG_ON(obj->pages == NULL);
2978 ret = i915_gem_object_finish_gpu(obj);
2981 /* Continue on if we fail due to EIO, the GPU is hung so we
2982 * should be safe and we need to cleanup or else we might
2983 * cause memory corruption through use-after-free.
2986 if (i915_is_ggtt(vma->vm) &&
2987 vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL) {
2988 i915_gem_object_finish_gtt(obj);
2990 /* release the fence reg _after_ flushing */
2991 ret = i915_gem_object_put_fence(obj);
2996 trace_i915_vma_unbind(vma);
2998 vma->unbind_vma(vma);
3000 list_del_init(&vma->mm_list);
3001 if (i915_is_ggtt(vma->vm)) {
3002 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL) {
3003 obj->map_and_fenceable = false;
3004 } else if (vma->ggtt_view.pages) {
3005 sg_free_table(vma->ggtt_view.pages);
3006 kfree(vma->ggtt_view.pages);
3008 vma->ggtt_view.pages = NULL;
3011 drm_mm_remove_node(&vma->node);
3012 i915_gem_vma_destroy(vma);
3014 /* Since the unbound list is global, only move to that list if
3015 * no more VMAs exist. */
3016 if (list_empty(&obj->vma_list)) {
3017 /* Throw away the active reference before
3018 * moving to the unbound list. */
3019 i915_gem_object_retire(obj);
3021 i915_gem_gtt_finish_object(obj);
3022 list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list);
3025 /* And finally now the object is completely decoupled from this vma,
3026 * we can drop its hold on the backing storage and allow it to be
3027 * reaped by the shrinker.
3029 i915_gem_object_unpin_pages(obj);
3034 int i915_gpu_idle(struct drm_device *dev)
3036 struct drm_i915_private *dev_priv = dev->dev_private;
3037 struct intel_engine_cs *ring;
3040 /* Flush everything onto the inactive list. */
3041 for_each_ring(ring, dev_priv, i) {
3042 if (!i915.enable_execlists) {
3043 ret = i915_switch_context(ring, ring->default_context);
3048 ret = intel_ring_idle(ring);
3056 static void i965_write_fence_reg(struct drm_device *dev, int reg,
3057 struct drm_i915_gem_object *obj)
3059 struct drm_i915_private *dev_priv = dev->dev_private;
3061 int fence_pitch_shift;
3063 if (INTEL_INFO(dev)->gen >= 6) {
3064 fence_reg = FENCE_REG_SANDYBRIDGE_0;
3065 fence_pitch_shift = SANDYBRIDGE_FENCE_PITCH_SHIFT;
3067 fence_reg = FENCE_REG_965_0;
3068 fence_pitch_shift = I965_FENCE_PITCH_SHIFT;
3071 fence_reg += reg * 8;
3073 /* To w/a incoherency with non-atomic 64-bit register updates,
3074 * we split the 64-bit update into two 32-bit writes. In order
3075 * for a partial fence not to be evaluated between writes, we
3076 * precede the update with write to turn off the fence register,
3077 * and only enable the fence as the last step.
3079 * For extra levels of paranoia, we make sure each step lands
3080 * before applying the next step.
3082 I915_WRITE(fence_reg, 0);
3083 POSTING_READ(fence_reg);
3086 u32 size = i915_gem_obj_ggtt_size(obj);
3089 /* Adjust fence size to match tiled area */
3090 if (obj->tiling_mode != I915_TILING_NONE) {
3091 uint32_t row_size = obj->stride *
3092 (obj->tiling_mode == I915_TILING_Y ? 32 : 8);
3093 size = (size / row_size) * row_size;
3096 val = (uint64_t)((i915_gem_obj_ggtt_offset(obj) + size - 4096) &
3098 val |= i915_gem_obj_ggtt_offset(obj) & 0xfffff000;
3099 val |= (uint64_t)((obj->stride / 128) - 1) << fence_pitch_shift;
3100 if (obj->tiling_mode == I915_TILING_Y)
3101 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
3102 val |= I965_FENCE_REG_VALID;
3104 I915_WRITE(fence_reg + 4, val >> 32);
3105 POSTING_READ(fence_reg + 4);
3107 I915_WRITE(fence_reg + 0, val);
3108 POSTING_READ(fence_reg);
3110 I915_WRITE(fence_reg + 4, 0);
3111 POSTING_READ(fence_reg + 4);
3115 static void i915_write_fence_reg(struct drm_device *dev, int reg,
3116 struct drm_i915_gem_object *obj)
3118 struct drm_i915_private *dev_priv = dev->dev_private;
3122 u32 size = i915_gem_obj_ggtt_size(obj);
3126 WARN((i915_gem_obj_ggtt_offset(obj) & ~I915_FENCE_START_MASK) ||
3127 (size & -size) != size ||
3128 (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
3129 "object 0x%08lx [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
3130 i915_gem_obj_ggtt_offset(obj), obj->map_and_fenceable, size);
3132 if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
3137 /* Note: pitch better be a power of two tile widths */
3138 pitch_val = obj->stride / tile_width;
3139 pitch_val = ffs(pitch_val) - 1;
3141 val = i915_gem_obj_ggtt_offset(obj);
3142 if (obj->tiling_mode == I915_TILING_Y)
3143 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
3144 val |= I915_FENCE_SIZE_BITS(size);
3145 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
3146 val |= I830_FENCE_REG_VALID;
3151 reg = FENCE_REG_830_0 + reg * 4;
3153 reg = FENCE_REG_945_8 + (reg - 8) * 4;
3155 I915_WRITE(reg, val);
3159 static void i830_write_fence_reg(struct drm_device *dev, int reg,
3160 struct drm_i915_gem_object *obj)
3162 struct drm_i915_private *dev_priv = dev->dev_private;
3166 u32 size = i915_gem_obj_ggtt_size(obj);
3169 WARN((i915_gem_obj_ggtt_offset(obj) & ~I830_FENCE_START_MASK) ||
3170 (size & -size) != size ||
3171 (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
3172 "object 0x%08lx not 512K or pot-size 0x%08x aligned\n",
3173 i915_gem_obj_ggtt_offset(obj), size);
3175 pitch_val = obj->stride / 128;
3176 pitch_val = ffs(pitch_val) - 1;
3178 val = i915_gem_obj_ggtt_offset(obj);
3179 if (obj->tiling_mode == I915_TILING_Y)
3180 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
3181 val |= I830_FENCE_SIZE_BITS(size);
3182 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
3183 val |= I830_FENCE_REG_VALID;
3187 I915_WRITE(FENCE_REG_830_0 + reg * 4, val);
3188 POSTING_READ(FENCE_REG_830_0 + reg * 4);
3191 inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object *obj)
3193 return obj && obj->base.read_domains & I915_GEM_DOMAIN_GTT;
3196 static void i915_gem_write_fence(struct drm_device *dev, int reg,
3197 struct drm_i915_gem_object *obj)
3199 struct drm_i915_private *dev_priv = dev->dev_private;
3201 /* Ensure that all CPU reads are completed before installing a fence
3202 * and all writes before removing the fence.
3204 if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj))
3207 WARN(obj && (!obj->stride || !obj->tiling_mode),
3208 "bogus fence setup with stride: 0x%x, tiling mode: %i\n",
3209 obj->stride, obj->tiling_mode);
3212 i830_write_fence_reg(dev, reg, obj);
3213 else if (IS_GEN3(dev))
3214 i915_write_fence_reg(dev, reg, obj);
3215 else if (INTEL_INFO(dev)->gen >= 4)
3216 i965_write_fence_reg(dev, reg, obj);
3218 /* And similarly be paranoid that no direct access to this region
3219 * is reordered to before the fence is installed.
3221 if (i915_gem_object_needs_mb(obj))
3225 static inline int fence_number(struct drm_i915_private *dev_priv,
3226 struct drm_i915_fence_reg *fence)
3228 return fence - dev_priv->fence_regs;
3231 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
3232 struct drm_i915_fence_reg *fence,
3235 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3236 int reg = fence_number(dev_priv, fence);
3238 i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL);
3241 obj->fence_reg = reg;
3243 list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list);
3245 obj->fence_reg = I915_FENCE_REG_NONE;
3247 list_del_init(&fence->lru_list);
3249 obj->fence_dirty = false;
3253 i915_gem_object_wait_fence(struct drm_i915_gem_object *obj)
3255 if (obj->last_fenced_req) {
3256 int ret = i915_wait_request(obj->last_fenced_req);
3260 i915_gem_request_assign(&obj->last_fenced_req, NULL);
3267 i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
3269 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3270 struct drm_i915_fence_reg *fence;
3273 ret = i915_gem_object_wait_fence(obj);
3277 if (obj->fence_reg == I915_FENCE_REG_NONE)
3280 fence = &dev_priv->fence_regs[obj->fence_reg];
3282 if (WARN_ON(fence->pin_count))
3285 i915_gem_object_fence_lost(obj);
3286 i915_gem_object_update_fence(obj, fence, false);
3291 static struct drm_i915_fence_reg *
3292 i915_find_fence_reg(struct drm_device *dev)
3294 struct drm_i915_private *dev_priv = dev->dev_private;
3295 struct drm_i915_fence_reg *reg, *avail;
3298 /* First try to find a free reg */
3300 for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
3301 reg = &dev_priv->fence_regs[i];
3305 if (!reg->pin_count)
3312 /* None available, try to steal one or wait for a user to finish */
3313 list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
3321 /* Wait for completion of pending flips which consume fences */
3322 if (intel_has_pending_fb_unpin(dev))
3323 return ERR_PTR(-EAGAIN);
3325 return ERR_PTR(-EDEADLK);
3329 * i915_gem_object_get_fence - set up fencing for an object
3330 * @obj: object to map through a fence reg
3332 * When mapping objects through the GTT, userspace wants to be able to write
3333 * to them without having to worry about swizzling if the object is tiled.
3334 * This function walks the fence regs looking for a free one for @obj,
3335 * stealing one if it can't find any.
3337 * It then sets up the reg based on the object's properties: address, pitch
3338 * and tiling format.
3340 * For an untiled surface, this removes any existing fence.
3343 i915_gem_object_get_fence(struct drm_i915_gem_object *obj)
3345 struct drm_device *dev = obj->base.dev;
3346 struct drm_i915_private *dev_priv = dev->dev_private;
3347 bool enable = obj->tiling_mode != I915_TILING_NONE;
3348 struct drm_i915_fence_reg *reg;
3351 /* Have we updated the tiling parameters upon the object and so
3352 * will need to serialise the write to the associated fence register?
3354 if (obj->fence_dirty) {
3355 ret = i915_gem_object_wait_fence(obj);
3360 /* Just update our place in the LRU if our fence is getting reused. */
3361 if (obj->fence_reg != I915_FENCE_REG_NONE) {
3362 reg = &dev_priv->fence_regs[obj->fence_reg];
3363 if (!obj->fence_dirty) {
3364 list_move_tail(®->lru_list,
3365 &dev_priv->mm.fence_list);
3368 } else if (enable) {
3369 if (WARN_ON(!obj->map_and_fenceable))
3372 reg = i915_find_fence_reg(dev);
3374 return PTR_ERR(reg);
3377 struct drm_i915_gem_object *old = reg->obj;
3379 ret = i915_gem_object_wait_fence(old);
3383 i915_gem_object_fence_lost(old);
3388 i915_gem_object_update_fence(obj, reg, enable);
3393 static bool i915_gem_valid_gtt_space(struct i915_vma *vma,
3394 unsigned long cache_level)
3396 struct drm_mm_node *gtt_space = &vma->node;
3397 struct drm_mm_node *other;
3400 * On some machines we have to be careful when putting differing types
3401 * of snoopable memory together to avoid the prefetcher crossing memory
3402 * domains and dying. During vm initialisation, we decide whether or not
3403 * these constraints apply and set the drm_mm.color_adjust
3406 if (vma->vm->mm.color_adjust == NULL)
3409 if (!drm_mm_node_allocated(gtt_space))
3412 if (list_empty(>t_space->node_list))
3415 other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list);
3416 if (other->allocated && !other->hole_follows && other->color != cache_level)
3419 other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list);
3420 if (other->allocated && !gtt_space->hole_follows && other->color != cache_level)
3427 * Finds free space in the GTT aperture and binds the object there.
3429 static struct i915_vma *
3430 i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj,
3431 struct i915_address_space *vm,
3432 const struct i915_ggtt_view *ggtt_view,
3436 struct drm_device *dev = obj->base.dev;
3437 struct drm_i915_private *dev_priv = dev->dev_private;
3438 u32 size, fence_size, fence_alignment, unfenced_alignment;
3439 unsigned long start =
3440 flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
3442 flags & PIN_MAPPABLE ? dev_priv->gtt.mappable_end : vm->total;
3443 struct i915_vma *vma;
3446 if(WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
3447 return ERR_PTR(-EINVAL);
3449 fence_size = i915_gem_get_gtt_size(dev,
3452 fence_alignment = i915_gem_get_gtt_alignment(dev,
3454 obj->tiling_mode, true);
3455 unfenced_alignment =
3456 i915_gem_get_gtt_alignment(dev,
3458 obj->tiling_mode, false);
3461 alignment = flags & PIN_MAPPABLE ? fence_alignment :
3463 if (flags & PIN_MAPPABLE && alignment & (fence_alignment - 1)) {
3464 DRM_DEBUG("Invalid object alignment requested %u\n", alignment);
3465 return ERR_PTR(-EINVAL);
3468 size = flags & PIN_MAPPABLE ? fence_size : obj->base.size;
3470 /* If the object is bigger than the entire aperture, reject it early
3471 * before evicting everything in a vain attempt to find space.
3473 if (obj->base.size > end) {
3474 DRM_DEBUG("Attempting to bind an object larger than the aperture: object=%zd > %s aperture=%lu\n",
3476 flags & PIN_MAPPABLE ? "mappable" : "total",
3478 return ERR_PTR(-E2BIG);
3481 ret = i915_gem_object_get_pages(obj);
3483 return ERR_PTR(ret);
3485 i915_gem_object_pin_pages(obj);
3487 vma = ggtt_view ? i915_gem_obj_lookup_or_create_ggtt_vma(obj, ggtt_view) :
3488 i915_gem_obj_lookup_or_create_vma(obj, vm);
3494 ret = drm_mm_insert_node_in_range_generic(&vm->mm, &vma->node,
3498 DRM_MM_SEARCH_DEFAULT,
3499 DRM_MM_CREATE_DEFAULT);
3501 ret = i915_gem_evict_something(dev, vm, size, alignment,
3510 if (WARN_ON(!i915_gem_valid_gtt_space(vma, obj->cache_level))) {
3512 goto err_remove_node;
3515 ret = i915_gem_gtt_prepare_object(obj);
3517 goto err_remove_node;
3519 /* allocate before insert / bind */
3520 if (vma->vm->allocate_va_range) {
3521 trace_i915_va_alloc(vma->vm, vma->node.start, vma->node.size,
3522 VM_TO_TRACE_NAME(vma->vm));
3523 ret = vma->vm->allocate_va_range(vma->vm,
3527 goto err_remove_node;
3530 trace_i915_vma_bind(vma, flags);
3531 ret = i915_vma_bind(vma, obj->cache_level,
3532 flags & PIN_GLOBAL ? GLOBAL_BIND : 0);
3534 goto err_finish_gtt;
3536 list_move_tail(&obj->global_list, &dev_priv->mm.bound_list);
3537 list_add_tail(&vma->mm_list, &vm->inactive_list);
3542 i915_gem_gtt_finish_object(obj);
3544 drm_mm_remove_node(&vma->node);
3546 i915_gem_vma_destroy(vma);
3549 i915_gem_object_unpin_pages(obj);
3554 i915_gem_clflush_object(struct drm_i915_gem_object *obj,
3557 /* If we don't have a page list set up, then we're not pinned
3558 * to GPU, and we can ignore the cache flush because it'll happen
3559 * again at bind time.
3561 if (obj->pages == NULL)
3565 * Stolen memory is always coherent with the GPU as it is explicitly
3566 * marked as wc by the system, or the system is cache-coherent.
3568 if (obj->stolen || obj->phys_handle)
3571 /* If the GPU is snooping the contents of the CPU cache,
3572 * we do not need to manually clear the CPU cache lines. However,
3573 * the caches are only snooped when the render cache is
3574 * flushed/invalidated. As we always have to emit invalidations
3575 * and flushes when moving into and out of the RENDER domain, correct
3576 * snooping behaviour occurs naturally as the result of our domain
3579 if (!force && cpu_cache_is_coherent(obj->base.dev, obj->cache_level)) {
3580 obj->cache_dirty = true;
3584 trace_i915_gem_object_clflush(obj);
3585 drm_clflush_sg(obj->pages);
3586 obj->cache_dirty = false;
3591 /** Flushes the GTT write domain for the object if it's dirty. */
3593 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
3595 uint32_t old_write_domain;
3597 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
3600 /* No actual flushing is required for the GTT write domain. Writes
3601 * to it immediately go to main memory as far as we know, so there's
3602 * no chipset flush. It also doesn't land in render cache.
3604 * However, we do have to enforce the order so that all writes through
3605 * the GTT land before any writes to the device, such as updates to
3610 old_write_domain = obj->base.write_domain;
3611 obj->base.write_domain = 0;
3613 intel_fb_obj_flush(obj, false);
3615 trace_i915_gem_object_change_domain(obj,
3616 obj->base.read_domains,
3620 /** Flushes the CPU write domain for the object if it's dirty. */
3622 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj)
3624 uint32_t old_write_domain;
3626 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
3629 if (i915_gem_clflush_object(obj, obj->pin_display))
3630 i915_gem_chipset_flush(obj->base.dev);
3632 old_write_domain = obj->base.write_domain;
3633 obj->base.write_domain = 0;
3635 intel_fb_obj_flush(obj, false);
3637 trace_i915_gem_object_change_domain(obj,
3638 obj->base.read_domains,
3643 * Moves a single object to the GTT read, and possibly write domain.
3645 * This function returns when the move is complete, including waiting on
3649 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
3651 uint32_t old_write_domain, old_read_domains;
3652 struct i915_vma *vma;
3655 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
3658 ret = i915_gem_object_wait_rendering(obj, !write);
3662 i915_gem_object_retire(obj);
3664 /* Flush and acquire obj->pages so that we are coherent through
3665 * direct access in memory with previous cached writes through
3666 * shmemfs and that our cache domain tracking remains valid.
3667 * For example, if the obj->filp was moved to swap without us
3668 * being notified and releasing the pages, we would mistakenly
3669 * continue to assume that the obj remained out of the CPU cached
3672 ret = i915_gem_object_get_pages(obj);
3676 i915_gem_object_flush_cpu_write_domain(obj);
3678 /* Serialise direct access to this object with the barriers for
3679 * coherent writes from the GPU, by effectively invalidating the
3680 * GTT domain upon first access.
3682 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
3685 old_write_domain = obj->base.write_domain;
3686 old_read_domains = obj->base.read_domains;
3688 /* It should now be out of any other write domains, and we can update
3689 * the domain values for our changes.
3691 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3692 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3694 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
3695 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
3700 intel_fb_obj_invalidate(obj, NULL, ORIGIN_GTT);
3702 trace_i915_gem_object_change_domain(obj,
3706 /* And bump the LRU for this access */
3707 vma = i915_gem_obj_to_ggtt(obj);
3708 if (vma && drm_mm_node_allocated(&vma->node) && !obj->active)
3709 list_move_tail(&vma->mm_list,
3710 &to_i915(obj->base.dev)->gtt.base.inactive_list);
3715 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
3716 enum i915_cache_level cache_level)
3718 struct drm_device *dev = obj->base.dev;
3719 struct i915_vma *vma, *next;
3722 if (obj->cache_level == cache_level)
3725 if (i915_gem_obj_is_pinned(obj)) {
3726 DRM_DEBUG("can not change the cache level of pinned objects\n");
3730 list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) {
3731 if (!i915_gem_valid_gtt_space(vma, cache_level)) {
3732 ret = i915_vma_unbind(vma);
3738 if (i915_gem_obj_bound_any(obj)) {
3739 ret = i915_gem_object_finish_gpu(obj);
3743 i915_gem_object_finish_gtt(obj);
3745 /* Before SandyBridge, you could not use tiling or fence
3746 * registers with snooped memory, so relinquish any fences
3747 * currently pointing to our region in the aperture.
3749 if (INTEL_INFO(dev)->gen < 6) {
3750 ret = i915_gem_object_put_fence(obj);
3755 list_for_each_entry(vma, &obj->vma_list, vma_link)
3756 if (drm_mm_node_allocated(&vma->node)) {
3757 ret = i915_vma_bind(vma, cache_level,
3758 vma->bound & GLOBAL_BIND);
3764 list_for_each_entry(vma, &obj->vma_list, vma_link)
3765 vma->node.color = cache_level;
3766 obj->cache_level = cache_level;
3768 if (obj->cache_dirty &&
3769 obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
3770 cpu_write_needs_clflush(obj)) {
3771 if (i915_gem_clflush_object(obj, true))
3772 i915_gem_chipset_flush(obj->base.dev);
3778 int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data,
3779 struct drm_file *file)
3781 struct drm_i915_gem_caching *args = data;
3782 struct drm_i915_gem_object *obj;
3785 ret = i915_mutex_lock_interruptible(dev);
3789 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3790 if (&obj->base == NULL) {
3795 switch (obj->cache_level) {
3796 case I915_CACHE_LLC:
3797 case I915_CACHE_L3_LLC:
3798 args->caching = I915_CACHING_CACHED;
3802 args->caching = I915_CACHING_DISPLAY;
3806 args->caching = I915_CACHING_NONE;
3810 drm_gem_object_unreference(&obj->base);
3812 mutex_unlock(&dev->struct_mutex);
3816 int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
3817 struct drm_file *file)
3819 struct drm_i915_gem_caching *args = data;
3820 struct drm_i915_gem_object *obj;
3821 enum i915_cache_level level;
3824 switch (args->caching) {
3825 case I915_CACHING_NONE:
3826 level = I915_CACHE_NONE;
3828 case I915_CACHING_CACHED:
3829 level = I915_CACHE_LLC;
3831 case I915_CACHING_DISPLAY:
3832 level = HAS_WT(dev) ? I915_CACHE_WT : I915_CACHE_NONE;
3838 ret = i915_mutex_lock_interruptible(dev);
3842 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3843 if (&obj->base == NULL) {
3848 ret = i915_gem_object_set_cache_level(obj, level);
3850 drm_gem_object_unreference(&obj->base);
3852 mutex_unlock(&dev->struct_mutex);
3856 static bool is_pin_display(struct drm_i915_gem_object *obj)
3858 struct i915_vma *vma;
3860 vma = i915_gem_obj_to_ggtt(obj);
3864 /* There are 2 sources that pin objects:
3865 * 1. The display engine (scanouts, sprites, cursors);
3866 * 2. Reservations for execbuffer;
3868 * We can ignore reservations as we hold the struct_mutex and
3869 * are only called outside of the reservation path.
3871 return vma->pin_count;
3875 * Prepare buffer for display plane (scanout, cursors, etc).
3876 * Can be called from an uninterruptible phase (modesetting) and allows
3877 * any flushes to be pipelined (for pageflips).
3880 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
3882 struct intel_engine_cs *pipelined,
3883 const struct i915_ggtt_view *view)
3885 u32 old_read_domains, old_write_domain;
3886 bool was_pin_display;
3889 if (pipelined != i915_gem_request_get_ring(obj->last_read_req)) {
3890 ret = i915_gem_object_sync(obj, pipelined);
3895 /* Mark the pin_display early so that we account for the
3896 * display coherency whilst setting up the cache domains.
3898 was_pin_display = obj->pin_display;
3899 obj->pin_display = true;
3901 /* The display engine is not coherent with the LLC cache on gen6. As
3902 * a result, we make sure that the pinning that is about to occur is
3903 * done with uncached PTEs. This is lowest common denominator for all
3906 * However for gen6+, we could do better by using the GFDT bit instead
3907 * of uncaching, which would allow us to flush all the LLC-cached data
3908 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3910 ret = i915_gem_object_set_cache_level(obj,
3911 HAS_WT(obj->base.dev) ? I915_CACHE_WT : I915_CACHE_NONE);
3913 goto err_unpin_display;
3915 /* As the user may map the buffer once pinned in the display plane
3916 * (e.g. libkms for the bootup splash), we have to ensure that we
3917 * always use map_and_fenceable for all scanout buffers.
3919 ret = i915_gem_object_ggtt_pin(obj, view, alignment,
3920 view->type == I915_GGTT_VIEW_NORMAL ?
3923 goto err_unpin_display;
3925 i915_gem_object_flush_cpu_write_domain(obj);
3927 old_write_domain = obj->base.write_domain;
3928 old_read_domains = obj->base.read_domains;
3930 /* It should now be out of any other write domains, and we can update
3931 * the domain values for our changes.
3933 obj->base.write_domain = 0;
3934 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3936 trace_i915_gem_object_change_domain(obj,
3943 WARN_ON(was_pin_display != is_pin_display(obj));
3944 obj->pin_display = was_pin_display;
3949 i915_gem_object_unpin_from_display_plane(struct drm_i915_gem_object *obj,
3950 const struct i915_ggtt_view *view)
3952 i915_gem_object_ggtt_unpin_view(obj, view);
3954 obj->pin_display = is_pin_display(obj);
3958 i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
3962 if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
3965 ret = i915_gem_object_wait_rendering(obj, false);
3969 /* Ensure that we invalidate the GPU's caches and TLBs. */
3970 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
3975 * Moves a single object to the CPU read, and possibly write domain.
3977 * This function returns when the move is complete, including waiting on
3981 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
3983 uint32_t old_write_domain, old_read_domains;
3986 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
3989 ret = i915_gem_object_wait_rendering(obj, !write);
3993 i915_gem_object_retire(obj);
3994 i915_gem_object_flush_gtt_write_domain(obj);
3996 old_write_domain = obj->base.write_domain;
3997 old_read_domains = obj->base.read_domains;
3999 /* Flush the CPU cache if it's still invalid. */
4000 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
4001 i915_gem_clflush_object(obj, false);
4003 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
4006 /* It should now be out of any other write domains, and we can update
4007 * the domain values for our changes.
4009 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
4011 /* If we're writing through the CPU, then the GPU read domains will
4012 * need to be invalidated at next use.
4015 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4016 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4020 intel_fb_obj_invalidate(obj, NULL, ORIGIN_CPU);
4022 trace_i915_gem_object_change_domain(obj,
4029 /* Throttle our rendering by waiting until the ring has completed our requests
4030 * emitted over 20 msec ago.
4032 * Note that if we were to use the current jiffies each time around the loop,
4033 * we wouldn't escape the function with any frames outstanding if the time to
4034 * render a frame was over 20ms.
4036 * This should get us reasonable parallelism between CPU and GPU but also
4037 * relatively low latency when blocking on a particular request to finish.
4040 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
4042 struct drm_i915_private *dev_priv = dev->dev_private;
4043 struct drm_i915_file_private *file_priv = file->driver_priv;
4044 unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
4045 struct drm_i915_gem_request *request, *target = NULL;
4046 unsigned reset_counter;
4049 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
4053 ret = i915_gem_check_wedge(&dev_priv->gpu_error, false);
4057 spin_lock(&file_priv->mm.lock);
4058 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
4059 if (time_after_eq(request->emitted_jiffies, recent_enough))
4064 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
4066 i915_gem_request_reference(target);
4067 spin_unlock(&file_priv->mm.lock);
4072 ret = __i915_wait_request(target, reset_counter, true, NULL, NULL);
4074 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
4076 mutex_lock(&dev->struct_mutex);
4077 i915_gem_request_unreference(target);
4078 mutex_unlock(&dev->struct_mutex);
4084 i915_vma_misplaced(struct i915_vma *vma, uint32_t alignment, uint64_t flags)
4086 struct drm_i915_gem_object *obj = vma->obj;
4089 vma->node.start & (alignment - 1))
4092 if (flags & PIN_MAPPABLE && !obj->map_and_fenceable)
4095 if (flags & PIN_OFFSET_BIAS &&
4096 vma->node.start < (flags & PIN_OFFSET_MASK))
4103 i915_gem_object_do_pin(struct drm_i915_gem_object *obj,
4104 struct i915_address_space *vm,
4105 const struct i915_ggtt_view *ggtt_view,
4109 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
4110 struct i915_vma *vma;
4114 if (WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base))
4117 if (WARN_ON(flags & (PIN_GLOBAL | PIN_MAPPABLE) && !i915_is_ggtt(vm)))
4120 if (WARN_ON((flags & (PIN_MAPPABLE | PIN_GLOBAL)) == PIN_MAPPABLE))
4123 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
4126 vma = ggtt_view ? i915_gem_obj_to_ggtt_view(obj, ggtt_view) :
4127 i915_gem_obj_to_vma(obj, vm);
4130 return PTR_ERR(vma);
4133 if (WARN_ON(vma->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT))
4136 if (i915_vma_misplaced(vma, alignment, flags)) {
4137 unsigned long offset;
4138 offset = ggtt_view ? i915_gem_obj_ggtt_offset_view(obj, ggtt_view) :
4139 i915_gem_obj_offset(obj, vm);
4140 WARN(vma->pin_count,
4141 "bo is already pinned in %s with incorrect alignment:"
4142 " offset=%lx, req.alignment=%x, req.map_and_fenceable=%d,"
4143 " obj->map_and_fenceable=%d\n",
4144 ggtt_view ? "ggtt" : "ppgtt",
4147 !!(flags & PIN_MAPPABLE),
4148 obj->map_and_fenceable);
4149 ret = i915_vma_unbind(vma);
4157 bound = vma ? vma->bound : 0;
4158 if (vma == NULL || !drm_mm_node_allocated(&vma->node)) {
4159 /* In true PPGTT, bind has possibly changed PDEs, which
4160 * means we must do a context switch before the GPU can
4161 * accurately read some of the VMAs.
4163 vma = i915_gem_object_bind_to_vm(obj, vm, ggtt_view, alignment,
4166 return PTR_ERR(vma);
4169 if (flags & PIN_GLOBAL && !(vma->bound & GLOBAL_BIND)) {
4170 ret = i915_vma_bind(vma, obj->cache_level, GLOBAL_BIND);
4175 if ((bound ^ vma->bound) & GLOBAL_BIND) {
4176 bool mappable, fenceable;
4177 u32 fence_size, fence_alignment;
4179 fence_size = i915_gem_get_gtt_size(obj->base.dev,
4182 fence_alignment = i915_gem_get_gtt_alignment(obj->base.dev,
4187 fenceable = (vma->node.size == fence_size &&
4188 (vma->node.start & (fence_alignment - 1)) == 0);
4190 mappable = (vma->node.start + fence_size <=
4191 dev_priv->gtt.mappable_end);
4193 obj->map_and_fenceable = mappable && fenceable;
4196 WARN_ON(flags & PIN_MAPPABLE && !obj->map_and_fenceable);
4199 if (flags & PIN_MAPPABLE)
4200 obj->pin_mappable |= true;
4206 i915_gem_object_pin(struct drm_i915_gem_object *obj,
4207 struct i915_address_space *vm,
4211 return i915_gem_object_do_pin(obj, vm,
4212 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL,
4217 i915_gem_object_ggtt_pin(struct drm_i915_gem_object *obj,
4218 const struct i915_ggtt_view *view,
4222 if (WARN_ONCE(!view, "no view specified"))
4225 return i915_gem_object_do_pin(obj, i915_obj_to_ggtt(obj), view,
4226 alignment, flags | PIN_GLOBAL);
4230 i915_gem_object_ggtt_unpin_view(struct drm_i915_gem_object *obj,
4231 const struct i915_ggtt_view *view)
4233 struct i915_vma *vma = i915_gem_obj_to_ggtt_view(obj, view);
4236 WARN_ON(vma->pin_count == 0);
4237 WARN_ON(!i915_gem_obj_ggtt_bound_view(obj, view));
4239 if (--vma->pin_count == 0 && view->type == I915_GGTT_VIEW_NORMAL)
4240 obj->pin_mappable = false;
4244 i915_gem_object_pin_fence(struct drm_i915_gem_object *obj)
4246 if (obj->fence_reg != I915_FENCE_REG_NONE) {
4247 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
4248 struct i915_vma *ggtt_vma = i915_gem_obj_to_ggtt(obj);
4250 WARN_ON(!ggtt_vma ||
4251 dev_priv->fence_regs[obj->fence_reg].pin_count >
4252 ggtt_vma->pin_count);
4253 dev_priv->fence_regs[obj->fence_reg].pin_count++;
4260 i915_gem_object_unpin_fence(struct drm_i915_gem_object *obj)
4262 if (obj->fence_reg != I915_FENCE_REG_NONE) {
4263 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
4264 WARN_ON(dev_priv->fence_regs[obj->fence_reg].pin_count <= 0);
4265 dev_priv->fence_regs[obj->fence_reg].pin_count--;
4270 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
4271 struct drm_file *file)
4273 struct drm_i915_gem_busy *args = data;
4274 struct drm_i915_gem_object *obj;
4277 ret = i915_mutex_lock_interruptible(dev);
4281 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
4282 if (&obj->base == NULL) {
4287 /* Count all active objects as busy, even if they are currently not used
4288 * by the gpu. Users of this interface expect objects to eventually
4289 * become non-busy without any further actions, therefore emit any
4290 * necessary flushes here.
4292 ret = i915_gem_object_flush_active(obj);
4294 args->busy = obj->active;
4295 if (obj->last_read_req) {
4296 struct intel_engine_cs *ring;
4297 BUILD_BUG_ON(I915_NUM_RINGS > 16);
4298 ring = i915_gem_request_get_ring(obj->last_read_req);
4299 args->busy |= intel_ring_flag(ring) << 16;
4302 drm_gem_object_unreference(&obj->base);
4304 mutex_unlock(&dev->struct_mutex);
4309 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
4310 struct drm_file *file_priv)
4312 return i915_gem_ring_throttle(dev, file_priv);
4316 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
4317 struct drm_file *file_priv)
4319 struct drm_i915_private *dev_priv = dev->dev_private;
4320 struct drm_i915_gem_madvise *args = data;
4321 struct drm_i915_gem_object *obj;
4324 switch (args->madv) {
4325 case I915_MADV_DONTNEED:
4326 case I915_MADV_WILLNEED:
4332 ret = i915_mutex_lock_interruptible(dev);
4336 obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
4337 if (&obj->base == NULL) {
4342 if (i915_gem_obj_is_pinned(obj)) {
4348 obj->tiling_mode != I915_TILING_NONE &&
4349 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
4350 if (obj->madv == I915_MADV_WILLNEED)
4351 i915_gem_object_unpin_pages(obj);
4352 if (args->madv == I915_MADV_WILLNEED)
4353 i915_gem_object_pin_pages(obj);
4356 if (obj->madv != __I915_MADV_PURGED)
4357 obj->madv = args->madv;
4359 /* if the object is no longer attached, discard its backing storage */
4360 if (obj->madv == I915_MADV_DONTNEED && obj->pages == NULL)
4361 i915_gem_object_truncate(obj);
4363 args->retained = obj->madv != __I915_MADV_PURGED;
4366 drm_gem_object_unreference(&obj->base);
4368 mutex_unlock(&dev->struct_mutex);
4372 void i915_gem_object_init(struct drm_i915_gem_object *obj,
4373 const struct drm_i915_gem_object_ops *ops)
4375 INIT_LIST_HEAD(&obj->global_list);
4376 INIT_LIST_HEAD(&obj->ring_list);
4377 INIT_LIST_HEAD(&obj->obj_exec_link);
4378 INIT_LIST_HEAD(&obj->vma_list);
4379 INIT_LIST_HEAD(&obj->batch_pool_list);
4383 obj->fence_reg = I915_FENCE_REG_NONE;
4384 obj->madv = I915_MADV_WILLNEED;
4386 i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size);
4389 static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
4390 .get_pages = i915_gem_object_get_pages_gtt,
4391 .put_pages = i915_gem_object_put_pages_gtt,
4394 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
4397 struct drm_i915_gem_object *obj;
4398 struct address_space *mapping;
4401 obj = i915_gem_object_alloc(dev);
4405 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
4406 i915_gem_object_free(obj);
4410 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
4411 if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) {
4412 /* 965gm cannot relocate objects above 4GiB. */
4413 mask &= ~__GFP_HIGHMEM;
4414 mask |= __GFP_DMA32;
4417 mapping = file_inode(obj->base.filp)->i_mapping;
4418 mapping_set_gfp_mask(mapping, mask);
4420 i915_gem_object_init(obj, &i915_gem_object_ops);
4422 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4423 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4426 /* On some devices, we can have the GPU use the LLC (the CPU
4427 * cache) for about a 10% performance improvement
4428 * compared to uncached. Graphics requests other than
4429 * display scanout are coherent with the CPU in
4430 * accessing this cache. This means in this mode we
4431 * don't need to clflush on the CPU side, and on the
4432 * GPU side we only need to flush internal caches to
4433 * get data visible to the CPU.
4435 * However, we maintain the display planes as UC, and so
4436 * need to rebind when first used as such.
4438 obj->cache_level = I915_CACHE_LLC;
4440 obj->cache_level = I915_CACHE_NONE;
4442 trace_i915_gem_object_create(obj);
4447 static bool discard_backing_storage(struct drm_i915_gem_object *obj)
4449 /* If we are the last user of the backing storage (be it shmemfs
4450 * pages or stolen etc), we know that the pages are going to be
4451 * immediately released. In this case, we can then skip copying
4452 * back the contents from the GPU.
4455 if (obj->madv != I915_MADV_WILLNEED)
4458 if (obj->base.filp == NULL)
4461 /* At first glance, this looks racy, but then again so would be
4462 * userspace racing mmap against close. However, the first external
4463 * reference to the filp can only be obtained through the
4464 * i915_gem_mmap_ioctl() which safeguards us against the user
4465 * acquiring such a reference whilst we are in the middle of
4466 * freeing the object.
4468 return atomic_long_read(&obj->base.filp->f_count) == 1;
4471 void i915_gem_free_object(struct drm_gem_object *gem_obj)
4473 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
4474 struct drm_device *dev = obj->base.dev;
4475 struct drm_i915_private *dev_priv = dev->dev_private;
4476 struct i915_vma *vma, *next;
4478 intel_runtime_pm_get(dev_priv);
4480 trace_i915_gem_object_destroy(obj);
4482 list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) {
4486 ret = i915_vma_unbind(vma);
4487 if (WARN_ON(ret == -ERESTARTSYS)) {
4488 bool was_interruptible;
4490 was_interruptible = dev_priv->mm.interruptible;
4491 dev_priv->mm.interruptible = false;
4493 WARN_ON(i915_vma_unbind(vma));
4495 dev_priv->mm.interruptible = was_interruptible;
4499 /* Stolen objects don't hold a ref, but do hold pin count. Fix that up
4500 * before progressing. */
4502 i915_gem_object_unpin_pages(obj);
4504 WARN_ON(obj->frontbuffer_bits);
4506 if (obj->pages && obj->madv == I915_MADV_WILLNEED &&
4507 dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES &&
4508 obj->tiling_mode != I915_TILING_NONE)
4509 i915_gem_object_unpin_pages(obj);
4511 if (WARN_ON(obj->pages_pin_count))
4512 obj->pages_pin_count = 0;
4513 if (discard_backing_storage(obj))
4514 obj->madv = I915_MADV_DONTNEED;
4515 i915_gem_object_put_pages(obj);
4516 i915_gem_object_free_mmap_offset(obj);
4520 if (obj->base.import_attach)
4521 drm_prime_gem_destroy(&obj->base, NULL);
4523 if (obj->ops->release)
4524 obj->ops->release(obj);
4526 drm_gem_object_release(&obj->base);
4527 i915_gem_info_remove_obj(dev_priv, obj->base.size);
4530 i915_gem_object_free(obj);
4532 intel_runtime_pm_put(dev_priv);
4535 struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj,
4536 struct i915_address_space *vm)
4538 struct i915_vma *vma;
4539 list_for_each_entry(vma, &obj->vma_list, vma_link) {
4540 if (i915_is_ggtt(vma->vm) &&
4541 vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL)
4549 struct i915_vma *i915_gem_obj_to_ggtt_view(struct drm_i915_gem_object *obj,
4550 const struct i915_ggtt_view *view)
4552 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj);
4553 struct i915_vma *vma;
4555 if (WARN_ONCE(!view, "no view specified"))
4556 return ERR_PTR(-EINVAL);
4558 list_for_each_entry(vma, &obj->vma_list, vma_link)
4559 if (vma->vm == ggtt &&
4560 i915_ggtt_view_equal(&vma->ggtt_view, view))
4565 void i915_gem_vma_destroy(struct i915_vma *vma)
4567 struct i915_address_space *vm = NULL;
4568 WARN_ON(vma->node.allocated);
4570 /* Keep the vma as a placeholder in the execbuffer reservation lists */
4571 if (!list_empty(&vma->exec_list))
4576 if (!i915_is_ggtt(vm))
4577 i915_ppgtt_put(i915_vm_to_ppgtt(vm));
4579 list_del(&vma->vma_link);
4585 i915_gem_stop_ringbuffers(struct drm_device *dev)
4587 struct drm_i915_private *dev_priv = dev->dev_private;
4588 struct intel_engine_cs *ring;
4591 for_each_ring(ring, dev_priv, i)
4592 dev_priv->gt.stop_ring(ring);
4596 i915_gem_suspend(struct drm_device *dev)
4598 struct drm_i915_private *dev_priv = dev->dev_private;
4601 mutex_lock(&dev->struct_mutex);
4602 ret = i915_gpu_idle(dev);
4606 i915_gem_retire_requests(dev);
4608 i915_gem_stop_ringbuffers(dev);
4609 mutex_unlock(&dev->struct_mutex);
4611 cancel_delayed_work_sync(&dev_priv->gpu_error.hangcheck_work);
4612 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
4613 flush_delayed_work(&dev_priv->mm.idle_work);
4615 /* Assert that we sucessfully flushed all the work and
4616 * reset the GPU back to its idle, low power state.
4618 WARN_ON(dev_priv->mm.busy);
4623 mutex_unlock(&dev->struct_mutex);
4627 int i915_gem_l3_remap(struct intel_engine_cs *ring, int slice)
4629 struct drm_device *dev = ring->dev;
4630 struct drm_i915_private *dev_priv = dev->dev_private;
4631 u32 reg_base = GEN7_L3LOG_BASE + (slice * 0x200);
4632 u32 *remap_info = dev_priv->l3_parity.remap_info[slice];
4635 if (!HAS_L3_DPF(dev) || !remap_info)
4638 ret = intel_ring_begin(ring, GEN7_L3LOG_SIZE / 4 * 3);
4643 * Note: We do not worry about the concurrent register cacheline hang
4644 * here because no other code should access these registers other than
4645 * at initialization time.
4647 for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) {
4648 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
4649 intel_ring_emit(ring, reg_base + i);
4650 intel_ring_emit(ring, remap_info[i/4]);
4653 intel_ring_advance(ring);
4658 void i915_gem_init_swizzling(struct drm_device *dev)
4660 struct drm_i915_private *dev_priv = dev->dev_private;
4662 if (INTEL_INFO(dev)->gen < 5 ||
4663 dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
4666 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
4667 DISP_TILE_SURFACE_SWIZZLING);
4672 I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
4674 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
4675 else if (IS_GEN7(dev))
4676 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
4677 else if (IS_GEN8(dev))
4678 I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW));
4684 intel_enable_blt(struct drm_device *dev)
4689 /* The blitter was dysfunctional on early prototypes */
4690 if (IS_GEN6(dev) && dev->pdev->revision < 8) {
4691 DRM_INFO("BLT not supported on this pre-production hardware;"
4692 " graphics performance will be degraded.\n");
4699 static void init_unused_ring(struct drm_device *dev, u32 base)
4701 struct drm_i915_private *dev_priv = dev->dev_private;
4703 I915_WRITE(RING_CTL(base), 0);
4704 I915_WRITE(RING_HEAD(base), 0);
4705 I915_WRITE(RING_TAIL(base), 0);
4706 I915_WRITE(RING_START(base), 0);
4709 static void init_unused_rings(struct drm_device *dev)
4712 init_unused_ring(dev, PRB1_BASE);
4713 init_unused_ring(dev, SRB0_BASE);
4714 init_unused_ring(dev, SRB1_BASE);
4715 init_unused_ring(dev, SRB2_BASE);
4716 init_unused_ring(dev, SRB3_BASE);
4717 } else if (IS_GEN2(dev)) {
4718 init_unused_ring(dev, SRB0_BASE);
4719 init_unused_ring(dev, SRB1_BASE);
4720 } else if (IS_GEN3(dev)) {
4721 init_unused_ring(dev, PRB1_BASE);
4722 init_unused_ring(dev, PRB2_BASE);
4726 int i915_gem_init_rings(struct drm_device *dev)
4728 struct drm_i915_private *dev_priv = dev->dev_private;
4731 ret = intel_init_render_ring_buffer(dev);
4736 ret = intel_init_bsd_ring_buffer(dev);
4738 goto cleanup_render_ring;
4741 if (intel_enable_blt(dev)) {
4742 ret = intel_init_blt_ring_buffer(dev);
4744 goto cleanup_bsd_ring;
4747 if (HAS_VEBOX(dev)) {
4748 ret = intel_init_vebox_ring_buffer(dev);
4750 goto cleanup_blt_ring;
4753 if (HAS_BSD2(dev)) {
4754 ret = intel_init_bsd2_ring_buffer(dev);
4756 goto cleanup_vebox_ring;
4759 ret = i915_gem_set_seqno(dev, ((u32)~0 - 0x1000));
4761 goto cleanup_bsd2_ring;
4766 intel_cleanup_ring_buffer(&dev_priv->ring[VCS2]);
4768 intel_cleanup_ring_buffer(&dev_priv->ring[VECS]);
4770 intel_cleanup_ring_buffer(&dev_priv->ring[BCS]);
4772 intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
4773 cleanup_render_ring:
4774 intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
4780 i915_gem_init_hw(struct drm_device *dev)
4782 struct drm_i915_private *dev_priv = dev->dev_private;
4783 struct intel_engine_cs *ring;
4786 if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt())
4789 /* Double layer security blanket, see i915_gem_init() */
4790 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4792 if (dev_priv->ellc_size)
4793 I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf));
4795 if (IS_HASWELL(dev))
4796 I915_WRITE(MI_PREDICATE_RESULT_2, IS_HSW_GT3(dev) ?
4797 LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED);
4799 if (HAS_PCH_NOP(dev)) {
4800 if (IS_IVYBRIDGE(dev)) {
4801 u32 temp = I915_READ(GEN7_MSG_CTL);
4802 temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK);
4803 I915_WRITE(GEN7_MSG_CTL, temp);
4804 } else if (INTEL_INFO(dev)->gen >= 7) {
4805 u32 temp = I915_READ(HSW_NDE_RSTWRN_OPT);
4806 temp &= ~RESET_PCH_HANDSHAKE_ENABLE;
4807 I915_WRITE(HSW_NDE_RSTWRN_OPT, temp);
4811 i915_gem_init_swizzling(dev);
4814 * At least 830 can leave some of the unused rings
4815 * "active" (ie. head != tail) after resume which
4816 * will prevent c3 entry. Makes sure all unused rings
4819 init_unused_rings(dev);
4821 for_each_ring(ring, dev_priv, i) {
4822 ret = ring->init_hw(ring);
4827 for (i = 0; i < NUM_L3_SLICES(dev); i++)
4828 i915_gem_l3_remap(&dev_priv->ring[RCS], i);
4830 ret = i915_ppgtt_init_hw(dev);
4831 if (ret && ret != -EIO) {
4832 DRM_ERROR("PPGTT enable failed %d\n", ret);
4833 i915_gem_cleanup_ringbuffer(dev);
4836 ret = i915_gem_context_enable(dev_priv);
4837 if (ret && ret != -EIO) {
4838 DRM_ERROR("Context enable failed %d\n", ret);
4839 i915_gem_cleanup_ringbuffer(dev);
4845 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4849 int i915_gem_init(struct drm_device *dev)
4851 struct drm_i915_private *dev_priv = dev->dev_private;
4854 i915.enable_execlists = intel_sanitize_enable_execlists(dev,
4855 i915.enable_execlists);
4857 mutex_lock(&dev->struct_mutex);
4859 if (IS_VALLEYVIEW(dev)) {
4860 /* VLVA0 (potential hack), BIOS isn't actually waking us */
4861 I915_WRITE(VLV_GTLC_WAKE_CTRL, VLV_GTLC_ALLOWWAKEREQ);
4862 if (wait_for((I915_READ(VLV_GTLC_PW_STATUS) &
4863 VLV_GTLC_ALLOWWAKEACK), 10))
4864 DRM_DEBUG_DRIVER("allow wake ack timed out\n");
4867 if (!i915.enable_execlists) {
4868 dev_priv->gt.do_execbuf = i915_gem_ringbuffer_submission;
4869 dev_priv->gt.init_rings = i915_gem_init_rings;
4870 dev_priv->gt.cleanup_ring = intel_cleanup_ring_buffer;
4871 dev_priv->gt.stop_ring = intel_stop_ring_buffer;
4873 dev_priv->gt.do_execbuf = intel_execlists_submission;
4874 dev_priv->gt.init_rings = intel_logical_rings_init;
4875 dev_priv->gt.cleanup_ring = intel_logical_ring_cleanup;
4876 dev_priv->gt.stop_ring = intel_logical_ring_stop;
4879 /* This is just a security blanket to placate dragons.
4880 * On some systems, we very sporadically observe that the first TLBs
4881 * used by the CS may be stale, despite us poking the TLB reset. If
4882 * we hold the forcewake during initialisation these problems
4883 * just magically go away.
4885 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4887 ret = i915_gem_init_userptr(dev);
4891 i915_gem_init_global_gtt(dev);
4893 ret = i915_gem_context_init(dev);
4897 ret = dev_priv->gt.init_rings(dev);
4901 ret = i915_gem_init_hw(dev);
4903 /* Allow ring initialisation to fail by marking the GPU as
4904 * wedged. But we only want to do this where the GPU is angry,
4905 * for all other failure, such as an allocation failure, bail.
4907 DRM_ERROR("Failed to initialize GPU, declaring it wedged\n");
4908 atomic_set_mask(I915_WEDGED, &dev_priv->gpu_error.reset_counter);
4913 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4914 mutex_unlock(&dev->struct_mutex);
4920 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
4922 struct drm_i915_private *dev_priv = dev->dev_private;
4923 struct intel_engine_cs *ring;
4926 for_each_ring(ring, dev_priv, i)
4927 dev_priv->gt.cleanup_ring(ring);
4931 init_ring_lists(struct intel_engine_cs *ring)
4933 INIT_LIST_HEAD(&ring->active_list);
4934 INIT_LIST_HEAD(&ring->request_list);
4937 void i915_init_vm(struct drm_i915_private *dev_priv,
4938 struct i915_address_space *vm)
4940 if (!i915_is_ggtt(vm))
4941 drm_mm_init(&vm->mm, vm->start, vm->total);
4942 vm->dev = dev_priv->dev;
4943 INIT_LIST_HEAD(&vm->active_list);
4944 INIT_LIST_HEAD(&vm->inactive_list);
4945 INIT_LIST_HEAD(&vm->global_link);
4946 list_add_tail(&vm->global_link, &dev_priv->vm_list);
4950 i915_gem_load(struct drm_device *dev)
4952 struct drm_i915_private *dev_priv = dev->dev_private;
4956 kmem_cache_create("i915_gem_object",
4957 sizeof(struct drm_i915_gem_object), 0,
4961 INIT_LIST_HEAD(&dev_priv->vm_list);
4962 i915_init_vm(dev_priv, &dev_priv->gtt.base);
4964 INIT_LIST_HEAD(&dev_priv->context_list);
4965 INIT_LIST_HEAD(&dev_priv->mm.unbound_list);
4966 INIT_LIST_HEAD(&dev_priv->mm.bound_list);
4967 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4968 for (i = 0; i < I915_NUM_RINGS; i++)
4969 init_ring_lists(&dev_priv->ring[i]);
4970 for (i = 0; i < I915_MAX_NUM_FENCES; i++)
4971 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
4972 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
4973 i915_gem_retire_work_handler);
4974 INIT_DELAYED_WORK(&dev_priv->mm.idle_work,
4975 i915_gem_idle_work_handler);
4976 init_waitqueue_head(&dev_priv->gpu_error.reset_queue);
4978 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
4980 if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev))
4981 dev_priv->num_fence_regs = 32;
4982 else if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
4983 dev_priv->num_fence_regs = 16;
4985 dev_priv->num_fence_regs = 8;
4987 if (intel_vgpu_active(dev))
4988 dev_priv->num_fence_regs =
4989 I915_READ(vgtif_reg(avail_rs.fence_num));
4991 /* Initialize fence registers to zero */
4992 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4993 i915_gem_restore_fences(dev);
4995 i915_gem_detect_bit_6_swizzle(dev);
4996 init_waitqueue_head(&dev_priv->pending_flip_queue);
4998 dev_priv->mm.interruptible = true;
5000 i915_gem_shrinker_init(dev_priv);
5002 i915_gem_batch_pool_init(dev, &dev_priv->mm.batch_pool);
5004 mutex_init(&dev_priv->fb_tracking.lock);
5007 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
5009 struct drm_i915_file_private *file_priv = file->driver_priv;
5011 cancel_delayed_work_sync(&file_priv->mm.idle_work);
5013 /* Clean up our request list when the client is going away, so that
5014 * later retire_requests won't dereference our soon-to-be-gone
5017 spin_lock(&file_priv->mm.lock);
5018 while (!list_empty(&file_priv->mm.request_list)) {
5019 struct drm_i915_gem_request *request;
5021 request = list_first_entry(&file_priv->mm.request_list,
5022 struct drm_i915_gem_request,
5024 list_del(&request->client_list);
5025 request->file_priv = NULL;
5027 spin_unlock(&file_priv->mm.lock);
5031 i915_gem_file_idle_work_handler(struct work_struct *work)
5033 struct drm_i915_file_private *file_priv =
5034 container_of(work, typeof(*file_priv), mm.idle_work.work);
5036 atomic_set(&file_priv->rps_wait_boost, false);
5039 int i915_gem_open(struct drm_device *dev, struct drm_file *file)
5041 struct drm_i915_file_private *file_priv;
5044 DRM_DEBUG_DRIVER("\n");
5046 file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL);
5050 file->driver_priv = file_priv;
5051 file_priv->dev_priv = dev->dev_private;
5052 file_priv->file = file;
5054 spin_lock_init(&file_priv->mm.lock);
5055 INIT_LIST_HEAD(&file_priv->mm.request_list);
5056 INIT_DELAYED_WORK(&file_priv->mm.idle_work,
5057 i915_gem_file_idle_work_handler);
5059 ret = i915_gem_context_open(dev, file);
5067 * i915_gem_track_fb - update frontbuffer tracking
5068 * old: current GEM buffer for the frontbuffer slots
5069 * new: new GEM buffer for the frontbuffer slots
5070 * frontbuffer_bits: bitmask of frontbuffer slots
5072 * This updates the frontbuffer tracking bits @frontbuffer_bits by clearing them
5073 * from @old and setting them in @new. Both @old and @new can be NULL.
5075 void i915_gem_track_fb(struct drm_i915_gem_object *old,
5076 struct drm_i915_gem_object *new,
5077 unsigned frontbuffer_bits)
5080 WARN_ON(!mutex_is_locked(&old->base.dev->struct_mutex));
5081 WARN_ON(!(old->frontbuffer_bits & frontbuffer_bits));
5082 old->frontbuffer_bits &= ~frontbuffer_bits;
5086 WARN_ON(!mutex_is_locked(&new->base.dev->struct_mutex));
5087 WARN_ON(new->frontbuffer_bits & frontbuffer_bits);
5088 new->frontbuffer_bits |= frontbuffer_bits;
5092 /* All the new VM stuff */
5094 i915_gem_obj_offset(struct drm_i915_gem_object *o,
5095 struct i915_address_space *vm)
5097 struct drm_i915_private *dev_priv = o->base.dev->dev_private;
5098 struct i915_vma *vma;
5100 WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base);
5102 list_for_each_entry(vma, &o->vma_list, vma_link) {
5103 if (i915_is_ggtt(vma->vm) &&
5104 vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL)
5107 return vma->node.start;
5110 WARN(1, "%s vma for this object not found.\n",
5111 i915_is_ggtt(vm) ? "global" : "ppgtt");
5116 i915_gem_obj_ggtt_offset_view(struct drm_i915_gem_object *o,
5117 const struct i915_ggtt_view *view)
5119 struct i915_address_space *ggtt = i915_obj_to_ggtt(o);
5120 struct i915_vma *vma;
5122 list_for_each_entry(vma, &o->vma_list, vma_link)
5123 if (vma->vm == ggtt &&
5124 i915_ggtt_view_equal(&vma->ggtt_view, view))
5125 return vma->node.start;
5127 WARN(1, "global vma for this object not found.\n");
5131 bool i915_gem_obj_bound(struct drm_i915_gem_object *o,
5132 struct i915_address_space *vm)
5134 struct i915_vma *vma;
5136 list_for_each_entry(vma, &o->vma_list, vma_link) {
5137 if (i915_is_ggtt(vma->vm) &&
5138 vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL)
5140 if (vma->vm == vm && drm_mm_node_allocated(&vma->node))
5147 bool i915_gem_obj_ggtt_bound_view(struct drm_i915_gem_object *o,
5148 const struct i915_ggtt_view *view)
5150 struct i915_address_space *ggtt = i915_obj_to_ggtt(o);
5151 struct i915_vma *vma;
5153 list_for_each_entry(vma, &o->vma_list, vma_link)
5154 if (vma->vm == ggtt &&
5155 i915_ggtt_view_equal(&vma->ggtt_view, view) &&
5156 drm_mm_node_allocated(&vma->node))
5162 bool i915_gem_obj_bound_any(struct drm_i915_gem_object *o)
5164 struct i915_vma *vma;
5166 list_for_each_entry(vma, &o->vma_list, vma_link)
5167 if (drm_mm_node_allocated(&vma->node))
5173 unsigned long i915_gem_obj_size(struct drm_i915_gem_object *o,
5174 struct i915_address_space *vm)
5176 struct drm_i915_private *dev_priv = o->base.dev->dev_private;
5177 struct i915_vma *vma;
5179 WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base);
5181 BUG_ON(list_empty(&o->vma_list));
5183 list_for_each_entry(vma, &o->vma_list, vma_link) {
5184 if (i915_is_ggtt(vma->vm) &&
5185 vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL)
5188 return vma->node.size;
5193 bool i915_gem_obj_is_pinned(struct drm_i915_gem_object *obj)
5195 struct i915_vma *vma;
5196 list_for_each_entry(vma, &obj->vma_list, vma_link) {
5197 if (i915_is_ggtt(vma->vm) &&
5198 vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL)
5200 if (vma->pin_count > 0)