2 * Copyright © 2008 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eric Anholt <eric@anholt.net>
29 #include <drm/drm_vma_manager.h>
30 #include <drm/i915_drm.h>
32 #include "i915_trace.h"
33 #include "intel_drv.h"
34 #include <linux/shmem_fs.h>
35 #include <linux/slab.h>
36 #include <linux/swap.h>
37 #include <linux/pci.h>
38 #include <linux/dma-buf.h>
40 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);
41 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj);
42 static __must_check int i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
44 bool map_and_fenceable,
46 static int i915_gem_phys_pwrite(struct drm_device *dev,
47 struct drm_i915_gem_object *obj,
48 struct drm_i915_gem_pwrite *args,
49 struct drm_file *file);
51 static void i915_gem_write_fence(struct drm_device *dev, int reg,
52 struct drm_i915_gem_object *obj);
53 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
54 struct drm_i915_fence_reg *fence,
57 static unsigned long i915_gem_inactive_count(struct shrinker *shrinker,
58 struct shrink_control *sc);
59 static unsigned long i915_gem_inactive_scan(struct shrinker *shrinker,
60 struct shrink_control *sc);
61 static long i915_gem_purge(struct drm_i915_private *dev_priv, long target);
62 static long i915_gem_shrink_all(struct drm_i915_private *dev_priv);
63 static void i915_gem_object_truncate(struct drm_i915_gem_object *obj);
65 static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object *obj)
68 i915_gem_release_mmap(obj);
70 /* As we do not have an associated fence register, we will force
71 * a tiling change if we ever need to acquire one.
73 obj->fence_dirty = false;
74 obj->fence_reg = I915_FENCE_REG_NONE;
77 /* some bookkeeping */
78 static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
81 spin_lock(&dev_priv->mm.object_stat_lock);
82 dev_priv->mm.object_count++;
83 dev_priv->mm.object_memory += size;
84 spin_unlock(&dev_priv->mm.object_stat_lock);
87 static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
90 spin_lock(&dev_priv->mm.object_stat_lock);
91 dev_priv->mm.object_count--;
92 dev_priv->mm.object_memory -= size;
93 spin_unlock(&dev_priv->mm.object_stat_lock);
97 i915_gem_wait_for_error(struct i915_gpu_error *error)
101 #define EXIT_COND (!i915_reset_in_progress(error) || \
102 i915_terminally_wedged(error))
107 * Only wait 10 seconds for the gpu reset to complete to avoid hanging
108 * userspace. If it takes that long something really bad is going on and
109 * we should simply try to bail out and fail as gracefully as possible.
111 ret = wait_event_interruptible_timeout(error->reset_queue,
115 DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
117 } else if (ret < 0) {
125 int i915_mutex_lock_interruptible(struct drm_device *dev)
127 struct drm_i915_private *dev_priv = dev->dev_private;
130 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
134 ret = mutex_lock_interruptible(&dev->struct_mutex);
138 WARN_ON(i915_verify_lists(dev));
143 i915_gem_object_is_inactive(struct drm_i915_gem_object *obj)
145 return i915_gem_obj_ggtt_bound(obj) && !obj->active;
149 i915_gem_init_ioctl(struct drm_device *dev, void *data,
150 struct drm_file *file)
152 struct drm_i915_private *dev_priv = dev->dev_private;
153 struct drm_i915_gem_init *args = data;
155 if (drm_core_check_feature(dev, DRIVER_MODESET))
158 if (args->gtt_start >= args->gtt_end ||
159 (args->gtt_end | args->gtt_start) & (PAGE_SIZE - 1))
162 /* GEM with user mode setting was never supported on ilk and later. */
163 if (INTEL_INFO(dev)->gen >= 5)
166 mutex_lock(&dev->struct_mutex);
167 i915_gem_setup_global_gtt(dev, args->gtt_start, args->gtt_end,
169 dev_priv->gtt.mappable_end = args->gtt_end;
170 mutex_unlock(&dev->struct_mutex);
176 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
177 struct drm_file *file)
179 struct drm_i915_private *dev_priv = dev->dev_private;
180 struct drm_i915_gem_get_aperture *args = data;
181 struct drm_i915_gem_object *obj;
185 mutex_lock(&dev->struct_mutex);
186 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list)
188 pinned += i915_gem_obj_ggtt_size(obj);
189 mutex_unlock(&dev->struct_mutex);
191 args->aper_size = dev_priv->gtt.base.total;
192 args->aper_available_size = args->aper_size - pinned;
197 void *i915_gem_object_alloc(struct drm_device *dev)
199 struct drm_i915_private *dev_priv = dev->dev_private;
200 return kmem_cache_alloc(dev_priv->slab, GFP_KERNEL | __GFP_ZERO);
203 void i915_gem_object_free(struct drm_i915_gem_object *obj)
205 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
206 kmem_cache_free(dev_priv->slab, obj);
210 i915_gem_create(struct drm_file *file,
211 struct drm_device *dev,
215 struct drm_i915_gem_object *obj;
219 size = roundup(size, PAGE_SIZE);
223 /* Allocate the new object */
224 obj = i915_gem_alloc_object(dev, size);
228 ret = drm_gem_handle_create(file, &obj->base, &handle);
229 /* drop reference from allocate - handle holds it now */
230 drm_gem_object_unreference_unlocked(&obj->base);
239 i915_gem_dumb_create(struct drm_file *file,
240 struct drm_device *dev,
241 struct drm_mode_create_dumb *args)
243 /* have to work out size/pitch and return them */
244 args->pitch = ALIGN(args->width * ((args->bpp + 7) / 8), 64);
245 args->size = args->pitch * args->height;
246 return i915_gem_create(file, dev,
247 args->size, &args->handle);
250 int i915_gem_dumb_destroy(struct drm_file *file,
251 struct drm_device *dev,
254 return drm_gem_handle_delete(file, handle);
258 * Creates a new mm object and returns a handle to it.
261 i915_gem_create_ioctl(struct drm_device *dev, void *data,
262 struct drm_file *file)
264 struct drm_i915_gem_create *args = data;
266 return i915_gem_create(file, dev,
267 args->size, &args->handle);
271 __copy_to_user_swizzled(char __user *cpu_vaddr,
272 const char *gpu_vaddr, int gpu_offset,
275 int ret, cpu_offset = 0;
278 int cacheline_end = ALIGN(gpu_offset + 1, 64);
279 int this_length = min(cacheline_end - gpu_offset, length);
280 int swizzled_gpu_offset = gpu_offset ^ 64;
282 ret = __copy_to_user(cpu_vaddr + cpu_offset,
283 gpu_vaddr + swizzled_gpu_offset,
288 cpu_offset += this_length;
289 gpu_offset += this_length;
290 length -= this_length;
297 __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset,
298 const char __user *cpu_vaddr,
301 int ret, cpu_offset = 0;
304 int cacheline_end = ALIGN(gpu_offset + 1, 64);
305 int this_length = min(cacheline_end - gpu_offset, length);
306 int swizzled_gpu_offset = gpu_offset ^ 64;
308 ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset,
309 cpu_vaddr + cpu_offset,
314 cpu_offset += this_length;
315 gpu_offset += this_length;
316 length -= this_length;
322 /* Per-page copy function for the shmem pread fastpath.
323 * Flushes invalid cachelines before reading the target if
324 * needs_clflush is set. */
326 shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length,
327 char __user *user_data,
328 bool page_do_bit17_swizzling, bool needs_clflush)
333 if (unlikely(page_do_bit17_swizzling))
336 vaddr = kmap_atomic(page);
338 drm_clflush_virt_range(vaddr + shmem_page_offset,
340 ret = __copy_to_user_inatomic(user_data,
341 vaddr + shmem_page_offset,
343 kunmap_atomic(vaddr);
345 return ret ? -EFAULT : 0;
349 shmem_clflush_swizzled_range(char *addr, unsigned long length,
352 if (unlikely(swizzled)) {
353 unsigned long start = (unsigned long) addr;
354 unsigned long end = (unsigned long) addr + length;
356 /* For swizzling simply ensure that we always flush both
357 * channels. Lame, but simple and it works. Swizzled
358 * pwrite/pread is far from a hotpath - current userspace
359 * doesn't use it at all. */
360 start = round_down(start, 128);
361 end = round_up(end, 128);
363 drm_clflush_virt_range((void *)start, end - start);
365 drm_clflush_virt_range(addr, length);
370 /* Only difference to the fast-path function is that this can handle bit17
371 * and uses non-atomic copy and kmap functions. */
373 shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length,
374 char __user *user_data,
375 bool page_do_bit17_swizzling, bool needs_clflush)
382 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
384 page_do_bit17_swizzling);
386 if (page_do_bit17_swizzling)
387 ret = __copy_to_user_swizzled(user_data,
388 vaddr, shmem_page_offset,
391 ret = __copy_to_user(user_data,
392 vaddr + shmem_page_offset,
396 return ret ? - EFAULT : 0;
400 i915_gem_shmem_pread(struct drm_device *dev,
401 struct drm_i915_gem_object *obj,
402 struct drm_i915_gem_pread *args,
403 struct drm_file *file)
405 char __user *user_data;
408 int shmem_page_offset, page_length, ret = 0;
409 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
411 int needs_clflush = 0;
412 struct sg_page_iter sg_iter;
414 user_data = to_user_ptr(args->data_ptr);
417 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
419 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) {
420 /* If we're not in the cpu read domain, set ourself into the gtt
421 * read domain and manually flush cachelines (if required). This
422 * optimizes for the case when the gpu will dirty the data
423 * anyway again before the next pread happens. */
424 if (obj->cache_level == I915_CACHE_NONE)
426 if (i915_gem_obj_ggtt_bound(obj)) {
427 ret = i915_gem_object_set_to_gtt_domain(obj, false);
433 ret = i915_gem_object_get_pages(obj);
437 i915_gem_object_pin_pages(obj);
439 offset = args->offset;
441 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
442 offset >> PAGE_SHIFT) {
443 struct page *page = sg_page_iter_page(&sg_iter);
448 /* Operation in this page
450 * shmem_page_offset = offset within page in shmem file
451 * page_length = bytes to copy for this page
453 shmem_page_offset = offset_in_page(offset);
454 page_length = remain;
455 if ((shmem_page_offset + page_length) > PAGE_SIZE)
456 page_length = PAGE_SIZE - shmem_page_offset;
458 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
459 (page_to_phys(page) & (1 << 17)) != 0;
461 ret = shmem_pread_fast(page, shmem_page_offset, page_length,
462 user_data, page_do_bit17_swizzling,
467 mutex_unlock(&dev->struct_mutex);
469 if (likely(!i915_prefault_disable) && !prefaulted) {
470 ret = fault_in_multipages_writeable(user_data, remain);
471 /* Userspace is tricking us, but we've already clobbered
472 * its pages with the prefault and promised to write the
473 * data up to the first fault. Hence ignore any errors
474 * and just continue. */
479 ret = shmem_pread_slow(page, shmem_page_offset, page_length,
480 user_data, page_do_bit17_swizzling,
483 mutex_lock(&dev->struct_mutex);
486 mark_page_accessed(page);
491 remain -= page_length;
492 user_data += page_length;
493 offset += page_length;
497 i915_gem_object_unpin_pages(obj);
503 * Reads data from the object referenced by handle.
505 * On error, the contents of *data are undefined.
508 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
509 struct drm_file *file)
511 struct drm_i915_gem_pread *args = data;
512 struct drm_i915_gem_object *obj;
518 if (!access_ok(VERIFY_WRITE,
519 to_user_ptr(args->data_ptr),
523 ret = i915_mutex_lock_interruptible(dev);
527 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
528 if (&obj->base == NULL) {
533 /* Bounds check source. */
534 if (args->offset > obj->base.size ||
535 args->size > obj->base.size - args->offset) {
540 /* prime objects have no backing filp to GEM pread/pwrite
543 if (!obj->base.filp) {
548 trace_i915_gem_object_pread(obj, args->offset, args->size);
550 ret = i915_gem_shmem_pread(dev, obj, args, file);
553 drm_gem_object_unreference(&obj->base);
555 mutex_unlock(&dev->struct_mutex);
559 /* This is the fast write path which cannot handle
560 * page faults in the source data
564 fast_user_write(struct io_mapping *mapping,
565 loff_t page_base, int page_offset,
566 char __user *user_data,
569 void __iomem *vaddr_atomic;
571 unsigned long unwritten;
573 vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
574 /* We can use the cpu mem copy function because this is X86. */
575 vaddr = (void __force*)vaddr_atomic + page_offset;
576 unwritten = __copy_from_user_inatomic_nocache(vaddr,
578 io_mapping_unmap_atomic(vaddr_atomic);
583 * This is the fast pwrite path, where we copy the data directly from the
584 * user into the GTT, uncached.
587 i915_gem_gtt_pwrite_fast(struct drm_device *dev,
588 struct drm_i915_gem_object *obj,
589 struct drm_i915_gem_pwrite *args,
590 struct drm_file *file)
592 drm_i915_private_t *dev_priv = dev->dev_private;
594 loff_t offset, page_base;
595 char __user *user_data;
596 int page_offset, page_length, ret;
598 ret = i915_gem_object_pin(obj, 0, true, true);
602 ret = i915_gem_object_set_to_gtt_domain(obj, true);
606 ret = i915_gem_object_put_fence(obj);
610 user_data = to_user_ptr(args->data_ptr);
613 offset = i915_gem_obj_ggtt_offset(obj) + args->offset;
616 /* Operation in this page
618 * page_base = page offset within aperture
619 * page_offset = offset within page
620 * page_length = bytes to copy for this page
622 page_base = offset & PAGE_MASK;
623 page_offset = offset_in_page(offset);
624 page_length = remain;
625 if ((page_offset + remain) > PAGE_SIZE)
626 page_length = PAGE_SIZE - page_offset;
628 /* If we get a fault while copying data, then (presumably) our
629 * source page isn't available. Return the error and we'll
630 * retry in the slow path.
632 if (fast_user_write(dev_priv->gtt.mappable, page_base,
633 page_offset, user_data, page_length)) {
638 remain -= page_length;
639 user_data += page_length;
640 offset += page_length;
644 i915_gem_object_unpin(obj);
649 /* Per-page copy function for the shmem pwrite fastpath.
650 * Flushes invalid cachelines before writing to the target if
651 * needs_clflush_before is set and flushes out any written cachelines after
652 * writing if needs_clflush is set. */
654 shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
655 char __user *user_data,
656 bool page_do_bit17_swizzling,
657 bool needs_clflush_before,
658 bool needs_clflush_after)
663 if (unlikely(page_do_bit17_swizzling))
666 vaddr = kmap_atomic(page);
667 if (needs_clflush_before)
668 drm_clflush_virt_range(vaddr + shmem_page_offset,
670 ret = __copy_from_user_inatomic_nocache(vaddr + shmem_page_offset,
673 if (needs_clflush_after)
674 drm_clflush_virt_range(vaddr + shmem_page_offset,
676 kunmap_atomic(vaddr);
678 return ret ? -EFAULT : 0;
681 /* Only difference to the fast-path function is that this can handle bit17
682 * and uses non-atomic copy and kmap functions. */
684 shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
685 char __user *user_data,
686 bool page_do_bit17_swizzling,
687 bool needs_clflush_before,
688 bool needs_clflush_after)
694 if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
695 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
697 page_do_bit17_swizzling);
698 if (page_do_bit17_swizzling)
699 ret = __copy_from_user_swizzled(vaddr, shmem_page_offset,
703 ret = __copy_from_user(vaddr + shmem_page_offset,
706 if (needs_clflush_after)
707 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
709 page_do_bit17_swizzling);
712 return ret ? -EFAULT : 0;
716 i915_gem_shmem_pwrite(struct drm_device *dev,
717 struct drm_i915_gem_object *obj,
718 struct drm_i915_gem_pwrite *args,
719 struct drm_file *file)
723 char __user *user_data;
724 int shmem_page_offset, page_length, ret = 0;
725 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
726 int hit_slowpath = 0;
727 int needs_clflush_after = 0;
728 int needs_clflush_before = 0;
729 struct sg_page_iter sg_iter;
731 user_data = to_user_ptr(args->data_ptr);
734 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
736 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
737 /* If we're not in the cpu write domain, set ourself into the gtt
738 * write domain and manually flush cachelines (if required). This
739 * optimizes for the case when the gpu will use the data
740 * right away and we therefore have to clflush anyway. */
741 if (obj->cache_level == I915_CACHE_NONE)
742 needs_clflush_after = 1;
743 if (i915_gem_obj_ggtt_bound(obj)) {
744 ret = i915_gem_object_set_to_gtt_domain(obj, true);
749 /* Same trick applies for invalidate partially written cachelines before
751 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)
752 && obj->cache_level == I915_CACHE_NONE)
753 needs_clflush_before = 1;
755 ret = i915_gem_object_get_pages(obj);
759 i915_gem_object_pin_pages(obj);
761 offset = args->offset;
764 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
765 offset >> PAGE_SHIFT) {
766 struct page *page = sg_page_iter_page(&sg_iter);
767 int partial_cacheline_write;
772 /* Operation in this page
774 * shmem_page_offset = offset within page in shmem file
775 * page_length = bytes to copy for this page
777 shmem_page_offset = offset_in_page(offset);
779 page_length = remain;
780 if ((shmem_page_offset + page_length) > PAGE_SIZE)
781 page_length = PAGE_SIZE - shmem_page_offset;
783 /* If we don't overwrite a cacheline completely we need to be
784 * careful to have up-to-date data by first clflushing. Don't
785 * overcomplicate things and flush the entire patch. */
786 partial_cacheline_write = needs_clflush_before &&
787 ((shmem_page_offset | page_length)
788 & (boot_cpu_data.x86_clflush_size - 1));
790 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
791 (page_to_phys(page) & (1 << 17)) != 0;
793 ret = shmem_pwrite_fast(page, shmem_page_offset, page_length,
794 user_data, page_do_bit17_swizzling,
795 partial_cacheline_write,
796 needs_clflush_after);
801 mutex_unlock(&dev->struct_mutex);
802 ret = shmem_pwrite_slow(page, shmem_page_offset, page_length,
803 user_data, page_do_bit17_swizzling,
804 partial_cacheline_write,
805 needs_clflush_after);
807 mutex_lock(&dev->struct_mutex);
810 set_page_dirty(page);
811 mark_page_accessed(page);
816 remain -= page_length;
817 user_data += page_length;
818 offset += page_length;
822 i915_gem_object_unpin_pages(obj);
826 * Fixup: Flush cpu caches in case we didn't flush the dirty
827 * cachelines in-line while writing and the object moved
828 * out of the cpu write domain while we've dropped the lock.
830 if (!needs_clflush_after &&
831 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
832 i915_gem_clflush_object(obj);
833 i915_gem_chipset_flush(dev);
837 if (needs_clflush_after)
838 i915_gem_chipset_flush(dev);
844 * Writes data to the object referenced by handle.
846 * On error, the contents of the buffer that were to be modified are undefined.
849 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
850 struct drm_file *file)
852 struct drm_i915_gem_pwrite *args = data;
853 struct drm_i915_gem_object *obj;
859 if (!access_ok(VERIFY_READ,
860 to_user_ptr(args->data_ptr),
864 if (likely(!i915_prefault_disable)) {
865 ret = fault_in_multipages_readable(to_user_ptr(args->data_ptr),
871 ret = i915_mutex_lock_interruptible(dev);
875 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
876 if (&obj->base == NULL) {
881 /* Bounds check destination. */
882 if (args->offset > obj->base.size ||
883 args->size > obj->base.size - args->offset) {
888 /* prime objects have no backing filp to GEM pread/pwrite
891 if (!obj->base.filp) {
896 trace_i915_gem_object_pwrite(obj, args->offset, args->size);
899 /* We can only do the GTT pwrite on untiled buffers, as otherwise
900 * it would end up going through the fenced access, and we'll get
901 * different detiling behavior between reading and writing.
902 * pread/pwrite currently are reading and writing from the CPU
903 * perspective, requiring manual detiling by the client.
906 ret = i915_gem_phys_pwrite(dev, obj, args, file);
910 if (obj->cache_level == I915_CACHE_NONE &&
911 obj->tiling_mode == I915_TILING_NONE &&
912 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
913 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
914 /* Note that the gtt paths might fail with non-page-backed user
915 * pointers (e.g. gtt mappings when moving data between
916 * textures). Fallback to the shmem path in that case. */
919 if (ret == -EFAULT || ret == -ENOSPC)
920 ret = i915_gem_shmem_pwrite(dev, obj, args, file);
923 drm_gem_object_unreference(&obj->base);
925 mutex_unlock(&dev->struct_mutex);
930 i915_gem_check_wedge(struct i915_gpu_error *error,
933 if (i915_reset_in_progress(error)) {
934 /* Non-interruptible callers can't handle -EAGAIN, hence return
935 * -EIO unconditionally for these. */
939 /* Recovery complete, but the reset failed ... */
940 if (i915_terminally_wedged(error))
950 * Compare seqno against outstanding lazy request. Emit a request if they are
954 i915_gem_check_olr(struct intel_ring_buffer *ring, u32 seqno)
958 BUG_ON(!mutex_is_locked(&ring->dev->struct_mutex));
961 if (seqno == ring->outstanding_lazy_request)
962 ret = i915_add_request(ring, NULL);
968 * __wait_seqno - wait until execution of seqno has finished
969 * @ring: the ring expected to report seqno
971 * @reset_counter: reset sequence associated with the given seqno
972 * @interruptible: do an interruptible wait (normally yes)
973 * @timeout: in - how long to wait (NULL forever); out - how much time remaining
975 * Note: It is of utmost importance that the passed in seqno and reset_counter
976 * values have been read by the caller in an smp safe manner. Where read-side
977 * locks are involved, it is sufficient to read the reset_counter before
978 * unlocking the lock that protects the seqno. For lockless tricks, the
979 * reset_counter _must_ be read before, and an appropriate smp_rmb must be
982 * Returns 0 if the seqno was found within the alloted time. Else returns the
983 * errno with remaining time filled in timeout argument.
985 static int __wait_seqno(struct intel_ring_buffer *ring, u32 seqno,
986 unsigned reset_counter,
987 bool interruptible, struct timespec *timeout)
989 drm_i915_private_t *dev_priv = ring->dev->dev_private;
990 struct timespec before, now, wait_time={1,0};
991 unsigned long timeout_jiffies;
993 bool wait_forever = true;
996 if (i915_seqno_passed(ring->get_seqno(ring, true), seqno))
999 trace_i915_gem_request_wait_begin(ring, seqno);
1001 if (timeout != NULL) {
1002 wait_time = *timeout;
1003 wait_forever = false;
1006 timeout_jiffies = timespec_to_jiffies_timeout(&wait_time);
1008 if (WARN_ON(!ring->irq_get(ring)))
1011 /* Record current time in case interrupted by signal, or wedged * */
1012 getrawmonotonic(&before);
1015 (i915_seqno_passed(ring->get_seqno(ring, false), seqno) || \
1016 i915_reset_in_progress(&dev_priv->gpu_error) || \
1017 reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter))
1020 end = wait_event_interruptible_timeout(ring->irq_queue,
1024 end = wait_event_timeout(ring->irq_queue, EXIT_COND,
1027 /* We need to check whether any gpu reset happened in between
1028 * the caller grabbing the seqno and now ... */
1029 if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter))
1032 /* ... but upgrade the -EGAIN to an -EIO if the gpu is truely
1034 ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1037 } while (end == 0 && wait_forever);
1039 getrawmonotonic(&now);
1041 ring->irq_put(ring);
1042 trace_i915_gem_request_wait_end(ring, seqno);
1046 struct timespec sleep_time = timespec_sub(now, before);
1047 *timeout = timespec_sub(*timeout, sleep_time);
1048 if (!timespec_valid(timeout)) /* i.e. negative time remains */
1049 set_normalized_timespec(timeout, 0, 0);
1054 case -EAGAIN: /* Wedged */
1055 case -ERESTARTSYS: /* Signal */
1057 case 0: /* Timeout */
1059 default: /* Completed */
1060 WARN_ON(end < 0); /* We're not aware of other errors */
1066 * Waits for a sequence number to be signaled, and cleans up the
1067 * request and object lists appropriately for that event.
1070 i915_wait_seqno(struct intel_ring_buffer *ring, uint32_t seqno)
1072 struct drm_device *dev = ring->dev;
1073 struct drm_i915_private *dev_priv = dev->dev_private;
1074 bool interruptible = dev_priv->mm.interruptible;
1077 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1080 ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1084 ret = i915_gem_check_olr(ring, seqno);
1088 return __wait_seqno(ring, seqno,
1089 atomic_read(&dev_priv->gpu_error.reset_counter),
1090 interruptible, NULL);
1094 i915_gem_object_wait_rendering__tail(struct drm_i915_gem_object *obj,
1095 struct intel_ring_buffer *ring)
1097 i915_gem_retire_requests_ring(ring);
1099 /* Manually manage the write flush as we may have not yet
1100 * retired the buffer.
1102 * Note that the last_write_seqno is always the earlier of
1103 * the two (read/write) seqno, so if we haved successfully waited,
1104 * we know we have passed the last write.
1106 obj->last_write_seqno = 0;
1107 obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
1113 * Ensures that all rendering to the object has completed and the object is
1114 * safe to unbind from the GTT or access from the CPU.
1116 static __must_check int
1117 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
1120 struct intel_ring_buffer *ring = obj->ring;
1124 seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno;
1128 ret = i915_wait_seqno(ring, seqno);
1132 return i915_gem_object_wait_rendering__tail(obj, ring);
1135 /* A nonblocking variant of the above wait. This is a highly dangerous routine
1136 * as the object state may change during this call.
1138 static __must_check int
1139 i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
1142 struct drm_device *dev = obj->base.dev;
1143 struct drm_i915_private *dev_priv = dev->dev_private;
1144 struct intel_ring_buffer *ring = obj->ring;
1145 unsigned reset_counter;
1149 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1150 BUG_ON(!dev_priv->mm.interruptible);
1152 seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno;
1156 ret = i915_gem_check_wedge(&dev_priv->gpu_error, true);
1160 ret = i915_gem_check_olr(ring, seqno);
1164 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
1165 mutex_unlock(&dev->struct_mutex);
1166 ret = __wait_seqno(ring, seqno, reset_counter, true, NULL);
1167 mutex_lock(&dev->struct_mutex);
1171 return i915_gem_object_wait_rendering__tail(obj, ring);
1175 * Called when user space prepares to use an object with the CPU, either
1176 * through the mmap ioctl's mapping or a GTT mapping.
1179 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1180 struct drm_file *file)
1182 struct drm_i915_gem_set_domain *args = data;
1183 struct drm_i915_gem_object *obj;
1184 uint32_t read_domains = args->read_domains;
1185 uint32_t write_domain = args->write_domain;
1188 /* Only handle setting domains to types used by the CPU. */
1189 if (write_domain & I915_GEM_GPU_DOMAINS)
1192 if (read_domains & I915_GEM_GPU_DOMAINS)
1195 /* Having something in the write domain implies it's in the read
1196 * domain, and only that read domain. Enforce that in the request.
1198 if (write_domain != 0 && read_domains != write_domain)
1201 ret = i915_mutex_lock_interruptible(dev);
1205 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1206 if (&obj->base == NULL) {
1211 /* Try to flush the object off the GPU without holding the lock.
1212 * We will repeat the flush holding the lock in the normal manner
1213 * to catch cases where we are gazumped.
1215 ret = i915_gem_object_wait_rendering__nonblocking(obj, !write_domain);
1219 if (read_domains & I915_GEM_DOMAIN_GTT) {
1220 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1222 /* Silently promote "you're not bound, there was nothing to do"
1223 * to success, since the client was just asking us to
1224 * make sure everything was done.
1229 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1233 drm_gem_object_unreference(&obj->base);
1235 mutex_unlock(&dev->struct_mutex);
1240 * Called when user space has done writes to this buffer
1243 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1244 struct drm_file *file)
1246 struct drm_i915_gem_sw_finish *args = data;
1247 struct drm_i915_gem_object *obj;
1250 ret = i915_mutex_lock_interruptible(dev);
1254 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1255 if (&obj->base == NULL) {
1260 /* Pinned buffers may be scanout, so flush the cache */
1262 i915_gem_object_flush_cpu_write_domain(obj);
1264 drm_gem_object_unreference(&obj->base);
1266 mutex_unlock(&dev->struct_mutex);
1271 * Maps the contents of an object, returning the address it is mapped
1274 * While the mapping holds a reference on the contents of the object, it doesn't
1275 * imply a ref on the object itself.
1278 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1279 struct drm_file *file)
1281 struct drm_i915_gem_mmap *args = data;
1282 struct drm_gem_object *obj;
1285 obj = drm_gem_object_lookup(dev, file, args->handle);
1289 /* prime objects have no backing filp to GEM mmap
1293 drm_gem_object_unreference_unlocked(obj);
1297 addr = vm_mmap(obj->filp, 0, args->size,
1298 PROT_READ | PROT_WRITE, MAP_SHARED,
1300 drm_gem_object_unreference_unlocked(obj);
1301 if (IS_ERR((void *)addr))
1304 args->addr_ptr = (uint64_t) addr;
1310 * i915_gem_fault - fault a page into the GTT
1311 * vma: VMA in question
1314 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1315 * from userspace. The fault handler takes care of binding the object to
1316 * the GTT (if needed), allocating and programming a fence register (again,
1317 * only if needed based on whether the old reg is still valid or the object
1318 * is tiled) and inserting a new PTE into the faulting process.
1320 * Note that the faulting process may involve evicting existing objects
1321 * from the GTT and/or fence registers to make room. So performance may
1322 * suffer if the GTT working set is large or there are few fence registers
1325 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1327 struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1328 struct drm_device *dev = obj->base.dev;
1329 drm_i915_private_t *dev_priv = dev->dev_private;
1330 pgoff_t page_offset;
1333 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1335 /* We don't use vmf->pgoff since that has the fake offset */
1336 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1339 ret = i915_mutex_lock_interruptible(dev);
1343 trace_i915_gem_object_fault(obj, page_offset, true, write);
1345 /* Access to snoopable pages through the GTT is incoherent. */
1346 if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) {
1351 /* Now bind it into the GTT if needed */
1352 ret = i915_gem_object_pin(obj, 0, true, false);
1356 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1360 ret = i915_gem_object_get_fence(obj);
1364 obj->fault_mappable = true;
1366 pfn = dev_priv->gtt.mappable_base + i915_gem_obj_ggtt_offset(obj);
1370 /* Finally, remap it using the new GTT offset */
1371 ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1373 i915_gem_object_unpin(obj);
1375 mutex_unlock(&dev->struct_mutex);
1379 /* If this -EIO is due to a gpu hang, give the reset code a
1380 * chance to clean up the mess. Otherwise return the proper
1382 if (i915_terminally_wedged(&dev_priv->gpu_error))
1383 return VM_FAULT_SIGBUS;
1385 /* Give the error handler a chance to run and move the
1386 * objects off the GPU active list. Next time we service the
1387 * fault, we should be able to transition the page into the
1388 * GTT without touching the GPU (and so avoid further
1389 * EIO/EGAIN). If the GPU is wedged, then there is no issue
1390 * with coherency, just lost writes.
1398 * EBUSY is ok: this just means that another thread
1399 * already did the job.
1401 return VM_FAULT_NOPAGE;
1403 return VM_FAULT_OOM;
1405 return VM_FAULT_SIGBUS;
1407 WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret);
1408 return VM_FAULT_SIGBUS;
1413 * i915_gem_release_mmap - remove physical page mappings
1414 * @obj: obj in question
1416 * Preserve the reservation of the mmapping with the DRM core code, but
1417 * relinquish ownership of the pages back to the system.
1419 * It is vital that we remove the page mapping if we have mapped a tiled
1420 * object through the GTT and then lose the fence register due to
1421 * resource pressure. Similarly if the object has been moved out of the
1422 * aperture, than pages mapped into userspace must be revoked. Removing the
1423 * mapping will then trigger a page fault on the next user access, allowing
1424 * fixup by i915_gem_fault().
1427 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1429 if (!obj->fault_mappable)
1432 drm_vma_node_unmap(&obj->base.vma_node, obj->base.dev->dev_mapping);
1433 obj->fault_mappable = false;
1437 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1441 if (INTEL_INFO(dev)->gen >= 4 ||
1442 tiling_mode == I915_TILING_NONE)
1445 /* Previous chips need a power-of-two fence region when tiling */
1446 if (INTEL_INFO(dev)->gen == 3)
1447 gtt_size = 1024*1024;
1449 gtt_size = 512*1024;
1451 while (gtt_size < size)
1458 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1459 * @obj: object to check
1461 * Return the required GTT alignment for an object, taking into account
1462 * potential fence register mapping.
1465 i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size,
1466 int tiling_mode, bool fenced)
1469 * Minimum alignment is 4k (GTT page size), but might be greater
1470 * if a fence register is needed for the object.
1472 if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) ||
1473 tiling_mode == I915_TILING_NONE)
1477 * Previous chips need to be aligned to the size of the smallest
1478 * fence register that can contain the object.
1480 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1483 static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj)
1485 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1488 if (drm_vma_node_has_offset(&obj->base.vma_node))
1491 dev_priv->mm.shrinker_no_lock_stealing = true;
1493 ret = drm_gem_create_mmap_offset(&obj->base);
1497 /* Badly fragmented mmap space? The only way we can recover
1498 * space is by destroying unwanted objects. We can't randomly release
1499 * mmap_offsets as userspace expects them to be persistent for the
1500 * lifetime of the objects. The closest we can is to release the
1501 * offsets on purgeable objects by truncating it and marking it purged,
1502 * which prevents userspace from ever using that object again.
1504 i915_gem_purge(dev_priv, obj->base.size >> PAGE_SHIFT);
1505 ret = drm_gem_create_mmap_offset(&obj->base);
1509 i915_gem_shrink_all(dev_priv);
1510 ret = drm_gem_create_mmap_offset(&obj->base);
1512 dev_priv->mm.shrinker_no_lock_stealing = false;
1517 static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)
1519 drm_gem_free_mmap_offset(&obj->base);
1523 i915_gem_mmap_gtt(struct drm_file *file,
1524 struct drm_device *dev,
1528 struct drm_i915_private *dev_priv = dev->dev_private;
1529 struct drm_i915_gem_object *obj;
1532 ret = i915_mutex_lock_interruptible(dev);
1536 obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
1537 if (&obj->base == NULL) {
1542 if (obj->base.size > dev_priv->gtt.mappable_end) {
1547 if (obj->madv != I915_MADV_WILLNEED) {
1548 DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1553 ret = i915_gem_object_create_mmap_offset(obj);
1557 *offset = drm_vma_node_offset_addr(&obj->base.vma_node);
1560 drm_gem_object_unreference(&obj->base);
1562 mutex_unlock(&dev->struct_mutex);
1567 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1569 * @data: GTT mapping ioctl data
1570 * @file: GEM object info
1572 * Simply returns the fake offset to userspace so it can mmap it.
1573 * The mmap call will end up in drm_gem_mmap(), which will set things
1574 * up so we can get faults in the handler above.
1576 * The fault handler will take care of binding the object into the GTT
1577 * (since it may have been evicted to make room for something), allocating
1578 * a fence register, and mapping the appropriate aperture address into
1582 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1583 struct drm_file *file)
1585 struct drm_i915_gem_mmap_gtt *args = data;
1587 return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
1590 /* Immediately discard the backing storage */
1592 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
1594 struct inode *inode;
1596 i915_gem_object_free_mmap_offset(obj);
1598 if (obj->base.filp == NULL)
1601 /* Our goal here is to return as much of the memory as
1602 * is possible back to the system as we are called from OOM.
1603 * To do this we must instruct the shmfs to drop all of its
1604 * backing pages, *now*.
1606 inode = file_inode(obj->base.filp);
1607 shmem_truncate_range(inode, 0, (loff_t)-1);
1609 obj->madv = __I915_MADV_PURGED;
1613 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj)
1615 return obj->madv == I915_MADV_DONTNEED;
1619 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
1621 struct sg_page_iter sg_iter;
1624 BUG_ON(obj->madv == __I915_MADV_PURGED);
1626 ret = i915_gem_object_set_to_cpu_domain(obj, true);
1628 /* In the event of a disaster, abandon all caches and
1629 * hope for the best.
1631 WARN_ON(ret != -EIO);
1632 i915_gem_clflush_object(obj);
1633 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
1636 if (i915_gem_object_needs_bit17_swizzle(obj))
1637 i915_gem_object_save_bit_17_swizzle(obj);
1639 if (obj->madv == I915_MADV_DONTNEED)
1642 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
1643 struct page *page = sg_page_iter_page(&sg_iter);
1646 set_page_dirty(page);
1648 if (obj->madv == I915_MADV_WILLNEED)
1649 mark_page_accessed(page);
1651 page_cache_release(page);
1655 sg_free_table(obj->pages);
1660 i915_gem_object_put_pages(struct drm_i915_gem_object *obj)
1662 const struct drm_i915_gem_object_ops *ops = obj->ops;
1664 if (obj->pages == NULL)
1667 BUG_ON(i915_gem_obj_ggtt_bound(obj));
1669 if (obj->pages_pin_count)
1672 /* ->put_pages might need to allocate memory for the bit17 swizzle
1673 * array, hence protect them from being reaped by removing them from gtt
1675 list_del(&obj->global_list);
1677 ops->put_pages(obj);
1680 if (i915_gem_object_is_purgeable(obj))
1681 i915_gem_object_truncate(obj);
1687 __i915_gem_shrink(struct drm_i915_private *dev_priv, long target,
1688 bool purgeable_only)
1690 struct drm_i915_gem_object *obj, *next;
1691 struct i915_address_space *vm = &dev_priv->gtt.base;
1694 list_for_each_entry_safe(obj, next,
1695 &dev_priv->mm.unbound_list,
1697 if ((i915_gem_object_is_purgeable(obj) || !purgeable_only) &&
1698 i915_gem_object_put_pages(obj) == 0) {
1699 count += obj->base.size >> PAGE_SHIFT;
1700 if (count >= target)
1705 list_for_each_entry_safe(obj, next, &vm->inactive_list, mm_list) {
1706 if ((i915_gem_object_is_purgeable(obj) || !purgeable_only) &&
1707 i915_gem_object_unbind(obj) == 0 &&
1708 i915_gem_object_put_pages(obj) == 0) {
1709 count += obj->base.size >> PAGE_SHIFT;
1710 if (count >= target)
1719 i915_gem_purge(struct drm_i915_private *dev_priv, long target)
1721 return __i915_gem_shrink(dev_priv, target, true);
1725 i915_gem_shrink_all(struct drm_i915_private *dev_priv)
1727 struct drm_i915_gem_object *obj, *next;
1730 i915_gem_evict_everything(dev_priv->dev);
1732 list_for_each_entry_safe(obj, next, &dev_priv->mm.unbound_list,
1734 if (obj->pages_pin_count == 0)
1735 freed += obj->base.size >> PAGE_SHIFT;
1736 i915_gem_object_put_pages(obj);
1742 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
1744 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1746 struct address_space *mapping;
1747 struct sg_table *st;
1748 struct scatterlist *sg;
1749 struct sg_page_iter sg_iter;
1751 unsigned long last_pfn = 0; /* suppress gcc warning */
1754 /* Assert that the object is not currently in any GPU domain. As it
1755 * wasn't in the GTT, there shouldn't be any way it could have been in
1758 BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
1759 BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
1761 st = kmalloc(sizeof(*st), GFP_KERNEL);
1765 page_count = obj->base.size / PAGE_SIZE;
1766 if (sg_alloc_table(st, page_count, GFP_KERNEL)) {
1772 /* Get the list of pages out of our struct file. They'll be pinned
1773 * at this point until we release them.
1775 * Fail silently without starting the shrinker
1777 mapping = file_inode(obj->base.filp)->i_mapping;
1778 gfp = mapping_gfp_mask(mapping);
1779 gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD;
1780 gfp &= ~(__GFP_IO | __GFP_WAIT);
1783 for (i = 0; i < page_count; i++) {
1784 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
1786 i915_gem_purge(dev_priv, page_count);
1787 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
1790 /* We've tried hard to allocate the memory by reaping
1791 * our own buffer, now let the real VM do its job and
1792 * go down in flames if truly OOM.
1794 gfp &= ~(__GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD);
1795 gfp |= __GFP_IO | __GFP_WAIT;
1797 i915_gem_shrink_all(dev_priv);
1798 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
1802 gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD;
1803 gfp &= ~(__GFP_IO | __GFP_WAIT);
1805 #ifdef CONFIG_SWIOTLB
1806 if (swiotlb_nr_tbl()) {
1808 sg_set_page(sg, page, PAGE_SIZE, 0);
1813 if (!i || page_to_pfn(page) != last_pfn + 1) {
1817 sg_set_page(sg, page, PAGE_SIZE, 0);
1819 sg->length += PAGE_SIZE;
1821 last_pfn = page_to_pfn(page);
1823 #ifdef CONFIG_SWIOTLB
1824 if (!swiotlb_nr_tbl())
1829 if (i915_gem_object_needs_bit17_swizzle(obj))
1830 i915_gem_object_do_bit_17_swizzle(obj);
1836 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0)
1837 page_cache_release(sg_page_iter_page(&sg_iter));
1840 return PTR_ERR(page);
1843 /* Ensure that the associated pages are gathered from the backing storage
1844 * and pinned into our object. i915_gem_object_get_pages() may be called
1845 * multiple times before they are released by a single call to
1846 * i915_gem_object_put_pages() - once the pages are no longer referenced
1847 * either as a result of memory pressure (reaping pages under the shrinker)
1848 * or as the object is itself released.
1851 i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
1853 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1854 const struct drm_i915_gem_object_ops *ops = obj->ops;
1860 if (obj->madv != I915_MADV_WILLNEED) {
1861 DRM_ERROR("Attempting to obtain a purgeable object\n");
1865 BUG_ON(obj->pages_pin_count);
1867 ret = ops->get_pages(obj);
1871 list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list);
1876 i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
1877 struct intel_ring_buffer *ring)
1879 struct drm_device *dev = obj->base.dev;
1880 struct drm_i915_private *dev_priv = dev->dev_private;
1881 struct i915_address_space *vm = &dev_priv->gtt.base;
1882 u32 seqno = intel_ring_get_seqno(ring);
1884 BUG_ON(ring == NULL);
1885 if (obj->ring != ring && obj->last_write_seqno) {
1886 /* Keep the seqno relative to the current ring */
1887 obj->last_write_seqno = seqno;
1891 /* Add a reference if we're newly entering the active list. */
1893 drm_gem_object_reference(&obj->base);
1897 /* Move from whatever list we were on to the tail of execution. */
1898 list_move_tail(&obj->mm_list, &vm->active_list);
1899 list_move_tail(&obj->ring_list, &ring->active_list);
1901 obj->last_read_seqno = seqno;
1903 if (obj->fenced_gpu_access) {
1904 obj->last_fenced_seqno = seqno;
1906 /* Bump MRU to take account of the delayed flush */
1907 if (obj->fence_reg != I915_FENCE_REG_NONE) {
1908 struct drm_i915_fence_reg *reg;
1910 reg = &dev_priv->fence_regs[obj->fence_reg];
1911 list_move_tail(®->lru_list,
1912 &dev_priv->mm.fence_list);
1918 i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
1920 struct drm_device *dev = obj->base.dev;
1921 struct drm_i915_private *dev_priv = dev->dev_private;
1922 struct i915_address_space *vm = &dev_priv->gtt.base;
1924 BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS);
1925 BUG_ON(!obj->active);
1927 list_move_tail(&obj->mm_list, &vm->inactive_list);
1929 list_del_init(&obj->ring_list);
1932 obj->last_read_seqno = 0;
1933 obj->last_write_seqno = 0;
1934 obj->base.write_domain = 0;
1936 obj->last_fenced_seqno = 0;
1937 obj->fenced_gpu_access = false;
1940 drm_gem_object_unreference(&obj->base);
1942 WARN_ON(i915_verify_lists(dev));
1946 i915_gem_init_seqno(struct drm_device *dev, u32 seqno)
1948 struct drm_i915_private *dev_priv = dev->dev_private;
1949 struct intel_ring_buffer *ring;
1952 /* Carefully retire all requests without writing to the rings */
1953 for_each_ring(ring, dev_priv, i) {
1954 ret = intel_ring_idle(ring);
1958 i915_gem_retire_requests(dev);
1960 /* Finally reset hw state */
1961 for_each_ring(ring, dev_priv, i) {
1962 intel_ring_init_seqno(ring, seqno);
1964 for (j = 0; j < ARRAY_SIZE(ring->sync_seqno); j++)
1965 ring->sync_seqno[j] = 0;
1971 int i915_gem_set_seqno(struct drm_device *dev, u32 seqno)
1973 struct drm_i915_private *dev_priv = dev->dev_private;
1979 /* HWS page needs to be set less than what we
1980 * will inject to ring
1982 ret = i915_gem_init_seqno(dev, seqno - 1);
1986 /* Carefully set the last_seqno value so that wrap
1987 * detection still works
1989 dev_priv->next_seqno = seqno;
1990 dev_priv->last_seqno = seqno - 1;
1991 if (dev_priv->last_seqno == 0)
1992 dev_priv->last_seqno--;
1998 i915_gem_get_seqno(struct drm_device *dev, u32 *seqno)
2000 struct drm_i915_private *dev_priv = dev->dev_private;
2002 /* reserve 0 for non-seqno */
2003 if (dev_priv->next_seqno == 0) {
2004 int ret = i915_gem_init_seqno(dev, 0);
2008 dev_priv->next_seqno = 1;
2011 *seqno = dev_priv->last_seqno = dev_priv->next_seqno++;
2015 int __i915_add_request(struct intel_ring_buffer *ring,
2016 struct drm_file *file,
2017 struct drm_i915_gem_object *obj,
2020 drm_i915_private_t *dev_priv = ring->dev->dev_private;
2021 struct drm_i915_gem_request *request;
2022 u32 request_ring_position, request_start;
2026 request_start = intel_ring_get_tail(ring);
2028 * Emit any outstanding flushes - execbuf can fail to emit the flush
2029 * after having emitted the batchbuffer command. Hence we need to fix
2030 * things up similar to emitting the lazy request. The difference here
2031 * is that the flush _must_ happen before the next request, no matter
2034 ret = intel_ring_flush_all_caches(ring);
2038 request = kmalloc(sizeof(*request), GFP_KERNEL);
2039 if (request == NULL)
2043 /* Record the position of the start of the request so that
2044 * should we detect the updated seqno part-way through the
2045 * GPU processing the request, we never over-estimate the
2046 * position of the head.
2048 request_ring_position = intel_ring_get_tail(ring);
2050 ret = ring->add_request(ring);
2056 request->seqno = intel_ring_get_seqno(ring);
2057 request->ring = ring;
2058 request->head = request_start;
2059 request->tail = request_ring_position;
2060 request->ctx = ring->last_context;
2061 request->batch_obj = obj;
2063 /* Whilst this request exists, batch_obj will be on the
2064 * active_list, and so will hold the active reference. Only when this
2065 * request is retired will the the batch_obj be moved onto the
2066 * inactive_list and lose its active reference. Hence we do not need
2067 * to explicitly hold another reference here.
2071 i915_gem_context_reference(request->ctx);
2073 request->emitted_jiffies = jiffies;
2074 was_empty = list_empty(&ring->request_list);
2075 list_add_tail(&request->list, &ring->request_list);
2076 request->file_priv = NULL;
2079 struct drm_i915_file_private *file_priv = file->driver_priv;
2081 spin_lock(&file_priv->mm.lock);
2082 request->file_priv = file_priv;
2083 list_add_tail(&request->client_list,
2084 &file_priv->mm.request_list);
2085 spin_unlock(&file_priv->mm.lock);
2088 trace_i915_gem_request_add(ring, request->seqno);
2089 ring->outstanding_lazy_request = 0;
2091 if (!dev_priv->ums.mm_suspended) {
2092 i915_queue_hangcheck(ring->dev);
2095 queue_delayed_work(dev_priv->wq,
2096 &dev_priv->mm.retire_work,
2097 round_jiffies_up_relative(HZ));
2098 intel_mark_busy(dev_priv->dev);
2103 *out_seqno = request->seqno;
2108 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
2110 struct drm_i915_file_private *file_priv = request->file_priv;
2115 spin_lock(&file_priv->mm.lock);
2116 if (request->file_priv) {
2117 list_del(&request->client_list);
2118 request->file_priv = NULL;
2120 spin_unlock(&file_priv->mm.lock);
2123 static bool i915_head_inside_object(u32 acthd, struct drm_i915_gem_object *obj)
2125 if (acthd >= i915_gem_obj_ggtt_offset(obj) &&
2126 acthd < i915_gem_obj_ggtt_offset(obj) + obj->base.size)
2132 static bool i915_head_inside_request(const u32 acthd_unmasked,
2133 const u32 request_start,
2134 const u32 request_end)
2136 const u32 acthd = acthd_unmasked & HEAD_ADDR;
2138 if (request_start < request_end) {
2139 if (acthd >= request_start && acthd < request_end)
2141 } else if (request_start > request_end) {
2142 if (acthd >= request_start || acthd < request_end)
2149 static bool i915_request_guilty(struct drm_i915_gem_request *request,
2150 const u32 acthd, bool *inside)
2152 /* There is a possibility that unmasked head address
2153 * pointing inside the ring, matches the batch_obj address range.
2154 * However this is extremely unlikely.
2157 if (request->batch_obj) {
2158 if (i915_head_inside_object(acthd, request->batch_obj)) {
2164 if (i915_head_inside_request(acthd, request->head, request->tail)) {
2172 static void i915_set_reset_status(struct intel_ring_buffer *ring,
2173 struct drm_i915_gem_request *request,
2176 struct i915_ctx_hang_stats *hs = NULL;
2177 bool inside, guilty;
2179 /* Innocent until proven guilty */
2182 if (ring->hangcheck.action != wait &&
2183 i915_request_guilty(request, acthd, &inside)) {
2184 DRM_ERROR("%s hung %s bo (0x%lx ctx %d) at 0x%x\n",
2186 inside ? "inside" : "flushing",
2187 request->batch_obj ?
2188 i915_gem_obj_ggtt_offset(request->batch_obj) : 0,
2189 request->ctx ? request->ctx->id : 0,
2195 /* If contexts are disabled or this is the default context, use
2196 * file_priv->reset_state
2198 if (request->ctx && request->ctx->id != DEFAULT_CONTEXT_ID)
2199 hs = &request->ctx->hang_stats;
2200 else if (request->file_priv)
2201 hs = &request->file_priv->hang_stats;
2207 hs->batch_pending++;
2211 static void i915_gem_free_request(struct drm_i915_gem_request *request)
2213 list_del(&request->list);
2214 i915_gem_request_remove_from_client(request);
2217 i915_gem_context_unreference(request->ctx);
2222 static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv,
2223 struct intel_ring_buffer *ring)
2225 u32 completed_seqno;
2228 acthd = intel_ring_get_active_head(ring);
2229 completed_seqno = ring->get_seqno(ring, false);
2231 while (!list_empty(&ring->request_list)) {
2232 struct drm_i915_gem_request *request;
2234 request = list_first_entry(&ring->request_list,
2235 struct drm_i915_gem_request,
2238 if (request->seqno > completed_seqno)
2239 i915_set_reset_status(ring, request, acthd);
2241 i915_gem_free_request(request);
2244 while (!list_empty(&ring->active_list)) {
2245 struct drm_i915_gem_object *obj;
2247 obj = list_first_entry(&ring->active_list,
2248 struct drm_i915_gem_object,
2251 i915_gem_object_move_to_inactive(obj);
2255 void i915_gem_restore_fences(struct drm_device *dev)
2257 struct drm_i915_private *dev_priv = dev->dev_private;
2260 for (i = 0; i < dev_priv->num_fence_regs; i++) {
2261 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
2264 * Commit delayed tiling changes if we have an object still
2265 * attached to the fence, otherwise just clear the fence.
2268 i915_gem_object_update_fence(reg->obj, reg,
2269 reg->obj->tiling_mode);
2271 i915_gem_write_fence(dev, i, NULL);
2276 void i915_gem_reset(struct drm_device *dev)
2278 struct drm_i915_private *dev_priv = dev->dev_private;
2279 struct i915_address_space *vm = &dev_priv->gtt.base;
2280 struct drm_i915_gem_object *obj;
2281 struct intel_ring_buffer *ring;
2284 for_each_ring(ring, dev_priv, i)
2285 i915_gem_reset_ring_lists(dev_priv, ring);
2287 /* Move everything out of the GPU domains to ensure we do any
2288 * necessary invalidation upon reuse.
2290 list_for_each_entry(obj, &vm->inactive_list, mm_list)
2291 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
2293 i915_gem_restore_fences(dev);
2297 * This function clears the request list as sequence numbers are passed.
2300 i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
2304 if (list_empty(&ring->request_list))
2307 WARN_ON(i915_verify_lists(ring->dev));
2309 seqno = ring->get_seqno(ring, true);
2311 while (!list_empty(&ring->request_list)) {
2312 struct drm_i915_gem_request *request;
2314 request = list_first_entry(&ring->request_list,
2315 struct drm_i915_gem_request,
2318 if (!i915_seqno_passed(seqno, request->seqno))
2321 trace_i915_gem_request_retire(ring, request->seqno);
2322 /* We know the GPU must have read the request to have
2323 * sent us the seqno + interrupt, so use the position
2324 * of tail of the request to update the last known position
2327 ring->last_retired_head = request->tail;
2329 i915_gem_free_request(request);
2332 /* Move any buffers on the active list that are no longer referenced
2333 * by the ringbuffer to the flushing/inactive lists as appropriate.
2335 while (!list_empty(&ring->active_list)) {
2336 struct drm_i915_gem_object *obj;
2338 obj = list_first_entry(&ring->active_list,
2339 struct drm_i915_gem_object,
2342 if (!i915_seqno_passed(seqno, obj->last_read_seqno))
2345 i915_gem_object_move_to_inactive(obj);
2348 if (unlikely(ring->trace_irq_seqno &&
2349 i915_seqno_passed(seqno, ring->trace_irq_seqno))) {
2350 ring->irq_put(ring);
2351 ring->trace_irq_seqno = 0;
2354 WARN_ON(i915_verify_lists(ring->dev));
2358 i915_gem_retire_requests(struct drm_device *dev)
2360 drm_i915_private_t *dev_priv = dev->dev_private;
2361 struct intel_ring_buffer *ring;
2364 for_each_ring(ring, dev_priv, i)
2365 i915_gem_retire_requests_ring(ring);
2369 i915_gem_retire_work_handler(struct work_struct *work)
2371 drm_i915_private_t *dev_priv;
2372 struct drm_device *dev;
2373 struct intel_ring_buffer *ring;
2377 dev_priv = container_of(work, drm_i915_private_t,
2378 mm.retire_work.work);
2379 dev = dev_priv->dev;
2381 /* Come back later if the device is busy... */
2382 if (!mutex_trylock(&dev->struct_mutex)) {
2383 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
2384 round_jiffies_up_relative(HZ));
2388 i915_gem_retire_requests(dev);
2390 /* Send a periodic flush down the ring so we don't hold onto GEM
2391 * objects indefinitely.
2394 for_each_ring(ring, dev_priv, i) {
2395 if (ring->gpu_caches_dirty)
2396 i915_add_request(ring, NULL);
2398 idle &= list_empty(&ring->request_list);
2401 if (!dev_priv->ums.mm_suspended && !idle)
2402 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
2403 round_jiffies_up_relative(HZ));
2405 intel_mark_idle(dev);
2407 mutex_unlock(&dev->struct_mutex);
2411 * Ensures that an object will eventually get non-busy by flushing any required
2412 * write domains, emitting any outstanding lazy request and retiring and
2413 * completed requests.
2416 i915_gem_object_flush_active(struct drm_i915_gem_object *obj)
2421 ret = i915_gem_check_olr(obj->ring, obj->last_read_seqno);
2425 i915_gem_retire_requests_ring(obj->ring);
2432 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
2433 * @DRM_IOCTL_ARGS: standard ioctl arguments
2435 * Returns 0 if successful, else an error is returned with the remaining time in
2436 * the timeout parameter.
2437 * -ETIME: object is still busy after timeout
2438 * -ERESTARTSYS: signal interrupted the wait
2439 * -ENONENT: object doesn't exist
2440 * Also possible, but rare:
2441 * -EAGAIN: GPU wedged
2443 * -ENODEV: Internal IRQ fail
2444 * -E?: The add request failed
2446 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
2447 * non-zero timeout parameter the wait ioctl will wait for the given number of
2448 * nanoseconds on an object becoming unbusy. Since the wait itself does so
2449 * without holding struct_mutex the object may become re-busied before this
2450 * function completes. A similar but shorter * race condition exists in the busy
2454 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
2456 drm_i915_private_t *dev_priv = dev->dev_private;
2457 struct drm_i915_gem_wait *args = data;
2458 struct drm_i915_gem_object *obj;
2459 struct intel_ring_buffer *ring = NULL;
2460 struct timespec timeout_stack, *timeout = NULL;
2461 unsigned reset_counter;
2465 if (args->timeout_ns >= 0) {
2466 timeout_stack = ns_to_timespec(args->timeout_ns);
2467 timeout = &timeout_stack;
2470 ret = i915_mutex_lock_interruptible(dev);
2474 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle));
2475 if (&obj->base == NULL) {
2476 mutex_unlock(&dev->struct_mutex);
2480 /* Need to make sure the object gets inactive eventually. */
2481 ret = i915_gem_object_flush_active(obj);
2486 seqno = obj->last_read_seqno;
2493 /* Do this after OLR check to make sure we make forward progress polling
2494 * on this IOCTL with a 0 timeout (like busy ioctl)
2496 if (!args->timeout_ns) {
2501 drm_gem_object_unreference(&obj->base);
2502 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
2503 mutex_unlock(&dev->struct_mutex);
2505 ret = __wait_seqno(ring, seqno, reset_counter, true, timeout);
2507 args->timeout_ns = timespec_to_ns(timeout);
2511 drm_gem_object_unreference(&obj->base);
2512 mutex_unlock(&dev->struct_mutex);
2517 * i915_gem_object_sync - sync an object to a ring.
2519 * @obj: object which may be in use on another ring.
2520 * @to: ring we wish to use the object on. May be NULL.
2522 * This code is meant to abstract object synchronization with the GPU.
2523 * Calling with NULL implies synchronizing the object with the CPU
2524 * rather than a particular GPU ring.
2526 * Returns 0 if successful, else propagates up the lower layer error.
2529 i915_gem_object_sync(struct drm_i915_gem_object *obj,
2530 struct intel_ring_buffer *to)
2532 struct intel_ring_buffer *from = obj->ring;
2536 if (from == NULL || to == from)
2539 if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev))
2540 return i915_gem_object_wait_rendering(obj, false);
2542 idx = intel_ring_sync_index(from, to);
2544 seqno = obj->last_read_seqno;
2545 if (seqno <= from->sync_seqno[idx])
2548 ret = i915_gem_check_olr(obj->ring, seqno);
2552 ret = to->sync_to(to, from, seqno);
2554 /* We use last_read_seqno because sync_to()
2555 * might have just caused seqno wrap under
2558 from->sync_seqno[idx] = obj->last_read_seqno;
2563 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
2565 u32 old_write_domain, old_read_domains;
2567 /* Force a pagefault for domain tracking on next user access */
2568 i915_gem_release_mmap(obj);
2570 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
2573 /* Wait for any direct GTT access to complete */
2576 old_read_domains = obj->base.read_domains;
2577 old_write_domain = obj->base.write_domain;
2579 obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
2580 obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
2582 trace_i915_gem_object_change_domain(obj,
2588 * Unbinds an object from the GTT aperture.
2591 i915_gem_object_unbind(struct drm_i915_gem_object *obj)
2593 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
2594 struct i915_vma *vma;
2597 if (!i915_gem_obj_ggtt_bound(obj))
2603 BUG_ON(obj->pages == NULL);
2605 ret = i915_gem_object_finish_gpu(obj);
2608 /* Continue on if we fail due to EIO, the GPU is hung so we
2609 * should be safe and we need to cleanup or else we might
2610 * cause memory corruption through use-after-free.
2613 i915_gem_object_finish_gtt(obj);
2615 /* release the fence reg _after_ flushing */
2616 ret = i915_gem_object_put_fence(obj);
2620 trace_i915_gem_object_unbind(obj);
2622 if (obj->has_global_gtt_mapping)
2623 i915_gem_gtt_unbind_object(obj);
2624 if (obj->has_aliasing_ppgtt_mapping) {
2625 i915_ppgtt_unbind_object(dev_priv->mm.aliasing_ppgtt, obj);
2626 obj->has_aliasing_ppgtt_mapping = 0;
2628 i915_gem_gtt_finish_object(obj);
2629 i915_gem_object_unpin_pages(obj);
2631 list_del(&obj->mm_list);
2632 /* Avoid an unnecessary call to unbind on rebind. */
2633 obj->map_and_fenceable = true;
2635 vma = __i915_gem_obj_to_vma(obj);
2636 list_del(&vma->vma_link);
2637 drm_mm_remove_node(&vma->node);
2638 i915_gem_vma_destroy(vma);
2640 /* Since the unbound list is global, only move to that list if
2641 * no more VMAs exist.
2642 * NB: Until we have real VMAs there will only ever be one */
2643 WARN_ON(!list_empty(&obj->vma_list));
2644 if (list_empty(&obj->vma_list))
2645 list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list);
2650 int i915_gpu_idle(struct drm_device *dev)
2652 drm_i915_private_t *dev_priv = dev->dev_private;
2653 struct intel_ring_buffer *ring;
2656 /* Flush everything onto the inactive list. */
2657 for_each_ring(ring, dev_priv, i) {
2658 ret = i915_switch_context(ring, NULL, DEFAULT_CONTEXT_ID);
2662 ret = intel_ring_idle(ring);
2670 static void i965_write_fence_reg(struct drm_device *dev, int reg,
2671 struct drm_i915_gem_object *obj)
2673 drm_i915_private_t *dev_priv = dev->dev_private;
2675 int fence_pitch_shift;
2677 if (INTEL_INFO(dev)->gen >= 6) {
2678 fence_reg = FENCE_REG_SANDYBRIDGE_0;
2679 fence_pitch_shift = SANDYBRIDGE_FENCE_PITCH_SHIFT;
2681 fence_reg = FENCE_REG_965_0;
2682 fence_pitch_shift = I965_FENCE_PITCH_SHIFT;
2685 fence_reg += reg * 8;
2687 /* To w/a incoherency with non-atomic 64-bit register updates,
2688 * we split the 64-bit update into two 32-bit writes. In order
2689 * for a partial fence not to be evaluated between writes, we
2690 * precede the update with write to turn off the fence register,
2691 * and only enable the fence as the last step.
2693 * For extra levels of paranoia, we make sure each step lands
2694 * before applying the next step.
2696 I915_WRITE(fence_reg, 0);
2697 POSTING_READ(fence_reg);
2700 u32 size = i915_gem_obj_ggtt_size(obj);
2703 val = (uint64_t)((i915_gem_obj_ggtt_offset(obj) + size - 4096) &
2705 val |= i915_gem_obj_ggtt_offset(obj) & 0xfffff000;
2706 val |= (uint64_t)((obj->stride / 128) - 1) << fence_pitch_shift;
2707 if (obj->tiling_mode == I915_TILING_Y)
2708 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2709 val |= I965_FENCE_REG_VALID;
2711 I915_WRITE(fence_reg + 4, val >> 32);
2712 POSTING_READ(fence_reg + 4);
2714 I915_WRITE(fence_reg + 0, val);
2715 POSTING_READ(fence_reg);
2717 I915_WRITE(fence_reg + 4, 0);
2718 POSTING_READ(fence_reg + 4);
2722 static void i915_write_fence_reg(struct drm_device *dev, int reg,
2723 struct drm_i915_gem_object *obj)
2725 drm_i915_private_t *dev_priv = dev->dev_private;
2729 u32 size = i915_gem_obj_ggtt_size(obj);
2733 WARN((i915_gem_obj_ggtt_offset(obj) & ~I915_FENCE_START_MASK) ||
2734 (size & -size) != size ||
2735 (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
2736 "object 0x%08lx [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
2737 i915_gem_obj_ggtt_offset(obj), obj->map_and_fenceable, size);
2739 if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
2744 /* Note: pitch better be a power of two tile widths */
2745 pitch_val = obj->stride / tile_width;
2746 pitch_val = ffs(pitch_val) - 1;
2748 val = i915_gem_obj_ggtt_offset(obj);
2749 if (obj->tiling_mode == I915_TILING_Y)
2750 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2751 val |= I915_FENCE_SIZE_BITS(size);
2752 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2753 val |= I830_FENCE_REG_VALID;
2758 reg = FENCE_REG_830_0 + reg * 4;
2760 reg = FENCE_REG_945_8 + (reg - 8) * 4;
2762 I915_WRITE(reg, val);
2766 static void i830_write_fence_reg(struct drm_device *dev, int reg,
2767 struct drm_i915_gem_object *obj)
2769 drm_i915_private_t *dev_priv = dev->dev_private;
2773 u32 size = i915_gem_obj_ggtt_size(obj);
2776 WARN((i915_gem_obj_ggtt_offset(obj) & ~I830_FENCE_START_MASK) ||
2777 (size & -size) != size ||
2778 (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
2779 "object 0x%08lx not 512K or pot-size 0x%08x aligned\n",
2780 i915_gem_obj_ggtt_offset(obj), size);
2782 pitch_val = obj->stride / 128;
2783 pitch_val = ffs(pitch_val) - 1;
2785 val = i915_gem_obj_ggtt_offset(obj);
2786 if (obj->tiling_mode == I915_TILING_Y)
2787 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2788 val |= I830_FENCE_SIZE_BITS(size);
2789 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2790 val |= I830_FENCE_REG_VALID;
2794 I915_WRITE(FENCE_REG_830_0 + reg * 4, val);
2795 POSTING_READ(FENCE_REG_830_0 + reg * 4);
2798 inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object *obj)
2800 return obj && obj->base.read_domains & I915_GEM_DOMAIN_GTT;
2803 static void i915_gem_write_fence(struct drm_device *dev, int reg,
2804 struct drm_i915_gem_object *obj)
2806 struct drm_i915_private *dev_priv = dev->dev_private;
2808 /* Ensure that all CPU reads are completed before installing a fence
2809 * and all writes before removing the fence.
2811 if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj))
2814 WARN(obj && (!obj->stride || !obj->tiling_mode),
2815 "bogus fence setup with stride: 0x%x, tiling mode: %i\n",
2816 obj->stride, obj->tiling_mode);
2818 switch (INTEL_INFO(dev)->gen) {
2822 case 4: i965_write_fence_reg(dev, reg, obj); break;
2823 case 3: i915_write_fence_reg(dev, reg, obj); break;
2824 case 2: i830_write_fence_reg(dev, reg, obj); break;
2828 /* And similarly be paranoid that no direct access to this region
2829 * is reordered to before the fence is installed.
2831 if (i915_gem_object_needs_mb(obj))
2835 static inline int fence_number(struct drm_i915_private *dev_priv,
2836 struct drm_i915_fence_reg *fence)
2838 return fence - dev_priv->fence_regs;
2841 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
2842 struct drm_i915_fence_reg *fence,
2845 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2846 int reg = fence_number(dev_priv, fence);
2848 i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL);
2851 obj->fence_reg = reg;
2853 list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list);
2855 obj->fence_reg = I915_FENCE_REG_NONE;
2857 list_del_init(&fence->lru_list);
2859 obj->fence_dirty = false;
2863 i915_gem_object_wait_fence(struct drm_i915_gem_object *obj)
2865 if (obj->last_fenced_seqno) {
2866 int ret = i915_wait_seqno(obj->ring, obj->last_fenced_seqno);
2870 obj->last_fenced_seqno = 0;
2873 obj->fenced_gpu_access = false;
2878 i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
2880 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2881 struct drm_i915_fence_reg *fence;
2884 ret = i915_gem_object_wait_fence(obj);
2888 if (obj->fence_reg == I915_FENCE_REG_NONE)
2891 fence = &dev_priv->fence_regs[obj->fence_reg];
2893 i915_gem_object_fence_lost(obj);
2894 i915_gem_object_update_fence(obj, fence, false);
2899 static struct drm_i915_fence_reg *
2900 i915_find_fence_reg(struct drm_device *dev)
2902 struct drm_i915_private *dev_priv = dev->dev_private;
2903 struct drm_i915_fence_reg *reg, *avail;
2906 /* First try to find a free reg */
2908 for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2909 reg = &dev_priv->fence_regs[i];
2913 if (!reg->pin_count)
2920 /* None available, try to steal one or wait for a user to finish */
2921 list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
2932 * i915_gem_object_get_fence - set up fencing for an object
2933 * @obj: object to map through a fence reg
2935 * When mapping objects through the GTT, userspace wants to be able to write
2936 * to them without having to worry about swizzling if the object is tiled.
2937 * This function walks the fence regs looking for a free one for @obj,
2938 * stealing one if it can't find any.
2940 * It then sets up the reg based on the object's properties: address, pitch
2941 * and tiling format.
2943 * For an untiled surface, this removes any existing fence.
2946 i915_gem_object_get_fence(struct drm_i915_gem_object *obj)
2948 struct drm_device *dev = obj->base.dev;
2949 struct drm_i915_private *dev_priv = dev->dev_private;
2950 bool enable = obj->tiling_mode != I915_TILING_NONE;
2951 struct drm_i915_fence_reg *reg;
2954 /* Have we updated the tiling parameters upon the object and so
2955 * will need to serialise the write to the associated fence register?
2957 if (obj->fence_dirty) {
2958 ret = i915_gem_object_wait_fence(obj);
2963 /* Just update our place in the LRU if our fence is getting reused. */
2964 if (obj->fence_reg != I915_FENCE_REG_NONE) {
2965 reg = &dev_priv->fence_regs[obj->fence_reg];
2966 if (!obj->fence_dirty) {
2967 list_move_tail(®->lru_list,
2968 &dev_priv->mm.fence_list);
2971 } else if (enable) {
2972 reg = i915_find_fence_reg(dev);
2977 struct drm_i915_gem_object *old = reg->obj;
2979 ret = i915_gem_object_wait_fence(old);
2983 i915_gem_object_fence_lost(old);
2988 i915_gem_object_update_fence(obj, reg, enable);
2993 static bool i915_gem_valid_gtt_space(struct drm_device *dev,
2994 struct drm_mm_node *gtt_space,
2995 unsigned long cache_level)
2997 struct drm_mm_node *other;
2999 /* On non-LLC machines we have to be careful when putting differing
3000 * types of snoopable memory together to avoid the prefetcher
3001 * crossing memory domains and dying.
3006 if (!drm_mm_node_allocated(gtt_space))
3009 if (list_empty(>t_space->node_list))
3012 other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list);
3013 if (other->allocated && !other->hole_follows && other->color != cache_level)
3016 other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list);
3017 if (other->allocated && !gtt_space->hole_follows && other->color != cache_level)
3023 static void i915_gem_verify_gtt(struct drm_device *dev)
3026 struct drm_i915_private *dev_priv = dev->dev_private;
3027 struct drm_i915_gem_object *obj;
3030 list_for_each_entry(obj, &dev_priv->mm.gtt_list, global_list) {
3031 if (obj->gtt_space == NULL) {
3032 printk(KERN_ERR "object found on GTT list with no space reserved\n");
3037 if (obj->cache_level != obj->gtt_space->color) {
3038 printk(KERN_ERR "object reserved space [%08lx, %08lx] with wrong color, cache_level=%x, color=%lx\n",
3039 i915_gem_obj_ggtt_offset(obj),
3040 i915_gem_obj_ggtt_offset(obj) + i915_gem_obj_ggtt_size(obj),
3042 obj->gtt_space->color);
3047 if (!i915_gem_valid_gtt_space(dev,
3049 obj->cache_level)) {
3050 printk(KERN_ERR "invalid GTT space found at [%08lx, %08lx] - color=%x\n",
3051 i915_gem_obj_ggtt_offset(obj),
3052 i915_gem_obj_ggtt_offset(obj) + i915_gem_obj_ggtt_size(obj),
3064 * Finds free space in the GTT aperture and binds the object there.
3067 i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
3069 bool map_and_fenceable,
3072 struct drm_device *dev = obj->base.dev;
3073 drm_i915_private_t *dev_priv = dev->dev_private;
3074 struct i915_address_space *vm = &dev_priv->gtt.base;
3075 u32 size, fence_size, fence_alignment, unfenced_alignment;
3076 bool mappable, fenceable;
3077 size_t gtt_max = map_and_fenceable ?
3078 dev_priv->gtt.mappable_end : dev_priv->gtt.base.total;
3079 struct i915_vma *vma;
3082 if (WARN_ON(!list_empty(&obj->vma_list)))
3085 fence_size = i915_gem_get_gtt_size(dev,
3088 fence_alignment = i915_gem_get_gtt_alignment(dev,
3090 obj->tiling_mode, true);
3091 unfenced_alignment =
3092 i915_gem_get_gtt_alignment(dev,
3094 obj->tiling_mode, false);
3097 alignment = map_and_fenceable ? fence_alignment :
3099 if (map_and_fenceable && alignment & (fence_alignment - 1)) {
3100 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
3104 size = map_and_fenceable ? fence_size : obj->base.size;
3106 /* If the object is bigger than the entire aperture, reject it early
3107 * before evicting everything in a vain attempt to find space.
3109 if (obj->base.size > gtt_max) {
3110 DRM_ERROR("Attempting to bind an object larger than the aperture: object=%zd > %s aperture=%zu\n",
3112 map_and_fenceable ? "mappable" : "total",
3117 ret = i915_gem_object_get_pages(obj);
3121 i915_gem_object_pin_pages(obj);
3123 vma = i915_gem_vma_create(obj, &dev_priv->gtt.base);
3130 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
3133 obj->cache_level, 0, gtt_max);
3135 ret = i915_gem_evict_something(dev, size, alignment,
3144 if (WARN_ON(!i915_gem_valid_gtt_space(dev, &vma->node,
3145 obj->cache_level))) {
3147 goto err_remove_node;
3150 ret = i915_gem_gtt_prepare_object(obj);
3152 goto err_remove_node;
3154 list_move_tail(&obj->global_list, &dev_priv->mm.bound_list);
3155 list_add_tail(&obj->mm_list, &vm->inactive_list);
3156 list_add(&vma->vma_link, &obj->vma_list);
3159 i915_gem_obj_ggtt_size(obj) == fence_size &&
3160 (i915_gem_obj_ggtt_offset(obj) & (fence_alignment - 1)) == 0;
3162 mappable = i915_gem_obj_ggtt_offset(obj) + obj->base.size <=
3163 dev_priv->gtt.mappable_end;
3165 obj->map_and_fenceable = mappable && fenceable;
3167 trace_i915_gem_object_bind(obj, map_and_fenceable);
3168 i915_gem_verify_gtt(dev);
3172 drm_mm_remove_node(&vma->node);
3174 i915_gem_vma_destroy(vma);
3176 i915_gem_object_unpin_pages(obj);
3181 i915_gem_clflush_object(struct drm_i915_gem_object *obj)
3183 /* If we don't have a page list set up, then we're not pinned
3184 * to GPU, and we can ignore the cache flush because it'll happen
3185 * again at bind time.
3187 if (obj->pages == NULL)
3191 * Stolen memory is always coherent with the GPU as it is explicitly
3192 * marked as wc by the system, or the system is cache-coherent.
3197 /* If the GPU is snooping the contents of the CPU cache,
3198 * we do not need to manually clear the CPU cache lines. However,
3199 * the caches are only snooped when the render cache is
3200 * flushed/invalidated. As we always have to emit invalidations
3201 * and flushes when moving into and out of the RENDER domain, correct
3202 * snooping behaviour occurs naturally as the result of our domain
3205 if (obj->cache_level != I915_CACHE_NONE)
3208 trace_i915_gem_object_clflush(obj);
3210 drm_clflush_sg(obj->pages);
3213 /** Flushes the GTT write domain for the object if it's dirty. */
3215 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
3217 uint32_t old_write_domain;
3219 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
3222 /* No actual flushing is required for the GTT write domain. Writes
3223 * to it immediately go to main memory as far as we know, so there's
3224 * no chipset flush. It also doesn't land in render cache.
3226 * However, we do have to enforce the order so that all writes through
3227 * the GTT land before any writes to the device, such as updates to
3232 old_write_domain = obj->base.write_domain;
3233 obj->base.write_domain = 0;
3235 trace_i915_gem_object_change_domain(obj,
3236 obj->base.read_domains,
3240 /** Flushes the CPU write domain for the object if it's dirty. */
3242 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj)
3244 uint32_t old_write_domain;
3246 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
3249 i915_gem_clflush_object(obj);
3250 i915_gem_chipset_flush(obj->base.dev);
3251 old_write_domain = obj->base.write_domain;
3252 obj->base.write_domain = 0;
3254 trace_i915_gem_object_change_domain(obj,
3255 obj->base.read_domains,
3260 * Moves a single object to the GTT read, and possibly write domain.
3262 * This function returns when the move is complete, including waiting on
3266 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
3268 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
3269 uint32_t old_write_domain, old_read_domains;
3272 /* Not valid to be called on unbound objects. */
3273 if (!i915_gem_obj_ggtt_bound(obj))
3276 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
3279 ret = i915_gem_object_wait_rendering(obj, !write);
3283 i915_gem_object_flush_cpu_write_domain(obj);
3285 /* Serialise direct access to this object with the barriers for
3286 * coherent writes from the GPU, by effectively invalidating the
3287 * GTT domain upon first access.
3289 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
3292 old_write_domain = obj->base.write_domain;
3293 old_read_domains = obj->base.read_domains;
3295 /* It should now be out of any other write domains, and we can update
3296 * the domain values for our changes.
3298 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3299 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3301 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
3302 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
3306 trace_i915_gem_object_change_domain(obj,
3310 /* And bump the LRU for this access */
3311 if (i915_gem_object_is_inactive(obj))
3312 list_move_tail(&obj->mm_list,
3313 &dev_priv->gtt.base.inactive_list);
3318 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
3319 enum i915_cache_level cache_level)
3321 struct drm_device *dev = obj->base.dev;
3322 drm_i915_private_t *dev_priv = dev->dev_private;
3323 struct i915_vma *vma = __i915_gem_obj_to_vma(obj);
3326 if (obj->cache_level == cache_level)
3329 if (obj->pin_count) {
3330 DRM_DEBUG("can not change the cache level of pinned objects\n");
3334 if (vma && !i915_gem_valid_gtt_space(dev, &vma->node, cache_level)) {
3335 ret = i915_gem_object_unbind(obj);
3340 if (i915_gem_obj_ggtt_bound(obj)) {
3341 ret = i915_gem_object_finish_gpu(obj);
3345 i915_gem_object_finish_gtt(obj);
3347 /* Before SandyBridge, you could not use tiling or fence
3348 * registers with snooped memory, so relinquish any fences
3349 * currently pointing to our region in the aperture.
3351 if (INTEL_INFO(dev)->gen < 6) {
3352 ret = i915_gem_object_put_fence(obj);
3357 if (obj->has_global_gtt_mapping)
3358 i915_gem_gtt_bind_object(obj, cache_level);
3359 if (obj->has_aliasing_ppgtt_mapping)
3360 i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt,
3363 i915_gem_obj_ggtt_set_color(obj, cache_level);
3366 if (cache_level == I915_CACHE_NONE) {
3367 u32 old_read_domains, old_write_domain;
3369 /* If we're coming from LLC cached, then we haven't
3370 * actually been tracking whether the data is in the
3371 * CPU cache or not, since we only allow one bit set
3372 * in obj->write_domain and have been skipping the clflushes.
3373 * Just set it to the CPU cache for now.
3375 WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU);
3376 WARN_ON(obj->base.read_domains & ~I915_GEM_DOMAIN_CPU);
3378 old_read_domains = obj->base.read_domains;
3379 old_write_domain = obj->base.write_domain;
3381 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3382 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3384 trace_i915_gem_object_change_domain(obj,
3389 obj->cache_level = cache_level;
3390 i915_gem_verify_gtt(dev);
3394 int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data,
3395 struct drm_file *file)
3397 struct drm_i915_gem_caching *args = data;
3398 struct drm_i915_gem_object *obj;
3401 ret = i915_mutex_lock_interruptible(dev);
3405 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3406 if (&obj->base == NULL) {
3411 args->caching = obj->cache_level != I915_CACHE_NONE;
3413 drm_gem_object_unreference(&obj->base);
3415 mutex_unlock(&dev->struct_mutex);
3419 int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
3420 struct drm_file *file)
3422 struct drm_i915_gem_caching *args = data;
3423 struct drm_i915_gem_object *obj;
3424 enum i915_cache_level level;
3427 switch (args->caching) {
3428 case I915_CACHING_NONE:
3429 level = I915_CACHE_NONE;
3431 case I915_CACHING_CACHED:
3432 level = I915_CACHE_LLC;
3438 ret = i915_mutex_lock_interruptible(dev);
3442 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3443 if (&obj->base == NULL) {
3448 ret = i915_gem_object_set_cache_level(obj, level);
3450 drm_gem_object_unreference(&obj->base);
3452 mutex_unlock(&dev->struct_mutex);
3457 * Prepare buffer for display plane (scanout, cursors, etc).
3458 * Can be called from an uninterruptible phase (modesetting) and allows
3459 * any flushes to be pipelined (for pageflips).
3462 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
3464 struct intel_ring_buffer *pipelined)
3466 u32 old_read_domains, old_write_domain;
3469 if (pipelined != obj->ring) {
3470 ret = i915_gem_object_sync(obj, pipelined);
3475 /* The display engine is not coherent with the LLC cache on gen6. As
3476 * a result, we make sure that the pinning that is about to occur is
3477 * done with uncached PTEs. This is lowest common denominator for all
3480 * However for gen6+, we could do better by using the GFDT bit instead
3481 * of uncaching, which would allow us to flush all the LLC-cached data
3482 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3484 ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE);
3488 /* As the user may map the buffer once pinned in the display plane
3489 * (e.g. libkms for the bootup splash), we have to ensure that we
3490 * always use map_and_fenceable for all scanout buffers.
3492 ret = i915_gem_object_pin(obj, alignment, true, false);
3496 i915_gem_object_flush_cpu_write_domain(obj);
3498 old_write_domain = obj->base.write_domain;
3499 old_read_domains = obj->base.read_domains;
3501 /* It should now be out of any other write domains, and we can update
3502 * the domain values for our changes.
3504 obj->base.write_domain = 0;
3505 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3507 trace_i915_gem_object_change_domain(obj,
3515 i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
3519 if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
3522 ret = i915_gem_object_wait_rendering(obj, false);
3526 /* Ensure that we invalidate the GPU's caches and TLBs. */
3527 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
3532 * Moves a single object to the CPU read, and possibly write domain.
3534 * This function returns when the move is complete, including waiting on
3538 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
3540 uint32_t old_write_domain, old_read_domains;
3543 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
3546 ret = i915_gem_object_wait_rendering(obj, !write);
3550 i915_gem_object_flush_gtt_write_domain(obj);
3552 old_write_domain = obj->base.write_domain;
3553 old_read_domains = obj->base.read_domains;
3555 /* Flush the CPU cache if it's still invalid. */
3556 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
3557 i915_gem_clflush_object(obj);
3559 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3562 /* It should now be out of any other write domains, and we can update
3563 * the domain values for our changes.
3565 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3567 /* If we're writing through the CPU, then the GPU read domains will
3568 * need to be invalidated at next use.
3571 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3572 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3575 trace_i915_gem_object_change_domain(obj,
3582 /* Throttle our rendering by waiting until the ring has completed our requests
3583 * emitted over 20 msec ago.
3585 * Note that if we were to use the current jiffies each time around the loop,
3586 * we wouldn't escape the function with any frames outstanding if the time to
3587 * render a frame was over 20ms.
3589 * This should get us reasonable parallelism between CPU and GPU but also
3590 * relatively low latency when blocking on a particular request to finish.
3593 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
3595 struct drm_i915_private *dev_priv = dev->dev_private;
3596 struct drm_i915_file_private *file_priv = file->driver_priv;
3597 unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3598 struct drm_i915_gem_request *request;
3599 struct intel_ring_buffer *ring = NULL;
3600 unsigned reset_counter;
3604 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
3608 ret = i915_gem_check_wedge(&dev_priv->gpu_error, false);
3612 spin_lock(&file_priv->mm.lock);
3613 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
3614 if (time_after_eq(request->emitted_jiffies, recent_enough))
3617 ring = request->ring;
3618 seqno = request->seqno;
3620 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
3621 spin_unlock(&file_priv->mm.lock);
3626 ret = __wait_seqno(ring, seqno, reset_counter, true, NULL);
3628 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
3634 i915_gem_object_pin(struct drm_i915_gem_object *obj,
3636 bool map_and_fenceable,
3641 if (WARN_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT))
3644 if (i915_gem_obj_ggtt_bound(obj)) {
3645 if ((alignment && i915_gem_obj_ggtt_offset(obj) & (alignment - 1)) ||
3646 (map_and_fenceable && !obj->map_and_fenceable)) {
3647 WARN(obj->pin_count,
3648 "bo is already pinned with incorrect alignment:"
3649 " offset=%lx, req.alignment=%x, req.map_and_fenceable=%d,"
3650 " obj->map_and_fenceable=%d\n",
3651 i915_gem_obj_ggtt_offset(obj), alignment,
3653 obj->map_and_fenceable);
3654 ret = i915_gem_object_unbind(obj);
3660 if (!i915_gem_obj_ggtt_bound(obj)) {
3661 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3663 ret = i915_gem_object_bind_to_gtt(obj, alignment,
3669 if (!dev_priv->mm.aliasing_ppgtt)
3670 i915_gem_gtt_bind_object(obj, obj->cache_level);
3673 if (!obj->has_global_gtt_mapping && map_and_fenceable)
3674 i915_gem_gtt_bind_object(obj, obj->cache_level);
3677 obj->pin_mappable |= map_and_fenceable;
3683 i915_gem_object_unpin(struct drm_i915_gem_object *obj)
3685 BUG_ON(obj->pin_count == 0);
3686 BUG_ON(!i915_gem_obj_ggtt_bound(obj));
3688 if (--obj->pin_count == 0)
3689 obj->pin_mappable = false;
3693 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
3694 struct drm_file *file)
3696 struct drm_i915_gem_pin *args = data;
3697 struct drm_i915_gem_object *obj;
3700 ret = i915_mutex_lock_interruptible(dev);
3704 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3705 if (&obj->base == NULL) {
3710 if (obj->madv != I915_MADV_WILLNEED) {
3711 DRM_ERROR("Attempting to pin a purgeable buffer\n");
3716 if (obj->pin_filp != NULL && obj->pin_filp != file) {
3717 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
3723 if (obj->user_pin_count == 0) {
3724 ret = i915_gem_object_pin(obj, args->alignment, true, false);
3729 obj->user_pin_count++;
3730 obj->pin_filp = file;
3732 /* XXX - flush the CPU caches for pinned objects
3733 * as the X server doesn't manage domains yet
3735 i915_gem_object_flush_cpu_write_domain(obj);
3736 args->offset = i915_gem_obj_ggtt_offset(obj);
3738 drm_gem_object_unreference(&obj->base);
3740 mutex_unlock(&dev->struct_mutex);
3745 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
3746 struct drm_file *file)
3748 struct drm_i915_gem_pin *args = data;
3749 struct drm_i915_gem_object *obj;
3752 ret = i915_mutex_lock_interruptible(dev);
3756 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3757 if (&obj->base == NULL) {
3762 if (obj->pin_filp != file) {
3763 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
3768 obj->user_pin_count--;
3769 if (obj->user_pin_count == 0) {
3770 obj->pin_filp = NULL;
3771 i915_gem_object_unpin(obj);
3775 drm_gem_object_unreference(&obj->base);
3777 mutex_unlock(&dev->struct_mutex);
3782 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
3783 struct drm_file *file)
3785 struct drm_i915_gem_busy *args = data;
3786 struct drm_i915_gem_object *obj;
3789 ret = i915_mutex_lock_interruptible(dev);
3793 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3794 if (&obj->base == NULL) {
3799 /* Count all active objects as busy, even if they are currently not used
3800 * by the gpu. Users of this interface expect objects to eventually
3801 * become non-busy without any further actions, therefore emit any
3802 * necessary flushes here.
3804 ret = i915_gem_object_flush_active(obj);
3806 args->busy = obj->active;
3808 BUILD_BUG_ON(I915_NUM_RINGS > 16);
3809 args->busy |= intel_ring_flag(obj->ring) << 16;
3812 drm_gem_object_unreference(&obj->base);
3814 mutex_unlock(&dev->struct_mutex);
3819 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
3820 struct drm_file *file_priv)
3822 return i915_gem_ring_throttle(dev, file_priv);
3826 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
3827 struct drm_file *file_priv)
3829 struct drm_i915_gem_madvise *args = data;
3830 struct drm_i915_gem_object *obj;
3833 switch (args->madv) {
3834 case I915_MADV_DONTNEED:
3835 case I915_MADV_WILLNEED:
3841 ret = i915_mutex_lock_interruptible(dev);
3845 obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
3846 if (&obj->base == NULL) {
3851 if (obj->pin_count) {
3856 if (obj->madv != __I915_MADV_PURGED)
3857 obj->madv = args->madv;
3859 /* if the object is no longer attached, discard its backing storage */
3860 if (i915_gem_object_is_purgeable(obj) && obj->pages == NULL)
3861 i915_gem_object_truncate(obj);
3863 args->retained = obj->madv != __I915_MADV_PURGED;
3866 drm_gem_object_unreference(&obj->base);
3868 mutex_unlock(&dev->struct_mutex);
3872 void i915_gem_object_init(struct drm_i915_gem_object *obj,
3873 const struct drm_i915_gem_object_ops *ops)
3875 INIT_LIST_HEAD(&obj->mm_list);
3876 INIT_LIST_HEAD(&obj->global_list);
3877 INIT_LIST_HEAD(&obj->ring_list);
3878 INIT_LIST_HEAD(&obj->exec_list);
3879 INIT_LIST_HEAD(&obj->vma_list);
3883 obj->fence_reg = I915_FENCE_REG_NONE;
3884 obj->madv = I915_MADV_WILLNEED;
3885 /* Avoid an unnecessary call to unbind on the first bind. */
3886 obj->map_and_fenceable = true;
3888 i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size);
3891 static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
3892 .get_pages = i915_gem_object_get_pages_gtt,
3893 .put_pages = i915_gem_object_put_pages_gtt,
3896 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
3899 struct drm_i915_gem_object *obj;
3900 struct address_space *mapping;
3903 obj = i915_gem_object_alloc(dev);
3907 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
3908 i915_gem_object_free(obj);
3912 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
3913 if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) {
3914 /* 965gm cannot relocate objects above 4GiB. */
3915 mask &= ~__GFP_HIGHMEM;
3916 mask |= __GFP_DMA32;
3919 mapping = file_inode(obj->base.filp)->i_mapping;
3920 mapping_set_gfp_mask(mapping, mask);
3922 i915_gem_object_init(obj, &i915_gem_object_ops);
3924 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3925 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3928 /* On some devices, we can have the GPU use the LLC (the CPU
3929 * cache) for about a 10% performance improvement
3930 * compared to uncached. Graphics requests other than
3931 * display scanout are coherent with the CPU in
3932 * accessing this cache. This means in this mode we
3933 * don't need to clflush on the CPU side, and on the
3934 * GPU side we only need to flush internal caches to
3935 * get data visible to the CPU.
3937 * However, we maintain the display planes as UC, and so
3938 * need to rebind when first used as such.
3940 obj->cache_level = I915_CACHE_LLC;
3942 obj->cache_level = I915_CACHE_NONE;
3944 trace_i915_gem_object_create(obj);
3949 int i915_gem_init_object(struct drm_gem_object *obj)
3956 void i915_gem_free_object(struct drm_gem_object *gem_obj)
3958 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
3959 struct drm_device *dev = obj->base.dev;
3960 drm_i915_private_t *dev_priv = dev->dev_private;
3962 trace_i915_gem_object_destroy(obj);
3965 i915_gem_detach_phys_object(dev, obj);
3968 if (WARN_ON(i915_gem_object_unbind(obj) == -ERESTARTSYS)) {
3969 bool was_interruptible;
3971 was_interruptible = dev_priv->mm.interruptible;
3972 dev_priv->mm.interruptible = false;
3974 WARN_ON(i915_gem_object_unbind(obj));
3976 dev_priv->mm.interruptible = was_interruptible;
3979 /* Stolen objects don't hold a ref, but do hold pin count. Fix that up
3980 * before progressing. */
3982 i915_gem_object_unpin_pages(obj);
3984 if (WARN_ON(obj->pages_pin_count))
3985 obj->pages_pin_count = 0;
3986 i915_gem_object_put_pages(obj);
3987 i915_gem_object_free_mmap_offset(obj);
3988 i915_gem_object_release_stolen(obj);
3992 if (obj->base.import_attach)
3993 drm_prime_gem_destroy(&obj->base, NULL);
3995 drm_gem_object_release(&obj->base);
3996 i915_gem_info_remove_obj(dev_priv, obj->base.size);
3999 i915_gem_object_free(obj);
4002 struct i915_vma *i915_gem_vma_create(struct drm_i915_gem_object *obj,
4003 struct i915_address_space *vm)
4005 struct i915_vma *vma = kzalloc(sizeof(*vma), GFP_KERNEL);
4007 return ERR_PTR(-ENOMEM);
4009 INIT_LIST_HEAD(&vma->vma_link);
4016 void i915_gem_vma_destroy(struct i915_vma *vma)
4018 WARN_ON(vma->node.allocated);
4023 i915_gem_idle(struct drm_device *dev)
4025 drm_i915_private_t *dev_priv = dev->dev_private;
4028 if (dev_priv->ums.mm_suspended) {
4029 mutex_unlock(&dev->struct_mutex);
4033 ret = i915_gpu_idle(dev);
4035 mutex_unlock(&dev->struct_mutex);
4038 i915_gem_retire_requests(dev);
4040 /* Under UMS, be paranoid and evict. */
4041 if (!drm_core_check_feature(dev, DRIVER_MODESET))
4042 i915_gem_evict_everything(dev);
4044 del_timer_sync(&dev_priv->gpu_error.hangcheck_timer);
4046 i915_kernel_lost_context(dev);
4047 i915_gem_cleanup_ringbuffer(dev);
4049 /* Cancel the retire work handler, which should be idle now. */
4050 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
4055 void i915_gem_l3_remap(struct drm_device *dev)
4057 drm_i915_private_t *dev_priv = dev->dev_private;
4061 if (!HAS_L3_GPU_CACHE(dev))
4064 if (!dev_priv->l3_parity.remap_info)
4067 misccpctl = I915_READ(GEN7_MISCCPCTL);
4068 I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
4069 POSTING_READ(GEN7_MISCCPCTL);
4071 for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) {
4072 u32 remap = I915_READ(GEN7_L3LOG_BASE + i);
4073 if (remap && remap != dev_priv->l3_parity.remap_info[i/4])
4074 DRM_DEBUG("0x%x was already programmed to %x\n",
4075 GEN7_L3LOG_BASE + i, remap);
4076 if (remap && !dev_priv->l3_parity.remap_info[i/4])
4077 DRM_DEBUG_DRIVER("Clearing remapped register\n");
4078 I915_WRITE(GEN7_L3LOG_BASE + i, dev_priv->l3_parity.remap_info[i/4]);
4081 /* Make sure all the writes land before disabling dop clock gating */
4082 POSTING_READ(GEN7_L3LOG_BASE);
4084 I915_WRITE(GEN7_MISCCPCTL, misccpctl);
4087 void i915_gem_init_swizzling(struct drm_device *dev)
4089 drm_i915_private_t *dev_priv = dev->dev_private;
4091 if (INTEL_INFO(dev)->gen < 5 ||
4092 dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
4095 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
4096 DISP_TILE_SURFACE_SWIZZLING);
4101 I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
4103 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
4104 else if (IS_GEN7(dev))
4105 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
4111 intel_enable_blt(struct drm_device *dev)
4116 /* The blitter was dysfunctional on early prototypes */
4117 if (IS_GEN6(dev) && dev->pdev->revision < 8) {
4118 DRM_INFO("BLT not supported on this pre-production hardware;"
4119 " graphics performance will be degraded.\n");
4126 static int i915_gem_init_rings(struct drm_device *dev)
4128 struct drm_i915_private *dev_priv = dev->dev_private;
4131 ret = intel_init_render_ring_buffer(dev);
4136 ret = intel_init_bsd_ring_buffer(dev);
4138 goto cleanup_render_ring;
4141 if (intel_enable_blt(dev)) {
4142 ret = intel_init_blt_ring_buffer(dev);
4144 goto cleanup_bsd_ring;
4147 if (HAS_VEBOX(dev)) {
4148 ret = intel_init_vebox_ring_buffer(dev);
4150 goto cleanup_blt_ring;
4154 ret = i915_gem_set_seqno(dev, ((u32)~0 - 0x1000));
4156 goto cleanup_vebox_ring;
4161 intel_cleanup_ring_buffer(&dev_priv->ring[VECS]);
4163 intel_cleanup_ring_buffer(&dev_priv->ring[BCS]);
4165 intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
4166 cleanup_render_ring:
4167 intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
4173 i915_gem_init_hw(struct drm_device *dev)
4175 drm_i915_private_t *dev_priv = dev->dev_private;
4178 if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt())
4181 if (dev_priv->ellc_size)
4182 I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf));
4184 if (HAS_PCH_NOP(dev)) {
4185 u32 temp = I915_READ(GEN7_MSG_CTL);
4186 temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK);
4187 I915_WRITE(GEN7_MSG_CTL, temp);
4190 i915_gem_l3_remap(dev);
4192 i915_gem_init_swizzling(dev);
4194 ret = i915_gem_init_rings(dev);
4199 * XXX: There was some w/a described somewhere suggesting loading
4200 * contexts before PPGTT.
4202 i915_gem_context_init(dev);
4203 if (dev_priv->mm.aliasing_ppgtt) {
4204 ret = dev_priv->mm.aliasing_ppgtt->enable(dev);
4206 i915_gem_cleanup_aliasing_ppgtt(dev);
4207 DRM_INFO("PPGTT enable failed. This is not fatal, but unexpected\n");
4214 int i915_gem_init(struct drm_device *dev)
4216 struct drm_i915_private *dev_priv = dev->dev_private;
4219 mutex_lock(&dev->struct_mutex);
4221 if (IS_VALLEYVIEW(dev)) {
4222 /* VLVA0 (potential hack), BIOS isn't actually waking us */
4223 I915_WRITE(VLV_GTLC_WAKE_CTRL, 1);
4224 if (wait_for((I915_READ(VLV_GTLC_PW_STATUS) & 1) == 1, 10))
4225 DRM_DEBUG_DRIVER("allow wake ack timed out\n");
4228 i915_gem_init_global_gtt(dev);
4230 ret = i915_gem_init_hw(dev);
4231 mutex_unlock(&dev->struct_mutex);
4233 i915_gem_cleanup_aliasing_ppgtt(dev);
4237 /* Allow hardware batchbuffers unless told otherwise, but not for KMS. */
4238 if (!drm_core_check_feature(dev, DRIVER_MODESET))
4239 dev_priv->dri1.allow_batchbuffer = 1;
4244 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
4246 drm_i915_private_t *dev_priv = dev->dev_private;
4247 struct intel_ring_buffer *ring;
4250 for_each_ring(ring, dev_priv, i)
4251 intel_cleanup_ring_buffer(ring);
4255 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
4256 struct drm_file *file_priv)
4258 struct drm_i915_private *dev_priv = dev->dev_private;
4261 if (drm_core_check_feature(dev, DRIVER_MODESET))
4264 if (i915_reset_in_progress(&dev_priv->gpu_error)) {
4265 DRM_ERROR("Reenabling wedged hardware, good luck\n");
4266 atomic_set(&dev_priv->gpu_error.reset_counter, 0);
4269 mutex_lock(&dev->struct_mutex);
4270 dev_priv->ums.mm_suspended = 0;
4272 ret = i915_gem_init_hw(dev);
4274 mutex_unlock(&dev->struct_mutex);
4278 BUG_ON(!list_empty(&dev_priv->gtt.base.active_list));
4279 mutex_unlock(&dev->struct_mutex);
4281 ret = drm_irq_install(dev);
4283 goto cleanup_ringbuffer;
4288 mutex_lock(&dev->struct_mutex);
4289 i915_gem_cleanup_ringbuffer(dev);
4290 dev_priv->ums.mm_suspended = 1;
4291 mutex_unlock(&dev->struct_mutex);
4297 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
4298 struct drm_file *file_priv)
4300 struct drm_i915_private *dev_priv = dev->dev_private;
4303 if (drm_core_check_feature(dev, DRIVER_MODESET))
4306 drm_irq_uninstall(dev);
4308 mutex_lock(&dev->struct_mutex);
4309 ret = i915_gem_idle(dev);
4311 /* Hack! Don't let anybody do execbuf while we don't control the chip.
4312 * We need to replace this with a semaphore, or something.
4313 * And not confound ums.mm_suspended!
4316 dev_priv->ums.mm_suspended = 1;
4317 mutex_unlock(&dev->struct_mutex);
4323 i915_gem_lastclose(struct drm_device *dev)
4327 if (drm_core_check_feature(dev, DRIVER_MODESET))
4330 mutex_lock(&dev->struct_mutex);
4331 ret = i915_gem_idle(dev);
4333 DRM_ERROR("failed to idle hardware: %d\n", ret);
4334 mutex_unlock(&dev->struct_mutex);
4338 init_ring_lists(struct intel_ring_buffer *ring)
4340 INIT_LIST_HEAD(&ring->active_list);
4341 INIT_LIST_HEAD(&ring->request_list);
4345 i915_gem_load(struct drm_device *dev)
4347 drm_i915_private_t *dev_priv = dev->dev_private;
4351 kmem_cache_create("i915_gem_object",
4352 sizeof(struct drm_i915_gem_object), 0,
4356 INIT_LIST_HEAD(&dev_priv->gtt.base.active_list);
4357 INIT_LIST_HEAD(&dev_priv->gtt.base.inactive_list);
4358 INIT_LIST_HEAD(&dev_priv->mm.unbound_list);
4359 INIT_LIST_HEAD(&dev_priv->mm.bound_list);
4360 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4361 for (i = 0; i < I915_NUM_RINGS; i++)
4362 init_ring_lists(&dev_priv->ring[i]);
4363 for (i = 0; i < I915_MAX_NUM_FENCES; i++)
4364 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
4365 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
4366 i915_gem_retire_work_handler);
4367 init_waitqueue_head(&dev_priv->gpu_error.reset_queue);
4369 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
4371 I915_WRITE(MI_ARB_STATE,
4372 _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
4375 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
4377 /* Old X drivers will take 0-2 for front, back, depth buffers */
4378 if (!drm_core_check_feature(dev, DRIVER_MODESET))
4379 dev_priv->fence_reg_start = 3;
4381 if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev))
4382 dev_priv->num_fence_regs = 32;
4383 else if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
4384 dev_priv->num_fence_regs = 16;
4386 dev_priv->num_fence_regs = 8;
4388 /* Initialize fence registers to zero */
4389 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4390 i915_gem_restore_fences(dev);
4392 i915_gem_detect_bit_6_swizzle(dev);
4393 init_waitqueue_head(&dev_priv->pending_flip_queue);
4395 dev_priv->mm.interruptible = true;
4397 dev_priv->mm.inactive_shrinker.scan_objects = i915_gem_inactive_scan;
4398 dev_priv->mm.inactive_shrinker.count_objects = i915_gem_inactive_count;
4399 dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS;
4400 register_shrinker(&dev_priv->mm.inactive_shrinker);
4404 * Create a physically contiguous memory object for this object
4405 * e.g. for cursor + overlay regs
4407 static int i915_gem_init_phys_object(struct drm_device *dev,
4408 int id, int size, int align)
4410 drm_i915_private_t *dev_priv = dev->dev_private;
4411 struct drm_i915_gem_phys_object *phys_obj;
4414 if (dev_priv->mm.phys_objs[id - 1] || !size)
4417 phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL);
4423 phys_obj->handle = drm_pci_alloc(dev, size, align);
4424 if (!phys_obj->handle) {
4429 set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4432 dev_priv->mm.phys_objs[id - 1] = phys_obj;
4440 static void i915_gem_free_phys_object(struct drm_device *dev, int id)
4442 drm_i915_private_t *dev_priv = dev->dev_private;
4443 struct drm_i915_gem_phys_object *phys_obj;
4445 if (!dev_priv->mm.phys_objs[id - 1])
4448 phys_obj = dev_priv->mm.phys_objs[id - 1];
4449 if (phys_obj->cur_obj) {
4450 i915_gem_detach_phys_object(dev, phys_obj->cur_obj);
4454 set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4456 drm_pci_free(dev, phys_obj->handle);
4458 dev_priv->mm.phys_objs[id - 1] = NULL;
4461 void i915_gem_free_all_phys_object(struct drm_device *dev)
4465 for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++)
4466 i915_gem_free_phys_object(dev, i);
4469 void i915_gem_detach_phys_object(struct drm_device *dev,
4470 struct drm_i915_gem_object *obj)
4472 struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
4479 vaddr = obj->phys_obj->handle->vaddr;
4481 page_count = obj->base.size / PAGE_SIZE;
4482 for (i = 0; i < page_count; i++) {
4483 struct page *page = shmem_read_mapping_page(mapping, i);
4484 if (!IS_ERR(page)) {
4485 char *dst = kmap_atomic(page);
4486 memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE);
4489 drm_clflush_pages(&page, 1);
4491 set_page_dirty(page);
4492 mark_page_accessed(page);
4493 page_cache_release(page);
4496 i915_gem_chipset_flush(dev);
4498 obj->phys_obj->cur_obj = NULL;
4499 obj->phys_obj = NULL;
4503 i915_gem_attach_phys_object(struct drm_device *dev,
4504 struct drm_i915_gem_object *obj,
4508 struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
4509 drm_i915_private_t *dev_priv = dev->dev_private;
4514 if (id > I915_MAX_PHYS_OBJECT)
4517 if (obj->phys_obj) {
4518 if (obj->phys_obj->id == id)
4520 i915_gem_detach_phys_object(dev, obj);
4523 /* create a new object */
4524 if (!dev_priv->mm.phys_objs[id - 1]) {
4525 ret = i915_gem_init_phys_object(dev, id,
4526 obj->base.size, align);
4528 DRM_ERROR("failed to init phys object %d size: %zu\n",
4529 id, obj->base.size);
4534 /* bind to the object */
4535 obj->phys_obj = dev_priv->mm.phys_objs[id - 1];
4536 obj->phys_obj->cur_obj = obj;
4538 page_count = obj->base.size / PAGE_SIZE;
4540 for (i = 0; i < page_count; i++) {
4544 page = shmem_read_mapping_page(mapping, i);
4546 return PTR_ERR(page);
4548 src = kmap_atomic(page);
4549 dst = obj->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4550 memcpy(dst, src, PAGE_SIZE);
4553 mark_page_accessed(page);
4554 page_cache_release(page);
4561 i915_gem_phys_pwrite(struct drm_device *dev,
4562 struct drm_i915_gem_object *obj,
4563 struct drm_i915_gem_pwrite *args,
4564 struct drm_file *file_priv)
4566 void *vaddr = obj->phys_obj->handle->vaddr + args->offset;
4567 char __user *user_data = to_user_ptr(args->data_ptr);
4569 if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
4570 unsigned long unwritten;
4572 /* The physical object once assigned is fixed for the lifetime
4573 * of the obj, so we can safely drop the lock and continue
4576 mutex_unlock(&dev->struct_mutex);
4577 unwritten = copy_from_user(vaddr, user_data, args->size);
4578 mutex_lock(&dev->struct_mutex);
4583 i915_gem_chipset_flush(dev);
4587 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
4589 struct drm_i915_file_private *file_priv = file->driver_priv;
4591 /* Clean up our request list when the client is going away, so that
4592 * later retire_requests won't dereference our soon-to-be-gone
4595 spin_lock(&file_priv->mm.lock);
4596 while (!list_empty(&file_priv->mm.request_list)) {
4597 struct drm_i915_gem_request *request;
4599 request = list_first_entry(&file_priv->mm.request_list,
4600 struct drm_i915_gem_request,
4602 list_del(&request->client_list);
4603 request->file_priv = NULL;
4605 spin_unlock(&file_priv->mm.lock);
4608 static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task)
4610 if (!mutex_is_locked(mutex))
4613 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES)
4614 return mutex->owner == task;
4616 /* Since UP may be pre-empted, we cannot assume that we own the lock */
4621 static unsigned long
4622 i915_gem_inactive_count(struct shrinker *shrinker, struct shrink_control *sc)
4624 struct drm_i915_private *dev_priv =
4625 container_of(shrinker,
4626 struct drm_i915_private,
4627 mm.inactive_shrinker);
4628 struct drm_device *dev = dev_priv->dev;
4629 struct i915_address_space *vm = &dev_priv->gtt.base;
4630 struct drm_i915_gem_object *obj;
4632 unsigned long count;
4634 if (!mutex_trylock(&dev->struct_mutex)) {
4635 if (!mutex_is_locked_by(&dev->struct_mutex, current))
4638 if (dev_priv->mm.shrinker_no_lock_stealing)
4645 list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list)
4646 if (obj->pages_pin_count == 0)
4647 count += obj->base.size >> PAGE_SHIFT;
4648 list_for_each_entry(obj, &vm->inactive_list, mm_list)
4649 if (obj->pin_count == 0 && obj->pages_pin_count == 0)
4650 count += obj->base.size >> PAGE_SHIFT;
4653 mutex_unlock(&dev->struct_mutex);
4657 static unsigned long
4658 i915_gem_inactive_scan(struct shrinker *shrinker, struct shrink_control *sc)
4660 struct drm_i915_private *dev_priv =
4661 container_of(shrinker,
4662 struct drm_i915_private,
4663 mm.inactive_shrinker);
4664 struct drm_device *dev = dev_priv->dev;
4665 int nr_to_scan = sc->nr_to_scan;
4666 unsigned long freed;
4669 if (!mutex_trylock(&dev->struct_mutex)) {
4670 if (!mutex_is_locked_by(&dev->struct_mutex, current))
4673 if (dev_priv->mm.shrinker_no_lock_stealing)
4679 freed = i915_gem_purge(dev_priv, nr_to_scan);
4680 if (freed < nr_to_scan)
4681 freed += __i915_gem_shrink(dev_priv, nr_to_scan,
4683 if (freed < nr_to_scan)
4684 freed += i915_gem_shrink_all(dev_priv);
4687 mutex_unlock(&dev->struct_mutex);