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/i915_drm.h>
31 #include "i915_trace.h"
32 #include "intel_drv.h"
33 #include <linux/shmem_fs.h>
34 #include <linux/slab.h>
35 #include <linux/swap.h>
36 #include <linux/pci.h>
37 #include <linux/dma-buf.h>
39 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);
40 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj);
41 static __must_check int i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
43 bool map_and_fenceable,
45 static int i915_gem_phys_pwrite(struct drm_device *dev,
46 struct drm_i915_gem_object *obj,
47 struct drm_i915_gem_pwrite *args,
48 struct drm_file *file);
50 static void i915_gem_write_fence(struct drm_device *dev, int reg,
51 struct drm_i915_gem_object *obj);
52 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
53 struct drm_i915_fence_reg *fence,
56 static int i915_gem_inactive_shrink(struct shrinker *shrinker,
57 struct shrink_control *sc);
58 static long i915_gem_purge(struct drm_i915_private *dev_priv, long target);
59 static void i915_gem_shrink_all(struct drm_i915_private *dev_priv);
60 static void i915_gem_object_truncate(struct drm_i915_gem_object *obj);
62 static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object *obj)
65 i915_gem_release_mmap(obj);
67 /* As we do not have an associated fence register, we will force
68 * a tiling change if we ever need to acquire one.
70 obj->fence_dirty = false;
71 obj->fence_reg = I915_FENCE_REG_NONE;
74 /* some bookkeeping */
75 static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
78 dev_priv->mm.object_count++;
79 dev_priv->mm.object_memory += size;
82 static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
85 dev_priv->mm.object_count--;
86 dev_priv->mm.object_memory -= size;
90 i915_gem_wait_for_error(struct i915_gpu_error *error)
94 #define EXIT_COND (!i915_reset_in_progress(error) || \
95 i915_terminally_wedged(error))
100 * Only wait 10 seconds for the gpu reset to complete to avoid hanging
101 * userspace. If it takes that long something really bad is going on and
102 * we should simply try to bail out and fail as gracefully as possible.
104 ret = wait_event_interruptible_timeout(error->reset_queue,
108 DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
110 } else if (ret < 0) {
118 int i915_mutex_lock_interruptible(struct drm_device *dev)
120 struct drm_i915_private *dev_priv = dev->dev_private;
123 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
127 ret = mutex_lock_interruptible(&dev->struct_mutex);
131 WARN_ON(i915_verify_lists(dev));
136 i915_gem_object_is_inactive(struct drm_i915_gem_object *obj)
138 return i915_gem_obj_ggtt_bound(obj) && !obj->active;
142 i915_gem_init_ioctl(struct drm_device *dev, void *data,
143 struct drm_file *file)
145 struct drm_i915_private *dev_priv = dev->dev_private;
146 struct drm_i915_gem_init *args = data;
148 if (drm_core_check_feature(dev, DRIVER_MODESET))
151 if (args->gtt_start >= args->gtt_end ||
152 (args->gtt_end | args->gtt_start) & (PAGE_SIZE - 1))
155 /* GEM with user mode setting was never supported on ilk and later. */
156 if (INTEL_INFO(dev)->gen >= 5)
159 mutex_lock(&dev->struct_mutex);
160 i915_gem_setup_global_gtt(dev, args->gtt_start, args->gtt_end,
162 dev_priv->gtt.mappable_end = args->gtt_end;
163 mutex_unlock(&dev->struct_mutex);
169 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
170 struct drm_file *file)
172 struct drm_i915_private *dev_priv = dev->dev_private;
173 struct drm_i915_gem_get_aperture *args = data;
174 struct drm_i915_gem_object *obj;
178 mutex_lock(&dev->struct_mutex);
179 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list)
181 pinned += i915_gem_obj_ggtt_size(obj);
182 mutex_unlock(&dev->struct_mutex);
184 args->aper_size = dev_priv->gtt.base.total;
185 args->aper_available_size = args->aper_size - pinned;
190 void *i915_gem_object_alloc(struct drm_device *dev)
192 struct drm_i915_private *dev_priv = dev->dev_private;
193 return kmem_cache_alloc(dev_priv->slab, GFP_KERNEL | __GFP_ZERO);
196 void i915_gem_object_free(struct drm_i915_gem_object *obj)
198 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
199 kmem_cache_free(dev_priv->slab, obj);
203 i915_gem_create(struct drm_file *file,
204 struct drm_device *dev,
208 struct drm_i915_gem_object *obj;
212 size = roundup(size, PAGE_SIZE);
216 /* Allocate the new object */
217 obj = i915_gem_alloc_object(dev, size);
221 ret = drm_gem_handle_create(file, &obj->base, &handle);
222 /* drop reference from allocate - handle holds it now */
223 drm_gem_object_unreference_unlocked(&obj->base);
232 i915_gem_dumb_create(struct drm_file *file,
233 struct drm_device *dev,
234 struct drm_mode_create_dumb *args)
236 /* have to work out size/pitch and return them */
237 args->pitch = ALIGN(args->width * ((args->bpp + 7) / 8), 64);
238 args->size = args->pitch * args->height;
239 return i915_gem_create(file, dev,
240 args->size, &args->handle);
243 int i915_gem_dumb_destroy(struct drm_file *file,
244 struct drm_device *dev,
247 return drm_gem_handle_delete(file, handle);
251 * Creates a new mm object and returns a handle to it.
254 i915_gem_create_ioctl(struct drm_device *dev, void *data,
255 struct drm_file *file)
257 struct drm_i915_gem_create *args = data;
259 return i915_gem_create(file, dev,
260 args->size, &args->handle);
264 __copy_to_user_swizzled(char __user *cpu_vaddr,
265 const char *gpu_vaddr, int gpu_offset,
268 int ret, cpu_offset = 0;
271 int cacheline_end = ALIGN(gpu_offset + 1, 64);
272 int this_length = min(cacheline_end - gpu_offset, length);
273 int swizzled_gpu_offset = gpu_offset ^ 64;
275 ret = __copy_to_user(cpu_vaddr + cpu_offset,
276 gpu_vaddr + swizzled_gpu_offset,
281 cpu_offset += this_length;
282 gpu_offset += this_length;
283 length -= this_length;
290 __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset,
291 const char __user *cpu_vaddr,
294 int ret, cpu_offset = 0;
297 int cacheline_end = ALIGN(gpu_offset + 1, 64);
298 int this_length = min(cacheline_end - gpu_offset, length);
299 int swizzled_gpu_offset = gpu_offset ^ 64;
301 ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset,
302 cpu_vaddr + cpu_offset,
307 cpu_offset += this_length;
308 gpu_offset += this_length;
309 length -= this_length;
315 /* Per-page copy function for the shmem pread fastpath.
316 * Flushes invalid cachelines before reading the target if
317 * needs_clflush is set. */
319 shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length,
320 char __user *user_data,
321 bool page_do_bit17_swizzling, bool needs_clflush)
326 if (unlikely(page_do_bit17_swizzling))
329 vaddr = kmap_atomic(page);
331 drm_clflush_virt_range(vaddr + shmem_page_offset,
333 ret = __copy_to_user_inatomic(user_data,
334 vaddr + shmem_page_offset,
336 kunmap_atomic(vaddr);
338 return ret ? -EFAULT : 0;
342 shmem_clflush_swizzled_range(char *addr, unsigned long length,
345 if (unlikely(swizzled)) {
346 unsigned long start = (unsigned long) addr;
347 unsigned long end = (unsigned long) addr + length;
349 /* For swizzling simply ensure that we always flush both
350 * channels. Lame, but simple and it works. Swizzled
351 * pwrite/pread is far from a hotpath - current userspace
352 * doesn't use it at all. */
353 start = round_down(start, 128);
354 end = round_up(end, 128);
356 drm_clflush_virt_range((void *)start, end - start);
358 drm_clflush_virt_range(addr, length);
363 /* Only difference to the fast-path function is that this can handle bit17
364 * and uses non-atomic copy and kmap functions. */
366 shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length,
367 char __user *user_data,
368 bool page_do_bit17_swizzling, bool needs_clflush)
375 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
377 page_do_bit17_swizzling);
379 if (page_do_bit17_swizzling)
380 ret = __copy_to_user_swizzled(user_data,
381 vaddr, shmem_page_offset,
384 ret = __copy_to_user(user_data,
385 vaddr + shmem_page_offset,
389 return ret ? - EFAULT : 0;
393 i915_gem_shmem_pread(struct drm_device *dev,
394 struct drm_i915_gem_object *obj,
395 struct drm_i915_gem_pread *args,
396 struct drm_file *file)
398 char __user *user_data;
401 int shmem_page_offset, page_length, ret = 0;
402 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
404 int needs_clflush = 0;
405 struct sg_page_iter sg_iter;
407 user_data = to_user_ptr(args->data_ptr);
410 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
412 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) {
413 /* If we're not in the cpu read domain, set ourself into the gtt
414 * read domain and manually flush cachelines (if required). This
415 * optimizes for the case when the gpu will dirty the data
416 * anyway again before the next pread happens. */
417 if (obj->cache_level == I915_CACHE_NONE)
419 if (i915_gem_obj_ggtt_bound(obj)) {
420 ret = i915_gem_object_set_to_gtt_domain(obj, false);
426 ret = i915_gem_object_get_pages(obj);
430 i915_gem_object_pin_pages(obj);
432 offset = args->offset;
434 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
435 offset >> PAGE_SHIFT) {
436 struct page *page = sg_page_iter_page(&sg_iter);
441 /* Operation in this page
443 * shmem_page_offset = offset within page in shmem file
444 * page_length = bytes to copy for this page
446 shmem_page_offset = offset_in_page(offset);
447 page_length = remain;
448 if ((shmem_page_offset + page_length) > PAGE_SIZE)
449 page_length = PAGE_SIZE - shmem_page_offset;
451 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
452 (page_to_phys(page) & (1 << 17)) != 0;
454 ret = shmem_pread_fast(page, shmem_page_offset, page_length,
455 user_data, page_do_bit17_swizzling,
460 mutex_unlock(&dev->struct_mutex);
462 if (likely(!i915_prefault_disable) && !prefaulted) {
463 ret = fault_in_multipages_writeable(user_data, remain);
464 /* Userspace is tricking us, but we've already clobbered
465 * its pages with the prefault and promised to write the
466 * data up to the first fault. Hence ignore any errors
467 * and just continue. */
472 ret = shmem_pread_slow(page, shmem_page_offset, page_length,
473 user_data, page_do_bit17_swizzling,
476 mutex_lock(&dev->struct_mutex);
479 mark_page_accessed(page);
484 remain -= page_length;
485 user_data += page_length;
486 offset += page_length;
490 i915_gem_object_unpin_pages(obj);
496 * Reads data from the object referenced by handle.
498 * On error, the contents of *data are undefined.
501 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
502 struct drm_file *file)
504 struct drm_i915_gem_pread *args = data;
505 struct drm_i915_gem_object *obj;
511 if (!access_ok(VERIFY_WRITE,
512 to_user_ptr(args->data_ptr),
516 ret = i915_mutex_lock_interruptible(dev);
520 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
521 if (&obj->base == NULL) {
526 /* Bounds check source. */
527 if (args->offset > obj->base.size ||
528 args->size > obj->base.size - args->offset) {
533 /* prime objects have no backing filp to GEM pread/pwrite
536 if (!obj->base.filp) {
541 trace_i915_gem_object_pread(obj, args->offset, args->size);
543 ret = i915_gem_shmem_pread(dev, obj, args, file);
546 drm_gem_object_unreference(&obj->base);
548 mutex_unlock(&dev->struct_mutex);
552 /* This is the fast write path which cannot handle
553 * page faults in the source data
557 fast_user_write(struct io_mapping *mapping,
558 loff_t page_base, int page_offset,
559 char __user *user_data,
562 void __iomem *vaddr_atomic;
564 unsigned long unwritten;
566 vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
567 /* We can use the cpu mem copy function because this is X86. */
568 vaddr = (void __force*)vaddr_atomic + page_offset;
569 unwritten = __copy_from_user_inatomic_nocache(vaddr,
571 io_mapping_unmap_atomic(vaddr_atomic);
576 * This is the fast pwrite path, where we copy the data directly from the
577 * user into the GTT, uncached.
580 i915_gem_gtt_pwrite_fast(struct drm_device *dev,
581 struct drm_i915_gem_object *obj,
582 struct drm_i915_gem_pwrite *args,
583 struct drm_file *file)
585 drm_i915_private_t *dev_priv = dev->dev_private;
587 loff_t offset, page_base;
588 char __user *user_data;
589 int page_offset, page_length, ret;
591 ret = i915_gem_object_pin(obj, 0, true, true);
595 ret = i915_gem_object_set_to_gtt_domain(obj, true);
599 ret = i915_gem_object_put_fence(obj);
603 user_data = to_user_ptr(args->data_ptr);
606 offset = i915_gem_obj_ggtt_offset(obj) + args->offset;
609 /* Operation in this page
611 * page_base = page offset within aperture
612 * page_offset = offset within page
613 * page_length = bytes to copy for this page
615 page_base = offset & PAGE_MASK;
616 page_offset = offset_in_page(offset);
617 page_length = remain;
618 if ((page_offset + remain) > PAGE_SIZE)
619 page_length = PAGE_SIZE - page_offset;
621 /* If we get a fault while copying data, then (presumably) our
622 * source page isn't available. Return the error and we'll
623 * retry in the slow path.
625 if (fast_user_write(dev_priv->gtt.mappable, page_base,
626 page_offset, user_data, page_length)) {
631 remain -= page_length;
632 user_data += page_length;
633 offset += page_length;
637 i915_gem_object_unpin(obj);
642 /* Per-page copy function for the shmem pwrite fastpath.
643 * Flushes invalid cachelines before writing to the target if
644 * needs_clflush_before is set and flushes out any written cachelines after
645 * writing if needs_clflush is set. */
647 shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
648 char __user *user_data,
649 bool page_do_bit17_swizzling,
650 bool needs_clflush_before,
651 bool needs_clflush_after)
656 if (unlikely(page_do_bit17_swizzling))
659 vaddr = kmap_atomic(page);
660 if (needs_clflush_before)
661 drm_clflush_virt_range(vaddr + shmem_page_offset,
663 ret = __copy_from_user_inatomic_nocache(vaddr + shmem_page_offset,
666 if (needs_clflush_after)
667 drm_clflush_virt_range(vaddr + shmem_page_offset,
669 kunmap_atomic(vaddr);
671 return ret ? -EFAULT : 0;
674 /* Only difference to the fast-path function is that this can handle bit17
675 * and uses non-atomic copy and kmap functions. */
677 shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
678 char __user *user_data,
679 bool page_do_bit17_swizzling,
680 bool needs_clflush_before,
681 bool needs_clflush_after)
687 if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
688 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
690 page_do_bit17_swizzling);
691 if (page_do_bit17_swizzling)
692 ret = __copy_from_user_swizzled(vaddr, shmem_page_offset,
696 ret = __copy_from_user(vaddr + shmem_page_offset,
699 if (needs_clflush_after)
700 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
702 page_do_bit17_swizzling);
705 return ret ? -EFAULT : 0;
709 i915_gem_shmem_pwrite(struct drm_device *dev,
710 struct drm_i915_gem_object *obj,
711 struct drm_i915_gem_pwrite *args,
712 struct drm_file *file)
716 char __user *user_data;
717 int shmem_page_offset, page_length, ret = 0;
718 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
719 int hit_slowpath = 0;
720 int needs_clflush_after = 0;
721 int needs_clflush_before = 0;
722 struct sg_page_iter sg_iter;
724 user_data = to_user_ptr(args->data_ptr);
727 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
729 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
730 /* If we're not in the cpu write domain, set ourself into the gtt
731 * write domain and manually flush cachelines (if required). This
732 * optimizes for the case when the gpu will use the data
733 * right away and we therefore have to clflush anyway. */
734 if (obj->cache_level == I915_CACHE_NONE)
735 needs_clflush_after = 1;
736 if (i915_gem_obj_ggtt_bound(obj)) {
737 ret = i915_gem_object_set_to_gtt_domain(obj, true);
742 /* Same trick applies for invalidate partially written cachelines before
744 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)
745 && obj->cache_level == I915_CACHE_NONE)
746 needs_clflush_before = 1;
748 ret = i915_gem_object_get_pages(obj);
752 i915_gem_object_pin_pages(obj);
754 offset = args->offset;
757 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
758 offset >> PAGE_SHIFT) {
759 struct page *page = sg_page_iter_page(&sg_iter);
760 int partial_cacheline_write;
765 /* Operation in this page
767 * shmem_page_offset = offset within page in shmem file
768 * page_length = bytes to copy for this page
770 shmem_page_offset = offset_in_page(offset);
772 page_length = remain;
773 if ((shmem_page_offset + page_length) > PAGE_SIZE)
774 page_length = PAGE_SIZE - shmem_page_offset;
776 /* If we don't overwrite a cacheline completely we need to be
777 * careful to have up-to-date data by first clflushing. Don't
778 * overcomplicate things and flush the entire patch. */
779 partial_cacheline_write = needs_clflush_before &&
780 ((shmem_page_offset | page_length)
781 & (boot_cpu_data.x86_clflush_size - 1));
783 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
784 (page_to_phys(page) & (1 << 17)) != 0;
786 ret = shmem_pwrite_fast(page, shmem_page_offset, page_length,
787 user_data, page_do_bit17_swizzling,
788 partial_cacheline_write,
789 needs_clflush_after);
794 mutex_unlock(&dev->struct_mutex);
795 ret = shmem_pwrite_slow(page, shmem_page_offset, page_length,
796 user_data, page_do_bit17_swizzling,
797 partial_cacheline_write,
798 needs_clflush_after);
800 mutex_lock(&dev->struct_mutex);
803 set_page_dirty(page);
804 mark_page_accessed(page);
809 remain -= page_length;
810 user_data += page_length;
811 offset += page_length;
815 i915_gem_object_unpin_pages(obj);
819 * Fixup: Flush cpu caches in case we didn't flush the dirty
820 * cachelines in-line while writing and the object moved
821 * out of the cpu write domain while we've dropped the lock.
823 if (!needs_clflush_after &&
824 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
825 i915_gem_clflush_object(obj);
826 i915_gem_chipset_flush(dev);
830 if (needs_clflush_after)
831 i915_gem_chipset_flush(dev);
837 * Writes data to the object referenced by handle.
839 * On error, the contents of the buffer that were to be modified are undefined.
842 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
843 struct drm_file *file)
845 struct drm_i915_gem_pwrite *args = data;
846 struct drm_i915_gem_object *obj;
852 if (!access_ok(VERIFY_READ,
853 to_user_ptr(args->data_ptr),
857 if (likely(!i915_prefault_disable)) {
858 ret = fault_in_multipages_readable(to_user_ptr(args->data_ptr),
864 ret = i915_mutex_lock_interruptible(dev);
868 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
869 if (&obj->base == NULL) {
874 /* Bounds check destination. */
875 if (args->offset > obj->base.size ||
876 args->size > obj->base.size - args->offset) {
881 /* prime objects have no backing filp to GEM pread/pwrite
884 if (!obj->base.filp) {
889 trace_i915_gem_object_pwrite(obj, args->offset, args->size);
892 /* We can only do the GTT pwrite on untiled buffers, as otherwise
893 * it would end up going through the fenced access, and we'll get
894 * different detiling behavior between reading and writing.
895 * pread/pwrite currently are reading and writing from the CPU
896 * perspective, requiring manual detiling by the client.
899 ret = i915_gem_phys_pwrite(dev, obj, args, file);
903 if (obj->cache_level == I915_CACHE_NONE &&
904 obj->tiling_mode == I915_TILING_NONE &&
905 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
906 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
907 /* Note that the gtt paths might fail with non-page-backed user
908 * pointers (e.g. gtt mappings when moving data between
909 * textures). Fallback to the shmem path in that case. */
912 if (ret == -EFAULT || ret == -ENOSPC)
913 ret = i915_gem_shmem_pwrite(dev, obj, args, file);
916 drm_gem_object_unreference(&obj->base);
918 mutex_unlock(&dev->struct_mutex);
923 i915_gem_check_wedge(struct i915_gpu_error *error,
926 if (i915_reset_in_progress(error)) {
927 /* Non-interruptible callers can't handle -EAGAIN, hence return
928 * -EIO unconditionally for these. */
932 /* Recovery complete, but the reset failed ... */
933 if (i915_terminally_wedged(error))
943 * Compare seqno against outstanding lazy request. Emit a request if they are
947 i915_gem_check_olr(struct intel_ring_buffer *ring, u32 seqno)
951 BUG_ON(!mutex_is_locked(&ring->dev->struct_mutex));
954 if (seqno == ring->outstanding_lazy_request)
955 ret = i915_add_request(ring, NULL);
961 * __wait_seqno - wait until execution of seqno has finished
962 * @ring: the ring expected to report seqno
964 * @reset_counter: reset sequence associated with the given seqno
965 * @interruptible: do an interruptible wait (normally yes)
966 * @timeout: in - how long to wait (NULL forever); out - how much time remaining
968 * Note: It is of utmost importance that the passed in seqno and reset_counter
969 * values have been read by the caller in an smp safe manner. Where read-side
970 * locks are involved, it is sufficient to read the reset_counter before
971 * unlocking the lock that protects the seqno. For lockless tricks, the
972 * reset_counter _must_ be read before, and an appropriate smp_rmb must be
975 * Returns 0 if the seqno was found within the alloted time. Else returns the
976 * errno with remaining time filled in timeout argument.
978 static int __wait_seqno(struct intel_ring_buffer *ring, u32 seqno,
979 unsigned reset_counter,
980 bool interruptible, struct timespec *timeout)
982 drm_i915_private_t *dev_priv = ring->dev->dev_private;
983 struct timespec before, now, wait_time={1,0};
984 unsigned long timeout_jiffies;
986 bool wait_forever = true;
989 if (i915_seqno_passed(ring->get_seqno(ring, true), seqno))
992 trace_i915_gem_request_wait_begin(ring, seqno);
994 if (timeout != NULL) {
995 wait_time = *timeout;
996 wait_forever = false;
999 timeout_jiffies = timespec_to_jiffies_timeout(&wait_time);
1001 if (WARN_ON(!ring->irq_get(ring)))
1004 /* Record current time in case interrupted by signal, or wedged * */
1005 getrawmonotonic(&before);
1008 (i915_seqno_passed(ring->get_seqno(ring, false), seqno) || \
1009 i915_reset_in_progress(&dev_priv->gpu_error) || \
1010 reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter))
1013 end = wait_event_interruptible_timeout(ring->irq_queue,
1017 end = wait_event_timeout(ring->irq_queue, EXIT_COND,
1020 /* We need to check whether any gpu reset happened in between
1021 * the caller grabbing the seqno and now ... */
1022 if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter))
1025 /* ... but upgrade the -EGAIN to an -EIO if the gpu is truely
1027 ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1030 } while (end == 0 && wait_forever);
1032 getrawmonotonic(&now);
1034 ring->irq_put(ring);
1035 trace_i915_gem_request_wait_end(ring, seqno);
1039 struct timespec sleep_time = timespec_sub(now, before);
1040 *timeout = timespec_sub(*timeout, sleep_time);
1041 if (!timespec_valid(timeout)) /* i.e. negative time remains */
1042 set_normalized_timespec(timeout, 0, 0);
1047 case -EAGAIN: /* Wedged */
1048 case -ERESTARTSYS: /* Signal */
1050 case 0: /* Timeout */
1052 default: /* Completed */
1053 WARN_ON(end < 0); /* We're not aware of other errors */
1059 * Waits for a sequence number to be signaled, and cleans up the
1060 * request and object lists appropriately for that event.
1063 i915_wait_seqno(struct intel_ring_buffer *ring, uint32_t seqno)
1065 struct drm_device *dev = ring->dev;
1066 struct drm_i915_private *dev_priv = dev->dev_private;
1067 bool interruptible = dev_priv->mm.interruptible;
1070 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1073 ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1077 ret = i915_gem_check_olr(ring, seqno);
1081 return __wait_seqno(ring, seqno,
1082 atomic_read(&dev_priv->gpu_error.reset_counter),
1083 interruptible, NULL);
1087 i915_gem_object_wait_rendering__tail(struct drm_i915_gem_object *obj,
1088 struct intel_ring_buffer *ring)
1090 i915_gem_retire_requests_ring(ring);
1092 /* Manually manage the write flush as we may have not yet
1093 * retired the buffer.
1095 * Note that the last_write_seqno is always the earlier of
1096 * the two (read/write) seqno, so if we haved successfully waited,
1097 * we know we have passed the last write.
1099 obj->last_write_seqno = 0;
1100 obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
1106 * Ensures that all rendering to the object has completed and the object is
1107 * safe to unbind from the GTT or access from the CPU.
1109 static __must_check int
1110 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
1113 struct intel_ring_buffer *ring = obj->ring;
1117 seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno;
1121 ret = i915_wait_seqno(ring, seqno);
1125 return i915_gem_object_wait_rendering__tail(obj, ring);
1128 /* A nonblocking variant of the above wait. This is a highly dangerous routine
1129 * as the object state may change during this call.
1131 static __must_check int
1132 i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
1135 struct drm_device *dev = obj->base.dev;
1136 struct drm_i915_private *dev_priv = dev->dev_private;
1137 struct intel_ring_buffer *ring = obj->ring;
1138 unsigned reset_counter;
1142 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1143 BUG_ON(!dev_priv->mm.interruptible);
1145 seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno;
1149 ret = i915_gem_check_wedge(&dev_priv->gpu_error, true);
1153 ret = i915_gem_check_olr(ring, seqno);
1157 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
1158 mutex_unlock(&dev->struct_mutex);
1159 ret = __wait_seqno(ring, seqno, reset_counter, true, NULL);
1160 mutex_lock(&dev->struct_mutex);
1164 return i915_gem_object_wait_rendering__tail(obj, ring);
1168 * Called when user space prepares to use an object with the CPU, either
1169 * through the mmap ioctl's mapping or a GTT mapping.
1172 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1173 struct drm_file *file)
1175 struct drm_i915_gem_set_domain *args = data;
1176 struct drm_i915_gem_object *obj;
1177 uint32_t read_domains = args->read_domains;
1178 uint32_t write_domain = args->write_domain;
1181 /* Only handle setting domains to types used by the CPU. */
1182 if (write_domain & I915_GEM_GPU_DOMAINS)
1185 if (read_domains & I915_GEM_GPU_DOMAINS)
1188 /* Having something in the write domain implies it's in the read
1189 * domain, and only that read domain. Enforce that in the request.
1191 if (write_domain != 0 && read_domains != write_domain)
1194 ret = i915_mutex_lock_interruptible(dev);
1198 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1199 if (&obj->base == NULL) {
1204 /* Try to flush the object off the GPU without holding the lock.
1205 * We will repeat the flush holding the lock in the normal manner
1206 * to catch cases where we are gazumped.
1208 ret = i915_gem_object_wait_rendering__nonblocking(obj, !write_domain);
1212 if (read_domains & I915_GEM_DOMAIN_GTT) {
1213 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1215 /* Silently promote "you're not bound, there was nothing to do"
1216 * to success, since the client was just asking us to
1217 * make sure everything was done.
1222 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1226 drm_gem_object_unreference(&obj->base);
1228 mutex_unlock(&dev->struct_mutex);
1233 * Called when user space has done writes to this buffer
1236 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1237 struct drm_file *file)
1239 struct drm_i915_gem_sw_finish *args = data;
1240 struct drm_i915_gem_object *obj;
1243 ret = i915_mutex_lock_interruptible(dev);
1247 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1248 if (&obj->base == NULL) {
1253 /* Pinned buffers may be scanout, so flush the cache */
1255 i915_gem_object_flush_cpu_write_domain(obj);
1257 drm_gem_object_unreference(&obj->base);
1259 mutex_unlock(&dev->struct_mutex);
1264 * Maps the contents of an object, returning the address it is mapped
1267 * While the mapping holds a reference on the contents of the object, it doesn't
1268 * imply a ref on the object itself.
1271 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1272 struct drm_file *file)
1274 struct drm_i915_gem_mmap *args = data;
1275 struct drm_gem_object *obj;
1278 obj = drm_gem_object_lookup(dev, file, args->handle);
1282 /* prime objects have no backing filp to GEM mmap
1286 drm_gem_object_unreference_unlocked(obj);
1290 addr = vm_mmap(obj->filp, 0, args->size,
1291 PROT_READ | PROT_WRITE, MAP_SHARED,
1293 drm_gem_object_unreference_unlocked(obj);
1294 if (IS_ERR((void *)addr))
1297 args->addr_ptr = (uint64_t) addr;
1303 * i915_gem_fault - fault a page into the GTT
1304 * vma: VMA in question
1307 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1308 * from userspace. The fault handler takes care of binding the object to
1309 * the GTT (if needed), allocating and programming a fence register (again,
1310 * only if needed based on whether the old reg is still valid or the object
1311 * is tiled) and inserting a new PTE into the faulting process.
1313 * Note that the faulting process may involve evicting existing objects
1314 * from the GTT and/or fence registers to make room. So performance may
1315 * suffer if the GTT working set is large or there are few fence registers
1318 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1320 struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1321 struct drm_device *dev = obj->base.dev;
1322 drm_i915_private_t *dev_priv = dev->dev_private;
1323 pgoff_t page_offset;
1326 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1328 /* We don't use vmf->pgoff since that has the fake offset */
1329 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1332 ret = i915_mutex_lock_interruptible(dev);
1336 trace_i915_gem_object_fault(obj, page_offset, true, write);
1338 /* Access to snoopable pages through the GTT is incoherent. */
1339 if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) {
1344 /* Now bind it into the GTT if needed */
1345 ret = i915_gem_object_pin(obj, 0, true, false);
1349 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1353 ret = i915_gem_object_get_fence(obj);
1357 obj->fault_mappable = true;
1359 pfn = dev_priv->gtt.mappable_base + i915_gem_obj_ggtt_offset(obj);
1363 /* Finally, remap it using the new GTT offset */
1364 ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1366 i915_gem_object_unpin(obj);
1368 mutex_unlock(&dev->struct_mutex);
1372 /* If this -EIO is due to a gpu hang, give the reset code a
1373 * chance to clean up the mess. Otherwise return the proper
1375 if (i915_terminally_wedged(&dev_priv->gpu_error))
1376 return VM_FAULT_SIGBUS;
1378 /* Give the error handler a chance to run and move the
1379 * objects off the GPU active list. Next time we service the
1380 * fault, we should be able to transition the page into the
1381 * GTT without touching the GPU (and so avoid further
1382 * EIO/EGAIN). If the GPU is wedged, then there is no issue
1383 * with coherency, just lost writes.
1391 * EBUSY is ok: this just means that another thread
1392 * already did the job.
1394 return VM_FAULT_NOPAGE;
1396 return VM_FAULT_OOM;
1398 return VM_FAULT_SIGBUS;
1400 WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret);
1401 return VM_FAULT_SIGBUS;
1406 * i915_gem_release_mmap - remove physical page mappings
1407 * @obj: obj in question
1409 * Preserve the reservation of the mmapping with the DRM core code, but
1410 * relinquish ownership of the pages back to the system.
1412 * It is vital that we remove the page mapping if we have mapped a tiled
1413 * object through the GTT and then lose the fence register due to
1414 * resource pressure. Similarly if the object has been moved out of the
1415 * aperture, than pages mapped into userspace must be revoked. Removing the
1416 * mapping will then trigger a page fault on the next user access, allowing
1417 * fixup by i915_gem_fault().
1420 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1422 if (!obj->fault_mappable)
1425 if (obj->base.dev->dev_mapping)
1426 unmap_mapping_range(obj->base.dev->dev_mapping,
1427 (loff_t)obj->base.map_list.hash.key<<PAGE_SHIFT,
1430 obj->fault_mappable = false;
1434 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1438 if (INTEL_INFO(dev)->gen >= 4 ||
1439 tiling_mode == I915_TILING_NONE)
1442 /* Previous chips need a power-of-two fence region when tiling */
1443 if (INTEL_INFO(dev)->gen == 3)
1444 gtt_size = 1024*1024;
1446 gtt_size = 512*1024;
1448 while (gtt_size < size)
1455 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1456 * @obj: object to check
1458 * Return the required GTT alignment for an object, taking into account
1459 * potential fence register mapping.
1462 i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size,
1463 int tiling_mode, bool fenced)
1466 * Minimum alignment is 4k (GTT page size), but might be greater
1467 * if a fence register is needed for the object.
1469 if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) ||
1470 tiling_mode == I915_TILING_NONE)
1474 * Previous chips need to be aligned to the size of the smallest
1475 * fence register that can contain the object.
1477 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1480 static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj)
1482 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1485 if (obj->base.map_list.map)
1488 dev_priv->mm.shrinker_no_lock_stealing = true;
1490 ret = drm_gem_create_mmap_offset(&obj->base);
1494 /* Badly fragmented mmap space? The only way we can recover
1495 * space is by destroying unwanted objects. We can't randomly release
1496 * mmap_offsets as userspace expects them to be persistent for the
1497 * lifetime of the objects. The closest we can is to release the
1498 * offsets on purgeable objects by truncating it and marking it purged,
1499 * which prevents userspace from ever using that object again.
1501 i915_gem_purge(dev_priv, obj->base.size >> PAGE_SHIFT);
1502 ret = drm_gem_create_mmap_offset(&obj->base);
1506 i915_gem_shrink_all(dev_priv);
1507 ret = drm_gem_create_mmap_offset(&obj->base);
1509 dev_priv->mm.shrinker_no_lock_stealing = false;
1514 static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)
1516 if (!obj->base.map_list.map)
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 = (u64)obj->base.map_list.hash.key << PAGE_SHIFT;
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;
1729 i915_gem_evict_everything(dev_priv->dev);
1731 list_for_each_entry_safe(obj, next, &dev_priv->mm.unbound_list,
1733 i915_gem_object_put_pages(obj);
1737 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
1739 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1741 struct address_space *mapping;
1742 struct sg_table *st;
1743 struct scatterlist *sg;
1744 struct sg_page_iter sg_iter;
1746 unsigned long last_pfn = 0; /* suppress gcc warning */
1749 /* Assert that the object is not currently in any GPU domain. As it
1750 * wasn't in the GTT, there shouldn't be any way it could have been in
1753 BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
1754 BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
1756 st = kmalloc(sizeof(*st), GFP_KERNEL);
1760 page_count = obj->base.size / PAGE_SIZE;
1761 if (sg_alloc_table(st, page_count, GFP_KERNEL)) {
1767 /* Get the list of pages out of our struct file. They'll be pinned
1768 * at this point until we release them.
1770 * Fail silently without starting the shrinker
1772 mapping = file_inode(obj->base.filp)->i_mapping;
1773 gfp = mapping_gfp_mask(mapping);
1774 gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD;
1775 gfp &= ~(__GFP_IO | __GFP_WAIT);
1778 for (i = 0; i < page_count; i++) {
1779 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
1781 i915_gem_purge(dev_priv, page_count);
1782 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
1785 /* We've tried hard to allocate the memory by reaping
1786 * our own buffer, now let the real VM do its job and
1787 * go down in flames if truly OOM.
1789 gfp &= ~(__GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD);
1790 gfp |= __GFP_IO | __GFP_WAIT;
1792 i915_gem_shrink_all(dev_priv);
1793 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
1797 gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD;
1798 gfp &= ~(__GFP_IO | __GFP_WAIT);
1800 #ifdef CONFIG_SWIOTLB
1801 if (swiotlb_nr_tbl()) {
1803 sg_set_page(sg, page, PAGE_SIZE, 0);
1808 if (!i || page_to_pfn(page) != last_pfn + 1) {
1812 sg_set_page(sg, page, PAGE_SIZE, 0);
1814 sg->length += PAGE_SIZE;
1816 last_pfn = page_to_pfn(page);
1818 #ifdef CONFIG_SWIOTLB
1819 if (!swiotlb_nr_tbl())
1824 if (i915_gem_object_needs_bit17_swizzle(obj))
1825 i915_gem_object_do_bit_17_swizzle(obj);
1831 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0)
1832 page_cache_release(sg_page_iter_page(&sg_iter));
1835 return PTR_ERR(page);
1838 /* Ensure that the associated pages are gathered from the backing storage
1839 * and pinned into our object. i915_gem_object_get_pages() may be called
1840 * multiple times before they are released by a single call to
1841 * i915_gem_object_put_pages() - once the pages are no longer referenced
1842 * either as a result of memory pressure (reaping pages under the shrinker)
1843 * or as the object is itself released.
1846 i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
1848 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1849 const struct drm_i915_gem_object_ops *ops = obj->ops;
1855 if (obj->madv != I915_MADV_WILLNEED) {
1856 DRM_ERROR("Attempting to obtain a purgeable object\n");
1860 BUG_ON(obj->pages_pin_count);
1862 ret = ops->get_pages(obj);
1866 list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list);
1871 i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
1872 struct intel_ring_buffer *ring)
1874 struct drm_device *dev = obj->base.dev;
1875 struct drm_i915_private *dev_priv = dev->dev_private;
1876 struct i915_address_space *vm = &dev_priv->gtt.base;
1877 u32 seqno = intel_ring_get_seqno(ring);
1879 BUG_ON(ring == NULL);
1880 if (obj->ring != ring && obj->last_write_seqno) {
1881 /* Keep the seqno relative to the current ring */
1882 obj->last_write_seqno = seqno;
1886 /* Add a reference if we're newly entering the active list. */
1888 drm_gem_object_reference(&obj->base);
1892 /* Move from whatever list we were on to the tail of execution. */
1893 list_move_tail(&obj->mm_list, &vm->active_list);
1894 list_move_tail(&obj->ring_list, &ring->active_list);
1896 obj->last_read_seqno = seqno;
1898 if (obj->fenced_gpu_access) {
1899 obj->last_fenced_seqno = seqno;
1901 /* Bump MRU to take account of the delayed flush */
1902 if (obj->fence_reg != I915_FENCE_REG_NONE) {
1903 struct drm_i915_fence_reg *reg;
1905 reg = &dev_priv->fence_regs[obj->fence_reg];
1906 list_move_tail(®->lru_list,
1907 &dev_priv->mm.fence_list);
1913 i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
1915 struct drm_device *dev = obj->base.dev;
1916 struct drm_i915_private *dev_priv = dev->dev_private;
1917 struct i915_address_space *vm = &dev_priv->gtt.base;
1919 BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS);
1920 BUG_ON(!obj->active);
1922 list_move_tail(&obj->mm_list, &vm->inactive_list);
1924 list_del_init(&obj->ring_list);
1927 obj->last_read_seqno = 0;
1928 obj->last_write_seqno = 0;
1929 obj->base.write_domain = 0;
1931 obj->last_fenced_seqno = 0;
1932 obj->fenced_gpu_access = false;
1935 drm_gem_object_unreference(&obj->base);
1937 WARN_ON(i915_verify_lists(dev));
1941 i915_gem_init_seqno(struct drm_device *dev, u32 seqno)
1943 struct drm_i915_private *dev_priv = dev->dev_private;
1944 struct intel_ring_buffer *ring;
1947 /* Carefully retire all requests without writing to the rings */
1948 for_each_ring(ring, dev_priv, i) {
1949 ret = intel_ring_idle(ring);
1953 i915_gem_retire_requests(dev);
1955 /* Finally reset hw state */
1956 for_each_ring(ring, dev_priv, i) {
1957 intel_ring_init_seqno(ring, seqno);
1959 for (j = 0; j < ARRAY_SIZE(ring->sync_seqno); j++)
1960 ring->sync_seqno[j] = 0;
1966 int i915_gem_set_seqno(struct drm_device *dev, u32 seqno)
1968 struct drm_i915_private *dev_priv = dev->dev_private;
1974 /* HWS page needs to be set less than what we
1975 * will inject to ring
1977 ret = i915_gem_init_seqno(dev, seqno - 1);
1981 /* Carefully set the last_seqno value so that wrap
1982 * detection still works
1984 dev_priv->next_seqno = seqno;
1985 dev_priv->last_seqno = seqno - 1;
1986 if (dev_priv->last_seqno == 0)
1987 dev_priv->last_seqno--;
1993 i915_gem_get_seqno(struct drm_device *dev, u32 *seqno)
1995 struct drm_i915_private *dev_priv = dev->dev_private;
1997 /* reserve 0 for non-seqno */
1998 if (dev_priv->next_seqno == 0) {
1999 int ret = i915_gem_init_seqno(dev, 0);
2003 dev_priv->next_seqno = 1;
2006 *seqno = dev_priv->last_seqno = dev_priv->next_seqno++;
2010 int __i915_add_request(struct intel_ring_buffer *ring,
2011 struct drm_file *file,
2012 struct drm_i915_gem_object *obj,
2015 drm_i915_private_t *dev_priv = ring->dev->dev_private;
2016 struct drm_i915_gem_request *request;
2017 u32 request_ring_position, request_start;
2021 request_start = intel_ring_get_tail(ring);
2023 * Emit any outstanding flushes - execbuf can fail to emit the flush
2024 * after having emitted the batchbuffer command. Hence we need to fix
2025 * things up similar to emitting the lazy request. The difference here
2026 * is that the flush _must_ happen before the next request, no matter
2029 ret = intel_ring_flush_all_caches(ring);
2033 request = kmalloc(sizeof(*request), GFP_KERNEL);
2034 if (request == NULL)
2038 /* Record the position of the start of the request so that
2039 * should we detect the updated seqno part-way through the
2040 * GPU processing the request, we never over-estimate the
2041 * position of the head.
2043 request_ring_position = intel_ring_get_tail(ring);
2045 ret = ring->add_request(ring);
2051 request->seqno = intel_ring_get_seqno(ring);
2052 request->ring = ring;
2053 request->head = request_start;
2054 request->tail = request_ring_position;
2055 request->ctx = ring->last_context;
2056 request->batch_obj = obj;
2058 /* Whilst this request exists, batch_obj will be on the
2059 * active_list, and so will hold the active reference. Only when this
2060 * request is retired will the the batch_obj be moved onto the
2061 * inactive_list and lose its active reference. Hence we do not need
2062 * to explicitly hold another reference here.
2066 i915_gem_context_reference(request->ctx);
2068 request->emitted_jiffies = jiffies;
2069 was_empty = list_empty(&ring->request_list);
2070 list_add_tail(&request->list, &ring->request_list);
2071 request->file_priv = NULL;
2074 struct drm_i915_file_private *file_priv = file->driver_priv;
2076 spin_lock(&file_priv->mm.lock);
2077 request->file_priv = file_priv;
2078 list_add_tail(&request->client_list,
2079 &file_priv->mm.request_list);
2080 spin_unlock(&file_priv->mm.lock);
2083 trace_i915_gem_request_add(ring, request->seqno);
2084 ring->outstanding_lazy_request = 0;
2086 if (!dev_priv->ums.mm_suspended) {
2087 i915_queue_hangcheck(ring->dev);
2090 queue_delayed_work(dev_priv->wq,
2091 &dev_priv->mm.retire_work,
2092 round_jiffies_up_relative(HZ));
2093 intel_mark_busy(dev_priv->dev);
2098 *out_seqno = request->seqno;
2103 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
2105 struct drm_i915_file_private *file_priv = request->file_priv;
2110 spin_lock(&file_priv->mm.lock);
2111 if (request->file_priv) {
2112 list_del(&request->client_list);
2113 request->file_priv = NULL;
2115 spin_unlock(&file_priv->mm.lock);
2118 static bool i915_head_inside_object(u32 acthd, struct drm_i915_gem_object *obj)
2120 if (acthd >= i915_gem_obj_ggtt_offset(obj) &&
2121 acthd < i915_gem_obj_ggtt_offset(obj) + obj->base.size)
2127 static bool i915_head_inside_request(const u32 acthd_unmasked,
2128 const u32 request_start,
2129 const u32 request_end)
2131 const u32 acthd = acthd_unmasked & HEAD_ADDR;
2133 if (request_start < request_end) {
2134 if (acthd >= request_start && acthd < request_end)
2136 } else if (request_start > request_end) {
2137 if (acthd >= request_start || acthd < request_end)
2144 static bool i915_request_guilty(struct drm_i915_gem_request *request,
2145 const u32 acthd, bool *inside)
2147 /* There is a possibility that unmasked head address
2148 * pointing inside the ring, matches the batch_obj address range.
2149 * However this is extremely unlikely.
2152 if (request->batch_obj) {
2153 if (i915_head_inside_object(acthd, request->batch_obj)) {
2159 if (i915_head_inside_request(acthd, request->head, request->tail)) {
2167 static void i915_set_reset_status(struct intel_ring_buffer *ring,
2168 struct drm_i915_gem_request *request,
2171 struct i915_ctx_hang_stats *hs = NULL;
2172 bool inside, guilty;
2174 /* Innocent until proven guilty */
2177 if (ring->hangcheck.action != wait &&
2178 i915_request_guilty(request, acthd, &inside)) {
2179 DRM_ERROR("%s hung %s bo (0x%lx ctx %d) at 0x%x\n",
2181 inside ? "inside" : "flushing",
2182 request->batch_obj ?
2183 i915_gem_obj_ggtt_offset(request->batch_obj) : 0,
2184 request->ctx ? request->ctx->id : 0,
2190 /* If contexts are disabled or this is the default context, use
2191 * file_priv->reset_state
2193 if (request->ctx && request->ctx->id != DEFAULT_CONTEXT_ID)
2194 hs = &request->ctx->hang_stats;
2195 else if (request->file_priv)
2196 hs = &request->file_priv->hang_stats;
2202 hs->batch_pending++;
2206 static void i915_gem_free_request(struct drm_i915_gem_request *request)
2208 list_del(&request->list);
2209 i915_gem_request_remove_from_client(request);
2212 i915_gem_context_unreference(request->ctx);
2217 static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv,
2218 struct intel_ring_buffer *ring)
2220 u32 completed_seqno;
2223 acthd = intel_ring_get_active_head(ring);
2224 completed_seqno = ring->get_seqno(ring, false);
2226 while (!list_empty(&ring->request_list)) {
2227 struct drm_i915_gem_request *request;
2229 request = list_first_entry(&ring->request_list,
2230 struct drm_i915_gem_request,
2233 if (request->seqno > completed_seqno)
2234 i915_set_reset_status(ring, request, acthd);
2236 i915_gem_free_request(request);
2239 while (!list_empty(&ring->active_list)) {
2240 struct drm_i915_gem_object *obj;
2242 obj = list_first_entry(&ring->active_list,
2243 struct drm_i915_gem_object,
2246 i915_gem_object_move_to_inactive(obj);
2250 void i915_gem_restore_fences(struct drm_device *dev)
2252 struct drm_i915_private *dev_priv = dev->dev_private;
2255 for (i = 0; i < dev_priv->num_fence_regs; i++) {
2256 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
2259 * Commit delayed tiling changes if we have an object still
2260 * attached to the fence, otherwise just clear the fence.
2263 i915_gem_object_update_fence(reg->obj, reg,
2264 reg->obj->tiling_mode);
2266 i915_gem_write_fence(dev, i, NULL);
2271 void i915_gem_reset(struct drm_device *dev)
2273 struct drm_i915_private *dev_priv = dev->dev_private;
2274 struct i915_address_space *vm = &dev_priv->gtt.base;
2275 struct drm_i915_gem_object *obj;
2276 struct intel_ring_buffer *ring;
2279 for_each_ring(ring, dev_priv, i)
2280 i915_gem_reset_ring_lists(dev_priv, ring);
2282 /* Move everything out of the GPU domains to ensure we do any
2283 * necessary invalidation upon reuse.
2285 list_for_each_entry(obj, &vm->inactive_list, mm_list)
2286 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
2288 i915_gem_restore_fences(dev);
2292 * This function clears the request list as sequence numbers are passed.
2295 i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
2299 if (list_empty(&ring->request_list))
2302 WARN_ON(i915_verify_lists(ring->dev));
2304 seqno = ring->get_seqno(ring, true);
2306 while (!list_empty(&ring->request_list)) {
2307 struct drm_i915_gem_request *request;
2309 request = list_first_entry(&ring->request_list,
2310 struct drm_i915_gem_request,
2313 if (!i915_seqno_passed(seqno, request->seqno))
2316 trace_i915_gem_request_retire(ring, request->seqno);
2317 /* We know the GPU must have read the request to have
2318 * sent us the seqno + interrupt, so use the position
2319 * of tail of the request to update the last known position
2322 ring->last_retired_head = request->tail;
2324 i915_gem_free_request(request);
2327 /* Move any buffers on the active list that are no longer referenced
2328 * by the ringbuffer to the flushing/inactive lists as appropriate.
2330 while (!list_empty(&ring->active_list)) {
2331 struct drm_i915_gem_object *obj;
2333 obj = list_first_entry(&ring->active_list,
2334 struct drm_i915_gem_object,
2337 if (!i915_seqno_passed(seqno, obj->last_read_seqno))
2340 i915_gem_object_move_to_inactive(obj);
2343 if (unlikely(ring->trace_irq_seqno &&
2344 i915_seqno_passed(seqno, ring->trace_irq_seqno))) {
2345 ring->irq_put(ring);
2346 ring->trace_irq_seqno = 0;
2349 WARN_ON(i915_verify_lists(ring->dev));
2353 i915_gem_retire_requests(struct drm_device *dev)
2355 drm_i915_private_t *dev_priv = dev->dev_private;
2356 struct intel_ring_buffer *ring;
2359 for_each_ring(ring, dev_priv, i)
2360 i915_gem_retire_requests_ring(ring);
2364 i915_gem_retire_work_handler(struct work_struct *work)
2366 drm_i915_private_t *dev_priv;
2367 struct drm_device *dev;
2368 struct intel_ring_buffer *ring;
2372 dev_priv = container_of(work, drm_i915_private_t,
2373 mm.retire_work.work);
2374 dev = dev_priv->dev;
2376 /* Come back later if the device is busy... */
2377 if (!mutex_trylock(&dev->struct_mutex)) {
2378 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
2379 round_jiffies_up_relative(HZ));
2383 i915_gem_retire_requests(dev);
2385 /* Send a periodic flush down the ring so we don't hold onto GEM
2386 * objects indefinitely.
2389 for_each_ring(ring, dev_priv, i) {
2390 if (ring->gpu_caches_dirty)
2391 i915_add_request(ring, NULL);
2393 idle &= list_empty(&ring->request_list);
2396 if (!dev_priv->ums.mm_suspended && !idle)
2397 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
2398 round_jiffies_up_relative(HZ));
2400 intel_mark_idle(dev);
2402 mutex_unlock(&dev->struct_mutex);
2406 * Ensures that an object will eventually get non-busy by flushing any required
2407 * write domains, emitting any outstanding lazy request and retiring and
2408 * completed requests.
2411 i915_gem_object_flush_active(struct drm_i915_gem_object *obj)
2416 ret = i915_gem_check_olr(obj->ring, obj->last_read_seqno);
2420 i915_gem_retire_requests_ring(obj->ring);
2427 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
2428 * @DRM_IOCTL_ARGS: standard ioctl arguments
2430 * Returns 0 if successful, else an error is returned with the remaining time in
2431 * the timeout parameter.
2432 * -ETIME: object is still busy after timeout
2433 * -ERESTARTSYS: signal interrupted the wait
2434 * -ENONENT: object doesn't exist
2435 * Also possible, but rare:
2436 * -EAGAIN: GPU wedged
2438 * -ENODEV: Internal IRQ fail
2439 * -E?: The add request failed
2441 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
2442 * non-zero timeout parameter the wait ioctl will wait for the given number of
2443 * nanoseconds on an object becoming unbusy. Since the wait itself does so
2444 * without holding struct_mutex the object may become re-busied before this
2445 * function completes. A similar but shorter * race condition exists in the busy
2449 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
2451 drm_i915_private_t *dev_priv = dev->dev_private;
2452 struct drm_i915_gem_wait *args = data;
2453 struct drm_i915_gem_object *obj;
2454 struct intel_ring_buffer *ring = NULL;
2455 struct timespec timeout_stack, *timeout = NULL;
2456 unsigned reset_counter;
2460 if (args->timeout_ns >= 0) {
2461 timeout_stack = ns_to_timespec(args->timeout_ns);
2462 timeout = &timeout_stack;
2465 ret = i915_mutex_lock_interruptible(dev);
2469 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle));
2470 if (&obj->base == NULL) {
2471 mutex_unlock(&dev->struct_mutex);
2475 /* Need to make sure the object gets inactive eventually. */
2476 ret = i915_gem_object_flush_active(obj);
2481 seqno = obj->last_read_seqno;
2488 /* Do this after OLR check to make sure we make forward progress polling
2489 * on this IOCTL with a 0 timeout (like busy ioctl)
2491 if (!args->timeout_ns) {
2496 drm_gem_object_unreference(&obj->base);
2497 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
2498 mutex_unlock(&dev->struct_mutex);
2500 ret = __wait_seqno(ring, seqno, reset_counter, true, timeout);
2502 args->timeout_ns = timespec_to_ns(timeout);
2506 drm_gem_object_unreference(&obj->base);
2507 mutex_unlock(&dev->struct_mutex);
2512 * i915_gem_object_sync - sync an object to a ring.
2514 * @obj: object which may be in use on another ring.
2515 * @to: ring we wish to use the object on. May be NULL.
2517 * This code is meant to abstract object synchronization with the GPU.
2518 * Calling with NULL implies synchronizing the object with the CPU
2519 * rather than a particular GPU ring.
2521 * Returns 0 if successful, else propagates up the lower layer error.
2524 i915_gem_object_sync(struct drm_i915_gem_object *obj,
2525 struct intel_ring_buffer *to)
2527 struct intel_ring_buffer *from = obj->ring;
2531 if (from == NULL || to == from)
2534 if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev))
2535 return i915_gem_object_wait_rendering(obj, false);
2537 idx = intel_ring_sync_index(from, to);
2539 seqno = obj->last_read_seqno;
2540 if (seqno <= from->sync_seqno[idx])
2543 ret = i915_gem_check_olr(obj->ring, seqno);
2547 ret = to->sync_to(to, from, seqno);
2549 /* We use last_read_seqno because sync_to()
2550 * might have just caused seqno wrap under
2553 from->sync_seqno[idx] = obj->last_read_seqno;
2558 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
2560 u32 old_write_domain, old_read_domains;
2562 /* Force a pagefault for domain tracking on next user access */
2563 i915_gem_release_mmap(obj);
2565 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
2568 /* Wait for any direct GTT access to complete */
2571 old_read_domains = obj->base.read_domains;
2572 old_write_domain = obj->base.write_domain;
2574 obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
2575 obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
2577 trace_i915_gem_object_change_domain(obj,
2583 * Unbinds an object from the GTT aperture.
2586 i915_gem_object_unbind(struct drm_i915_gem_object *obj)
2588 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
2589 struct i915_vma *vma;
2592 if (!i915_gem_obj_ggtt_bound(obj))
2598 BUG_ON(obj->pages == NULL);
2600 ret = i915_gem_object_finish_gpu(obj);
2603 /* Continue on if we fail due to EIO, the GPU is hung so we
2604 * should be safe and we need to cleanup or else we might
2605 * cause memory corruption through use-after-free.
2608 i915_gem_object_finish_gtt(obj);
2610 /* release the fence reg _after_ flushing */
2611 ret = i915_gem_object_put_fence(obj);
2615 trace_i915_gem_object_unbind(obj);
2617 if (obj->has_global_gtt_mapping)
2618 i915_gem_gtt_unbind_object(obj);
2619 if (obj->has_aliasing_ppgtt_mapping) {
2620 i915_ppgtt_unbind_object(dev_priv->mm.aliasing_ppgtt, obj);
2621 obj->has_aliasing_ppgtt_mapping = 0;
2623 i915_gem_gtt_finish_object(obj);
2624 i915_gem_object_unpin_pages(obj);
2626 list_del(&obj->mm_list);
2627 /* Avoid an unnecessary call to unbind on rebind. */
2628 obj->map_and_fenceable = true;
2630 vma = __i915_gem_obj_to_vma(obj);
2631 list_del(&vma->vma_link);
2632 drm_mm_remove_node(&vma->node);
2633 i915_gem_vma_destroy(vma);
2635 /* Since the unbound list is global, only move to that list if
2636 * no more VMAs exist.
2637 * NB: Until we have real VMAs there will only ever be one */
2638 WARN_ON(!list_empty(&obj->vma_list));
2639 if (list_empty(&obj->vma_list))
2640 list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list);
2645 int i915_gpu_idle(struct drm_device *dev)
2647 drm_i915_private_t *dev_priv = dev->dev_private;
2648 struct intel_ring_buffer *ring;
2651 /* Flush everything onto the inactive list. */
2652 for_each_ring(ring, dev_priv, i) {
2653 ret = i915_switch_context(ring, NULL, DEFAULT_CONTEXT_ID);
2657 ret = intel_ring_idle(ring);
2665 static void i965_write_fence_reg(struct drm_device *dev, int reg,
2666 struct drm_i915_gem_object *obj)
2668 drm_i915_private_t *dev_priv = dev->dev_private;
2670 int fence_pitch_shift;
2672 if (INTEL_INFO(dev)->gen >= 6) {
2673 fence_reg = FENCE_REG_SANDYBRIDGE_0;
2674 fence_pitch_shift = SANDYBRIDGE_FENCE_PITCH_SHIFT;
2676 fence_reg = FENCE_REG_965_0;
2677 fence_pitch_shift = I965_FENCE_PITCH_SHIFT;
2680 fence_reg += reg * 8;
2682 /* To w/a incoherency with non-atomic 64-bit register updates,
2683 * we split the 64-bit update into two 32-bit writes. In order
2684 * for a partial fence not to be evaluated between writes, we
2685 * precede the update with write to turn off the fence register,
2686 * and only enable the fence as the last step.
2688 * For extra levels of paranoia, we make sure each step lands
2689 * before applying the next step.
2691 I915_WRITE(fence_reg, 0);
2692 POSTING_READ(fence_reg);
2695 u32 size = i915_gem_obj_ggtt_size(obj);
2698 val = (uint64_t)((i915_gem_obj_ggtt_offset(obj) + size - 4096) &
2700 val |= i915_gem_obj_ggtt_offset(obj) & 0xfffff000;
2701 val |= (uint64_t)((obj->stride / 128) - 1) << fence_pitch_shift;
2702 if (obj->tiling_mode == I915_TILING_Y)
2703 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2704 val |= I965_FENCE_REG_VALID;
2706 I915_WRITE(fence_reg + 4, val >> 32);
2707 POSTING_READ(fence_reg + 4);
2709 I915_WRITE(fence_reg + 0, val);
2710 POSTING_READ(fence_reg);
2712 I915_WRITE(fence_reg + 4, 0);
2713 POSTING_READ(fence_reg + 4);
2717 static void i915_write_fence_reg(struct drm_device *dev, int reg,
2718 struct drm_i915_gem_object *obj)
2720 drm_i915_private_t *dev_priv = dev->dev_private;
2724 u32 size = i915_gem_obj_ggtt_size(obj);
2728 WARN((i915_gem_obj_ggtt_offset(obj) & ~I915_FENCE_START_MASK) ||
2729 (size & -size) != size ||
2730 (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
2731 "object 0x%08lx [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
2732 i915_gem_obj_ggtt_offset(obj), obj->map_and_fenceable, size);
2734 if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
2739 /* Note: pitch better be a power of two tile widths */
2740 pitch_val = obj->stride / tile_width;
2741 pitch_val = ffs(pitch_val) - 1;
2743 val = i915_gem_obj_ggtt_offset(obj);
2744 if (obj->tiling_mode == I915_TILING_Y)
2745 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2746 val |= I915_FENCE_SIZE_BITS(size);
2747 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2748 val |= I830_FENCE_REG_VALID;
2753 reg = FENCE_REG_830_0 + reg * 4;
2755 reg = FENCE_REG_945_8 + (reg - 8) * 4;
2757 I915_WRITE(reg, val);
2761 static void i830_write_fence_reg(struct drm_device *dev, int reg,
2762 struct drm_i915_gem_object *obj)
2764 drm_i915_private_t *dev_priv = dev->dev_private;
2768 u32 size = i915_gem_obj_ggtt_size(obj);
2771 WARN((i915_gem_obj_ggtt_offset(obj) & ~I830_FENCE_START_MASK) ||
2772 (size & -size) != size ||
2773 (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
2774 "object 0x%08lx not 512K or pot-size 0x%08x aligned\n",
2775 i915_gem_obj_ggtt_offset(obj), size);
2777 pitch_val = obj->stride / 128;
2778 pitch_val = ffs(pitch_val) - 1;
2780 val = i915_gem_obj_ggtt_offset(obj);
2781 if (obj->tiling_mode == I915_TILING_Y)
2782 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2783 val |= I830_FENCE_SIZE_BITS(size);
2784 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2785 val |= I830_FENCE_REG_VALID;
2789 I915_WRITE(FENCE_REG_830_0 + reg * 4, val);
2790 POSTING_READ(FENCE_REG_830_0 + reg * 4);
2793 inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object *obj)
2795 return obj && obj->base.read_domains & I915_GEM_DOMAIN_GTT;
2798 static void i915_gem_write_fence(struct drm_device *dev, int reg,
2799 struct drm_i915_gem_object *obj)
2801 struct drm_i915_private *dev_priv = dev->dev_private;
2803 /* Ensure that all CPU reads are completed before installing a fence
2804 * and all writes before removing the fence.
2806 if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj))
2809 WARN(obj && (!obj->stride || !obj->tiling_mode),
2810 "bogus fence setup with stride: 0x%x, tiling mode: %i\n",
2811 obj->stride, obj->tiling_mode);
2813 switch (INTEL_INFO(dev)->gen) {
2817 case 4: i965_write_fence_reg(dev, reg, obj); break;
2818 case 3: i915_write_fence_reg(dev, reg, obj); break;
2819 case 2: i830_write_fence_reg(dev, reg, obj); break;
2823 /* And similarly be paranoid that no direct access to this region
2824 * is reordered to before the fence is installed.
2826 if (i915_gem_object_needs_mb(obj))
2830 static inline int fence_number(struct drm_i915_private *dev_priv,
2831 struct drm_i915_fence_reg *fence)
2833 return fence - dev_priv->fence_regs;
2836 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
2837 struct drm_i915_fence_reg *fence,
2840 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2841 int reg = fence_number(dev_priv, fence);
2843 i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL);
2846 obj->fence_reg = reg;
2848 list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list);
2850 obj->fence_reg = I915_FENCE_REG_NONE;
2852 list_del_init(&fence->lru_list);
2854 obj->fence_dirty = false;
2858 i915_gem_object_wait_fence(struct drm_i915_gem_object *obj)
2860 if (obj->last_fenced_seqno) {
2861 int ret = i915_wait_seqno(obj->ring, obj->last_fenced_seqno);
2865 obj->last_fenced_seqno = 0;
2868 obj->fenced_gpu_access = false;
2873 i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
2875 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2876 struct drm_i915_fence_reg *fence;
2879 ret = i915_gem_object_wait_fence(obj);
2883 if (obj->fence_reg == I915_FENCE_REG_NONE)
2886 fence = &dev_priv->fence_regs[obj->fence_reg];
2888 i915_gem_object_fence_lost(obj);
2889 i915_gem_object_update_fence(obj, fence, false);
2894 static struct drm_i915_fence_reg *
2895 i915_find_fence_reg(struct drm_device *dev)
2897 struct drm_i915_private *dev_priv = dev->dev_private;
2898 struct drm_i915_fence_reg *reg, *avail;
2901 /* First try to find a free reg */
2903 for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2904 reg = &dev_priv->fence_regs[i];
2908 if (!reg->pin_count)
2915 /* None available, try to steal one or wait for a user to finish */
2916 list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
2927 * i915_gem_object_get_fence - set up fencing for an object
2928 * @obj: object to map through a fence reg
2930 * When mapping objects through the GTT, userspace wants to be able to write
2931 * to them without having to worry about swizzling if the object is tiled.
2932 * This function walks the fence regs looking for a free one for @obj,
2933 * stealing one if it can't find any.
2935 * It then sets up the reg based on the object's properties: address, pitch
2936 * and tiling format.
2938 * For an untiled surface, this removes any existing fence.
2941 i915_gem_object_get_fence(struct drm_i915_gem_object *obj)
2943 struct drm_device *dev = obj->base.dev;
2944 struct drm_i915_private *dev_priv = dev->dev_private;
2945 bool enable = obj->tiling_mode != I915_TILING_NONE;
2946 struct drm_i915_fence_reg *reg;
2949 /* Have we updated the tiling parameters upon the object and so
2950 * will need to serialise the write to the associated fence register?
2952 if (obj->fence_dirty) {
2953 ret = i915_gem_object_wait_fence(obj);
2958 /* Just update our place in the LRU if our fence is getting reused. */
2959 if (obj->fence_reg != I915_FENCE_REG_NONE) {
2960 reg = &dev_priv->fence_regs[obj->fence_reg];
2961 if (!obj->fence_dirty) {
2962 list_move_tail(®->lru_list,
2963 &dev_priv->mm.fence_list);
2966 } else if (enable) {
2967 reg = i915_find_fence_reg(dev);
2972 struct drm_i915_gem_object *old = reg->obj;
2974 ret = i915_gem_object_wait_fence(old);
2978 i915_gem_object_fence_lost(old);
2983 i915_gem_object_update_fence(obj, reg, enable);
2988 static bool i915_gem_valid_gtt_space(struct drm_device *dev,
2989 struct drm_mm_node *gtt_space,
2990 unsigned long cache_level)
2992 struct drm_mm_node *other;
2994 /* On non-LLC machines we have to be careful when putting differing
2995 * types of snoopable memory together to avoid the prefetcher
2996 * crossing memory domains and dying.
3001 if (!drm_mm_node_allocated(gtt_space))
3004 if (list_empty(>t_space->node_list))
3007 other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list);
3008 if (other->allocated && !other->hole_follows && other->color != cache_level)
3011 other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list);
3012 if (other->allocated && !gtt_space->hole_follows && other->color != cache_level)
3018 static void i915_gem_verify_gtt(struct drm_device *dev)
3021 struct drm_i915_private *dev_priv = dev->dev_private;
3022 struct drm_i915_gem_object *obj;
3025 list_for_each_entry(obj, &dev_priv->mm.gtt_list, global_list) {
3026 if (obj->gtt_space == NULL) {
3027 printk(KERN_ERR "object found on GTT list with no space reserved\n");
3032 if (obj->cache_level != obj->gtt_space->color) {
3033 printk(KERN_ERR "object reserved space [%08lx, %08lx] with wrong color, cache_level=%x, color=%lx\n",
3034 i915_gem_obj_ggtt_offset(obj),
3035 i915_gem_obj_ggtt_offset(obj) + i915_gem_obj_ggtt_size(obj),
3037 obj->gtt_space->color);
3042 if (!i915_gem_valid_gtt_space(dev,
3044 obj->cache_level)) {
3045 printk(KERN_ERR "invalid GTT space found at [%08lx, %08lx] - color=%x\n",
3046 i915_gem_obj_ggtt_offset(obj),
3047 i915_gem_obj_ggtt_offset(obj) + i915_gem_obj_ggtt_size(obj),
3059 * Finds free space in the GTT aperture and binds the object there.
3062 i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
3064 bool map_and_fenceable,
3067 struct drm_device *dev = obj->base.dev;
3068 drm_i915_private_t *dev_priv = dev->dev_private;
3069 struct i915_address_space *vm = &dev_priv->gtt.base;
3070 u32 size, fence_size, fence_alignment, unfenced_alignment;
3071 bool mappable, fenceable;
3072 size_t gtt_max = map_and_fenceable ?
3073 dev_priv->gtt.mappable_end : dev_priv->gtt.base.total;
3074 struct i915_vma *vma;
3077 if (WARN_ON(!list_empty(&obj->vma_list)))
3080 fence_size = i915_gem_get_gtt_size(dev,
3083 fence_alignment = i915_gem_get_gtt_alignment(dev,
3085 obj->tiling_mode, true);
3086 unfenced_alignment =
3087 i915_gem_get_gtt_alignment(dev,
3089 obj->tiling_mode, false);
3092 alignment = map_and_fenceable ? fence_alignment :
3094 if (map_and_fenceable && alignment & (fence_alignment - 1)) {
3095 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
3099 size = map_and_fenceable ? fence_size : obj->base.size;
3101 /* If the object is bigger than the entire aperture, reject it early
3102 * before evicting everything in a vain attempt to find space.
3104 if (obj->base.size > gtt_max) {
3105 DRM_ERROR("Attempting to bind an object larger than the aperture: object=%zd > %s aperture=%zu\n",
3107 map_and_fenceable ? "mappable" : "total",
3112 ret = i915_gem_object_get_pages(obj);
3116 i915_gem_object_pin_pages(obj);
3118 vma = i915_gem_vma_create(obj, &dev_priv->gtt.base);
3125 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
3128 obj->cache_level, 0, gtt_max);
3130 ret = i915_gem_evict_something(dev, size, alignment,
3139 if (WARN_ON(!i915_gem_valid_gtt_space(dev, &vma->node,
3140 obj->cache_level))) {
3142 goto err_remove_node;
3145 ret = i915_gem_gtt_prepare_object(obj);
3147 goto err_remove_node;
3149 list_move_tail(&obj->global_list, &dev_priv->mm.bound_list);
3150 list_add_tail(&obj->mm_list, &vm->inactive_list);
3151 list_add(&vma->vma_link, &obj->vma_list);
3154 i915_gem_obj_ggtt_size(obj) == fence_size &&
3155 (i915_gem_obj_ggtt_offset(obj) & (fence_alignment - 1)) == 0;
3157 mappable = i915_gem_obj_ggtt_offset(obj) + obj->base.size <=
3158 dev_priv->gtt.mappable_end;
3160 obj->map_and_fenceable = mappable && fenceable;
3162 trace_i915_gem_object_bind(obj, map_and_fenceable);
3163 i915_gem_verify_gtt(dev);
3167 drm_mm_remove_node(&vma->node);
3169 i915_gem_vma_destroy(vma);
3171 i915_gem_object_unpin_pages(obj);
3176 i915_gem_clflush_object(struct drm_i915_gem_object *obj)
3178 /* If we don't have a page list set up, then we're not pinned
3179 * to GPU, and we can ignore the cache flush because it'll happen
3180 * again at bind time.
3182 if (obj->pages == NULL)
3186 * Stolen memory is always coherent with the GPU as it is explicitly
3187 * marked as wc by the system, or the system is cache-coherent.
3192 /* If the GPU is snooping the contents of the CPU cache,
3193 * we do not need to manually clear the CPU cache lines. However,
3194 * the caches are only snooped when the render cache is
3195 * flushed/invalidated. As we always have to emit invalidations
3196 * and flushes when moving into and out of the RENDER domain, correct
3197 * snooping behaviour occurs naturally as the result of our domain
3200 if (obj->cache_level != I915_CACHE_NONE)
3203 trace_i915_gem_object_clflush(obj);
3205 drm_clflush_sg(obj->pages);
3208 /** Flushes the GTT write domain for the object if it's dirty. */
3210 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
3212 uint32_t old_write_domain;
3214 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
3217 /* No actual flushing is required for the GTT write domain. Writes
3218 * to it immediately go to main memory as far as we know, so there's
3219 * no chipset flush. It also doesn't land in render cache.
3221 * However, we do have to enforce the order so that all writes through
3222 * the GTT land before any writes to the device, such as updates to
3227 old_write_domain = obj->base.write_domain;
3228 obj->base.write_domain = 0;
3230 trace_i915_gem_object_change_domain(obj,
3231 obj->base.read_domains,
3235 /** Flushes the CPU write domain for the object if it's dirty. */
3237 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj)
3239 uint32_t old_write_domain;
3241 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
3244 i915_gem_clflush_object(obj);
3245 i915_gem_chipset_flush(obj->base.dev);
3246 old_write_domain = obj->base.write_domain;
3247 obj->base.write_domain = 0;
3249 trace_i915_gem_object_change_domain(obj,
3250 obj->base.read_domains,
3255 * Moves a single object to the GTT read, and possibly write domain.
3257 * This function returns when the move is complete, including waiting on
3261 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
3263 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
3264 uint32_t old_write_domain, old_read_domains;
3267 /* Not valid to be called on unbound objects. */
3268 if (!i915_gem_obj_ggtt_bound(obj))
3271 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
3274 ret = i915_gem_object_wait_rendering(obj, !write);
3278 i915_gem_object_flush_cpu_write_domain(obj);
3280 /* Serialise direct access to this object with the barriers for
3281 * coherent writes from the GPU, by effectively invalidating the
3282 * GTT domain upon first access.
3284 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
3287 old_write_domain = obj->base.write_domain;
3288 old_read_domains = obj->base.read_domains;
3290 /* It should now be out of any other write domains, and we can update
3291 * the domain values for our changes.
3293 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3294 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3296 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
3297 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
3301 trace_i915_gem_object_change_domain(obj,
3305 /* And bump the LRU for this access */
3306 if (i915_gem_object_is_inactive(obj))
3307 list_move_tail(&obj->mm_list,
3308 &dev_priv->gtt.base.inactive_list);
3313 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
3314 enum i915_cache_level cache_level)
3316 struct drm_device *dev = obj->base.dev;
3317 drm_i915_private_t *dev_priv = dev->dev_private;
3318 struct i915_vma *vma = __i915_gem_obj_to_vma(obj);
3321 if (obj->cache_level == cache_level)
3324 if (obj->pin_count) {
3325 DRM_DEBUG("can not change the cache level of pinned objects\n");
3329 if (vma && !i915_gem_valid_gtt_space(dev, &vma->node, cache_level)) {
3330 ret = i915_gem_object_unbind(obj);
3335 if (i915_gem_obj_ggtt_bound(obj)) {
3336 ret = i915_gem_object_finish_gpu(obj);
3340 i915_gem_object_finish_gtt(obj);
3342 /* Before SandyBridge, you could not use tiling or fence
3343 * registers with snooped memory, so relinquish any fences
3344 * currently pointing to our region in the aperture.
3346 if (INTEL_INFO(dev)->gen < 6) {
3347 ret = i915_gem_object_put_fence(obj);
3352 if (obj->has_global_gtt_mapping)
3353 i915_gem_gtt_bind_object(obj, cache_level);
3354 if (obj->has_aliasing_ppgtt_mapping)
3355 i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt,
3358 i915_gem_obj_ggtt_set_color(obj, cache_level);
3361 if (cache_level == I915_CACHE_NONE) {
3362 u32 old_read_domains, old_write_domain;
3364 /* If we're coming from LLC cached, then we haven't
3365 * actually been tracking whether the data is in the
3366 * CPU cache or not, since we only allow one bit set
3367 * in obj->write_domain and have been skipping the clflushes.
3368 * Just set it to the CPU cache for now.
3370 WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU);
3371 WARN_ON(obj->base.read_domains & ~I915_GEM_DOMAIN_CPU);
3373 old_read_domains = obj->base.read_domains;
3374 old_write_domain = obj->base.write_domain;
3376 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3377 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3379 trace_i915_gem_object_change_domain(obj,
3384 obj->cache_level = cache_level;
3385 i915_gem_verify_gtt(dev);
3389 int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data,
3390 struct drm_file *file)
3392 struct drm_i915_gem_caching *args = data;
3393 struct drm_i915_gem_object *obj;
3396 ret = i915_mutex_lock_interruptible(dev);
3400 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3401 if (&obj->base == NULL) {
3406 args->caching = obj->cache_level != I915_CACHE_NONE;
3408 drm_gem_object_unreference(&obj->base);
3410 mutex_unlock(&dev->struct_mutex);
3414 int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
3415 struct drm_file *file)
3417 struct drm_i915_gem_caching *args = data;
3418 struct drm_i915_gem_object *obj;
3419 enum i915_cache_level level;
3422 switch (args->caching) {
3423 case I915_CACHING_NONE:
3424 level = I915_CACHE_NONE;
3426 case I915_CACHING_CACHED:
3427 level = I915_CACHE_LLC;
3433 ret = i915_mutex_lock_interruptible(dev);
3437 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3438 if (&obj->base == NULL) {
3443 ret = i915_gem_object_set_cache_level(obj, level);
3445 drm_gem_object_unreference(&obj->base);
3447 mutex_unlock(&dev->struct_mutex);
3452 * Prepare buffer for display plane (scanout, cursors, etc).
3453 * Can be called from an uninterruptible phase (modesetting) and allows
3454 * any flushes to be pipelined (for pageflips).
3457 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
3459 struct intel_ring_buffer *pipelined)
3461 u32 old_read_domains, old_write_domain;
3464 if (pipelined != obj->ring) {
3465 ret = i915_gem_object_sync(obj, pipelined);
3470 /* The display engine is not coherent with the LLC cache on gen6. As
3471 * a result, we make sure that the pinning that is about to occur is
3472 * done with uncached PTEs. This is lowest common denominator for all
3475 * However for gen6+, we could do better by using the GFDT bit instead
3476 * of uncaching, which would allow us to flush all the LLC-cached data
3477 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3479 ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE);
3483 /* As the user may map the buffer once pinned in the display plane
3484 * (e.g. libkms for the bootup splash), we have to ensure that we
3485 * always use map_and_fenceable for all scanout buffers.
3487 ret = i915_gem_object_pin(obj, alignment, true, false);
3491 i915_gem_object_flush_cpu_write_domain(obj);
3493 old_write_domain = obj->base.write_domain;
3494 old_read_domains = obj->base.read_domains;
3496 /* It should now be out of any other write domains, and we can update
3497 * the domain values for our changes.
3499 obj->base.write_domain = 0;
3500 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3502 trace_i915_gem_object_change_domain(obj,
3510 i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
3514 if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
3517 ret = i915_gem_object_wait_rendering(obj, false);
3521 /* Ensure that we invalidate the GPU's caches and TLBs. */
3522 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
3527 * Moves a single object to the CPU read, and possibly write domain.
3529 * This function returns when the move is complete, including waiting on
3533 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
3535 uint32_t old_write_domain, old_read_domains;
3538 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
3541 ret = i915_gem_object_wait_rendering(obj, !write);
3545 i915_gem_object_flush_gtt_write_domain(obj);
3547 old_write_domain = obj->base.write_domain;
3548 old_read_domains = obj->base.read_domains;
3550 /* Flush the CPU cache if it's still invalid. */
3551 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
3552 i915_gem_clflush_object(obj);
3554 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3557 /* It should now be out of any other write domains, and we can update
3558 * the domain values for our changes.
3560 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3562 /* If we're writing through the CPU, then the GPU read domains will
3563 * need to be invalidated at next use.
3566 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3567 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3570 trace_i915_gem_object_change_domain(obj,
3577 /* Throttle our rendering by waiting until the ring has completed our requests
3578 * emitted over 20 msec ago.
3580 * Note that if we were to use the current jiffies each time around the loop,
3581 * we wouldn't escape the function with any frames outstanding if the time to
3582 * render a frame was over 20ms.
3584 * This should get us reasonable parallelism between CPU and GPU but also
3585 * relatively low latency when blocking on a particular request to finish.
3588 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
3590 struct drm_i915_private *dev_priv = dev->dev_private;
3591 struct drm_i915_file_private *file_priv = file->driver_priv;
3592 unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3593 struct drm_i915_gem_request *request;
3594 struct intel_ring_buffer *ring = NULL;
3595 unsigned reset_counter;
3599 ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
3603 ret = i915_gem_check_wedge(&dev_priv->gpu_error, false);
3607 spin_lock(&file_priv->mm.lock);
3608 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
3609 if (time_after_eq(request->emitted_jiffies, recent_enough))
3612 ring = request->ring;
3613 seqno = request->seqno;
3615 reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
3616 spin_unlock(&file_priv->mm.lock);
3621 ret = __wait_seqno(ring, seqno, reset_counter, true, NULL);
3623 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
3629 i915_gem_object_pin(struct drm_i915_gem_object *obj,
3631 bool map_and_fenceable,
3636 if (WARN_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT))
3639 if (i915_gem_obj_ggtt_bound(obj)) {
3640 if ((alignment && i915_gem_obj_ggtt_offset(obj) & (alignment - 1)) ||
3641 (map_and_fenceable && !obj->map_and_fenceable)) {
3642 WARN(obj->pin_count,
3643 "bo is already pinned with incorrect alignment:"
3644 " offset=%lx, req.alignment=%x, req.map_and_fenceable=%d,"
3645 " obj->map_and_fenceable=%d\n",
3646 i915_gem_obj_ggtt_offset(obj), alignment,
3648 obj->map_and_fenceable);
3649 ret = i915_gem_object_unbind(obj);
3655 if (!i915_gem_obj_ggtt_bound(obj)) {
3656 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3658 ret = i915_gem_object_bind_to_gtt(obj, alignment,
3664 if (!dev_priv->mm.aliasing_ppgtt)
3665 i915_gem_gtt_bind_object(obj, obj->cache_level);
3668 if (!obj->has_global_gtt_mapping && map_and_fenceable)
3669 i915_gem_gtt_bind_object(obj, obj->cache_level);
3672 obj->pin_mappable |= map_and_fenceable;
3678 i915_gem_object_unpin(struct drm_i915_gem_object *obj)
3680 BUG_ON(obj->pin_count == 0);
3681 BUG_ON(!i915_gem_obj_ggtt_bound(obj));
3683 if (--obj->pin_count == 0)
3684 obj->pin_mappable = false;
3688 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
3689 struct drm_file *file)
3691 struct drm_i915_gem_pin *args = data;
3692 struct drm_i915_gem_object *obj;
3695 ret = i915_mutex_lock_interruptible(dev);
3699 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3700 if (&obj->base == NULL) {
3705 if (obj->madv != I915_MADV_WILLNEED) {
3706 DRM_ERROR("Attempting to pin a purgeable buffer\n");
3711 if (obj->pin_filp != NULL && obj->pin_filp != file) {
3712 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
3718 if (obj->user_pin_count == 0) {
3719 ret = i915_gem_object_pin(obj, args->alignment, true, false);
3724 obj->user_pin_count++;
3725 obj->pin_filp = file;
3727 /* XXX - flush the CPU caches for pinned objects
3728 * as the X server doesn't manage domains yet
3730 i915_gem_object_flush_cpu_write_domain(obj);
3731 args->offset = i915_gem_obj_ggtt_offset(obj);
3733 drm_gem_object_unreference(&obj->base);
3735 mutex_unlock(&dev->struct_mutex);
3740 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
3741 struct drm_file *file)
3743 struct drm_i915_gem_pin *args = data;
3744 struct drm_i915_gem_object *obj;
3747 ret = i915_mutex_lock_interruptible(dev);
3751 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3752 if (&obj->base == NULL) {
3757 if (obj->pin_filp != file) {
3758 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
3763 obj->user_pin_count--;
3764 if (obj->user_pin_count == 0) {
3765 obj->pin_filp = NULL;
3766 i915_gem_object_unpin(obj);
3770 drm_gem_object_unreference(&obj->base);
3772 mutex_unlock(&dev->struct_mutex);
3777 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
3778 struct drm_file *file)
3780 struct drm_i915_gem_busy *args = data;
3781 struct drm_i915_gem_object *obj;
3784 ret = i915_mutex_lock_interruptible(dev);
3788 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3789 if (&obj->base == NULL) {
3794 /* Count all active objects as busy, even if they are currently not used
3795 * by the gpu. Users of this interface expect objects to eventually
3796 * become non-busy without any further actions, therefore emit any
3797 * necessary flushes here.
3799 ret = i915_gem_object_flush_active(obj);
3801 args->busy = obj->active;
3803 BUILD_BUG_ON(I915_NUM_RINGS > 16);
3804 args->busy |= intel_ring_flag(obj->ring) << 16;
3807 drm_gem_object_unreference(&obj->base);
3809 mutex_unlock(&dev->struct_mutex);
3814 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
3815 struct drm_file *file_priv)
3817 return i915_gem_ring_throttle(dev, file_priv);
3821 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
3822 struct drm_file *file_priv)
3824 struct drm_i915_gem_madvise *args = data;
3825 struct drm_i915_gem_object *obj;
3828 switch (args->madv) {
3829 case I915_MADV_DONTNEED:
3830 case I915_MADV_WILLNEED:
3836 ret = i915_mutex_lock_interruptible(dev);
3840 obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
3841 if (&obj->base == NULL) {
3846 if (obj->pin_count) {
3851 if (obj->madv != __I915_MADV_PURGED)
3852 obj->madv = args->madv;
3854 /* if the object is no longer attached, discard its backing storage */
3855 if (i915_gem_object_is_purgeable(obj) && obj->pages == NULL)
3856 i915_gem_object_truncate(obj);
3858 args->retained = obj->madv != __I915_MADV_PURGED;
3861 drm_gem_object_unreference(&obj->base);
3863 mutex_unlock(&dev->struct_mutex);
3867 void i915_gem_object_init(struct drm_i915_gem_object *obj,
3868 const struct drm_i915_gem_object_ops *ops)
3870 INIT_LIST_HEAD(&obj->mm_list);
3871 INIT_LIST_HEAD(&obj->global_list);
3872 INIT_LIST_HEAD(&obj->ring_list);
3873 INIT_LIST_HEAD(&obj->exec_list);
3874 INIT_LIST_HEAD(&obj->vma_list);
3878 obj->fence_reg = I915_FENCE_REG_NONE;
3879 obj->madv = I915_MADV_WILLNEED;
3880 /* Avoid an unnecessary call to unbind on the first bind. */
3881 obj->map_and_fenceable = true;
3883 i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size);
3886 static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
3887 .get_pages = i915_gem_object_get_pages_gtt,
3888 .put_pages = i915_gem_object_put_pages_gtt,
3891 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
3894 struct drm_i915_gem_object *obj;
3895 struct address_space *mapping;
3898 obj = i915_gem_object_alloc(dev);
3902 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
3903 i915_gem_object_free(obj);
3907 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
3908 if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) {
3909 /* 965gm cannot relocate objects above 4GiB. */
3910 mask &= ~__GFP_HIGHMEM;
3911 mask |= __GFP_DMA32;
3914 mapping = file_inode(obj->base.filp)->i_mapping;
3915 mapping_set_gfp_mask(mapping, mask);
3917 i915_gem_object_init(obj, &i915_gem_object_ops);
3919 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3920 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3923 /* On some devices, we can have the GPU use the LLC (the CPU
3924 * cache) for about a 10% performance improvement
3925 * compared to uncached. Graphics requests other than
3926 * display scanout are coherent with the CPU in
3927 * accessing this cache. This means in this mode we
3928 * don't need to clflush on the CPU side, and on the
3929 * GPU side we only need to flush internal caches to
3930 * get data visible to the CPU.
3932 * However, we maintain the display planes as UC, and so
3933 * need to rebind when first used as such.
3935 obj->cache_level = I915_CACHE_LLC;
3937 obj->cache_level = I915_CACHE_NONE;
3939 trace_i915_gem_object_create(obj);
3944 int i915_gem_init_object(struct drm_gem_object *obj)
3951 void i915_gem_free_object(struct drm_gem_object *gem_obj)
3953 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
3954 struct drm_device *dev = obj->base.dev;
3955 drm_i915_private_t *dev_priv = dev->dev_private;
3957 trace_i915_gem_object_destroy(obj);
3960 i915_gem_detach_phys_object(dev, obj);
3963 if (WARN_ON(i915_gem_object_unbind(obj) == -ERESTARTSYS)) {
3964 bool was_interruptible;
3966 was_interruptible = dev_priv->mm.interruptible;
3967 dev_priv->mm.interruptible = false;
3969 WARN_ON(i915_gem_object_unbind(obj));
3971 dev_priv->mm.interruptible = was_interruptible;
3974 /* Stolen objects don't hold a ref, but do hold pin count. Fix that up
3975 * before progressing. */
3977 i915_gem_object_unpin_pages(obj);
3979 if (WARN_ON(obj->pages_pin_count))
3980 obj->pages_pin_count = 0;
3981 i915_gem_object_put_pages(obj);
3982 i915_gem_object_free_mmap_offset(obj);
3983 i915_gem_object_release_stolen(obj);
3987 if (obj->base.import_attach)
3988 drm_prime_gem_destroy(&obj->base, NULL);
3990 drm_gem_object_release(&obj->base);
3991 i915_gem_info_remove_obj(dev_priv, obj->base.size);
3994 i915_gem_object_free(obj);
3997 struct i915_vma *i915_gem_vma_create(struct drm_i915_gem_object *obj,
3998 struct i915_address_space *vm)
4000 struct i915_vma *vma = kzalloc(sizeof(*vma), GFP_KERNEL);
4002 return ERR_PTR(-ENOMEM);
4004 INIT_LIST_HEAD(&vma->vma_link);
4011 void i915_gem_vma_destroy(struct i915_vma *vma)
4013 WARN_ON(vma->node.allocated);
4018 i915_gem_idle(struct drm_device *dev)
4020 drm_i915_private_t *dev_priv = dev->dev_private;
4023 if (dev_priv->ums.mm_suspended) {
4024 mutex_unlock(&dev->struct_mutex);
4028 ret = i915_gpu_idle(dev);
4030 mutex_unlock(&dev->struct_mutex);
4033 i915_gem_retire_requests(dev);
4035 /* Under UMS, be paranoid and evict. */
4036 if (!drm_core_check_feature(dev, DRIVER_MODESET))
4037 i915_gem_evict_everything(dev);
4039 del_timer_sync(&dev_priv->gpu_error.hangcheck_timer);
4041 i915_kernel_lost_context(dev);
4042 i915_gem_cleanup_ringbuffer(dev);
4044 /* Cancel the retire work handler, which should be idle now. */
4045 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
4050 void i915_gem_l3_remap(struct drm_device *dev)
4052 drm_i915_private_t *dev_priv = dev->dev_private;
4056 if (!HAS_L3_GPU_CACHE(dev))
4059 if (!dev_priv->l3_parity.remap_info)
4062 misccpctl = I915_READ(GEN7_MISCCPCTL);
4063 I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
4064 POSTING_READ(GEN7_MISCCPCTL);
4066 for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) {
4067 u32 remap = I915_READ(GEN7_L3LOG_BASE + i);
4068 if (remap && remap != dev_priv->l3_parity.remap_info[i/4])
4069 DRM_DEBUG("0x%x was already programmed to %x\n",
4070 GEN7_L3LOG_BASE + i, remap);
4071 if (remap && !dev_priv->l3_parity.remap_info[i/4])
4072 DRM_DEBUG_DRIVER("Clearing remapped register\n");
4073 I915_WRITE(GEN7_L3LOG_BASE + i, dev_priv->l3_parity.remap_info[i/4]);
4076 /* Make sure all the writes land before disabling dop clock gating */
4077 POSTING_READ(GEN7_L3LOG_BASE);
4079 I915_WRITE(GEN7_MISCCPCTL, misccpctl);
4082 void i915_gem_init_swizzling(struct drm_device *dev)
4084 drm_i915_private_t *dev_priv = dev->dev_private;
4086 if (INTEL_INFO(dev)->gen < 5 ||
4087 dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
4090 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
4091 DISP_TILE_SURFACE_SWIZZLING);
4096 I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
4098 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
4099 else if (IS_GEN7(dev))
4100 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
4106 intel_enable_blt(struct drm_device *dev)
4111 /* The blitter was dysfunctional on early prototypes */
4112 if (IS_GEN6(dev) && dev->pdev->revision < 8) {
4113 DRM_INFO("BLT not supported on this pre-production hardware;"
4114 " graphics performance will be degraded.\n");
4121 static int i915_gem_init_rings(struct drm_device *dev)
4123 struct drm_i915_private *dev_priv = dev->dev_private;
4126 ret = intel_init_render_ring_buffer(dev);
4131 ret = intel_init_bsd_ring_buffer(dev);
4133 goto cleanup_render_ring;
4136 if (intel_enable_blt(dev)) {
4137 ret = intel_init_blt_ring_buffer(dev);
4139 goto cleanup_bsd_ring;
4142 if (HAS_VEBOX(dev)) {
4143 ret = intel_init_vebox_ring_buffer(dev);
4145 goto cleanup_blt_ring;
4149 ret = i915_gem_set_seqno(dev, ((u32)~0 - 0x1000));
4151 goto cleanup_vebox_ring;
4156 intel_cleanup_ring_buffer(&dev_priv->ring[VECS]);
4158 intel_cleanup_ring_buffer(&dev_priv->ring[BCS]);
4160 intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
4161 cleanup_render_ring:
4162 intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
4168 i915_gem_init_hw(struct drm_device *dev)
4170 drm_i915_private_t *dev_priv = dev->dev_private;
4173 if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt())
4176 if (dev_priv->ellc_size)
4177 I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf));
4179 if (HAS_PCH_NOP(dev)) {
4180 u32 temp = I915_READ(GEN7_MSG_CTL);
4181 temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK);
4182 I915_WRITE(GEN7_MSG_CTL, temp);
4185 i915_gem_l3_remap(dev);
4187 i915_gem_init_swizzling(dev);
4189 ret = i915_gem_init_rings(dev);
4194 * XXX: There was some w/a described somewhere suggesting loading
4195 * contexts before PPGTT.
4197 i915_gem_context_init(dev);
4198 if (dev_priv->mm.aliasing_ppgtt) {
4199 ret = dev_priv->mm.aliasing_ppgtt->enable(dev);
4201 i915_gem_cleanup_aliasing_ppgtt(dev);
4202 DRM_INFO("PPGTT enable failed. This is not fatal, but unexpected\n");
4209 int i915_gem_init(struct drm_device *dev)
4211 struct drm_i915_private *dev_priv = dev->dev_private;
4214 mutex_lock(&dev->struct_mutex);
4216 if (IS_VALLEYVIEW(dev)) {
4217 /* VLVA0 (potential hack), BIOS isn't actually waking us */
4218 I915_WRITE(VLV_GTLC_WAKE_CTRL, 1);
4219 if (wait_for((I915_READ(VLV_GTLC_PW_STATUS) & 1) == 1, 10))
4220 DRM_DEBUG_DRIVER("allow wake ack timed out\n");
4223 i915_gem_init_global_gtt(dev);
4225 ret = i915_gem_init_hw(dev);
4226 mutex_unlock(&dev->struct_mutex);
4228 i915_gem_cleanup_aliasing_ppgtt(dev);
4232 /* Allow hardware batchbuffers unless told otherwise, but not for KMS. */
4233 if (!drm_core_check_feature(dev, DRIVER_MODESET))
4234 dev_priv->dri1.allow_batchbuffer = 1;
4239 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
4241 drm_i915_private_t *dev_priv = dev->dev_private;
4242 struct intel_ring_buffer *ring;
4245 for_each_ring(ring, dev_priv, i)
4246 intel_cleanup_ring_buffer(ring);
4250 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
4251 struct drm_file *file_priv)
4253 struct drm_i915_private *dev_priv = dev->dev_private;
4256 if (drm_core_check_feature(dev, DRIVER_MODESET))
4259 if (i915_reset_in_progress(&dev_priv->gpu_error)) {
4260 DRM_ERROR("Reenabling wedged hardware, good luck\n");
4261 atomic_set(&dev_priv->gpu_error.reset_counter, 0);
4264 mutex_lock(&dev->struct_mutex);
4265 dev_priv->ums.mm_suspended = 0;
4267 ret = i915_gem_init_hw(dev);
4269 mutex_unlock(&dev->struct_mutex);
4273 BUG_ON(!list_empty(&dev_priv->gtt.base.active_list));
4274 mutex_unlock(&dev->struct_mutex);
4276 ret = drm_irq_install(dev);
4278 goto cleanup_ringbuffer;
4283 mutex_lock(&dev->struct_mutex);
4284 i915_gem_cleanup_ringbuffer(dev);
4285 dev_priv->ums.mm_suspended = 1;
4286 mutex_unlock(&dev->struct_mutex);
4292 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
4293 struct drm_file *file_priv)
4295 struct drm_i915_private *dev_priv = dev->dev_private;
4298 if (drm_core_check_feature(dev, DRIVER_MODESET))
4301 drm_irq_uninstall(dev);
4303 mutex_lock(&dev->struct_mutex);
4304 ret = i915_gem_idle(dev);
4306 /* Hack! Don't let anybody do execbuf while we don't control the chip.
4307 * We need to replace this with a semaphore, or something.
4308 * And not confound ums.mm_suspended!
4311 dev_priv->ums.mm_suspended = 1;
4312 mutex_unlock(&dev->struct_mutex);
4318 i915_gem_lastclose(struct drm_device *dev)
4322 if (drm_core_check_feature(dev, DRIVER_MODESET))
4325 mutex_lock(&dev->struct_mutex);
4326 ret = i915_gem_idle(dev);
4328 DRM_ERROR("failed to idle hardware: %d\n", ret);
4329 mutex_unlock(&dev->struct_mutex);
4333 init_ring_lists(struct intel_ring_buffer *ring)
4335 INIT_LIST_HEAD(&ring->active_list);
4336 INIT_LIST_HEAD(&ring->request_list);
4340 i915_gem_load(struct drm_device *dev)
4342 drm_i915_private_t *dev_priv = dev->dev_private;
4346 kmem_cache_create("i915_gem_object",
4347 sizeof(struct drm_i915_gem_object), 0,
4351 INIT_LIST_HEAD(&dev_priv->gtt.base.active_list);
4352 INIT_LIST_HEAD(&dev_priv->gtt.base.inactive_list);
4353 INIT_LIST_HEAD(&dev_priv->mm.unbound_list);
4354 INIT_LIST_HEAD(&dev_priv->mm.bound_list);
4355 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4356 for (i = 0; i < I915_NUM_RINGS; i++)
4357 init_ring_lists(&dev_priv->ring[i]);
4358 for (i = 0; i < I915_MAX_NUM_FENCES; i++)
4359 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
4360 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
4361 i915_gem_retire_work_handler);
4362 init_waitqueue_head(&dev_priv->gpu_error.reset_queue);
4364 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
4366 I915_WRITE(MI_ARB_STATE,
4367 _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
4370 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
4372 /* Old X drivers will take 0-2 for front, back, depth buffers */
4373 if (!drm_core_check_feature(dev, DRIVER_MODESET))
4374 dev_priv->fence_reg_start = 3;
4376 if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev))
4377 dev_priv->num_fence_regs = 32;
4378 else if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
4379 dev_priv->num_fence_regs = 16;
4381 dev_priv->num_fence_regs = 8;
4383 /* Initialize fence registers to zero */
4384 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4385 i915_gem_restore_fences(dev);
4387 i915_gem_detect_bit_6_swizzle(dev);
4388 init_waitqueue_head(&dev_priv->pending_flip_queue);
4390 dev_priv->mm.interruptible = true;
4392 dev_priv->mm.inactive_shrinker.shrink = i915_gem_inactive_shrink;
4393 dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS;
4394 register_shrinker(&dev_priv->mm.inactive_shrinker);
4398 * Create a physically contiguous memory object for this object
4399 * e.g. for cursor + overlay regs
4401 static int i915_gem_init_phys_object(struct drm_device *dev,
4402 int id, int size, int align)
4404 drm_i915_private_t *dev_priv = dev->dev_private;
4405 struct drm_i915_gem_phys_object *phys_obj;
4408 if (dev_priv->mm.phys_objs[id - 1] || !size)
4411 phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL);
4417 phys_obj->handle = drm_pci_alloc(dev, size, align);
4418 if (!phys_obj->handle) {
4423 set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4426 dev_priv->mm.phys_objs[id - 1] = phys_obj;
4434 static void i915_gem_free_phys_object(struct drm_device *dev, int id)
4436 drm_i915_private_t *dev_priv = dev->dev_private;
4437 struct drm_i915_gem_phys_object *phys_obj;
4439 if (!dev_priv->mm.phys_objs[id - 1])
4442 phys_obj = dev_priv->mm.phys_objs[id - 1];
4443 if (phys_obj->cur_obj) {
4444 i915_gem_detach_phys_object(dev, phys_obj->cur_obj);
4448 set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4450 drm_pci_free(dev, phys_obj->handle);
4452 dev_priv->mm.phys_objs[id - 1] = NULL;
4455 void i915_gem_free_all_phys_object(struct drm_device *dev)
4459 for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++)
4460 i915_gem_free_phys_object(dev, i);
4463 void i915_gem_detach_phys_object(struct drm_device *dev,
4464 struct drm_i915_gem_object *obj)
4466 struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
4473 vaddr = obj->phys_obj->handle->vaddr;
4475 page_count = obj->base.size / PAGE_SIZE;
4476 for (i = 0; i < page_count; i++) {
4477 struct page *page = shmem_read_mapping_page(mapping, i);
4478 if (!IS_ERR(page)) {
4479 char *dst = kmap_atomic(page);
4480 memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE);
4483 drm_clflush_pages(&page, 1);
4485 set_page_dirty(page);
4486 mark_page_accessed(page);
4487 page_cache_release(page);
4490 i915_gem_chipset_flush(dev);
4492 obj->phys_obj->cur_obj = NULL;
4493 obj->phys_obj = NULL;
4497 i915_gem_attach_phys_object(struct drm_device *dev,
4498 struct drm_i915_gem_object *obj,
4502 struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
4503 drm_i915_private_t *dev_priv = dev->dev_private;
4508 if (id > I915_MAX_PHYS_OBJECT)
4511 if (obj->phys_obj) {
4512 if (obj->phys_obj->id == id)
4514 i915_gem_detach_phys_object(dev, obj);
4517 /* create a new object */
4518 if (!dev_priv->mm.phys_objs[id - 1]) {
4519 ret = i915_gem_init_phys_object(dev, id,
4520 obj->base.size, align);
4522 DRM_ERROR("failed to init phys object %d size: %zu\n",
4523 id, obj->base.size);
4528 /* bind to the object */
4529 obj->phys_obj = dev_priv->mm.phys_objs[id - 1];
4530 obj->phys_obj->cur_obj = obj;
4532 page_count = obj->base.size / PAGE_SIZE;
4534 for (i = 0; i < page_count; i++) {
4538 page = shmem_read_mapping_page(mapping, i);
4540 return PTR_ERR(page);
4542 src = kmap_atomic(page);
4543 dst = obj->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4544 memcpy(dst, src, PAGE_SIZE);
4547 mark_page_accessed(page);
4548 page_cache_release(page);
4555 i915_gem_phys_pwrite(struct drm_device *dev,
4556 struct drm_i915_gem_object *obj,
4557 struct drm_i915_gem_pwrite *args,
4558 struct drm_file *file_priv)
4560 void *vaddr = obj->phys_obj->handle->vaddr + args->offset;
4561 char __user *user_data = to_user_ptr(args->data_ptr);
4563 if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
4564 unsigned long unwritten;
4566 /* The physical object once assigned is fixed for the lifetime
4567 * of the obj, so we can safely drop the lock and continue
4570 mutex_unlock(&dev->struct_mutex);
4571 unwritten = copy_from_user(vaddr, user_data, args->size);
4572 mutex_lock(&dev->struct_mutex);
4577 i915_gem_chipset_flush(dev);
4581 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
4583 struct drm_i915_file_private *file_priv = file->driver_priv;
4585 /* Clean up our request list when the client is going away, so that
4586 * later retire_requests won't dereference our soon-to-be-gone
4589 spin_lock(&file_priv->mm.lock);
4590 while (!list_empty(&file_priv->mm.request_list)) {
4591 struct drm_i915_gem_request *request;
4593 request = list_first_entry(&file_priv->mm.request_list,
4594 struct drm_i915_gem_request,
4596 list_del(&request->client_list);
4597 request->file_priv = NULL;
4599 spin_unlock(&file_priv->mm.lock);
4602 static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task)
4604 if (!mutex_is_locked(mutex))
4607 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES)
4608 return mutex->owner == task;
4610 /* Since UP may be pre-empted, we cannot assume that we own the lock */
4616 i915_gem_inactive_shrink(struct shrinker *shrinker, struct shrink_control *sc)
4618 struct drm_i915_private *dev_priv =
4619 container_of(shrinker,
4620 struct drm_i915_private,
4621 mm.inactive_shrinker);
4622 struct drm_device *dev = dev_priv->dev;
4623 struct i915_address_space *vm = &dev_priv->gtt.base;
4624 struct drm_i915_gem_object *obj;
4625 int nr_to_scan = sc->nr_to_scan;
4629 if (!mutex_trylock(&dev->struct_mutex)) {
4630 if (!mutex_is_locked_by(&dev->struct_mutex, current))
4633 if (dev_priv->mm.shrinker_no_lock_stealing)
4640 nr_to_scan -= i915_gem_purge(dev_priv, nr_to_scan);
4642 nr_to_scan -= __i915_gem_shrink(dev_priv, nr_to_scan,
4645 i915_gem_shrink_all(dev_priv);
4649 list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list)
4650 if (obj->pages_pin_count == 0)
4651 cnt += obj->base.size >> PAGE_SHIFT;
4652 list_for_each_entry(obj, &vm->inactive_list, mm_list)
4653 if (obj->pin_count == 0 && obj->pages_pin_count == 0)
4654 cnt += obj->base.size >> PAGE_SHIFT;
4657 mutex_unlock(&dev->struct_mutex);