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 drm_device *dev)
92 struct drm_i915_private *dev_priv = dev->dev_private;
93 struct completion *x = &dev_priv->error_completion;
97 if (!atomic_read(&dev_priv->mm.wedged))
101 * Only wait 10 seconds for the gpu reset to complete to avoid hanging
102 * userspace. If it takes that long something really bad is going on and
103 * we should simply try to bail out and fail as gracefully as possible.
105 ret = wait_for_completion_interruptible_timeout(x, 10*HZ);
107 DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
109 } else if (ret < 0) {
113 if (atomic_read(&dev_priv->mm.wedged)) {
114 /* GPU is hung, bump the completion count to account for
115 * the token we just consumed so that we never hit zero and
116 * end up waiting upon a subsequent completion event that
119 spin_lock_irqsave(&x->wait.lock, flags);
121 spin_unlock_irqrestore(&x->wait.lock, flags);
126 int i915_mutex_lock_interruptible(struct drm_device *dev)
130 ret = i915_gem_wait_for_error(dev);
134 ret = mutex_lock_interruptible(&dev->struct_mutex);
138 WARN_ON(i915_verify_lists(dev));
143 i915_gem_object_is_inactive(struct drm_i915_gem_object *obj)
145 return obj->gtt_space && !obj->active;
149 i915_gem_init_ioctl(struct drm_device *dev, void *data,
150 struct drm_file *file)
152 struct drm_i915_gem_init *args = data;
154 if (drm_core_check_feature(dev, DRIVER_MODESET))
157 if (args->gtt_start >= args->gtt_end ||
158 (args->gtt_end | args->gtt_start) & (PAGE_SIZE - 1))
161 /* GEM with user mode setting was never supported on ilk and later. */
162 if (INTEL_INFO(dev)->gen >= 5)
165 mutex_lock(&dev->struct_mutex);
166 i915_gem_init_global_gtt(dev, args->gtt_start,
167 args->gtt_end, args->gtt_end);
168 mutex_unlock(&dev->struct_mutex);
174 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
175 struct drm_file *file)
177 struct drm_i915_private *dev_priv = dev->dev_private;
178 struct drm_i915_gem_get_aperture *args = data;
179 struct drm_i915_gem_object *obj;
183 mutex_lock(&dev->struct_mutex);
184 list_for_each_entry(obj, &dev_priv->mm.bound_list, gtt_list)
186 pinned += obj->gtt_space->size;
187 mutex_unlock(&dev->struct_mutex);
189 args->aper_size = dev_priv->mm.gtt_total;
190 args->aper_available_size = args->aper_size - pinned;
196 i915_gem_create(struct drm_file *file,
197 struct drm_device *dev,
201 struct drm_i915_gem_object *obj;
205 size = roundup(size, PAGE_SIZE);
209 /* Allocate the new object */
210 obj = i915_gem_alloc_object(dev, size);
214 ret = drm_gem_handle_create(file, &obj->base, &handle);
216 drm_gem_object_release(&obj->base);
217 i915_gem_info_remove_obj(dev->dev_private, obj->base.size);
222 /* drop reference from allocate - handle holds it now */
223 drm_gem_object_unreference(&obj->base);
224 trace_i915_gem_object_create(obj);
231 i915_gem_dumb_create(struct drm_file *file,
232 struct drm_device *dev,
233 struct drm_mode_create_dumb *args)
235 /* have to work out size/pitch and return them */
236 args->pitch = ALIGN(args->width * ((args->bpp + 7) / 8), 64);
237 args->size = args->pitch * args->height;
238 return i915_gem_create(file, dev,
239 args->size, &args->handle);
242 int i915_gem_dumb_destroy(struct drm_file *file,
243 struct drm_device *dev,
246 return drm_gem_handle_delete(file, handle);
250 * Creates a new mm object and returns a handle to it.
253 i915_gem_create_ioctl(struct drm_device *dev, void *data,
254 struct drm_file *file)
256 struct drm_i915_gem_create *args = data;
258 return i915_gem_create(file, dev,
259 args->size, &args->handle);
262 static int i915_gem_object_needs_bit17_swizzle(struct drm_i915_gem_object *obj)
264 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
266 return dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_9_10_17 &&
267 obj->tiling_mode != I915_TILING_NONE;
271 __copy_to_user_swizzled(char __user *cpu_vaddr,
272 const char *gpu_vaddr, int gpu_offset,
275 int ret, cpu_offset = 0;
278 int cacheline_end = ALIGN(gpu_offset + 1, 64);
279 int this_length = min(cacheline_end - gpu_offset, length);
280 int swizzled_gpu_offset = gpu_offset ^ 64;
282 ret = __copy_to_user(cpu_vaddr + cpu_offset,
283 gpu_vaddr + swizzled_gpu_offset,
288 cpu_offset += this_length;
289 gpu_offset += this_length;
290 length -= this_length;
297 __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset,
298 const char __user *cpu_vaddr,
301 int ret, cpu_offset = 0;
304 int cacheline_end = ALIGN(gpu_offset + 1, 64);
305 int this_length = min(cacheline_end - gpu_offset, length);
306 int swizzled_gpu_offset = gpu_offset ^ 64;
308 ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset,
309 cpu_vaddr + cpu_offset,
314 cpu_offset += this_length;
315 gpu_offset += this_length;
316 length -= this_length;
322 /* Per-page copy function for the shmem pread fastpath.
323 * Flushes invalid cachelines before reading the target if
324 * needs_clflush is set. */
326 shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length,
327 char __user *user_data,
328 bool page_do_bit17_swizzling, bool needs_clflush)
333 if (unlikely(page_do_bit17_swizzling))
336 vaddr = kmap_atomic(page);
338 drm_clflush_virt_range(vaddr + shmem_page_offset,
340 ret = __copy_to_user_inatomic(user_data,
341 vaddr + shmem_page_offset,
343 kunmap_atomic(vaddr);
345 return ret ? -EFAULT : 0;
349 shmem_clflush_swizzled_range(char *addr, unsigned long length,
352 if (unlikely(swizzled)) {
353 unsigned long start = (unsigned long) addr;
354 unsigned long end = (unsigned long) addr + length;
356 /* For swizzling simply ensure that we always flush both
357 * channels. Lame, but simple and it works. Swizzled
358 * pwrite/pread is far from a hotpath - current userspace
359 * doesn't use it at all. */
360 start = round_down(start, 128);
361 end = round_up(end, 128);
363 drm_clflush_virt_range((void *)start, end - start);
365 drm_clflush_virt_range(addr, length);
370 /* Only difference to the fast-path function is that this can handle bit17
371 * and uses non-atomic copy and kmap functions. */
373 shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length,
374 char __user *user_data,
375 bool page_do_bit17_swizzling, bool needs_clflush)
382 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
384 page_do_bit17_swizzling);
386 if (page_do_bit17_swizzling)
387 ret = __copy_to_user_swizzled(user_data,
388 vaddr, shmem_page_offset,
391 ret = __copy_to_user(user_data,
392 vaddr + shmem_page_offset,
396 return ret ? - EFAULT : 0;
400 i915_gem_shmem_pread(struct drm_device *dev,
401 struct drm_i915_gem_object *obj,
402 struct drm_i915_gem_pread *args,
403 struct drm_file *file)
405 char __user *user_data;
408 int shmem_page_offset, page_length, ret = 0;
409 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
411 int needs_clflush = 0;
412 struct scatterlist *sg;
415 user_data = (char __user *) (uintptr_t) args->data_ptr;
418 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
420 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) {
421 /* If we're not in the cpu read domain, set ourself into the gtt
422 * read domain and manually flush cachelines (if required). This
423 * optimizes for the case when the gpu will dirty the data
424 * anyway again before the next pread happens. */
425 if (obj->cache_level == I915_CACHE_NONE)
427 if (obj->gtt_space) {
428 ret = i915_gem_object_set_to_gtt_domain(obj, false);
434 ret = i915_gem_object_get_pages(obj);
438 i915_gem_object_pin_pages(obj);
440 offset = args->offset;
442 for_each_sg(obj->pages->sgl, sg, obj->pages->nents, i) {
445 if (i < offset >> PAGE_SHIFT)
451 /* Operation in this page
453 * shmem_page_offset = offset within page in shmem file
454 * page_length = bytes to copy for this page
456 shmem_page_offset = offset_in_page(offset);
457 page_length = remain;
458 if ((shmem_page_offset + page_length) > PAGE_SIZE)
459 page_length = PAGE_SIZE - shmem_page_offset;
462 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
463 (page_to_phys(page) & (1 << 17)) != 0;
465 ret = shmem_pread_fast(page, shmem_page_offset, page_length,
466 user_data, page_do_bit17_swizzling,
471 mutex_unlock(&dev->struct_mutex);
474 ret = fault_in_multipages_writeable(user_data, remain);
475 /* Userspace is tricking us, but we've already clobbered
476 * its pages with the prefault and promised to write the
477 * data up to the first fault. Hence ignore any errors
478 * and just continue. */
483 ret = shmem_pread_slow(page, shmem_page_offset, page_length,
484 user_data, page_do_bit17_swizzling,
487 mutex_lock(&dev->struct_mutex);
490 mark_page_accessed(page);
495 remain -= page_length;
496 user_data += page_length;
497 offset += page_length;
501 i915_gem_object_unpin_pages(obj);
507 * Reads data from the object referenced by handle.
509 * On error, the contents of *data are undefined.
512 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
513 struct drm_file *file)
515 struct drm_i915_gem_pread *args = data;
516 struct drm_i915_gem_object *obj;
522 if (!access_ok(VERIFY_WRITE,
523 (char __user *)(uintptr_t)args->data_ptr,
527 ret = i915_mutex_lock_interruptible(dev);
531 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
532 if (&obj->base == NULL) {
537 /* Bounds check source. */
538 if (args->offset > obj->base.size ||
539 args->size > obj->base.size - args->offset) {
544 /* prime objects have no backing filp to GEM pread/pwrite
547 if (!obj->base.filp) {
552 trace_i915_gem_object_pread(obj, args->offset, args->size);
554 ret = i915_gem_shmem_pread(dev, obj, args, file);
557 drm_gem_object_unreference(&obj->base);
559 mutex_unlock(&dev->struct_mutex);
563 /* This is the fast write path which cannot handle
564 * page faults in the source data
568 fast_user_write(struct io_mapping *mapping,
569 loff_t page_base, int page_offset,
570 char __user *user_data,
573 void __iomem *vaddr_atomic;
575 unsigned long unwritten;
577 vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
578 /* We can use the cpu mem copy function because this is X86. */
579 vaddr = (void __force*)vaddr_atomic + page_offset;
580 unwritten = __copy_from_user_inatomic_nocache(vaddr,
582 io_mapping_unmap_atomic(vaddr_atomic);
587 * This is the fast pwrite path, where we copy the data directly from the
588 * user into the GTT, uncached.
591 i915_gem_gtt_pwrite_fast(struct drm_device *dev,
592 struct drm_i915_gem_object *obj,
593 struct drm_i915_gem_pwrite *args,
594 struct drm_file *file)
596 drm_i915_private_t *dev_priv = dev->dev_private;
598 loff_t offset, page_base;
599 char __user *user_data;
600 int page_offset, page_length, ret;
602 ret = i915_gem_object_pin(obj, 0, true, true);
606 ret = i915_gem_object_set_to_gtt_domain(obj, true);
610 ret = i915_gem_object_put_fence(obj);
614 user_data = (char __user *) (uintptr_t) args->data_ptr;
617 offset = obj->gtt_offset + args->offset;
620 /* Operation in this page
622 * page_base = page offset within aperture
623 * page_offset = offset within page
624 * page_length = bytes to copy for this page
626 page_base = offset & PAGE_MASK;
627 page_offset = offset_in_page(offset);
628 page_length = remain;
629 if ((page_offset + remain) > PAGE_SIZE)
630 page_length = PAGE_SIZE - page_offset;
632 /* If we get a fault while copying data, then (presumably) our
633 * source page isn't available. Return the error and we'll
634 * retry in the slow path.
636 if (fast_user_write(dev_priv->mm.gtt_mapping, page_base,
637 page_offset, user_data, page_length)) {
642 remain -= page_length;
643 user_data += page_length;
644 offset += page_length;
648 i915_gem_object_unpin(obj);
653 /* Per-page copy function for the shmem pwrite fastpath.
654 * Flushes invalid cachelines before writing to the target if
655 * needs_clflush_before is set and flushes out any written cachelines after
656 * writing if needs_clflush is set. */
658 shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
659 char __user *user_data,
660 bool page_do_bit17_swizzling,
661 bool needs_clflush_before,
662 bool needs_clflush_after)
667 if (unlikely(page_do_bit17_swizzling))
670 vaddr = kmap_atomic(page);
671 if (needs_clflush_before)
672 drm_clflush_virt_range(vaddr + shmem_page_offset,
674 ret = __copy_from_user_inatomic_nocache(vaddr + shmem_page_offset,
677 if (needs_clflush_after)
678 drm_clflush_virt_range(vaddr + shmem_page_offset,
680 kunmap_atomic(vaddr);
682 return ret ? -EFAULT : 0;
685 /* Only difference to the fast-path function is that this can handle bit17
686 * and uses non-atomic copy and kmap functions. */
688 shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
689 char __user *user_data,
690 bool page_do_bit17_swizzling,
691 bool needs_clflush_before,
692 bool needs_clflush_after)
698 if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
699 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
701 page_do_bit17_swizzling);
702 if (page_do_bit17_swizzling)
703 ret = __copy_from_user_swizzled(vaddr, shmem_page_offset,
707 ret = __copy_from_user(vaddr + shmem_page_offset,
710 if (needs_clflush_after)
711 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
713 page_do_bit17_swizzling);
716 return ret ? -EFAULT : 0;
720 i915_gem_shmem_pwrite(struct drm_device *dev,
721 struct drm_i915_gem_object *obj,
722 struct drm_i915_gem_pwrite *args,
723 struct drm_file *file)
727 char __user *user_data;
728 int shmem_page_offset, page_length, ret = 0;
729 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
730 int hit_slowpath = 0;
731 int needs_clflush_after = 0;
732 int needs_clflush_before = 0;
734 struct scatterlist *sg;
736 user_data = (char __user *) (uintptr_t) args->data_ptr;
739 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
741 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
742 /* If we're not in the cpu write domain, set ourself into the gtt
743 * write domain and manually flush cachelines (if required). This
744 * optimizes for the case when the gpu will use the data
745 * right away and we therefore have to clflush anyway. */
746 if (obj->cache_level == I915_CACHE_NONE)
747 needs_clflush_after = 1;
748 if (obj->gtt_space) {
749 ret = i915_gem_object_set_to_gtt_domain(obj, true);
754 /* Same trick applies for invalidate partially written cachelines before
756 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)
757 && obj->cache_level == I915_CACHE_NONE)
758 needs_clflush_before = 1;
760 ret = i915_gem_object_get_pages(obj);
764 i915_gem_object_pin_pages(obj);
766 offset = args->offset;
769 for_each_sg(obj->pages->sgl, sg, obj->pages->nents, i) {
771 int partial_cacheline_write;
773 if (i < offset >> PAGE_SHIFT)
779 /* Operation in this page
781 * shmem_page_offset = offset within page in shmem file
782 * page_length = bytes to copy for this page
784 shmem_page_offset = offset_in_page(offset);
786 page_length = remain;
787 if ((shmem_page_offset + page_length) > PAGE_SIZE)
788 page_length = PAGE_SIZE - shmem_page_offset;
790 /* If we don't overwrite a cacheline completely we need to be
791 * careful to have up-to-date data by first clflushing. Don't
792 * overcomplicate things and flush the entire patch. */
793 partial_cacheline_write = needs_clflush_before &&
794 ((shmem_page_offset | page_length)
795 & (boot_cpu_data.x86_clflush_size - 1));
798 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
799 (page_to_phys(page) & (1 << 17)) != 0;
801 ret = shmem_pwrite_fast(page, shmem_page_offset, page_length,
802 user_data, page_do_bit17_swizzling,
803 partial_cacheline_write,
804 needs_clflush_after);
809 mutex_unlock(&dev->struct_mutex);
810 ret = shmem_pwrite_slow(page, shmem_page_offset, page_length,
811 user_data, page_do_bit17_swizzling,
812 partial_cacheline_write,
813 needs_clflush_after);
815 mutex_lock(&dev->struct_mutex);
818 set_page_dirty(page);
819 mark_page_accessed(page);
824 remain -= page_length;
825 user_data += page_length;
826 offset += page_length;
830 i915_gem_object_unpin_pages(obj);
834 * Fixup: Flush cpu caches in case we didn't flush the dirty
835 * cachelines in-line while writing and the object moved
836 * out of the cpu write domain while we've dropped the lock.
838 if (!needs_clflush_after &&
839 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
840 i915_gem_clflush_object(obj);
841 i915_gem_chipset_flush(dev);
845 if (needs_clflush_after)
846 i915_gem_chipset_flush(dev);
852 * Writes data to the object referenced by handle.
854 * On error, the contents of the buffer that were to be modified are undefined.
857 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
858 struct drm_file *file)
860 struct drm_i915_gem_pwrite *args = data;
861 struct drm_i915_gem_object *obj;
867 if (!access_ok(VERIFY_READ,
868 (char __user *)(uintptr_t)args->data_ptr,
872 ret = fault_in_multipages_readable((char __user *)(uintptr_t)args->data_ptr,
877 ret = i915_mutex_lock_interruptible(dev);
881 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
882 if (&obj->base == NULL) {
887 /* Bounds check destination. */
888 if (args->offset > obj->base.size ||
889 args->size > obj->base.size - args->offset) {
894 /* prime objects have no backing filp to GEM pread/pwrite
897 if (!obj->base.filp) {
902 trace_i915_gem_object_pwrite(obj, args->offset, args->size);
905 /* We can only do the GTT pwrite on untiled buffers, as otherwise
906 * it would end up going through the fenced access, and we'll get
907 * different detiling behavior between reading and writing.
908 * pread/pwrite currently are reading and writing from the CPU
909 * perspective, requiring manual detiling by the client.
912 ret = i915_gem_phys_pwrite(dev, obj, args, file);
916 if (obj->cache_level == I915_CACHE_NONE &&
917 obj->tiling_mode == I915_TILING_NONE &&
918 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
919 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
920 /* Note that the gtt paths might fail with non-page-backed user
921 * pointers (e.g. gtt mappings when moving data between
922 * textures). Fallback to the shmem path in that case. */
925 if (ret == -EFAULT || ret == -ENOSPC)
926 ret = i915_gem_shmem_pwrite(dev, obj, args, file);
929 drm_gem_object_unreference(&obj->base);
931 mutex_unlock(&dev->struct_mutex);
936 i915_gem_check_wedge(struct drm_i915_private *dev_priv,
939 if (atomic_read(&dev_priv->mm.wedged)) {
940 struct completion *x = &dev_priv->error_completion;
941 bool recovery_complete;
944 /* Give the error handler a chance to run. */
945 spin_lock_irqsave(&x->wait.lock, flags);
946 recovery_complete = x->done > 0;
947 spin_unlock_irqrestore(&x->wait.lock, flags);
949 /* Non-interruptible callers can't handle -EAGAIN, hence return
950 * -EIO unconditionally for these. */
954 /* Recovery complete, but still wedged means reset failure. */
955 if (recovery_complete)
965 * Compare seqno against outstanding lazy request. Emit a request if they are
969 i915_gem_check_olr(struct intel_ring_buffer *ring, u32 seqno)
973 BUG_ON(!mutex_is_locked(&ring->dev->struct_mutex));
976 if (seqno == ring->outstanding_lazy_request)
977 ret = i915_add_request(ring, NULL, NULL);
983 * __wait_seqno - wait until execution of seqno has finished
984 * @ring: the ring expected to report seqno
986 * @interruptible: do an interruptible wait (normally yes)
987 * @timeout: in - how long to wait (NULL forever); out - how much time remaining
989 * Returns 0 if the seqno was found within the alloted time. Else returns the
990 * errno with remaining time filled in timeout argument.
992 static int __wait_seqno(struct intel_ring_buffer *ring, u32 seqno,
993 bool interruptible, struct timespec *timeout)
995 drm_i915_private_t *dev_priv = ring->dev->dev_private;
996 struct timespec before, now, wait_time={1,0};
997 unsigned long timeout_jiffies;
999 bool wait_forever = true;
1002 if (i915_seqno_passed(ring->get_seqno(ring, true), seqno))
1005 trace_i915_gem_request_wait_begin(ring, seqno);
1007 if (timeout != NULL) {
1008 wait_time = *timeout;
1009 wait_forever = false;
1012 timeout_jiffies = timespec_to_jiffies(&wait_time);
1014 if (WARN_ON(!ring->irq_get(ring)))
1017 /* Record current time in case interrupted by signal, or wedged * */
1018 getrawmonotonic(&before);
1021 (i915_seqno_passed(ring->get_seqno(ring, false), seqno) || \
1022 atomic_read(&dev_priv->mm.wedged))
1025 end = wait_event_interruptible_timeout(ring->irq_queue,
1029 end = wait_event_timeout(ring->irq_queue, EXIT_COND,
1032 ret = i915_gem_check_wedge(dev_priv, interruptible);
1035 } while (end == 0 && wait_forever);
1037 getrawmonotonic(&now);
1039 ring->irq_put(ring);
1040 trace_i915_gem_request_wait_end(ring, seqno);
1044 struct timespec sleep_time = timespec_sub(now, before);
1045 *timeout = timespec_sub(*timeout, sleep_time);
1050 case -EAGAIN: /* Wedged */
1051 case -ERESTARTSYS: /* Signal */
1053 case 0: /* Timeout */
1055 set_normalized_timespec(timeout, 0, 0);
1057 default: /* Completed */
1058 WARN_ON(end < 0); /* We're not aware of other errors */
1064 * Waits for a sequence number to be signaled, and cleans up the
1065 * request and object lists appropriately for that event.
1068 i915_wait_seqno(struct intel_ring_buffer *ring, uint32_t seqno)
1070 struct drm_device *dev = ring->dev;
1071 struct drm_i915_private *dev_priv = dev->dev_private;
1072 bool interruptible = dev_priv->mm.interruptible;
1075 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1078 ret = i915_gem_check_wedge(dev_priv, interruptible);
1082 ret = i915_gem_check_olr(ring, seqno);
1086 return __wait_seqno(ring, seqno, interruptible, NULL);
1090 * Ensures that all rendering to the object has completed and the object is
1091 * safe to unbind from the GTT or access from the CPU.
1093 static __must_check int
1094 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
1097 struct intel_ring_buffer *ring = obj->ring;
1101 seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno;
1105 ret = i915_wait_seqno(ring, seqno);
1109 i915_gem_retire_requests_ring(ring);
1111 /* Manually manage the write flush as we may have not yet
1112 * retired the buffer.
1114 if (obj->last_write_seqno &&
1115 i915_seqno_passed(seqno, obj->last_write_seqno)) {
1116 obj->last_write_seqno = 0;
1117 obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
1123 /* A nonblocking variant of the above wait. This is a highly dangerous routine
1124 * as the object state may change during this call.
1126 static __must_check int
1127 i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
1130 struct drm_device *dev = obj->base.dev;
1131 struct drm_i915_private *dev_priv = dev->dev_private;
1132 struct intel_ring_buffer *ring = obj->ring;
1136 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1137 BUG_ON(!dev_priv->mm.interruptible);
1139 seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno;
1143 ret = i915_gem_check_wedge(dev_priv, true);
1147 ret = i915_gem_check_olr(ring, seqno);
1151 mutex_unlock(&dev->struct_mutex);
1152 ret = __wait_seqno(ring, seqno, true, NULL);
1153 mutex_lock(&dev->struct_mutex);
1155 i915_gem_retire_requests_ring(ring);
1157 /* Manually manage the write flush as we may have not yet
1158 * retired the buffer.
1160 if (obj->last_write_seqno &&
1161 i915_seqno_passed(seqno, obj->last_write_seqno)) {
1162 obj->last_write_seqno = 0;
1163 obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
1170 * Called when user space prepares to use an object with the CPU, either
1171 * through the mmap ioctl's mapping or a GTT mapping.
1174 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1175 struct drm_file *file)
1177 struct drm_i915_gem_set_domain *args = data;
1178 struct drm_i915_gem_object *obj;
1179 uint32_t read_domains = args->read_domains;
1180 uint32_t write_domain = args->write_domain;
1183 /* Only handle setting domains to types used by the CPU. */
1184 if (write_domain & I915_GEM_GPU_DOMAINS)
1187 if (read_domains & I915_GEM_GPU_DOMAINS)
1190 /* Having something in the write domain implies it's in the read
1191 * domain, and only that read domain. Enforce that in the request.
1193 if (write_domain != 0 && read_domains != write_domain)
1196 ret = i915_mutex_lock_interruptible(dev);
1200 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1201 if (&obj->base == NULL) {
1206 /* Try to flush the object off the GPU without holding the lock.
1207 * We will repeat the flush holding the lock in the normal manner
1208 * to catch cases where we are gazumped.
1210 ret = i915_gem_object_wait_rendering__nonblocking(obj, !write_domain);
1214 if (read_domains & I915_GEM_DOMAIN_GTT) {
1215 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1217 /* Silently promote "you're not bound, there was nothing to do"
1218 * to success, since the client was just asking us to
1219 * make sure everything was done.
1224 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1228 drm_gem_object_unreference(&obj->base);
1230 mutex_unlock(&dev->struct_mutex);
1235 * Called when user space has done writes to this buffer
1238 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1239 struct drm_file *file)
1241 struct drm_i915_gem_sw_finish *args = data;
1242 struct drm_i915_gem_object *obj;
1245 ret = i915_mutex_lock_interruptible(dev);
1249 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1250 if (&obj->base == NULL) {
1255 /* Pinned buffers may be scanout, so flush the cache */
1257 i915_gem_object_flush_cpu_write_domain(obj);
1259 drm_gem_object_unreference(&obj->base);
1261 mutex_unlock(&dev->struct_mutex);
1266 * Maps the contents of an object, returning the address it is mapped
1269 * While the mapping holds a reference on the contents of the object, it doesn't
1270 * imply a ref on the object itself.
1273 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1274 struct drm_file *file)
1276 struct drm_i915_gem_mmap *args = data;
1277 struct drm_gem_object *obj;
1280 obj = drm_gem_object_lookup(dev, file, args->handle);
1284 /* prime objects have no backing filp to GEM mmap
1288 drm_gem_object_unreference_unlocked(obj);
1292 addr = vm_mmap(obj->filp, 0, args->size,
1293 PROT_READ | PROT_WRITE, MAP_SHARED,
1295 drm_gem_object_unreference_unlocked(obj);
1296 if (IS_ERR((void *)addr))
1299 args->addr_ptr = (uint64_t) addr;
1305 * i915_gem_fault - fault a page into the GTT
1306 * vma: VMA in question
1309 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1310 * from userspace. The fault handler takes care of binding the object to
1311 * the GTT (if needed), allocating and programming a fence register (again,
1312 * only if needed based on whether the old reg is still valid or the object
1313 * is tiled) and inserting a new PTE into the faulting process.
1315 * Note that the faulting process may involve evicting existing objects
1316 * from the GTT and/or fence registers to make room. So performance may
1317 * suffer if the GTT working set is large or there are few fence registers
1320 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1322 struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1323 struct drm_device *dev = obj->base.dev;
1324 drm_i915_private_t *dev_priv = dev->dev_private;
1325 pgoff_t page_offset;
1328 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1330 /* We don't use vmf->pgoff since that has the fake offset */
1331 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1334 ret = i915_mutex_lock_interruptible(dev);
1338 trace_i915_gem_object_fault(obj, page_offset, true, write);
1340 /* Now bind it into the GTT if needed */
1341 ret = i915_gem_object_pin(obj, 0, true, false);
1345 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1349 ret = i915_gem_object_get_fence(obj);
1353 obj->fault_mappable = true;
1355 pfn = ((dev_priv->mm.gtt_base_addr + obj->gtt_offset) >> PAGE_SHIFT) +
1358 /* Finally, remap it using the new GTT offset */
1359 ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1361 i915_gem_object_unpin(obj);
1363 mutex_unlock(&dev->struct_mutex);
1367 /* If this -EIO is due to a gpu hang, give the reset code a
1368 * chance to clean up the mess. Otherwise return the proper
1370 if (!atomic_read(&dev_priv->mm.wedged))
1371 return VM_FAULT_SIGBUS;
1373 /* Give the error handler a chance to run and move the
1374 * objects off the GPU active list. Next time we service the
1375 * fault, we should be able to transition the page into the
1376 * GTT without touching the GPU (and so avoid further
1377 * EIO/EGAIN). If the GPU is wedged, then there is no issue
1378 * with coherency, just lost writes.
1386 * EBUSY is ok: this just means that another thread
1387 * already did the job.
1389 return VM_FAULT_NOPAGE;
1391 return VM_FAULT_OOM;
1393 return VM_FAULT_SIGBUS;
1395 WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret);
1396 return VM_FAULT_SIGBUS;
1401 * i915_gem_release_mmap - remove physical page mappings
1402 * @obj: obj in question
1404 * Preserve the reservation of the mmapping with the DRM core code, but
1405 * relinquish ownership of the pages back to the system.
1407 * It is vital that we remove the page mapping if we have mapped a tiled
1408 * object through the GTT and then lose the fence register due to
1409 * resource pressure. Similarly if the object has been moved out of the
1410 * aperture, than pages mapped into userspace must be revoked. Removing the
1411 * mapping will then trigger a page fault on the next user access, allowing
1412 * fixup by i915_gem_fault().
1415 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1417 if (!obj->fault_mappable)
1420 if (obj->base.dev->dev_mapping)
1421 unmap_mapping_range(obj->base.dev->dev_mapping,
1422 (loff_t)obj->base.map_list.hash.key<<PAGE_SHIFT,
1425 obj->fault_mappable = false;
1429 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1433 if (INTEL_INFO(dev)->gen >= 4 ||
1434 tiling_mode == I915_TILING_NONE)
1437 /* Previous chips need a power-of-two fence region when tiling */
1438 if (INTEL_INFO(dev)->gen == 3)
1439 gtt_size = 1024*1024;
1441 gtt_size = 512*1024;
1443 while (gtt_size < size)
1450 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1451 * @obj: object to check
1453 * Return the required GTT alignment for an object, taking into account
1454 * potential fence register mapping.
1457 i915_gem_get_gtt_alignment(struct drm_device *dev,
1462 * Minimum alignment is 4k (GTT page size), but might be greater
1463 * if a fence register is needed for the object.
1465 if (INTEL_INFO(dev)->gen >= 4 ||
1466 tiling_mode == I915_TILING_NONE)
1470 * Previous chips need to be aligned to the size of the smallest
1471 * fence register that can contain the object.
1473 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1477 * i915_gem_get_unfenced_gtt_alignment - return required GTT alignment for an
1480 * @size: size of the object
1481 * @tiling_mode: tiling mode of the object
1483 * Return the required GTT alignment for an object, only taking into account
1484 * unfenced tiled surface requirements.
1487 i915_gem_get_unfenced_gtt_alignment(struct drm_device *dev,
1492 * Minimum alignment is 4k (GTT page size) for sane hw.
1494 if (INTEL_INFO(dev)->gen >= 4 || IS_G33(dev) ||
1495 tiling_mode == I915_TILING_NONE)
1498 /* Previous hardware however needs to be aligned to a power-of-two
1499 * tile height. The simplest method for determining this is to reuse
1500 * the power-of-tile object size.
1502 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1505 static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj)
1507 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1510 if (obj->base.map_list.map)
1513 ret = drm_gem_create_mmap_offset(&obj->base);
1517 /* Badly fragmented mmap space? The only way we can recover
1518 * space is by destroying unwanted objects. We can't randomly release
1519 * mmap_offsets as userspace expects them to be persistent for the
1520 * lifetime of the objects. The closest we can is to release the
1521 * offsets on purgeable objects by truncating it and marking it purged,
1522 * which prevents userspace from ever using that object again.
1524 i915_gem_purge(dev_priv, obj->base.size >> PAGE_SHIFT);
1525 ret = drm_gem_create_mmap_offset(&obj->base);
1529 i915_gem_shrink_all(dev_priv);
1530 return drm_gem_create_mmap_offset(&obj->base);
1533 static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)
1535 if (!obj->base.map_list.map)
1538 drm_gem_free_mmap_offset(&obj->base);
1542 i915_gem_mmap_gtt(struct drm_file *file,
1543 struct drm_device *dev,
1547 struct drm_i915_private *dev_priv = dev->dev_private;
1548 struct drm_i915_gem_object *obj;
1551 ret = i915_mutex_lock_interruptible(dev);
1555 obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
1556 if (&obj->base == NULL) {
1561 if (obj->base.size > dev_priv->mm.gtt_mappable_end) {
1566 if (obj->madv != I915_MADV_WILLNEED) {
1567 DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1572 ret = i915_gem_object_create_mmap_offset(obj);
1576 *offset = (u64)obj->base.map_list.hash.key << PAGE_SHIFT;
1579 drm_gem_object_unreference(&obj->base);
1581 mutex_unlock(&dev->struct_mutex);
1586 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1588 * @data: GTT mapping ioctl data
1589 * @file: GEM object info
1591 * Simply returns the fake offset to userspace so it can mmap it.
1592 * The mmap call will end up in drm_gem_mmap(), which will set things
1593 * up so we can get faults in the handler above.
1595 * The fault handler will take care of binding the object into the GTT
1596 * (since it may have been evicted to make room for something), allocating
1597 * a fence register, and mapping the appropriate aperture address into
1601 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1602 struct drm_file *file)
1604 struct drm_i915_gem_mmap_gtt *args = data;
1606 return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
1609 /* Immediately discard the backing storage */
1611 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
1613 struct inode *inode;
1615 i915_gem_object_free_mmap_offset(obj);
1617 if (obj->base.filp == NULL)
1620 /* Our goal here is to return as much of the memory as
1621 * is possible back to the system as we are called from OOM.
1622 * To do this we must instruct the shmfs to drop all of its
1623 * backing pages, *now*.
1625 inode = obj->base.filp->f_path.dentry->d_inode;
1626 shmem_truncate_range(inode, 0, (loff_t)-1);
1628 obj->madv = __I915_MADV_PURGED;
1632 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj)
1634 return obj->madv == I915_MADV_DONTNEED;
1638 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
1640 int page_count = obj->base.size / PAGE_SIZE;
1641 struct scatterlist *sg;
1644 BUG_ON(obj->madv == __I915_MADV_PURGED);
1646 ret = i915_gem_object_set_to_cpu_domain(obj, true);
1648 /* In the event of a disaster, abandon all caches and
1649 * hope for the best.
1651 WARN_ON(ret != -EIO);
1652 i915_gem_clflush_object(obj);
1653 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
1656 if (i915_gem_object_needs_bit17_swizzle(obj))
1657 i915_gem_object_save_bit_17_swizzle(obj);
1659 if (obj->madv == I915_MADV_DONTNEED)
1662 for_each_sg(obj->pages->sgl, sg, page_count, i) {
1663 struct page *page = sg_page(sg);
1666 set_page_dirty(page);
1668 if (obj->madv == I915_MADV_WILLNEED)
1669 mark_page_accessed(page);
1671 page_cache_release(page);
1675 sg_free_table(obj->pages);
1680 i915_gem_object_put_pages(struct drm_i915_gem_object *obj)
1682 const struct drm_i915_gem_object_ops *ops = obj->ops;
1684 if (obj->pages == NULL)
1687 BUG_ON(obj->gtt_space);
1689 if (obj->pages_pin_count)
1692 ops->put_pages(obj);
1695 list_del(&obj->gtt_list);
1696 if (i915_gem_object_is_purgeable(obj))
1697 i915_gem_object_truncate(obj);
1703 i915_gem_purge(struct drm_i915_private *dev_priv, long target)
1705 struct drm_i915_gem_object *obj, *next;
1708 list_for_each_entry_safe(obj, next,
1709 &dev_priv->mm.unbound_list,
1711 if (i915_gem_object_is_purgeable(obj) &&
1712 i915_gem_object_put_pages(obj) == 0) {
1713 count += obj->base.size >> PAGE_SHIFT;
1714 if (count >= target)
1719 list_for_each_entry_safe(obj, next,
1720 &dev_priv->mm.inactive_list,
1722 if (i915_gem_object_is_purgeable(obj) &&
1723 i915_gem_object_unbind(obj) == 0 &&
1724 i915_gem_object_put_pages(obj) == 0) {
1725 count += obj->base.size >> PAGE_SHIFT;
1726 if (count >= target)
1735 i915_gem_shrink_all(struct drm_i915_private *dev_priv)
1737 struct drm_i915_gem_object *obj, *next;
1739 i915_gem_evict_everything(dev_priv->dev);
1741 list_for_each_entry_safe(obj, next, &dev_priv->mm.unbound_list, gtt_list)
1742 i915_gem_object_put_pages(obj);
1746 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
1748 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1750 struct address_space *mapping;
1751 struct sg_table *st;
1752 struct scatterlist *sg;
1756 /* Assert that the object is not currently in any GPU domain. As it
1757 * wasn't in the GTT, there shouldn't be any way it could have been in
1760 BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
1761 BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
1763 st = kmalloc(sizeof(*st), GFP_KERNEL);
1767 page_count = obj->base.size / PAGE_SIZE;
1768 if (sg_alloc_table(st, page_count, GFP_KERNEL)) {
1774 /* Get the list of pages out of our struct file. They'll be pinned
1775 * at this point until we release them.
1777 * Fail silently without starting the shrinker
1779 mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
1780 gfp = mapping_gfp_mask(mapping);
1781 gfp |= __GFP_NORETRY | __GFP_NOWARN;
1782 gfp &= ~(__GFP_IO | __GFP_WAIT);
1783 for_each_sg(st->sgl, sg, page_count, i) {
1784 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
1786 i915_gem_purge(dev_priv, page_count);
1787 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
1790 /* We've tried hard to allocate the memory by reaping
1791 * our own buffer, now let the real VM do its job and
1792 * go down in flames if truly OOM.
1794 gfp &= ~(__GFP_NORETRY | __GFP_NOWARN);
1795 gfp |= __GFP_IO | __GFP_WAIT;
1797 i915_gem_shrink_all(dev_priv);
1798 page = shmem_read_mapping_page_gfp(mapping, i, gfp);
1802 gfp |= __GFP_NORETRY | __GFP_NOWARN;
1803 gfp &= ~(__GFP_IO | __GFP_WAIT);
1806 sg_set_page(sg, page, PAGE_SIZE, 0);
1811 if (i915_gem_object_needs_bit17_swizzle(obj))
1812 i915_gem_object_do_bit_17_swizzle(obj);
1817 for_each_sg(st->sgl, sg, i, page_count)
1818 page_cache_release(sg_page(sg));
1821 return PTR_ERR(page);
1824 /* Ensure that the associated pages are gathered from the backing storage
1825 * and pinned into our object. i915_gem_object_get_pages() may be called
1826 * multiple times before they are released by a single call to
1827 * i915_gem_object_put_pages() - once the pages are no longer referenced
1828 * either as a result of memory pressure (reaping pages under the shrinker)
1829 * or as the object is itself released.
1832 i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
1834 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1835 const struct drm_i915_gem_object_ops *ops = obj->ops;
1841 BUG_ON(obj->pages_pin_count);
1843 ret = ops->get_pages(obj);
1847 list_add_tail(&obj->gtt_list, &dev_priv->mm.unbound_list);
1852 i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
1853 struct intel_ring_buffer *ring)
1855 struct drm_device *dev = obj->base.dev;
1856 struct drm_i915_private *dev_priv = dev->dev_private;
1857 u32 seqno = intel_ring_get_seqno(ring);
1859 BUG_ON(ring == NULL);
1862 /* Add a reference if we're newly entering the active list. */
1864 drm_gem_object_reference(&obj->base);
1868 /* Move from whatever list we were on to the tail of execution. */
1869 list_move_tail(&obj->mm_list, &dev_priv->mm.active_list);
1870 list_move_tail(&obj->ring_list, &ring->active_list);
1872 obj->last_read_seqno = seqno;
1874 if (obj->fenced_gpu_access) {
1875 obj->last_fenced_seqno = seqno;
1877 /* Bump MRU to take account of the delayed flush */
1878 if (obj->fence_reg != I915_FENCE_REG_NONE) {
1879 struct drm_i915_fence_reg *reg;
1881 reg = &dev_priv->fence_regs[obj->fence_reg];
1882 list_move_tail(®->lru_list,
1883 &dev_priv->mm.fence_list);
1889 i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
1891 struct drm_device *dev = obj->base.dev;
1892 struct drm_i915_private *dev_priv = dev->dev_private;
1894 BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS);
1895 BUG_ON(!obj->active);
1897 if (obj->pin_count) /* are we a framebuffer? */
1898 intel_mark_fb_idle(obj);
1900 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1902 list_del_init(&obj->ring_list);
1905 obj->last_read_seqno = 0;
1906 obj->last_write_seqno = 0;
1907 obj->base.write_domain = 0;
1909 obj->last_fenced_seqno = 0;
1910 obj->fenced_gpu_access = false;
1913 drm_gem_object_unreference(&obj->base);
1915 WARN_ON(i915_verify_lists(dev));
1919 i915_gem_handle_seqno_wrap(struct drm_device *dev)
1921 struct drm_i915_private *dev_priv = dev->dev_private;
1922 struct intel_ring_buffer *ring;
1925 /* The hardware uses various monotonic 32-bit counters, if we
1926 * detect that they will wraparound we need to idle the GPU
1927 * and reset those counters.
1930 for_each_ring(ring, dev_priv, i) {
1931 for (j = 0; j < ARRAY_SIZE(ring->sync_seqno); j++)
1932 ret |= ring->sync_seqno[j] != 0;
1937 ret = i915_gpu_idle(dev);
1941 i915_gem_retire_requests(dev);
1942 for_each_ring(ring, dev_priv, i) {
1943 for (j = 0; j < ARRAY_SIZE(ring->sync_seqno); j++)
1944 ring->sync_seqno[j] = 0;
1951 i915_gem_get_seqno(struct drm_device *dev, u32 *seqno)
1953 struct drm_i915_private *dev_priv = dev->dev_private;
1955 /* reserve 0 for non-seqno */
1956 if (dev_priv->next_seqno == 0) {
1957 int ret = i915_gem_handle_seqno_wrap(dev);
1961 dev_priv->next_seqno = 1;
1964 *seqno = dev_priv->next_seqno++;
1969 i915_add_request(struct intel_ring_buffer *ring,
1970 struct drm_file *file,
1973 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1974 struct drm_i915_gem_request *request;
1975 u32 request_ring_position;
1980 * Emit any outstanding flushes - execbuf can fail to emit the flush
1981 * after having emitted the batchbuffer command. Hence we need to fix
1982 * things up similar to emitting the lazy request. The difference here
1983 * is that the flush _must_ happen before the next request, no matter
1986 ret = intel_ring_flush_all_caches(ring);
1990 request = kmalloc(sizeof(*request), GFP_KERNEL);
1991 if (request == NULL)
1995 /* Record the position of the start of the request so that
1996 * should we detect the updated seqno part-way through the
1997 * GPU processing the request, we never over-estimate the
1998 * position of the head.
2000 request_ring_position = intel_ring_get_tail(ring);
2002 ret = ring->add_request(ring);
2008 request->seqno = intel_ring_get_seqno(ring);
2009 request->ring = ring;
2010 request->tail = request_ring_position;
2011 request->emitted_jiffies = jiffies;
2012 was_empty = list_empty(&ring->request_list);
2013 list_add_tail(&request->list, &ring->request_list);
2014 request->file_priv = NULL;
2017 struct drm_i915_file_private *file_priv = file->driver_priv;
2019 spin_lock(&file_priv->mm.lock);
2020 request->file_priv = file_priv;
2021 list_add_tail(&request->client_list,
2022 &file_priv->mm.request_list);
2023 spin_unlock(&file_priv->mm.lock);
2026 trace_i915_gem_request_add(ring, request->seqno);
2027 ring->outstanding_lazy_request = 0;
2029 if (!dev_priv->mm.suspended) {
2030 if (i915_enable_hangcheck) {
2031 mod_timer(&dev_priv->hangcheck_timer,
2032 round_jiffies_up(jiffies + DRM_I915_HANGCHECK_JIFFIES));
2035 queue_delayed_work(dev_priv->wq,
2036 &dev_priv->mm.retire_work,
2037 round_jiffies_up_relative(HZ));
2038 intel_mark_busy(dev_priv->dev);
2043 *out_seqno = request->seqno;
2048 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
2050 struct drm_i915_file_private *file_priv = request->file_priv;
2055 spin_lock(&file_priv->mm.lock);
2056 if (request->file_priv) {
2057 list_del(&request->client_list);
2058 request->file_priv = NULL;
2060 spin_unlock(&file_priv->mm.lock);
2063 static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv,
2064 struct intel_ring_buffer *ring)
2066 while (!list_empty(&ring->request_list)) {
2067 struct drm_i915_gem_request *request;
2069 request = list_first_entry(&ring->request_list,
2070 struct drm_i915_gem_request,
2073 list_del(&request->list);
2074 i915_gem_request_remove_from_client(request);
2078 while (!list_empty(&ring->active_list)) {
2079 struct drm_i915_gem_object *obj;
2081 obj = list_first_entry(&ring->active_list,
2082 struct drm_i915_gem_object,
2085 i915_gem_object_move_to_inactive(obj);
2089 static void i915_gem_reset_fences(struct drm_device *dev)
2091 struct drm_i915_private *dev_priv = dev->dev_private;
2094 for (i = 0; i < dev_priv->num_fence_regs; i++) {
2095 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
2097 i915_gem_write_fence(dev, i, NULL);
2100 i915_gem_object_fence_lost(reg->obj);
2104 INIT_LIST_HEAD(®->lru_list);
2107 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
2110 void i915_gem_reset(struct drm_device *dev)
2112 struct drm_i915_private *dev_priv = dev->dev_private;
2113 struct drm_i915_gem_object *obj;
2114 struct intel_ring_buffer *ring;
2117 for_each_ring(ring, dev_priv, i)
2118 i915_gem_reset_ring_lists(dev_priv, ring);
2120 /* Move everything out of the GPU domains to ensure we do any
2121 * necessary invalidation upon reuse.
2123 list_for_each_entry(obj,
2124 &dev_priv->mm.inactive_list,
2127 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
2130 /* The fence registers are invalidated so clear them out */
2131 i915_gem_reset_fences(dev);
2135 * This function clears the request list as sequence numbers are passed.
2138 i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
2142 if (list_empty(&ring->request_list))
2145 WARN_ON(i915_verify_lists(ring->dev));
2147 seqno = ring->get_seqno(ring, true);
2149 while (!list_empty(&ring->request_list)) {
2150 struct drm_i915_gem_request *request;
2152 request = list_first_entry(&ring->request_list,
2153 struct drm_i915_gem_request,
2156 if (!i915_seqno_passed(seqno, request->seqno))
2159 trace_i915_gem_request_retire(ring, request->seqno);
2160 /* We know the GPU must have read the request to have
2161 * sent us the seqno + interrupt, so use the position
2162 * of tail of the request to update the last known position
2165 ring->last_retired_head = request->tail;
2167 list_del(&request->list);
2168 i915_gem_request_remove_from_client(request);
2172 /* Move any buffers on the active list that are no longer referenced
2173 * by the ringbuffer to the flushing/inactive lists as appropriate.
2175 while (!list_empty(&ring->active_list)) {
2176 struct drm_i915_gem_object *obj;
2178 obj = list_first_entry(&ring->active_list,
2179 struct drm_i915_gem_object,
2182 if (!i915_seqno_passed(seqno, obj->last_read_seqno))
2185 i915_gem_object_move_to_inactive(obj);
2188 if (unlikely(ring->trace_irq_seqno &&
2189 i915_seqno_passed(seqno, ring->trace_irq_seqno))) {
2190 ring->irq_put(ring);
2191 ring->trace_irq_seqno = 0;
2194 WARN_ON(i915_verify_lists(ring->dev));
2198 i915_gem_retire_requests(struct drm_device *dev)
2200 drm_i915_private_t *dev_priv = dev->dev_private;
2201 struct intel_ring_buffer *ring;
2204 for_each_ring(ring, dev_priv, i)
2205 i915_gem_retire_requests_ring(ring);
2209 i915_gem_retire_work_handler(struct work_struct *work)
2211 drm_i915_private_t *dev_priv;
2212 struct drm_device *dev;
2213 struct intel_ring_buffer *ring;
2217 dev_priv = container_of(work, drm_i915_private_t,
2218 mm.retire_work.work);
2219 dev = dev_priv->dev;
2221 /* Come back later if the device is busy... */
2222 if (!mutex_trylock(&dev->struct_mutex)) {
2223 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
2224 round_jiffies_up_relative(HZ));
2228 i915_gem_retire_requests(dev);
2230 /* Send a periodic flush down the ring so we don't hold onto GEM
2231 * objects indefinitely.
2234 for_each_ring(ring, dev_priv, i) {
2235 if (ring->gpu_caches_dirty)
2236 i915_add_request(ring, NULL, NULL);
2238 idle &= list_empty(&ring->request_list);
2241 if (!dev_priv->mm.suspended && !idle)
2242 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
2243 round_jiffies_up_relative(HZ));
2245 intel_mark_idle(dev);
2247 mutex_unlock(&dev->struct_mutex);
2251 * Ensures that an object will eventually get non-busy by flushing any required
2252 * write domains, emitting any outstanding lazy request and retiring and
2253 * completed requests.
2256 i915_gem_object_flush_active(struct drm_i915_gem_object *obj)
2261 ret = i915_gem_check_olr(obj->ring, obj->last_read_seqno);
2265 i915_gem_retire_requests_ring(obj->ring);
2272 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
2273 * @DRM_IOCTL_ARGS: standard ioctl arguments
2275 * Returns 0 if successful, else an error is returned with the remaining time in
2276 * the timeout parameter.
2277 * -ETIME: object is still busy after timeout
2278 * -ERESTARTSYS: signal interrupted the wait
2279 * -ENONENT: object doesn't exist
2280 * Also possible, but rare:
2281 * -EAGAIN: GPU wedged
2283 * -ENODEV: Internal IRQ fail
2284 * -E?: The add request failed
2286 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
2287 * non-zero timeout parameter the wait ioctl will wait for the given number of
2288 * nanoseconds on an object becoming unbusy. Since the wait itself does so
2289 * without holding struct_mutex the object may become re-busied before this
2290 * function completes. A similar but shorter * race condition exists in the busy
2294 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
2296 struct drm_i915_gem_wait *args = data;
2297 struct drm_i915_gem_object *obj;
2298 struct intel_ring_buffer *ring = NULL;
2299 struct timespec timeout_stack, *timeout = NULL;
2303 if (args->timeout_ns >= 0) {
2304 timeout_stack = ns_to_timespec(args->timeout_ns);
2305 timeout = &timeout_stack;
2308 ret = i915_mutex_lock_interruptible(dev);
2312 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle));
2313 if (&obj->base == NULL) {
2314 mutex_unlock(&dev->struct_mutex);
2318 /* Need to make sure the object gets inactive eventually. */
2319 ret = i915_gem_object_flush_active(obj);
2324 seqno = obj->last_read_seqno;
2331 /* Do this after OLR check to make sure we make forward progress polling
2332 * on this IOCTL with a 0 timeout (like busy ioctl)
2334 if (!args->timeout_ns) {
2339 drm_gem_object_unreference(&obj->base);
2340 mutex_unlock(&dev->struct_mutex);
2342 ret = __wait_seqno(ring, seqno, true, timeout);
2344 WARN_ON(!timespec_valid(timeout));
2345 args->timeout_ns = timespec_to_ns(timeout);
2350 drm_gem_object_unreference(&obj->base);
2351 mutex_unlock(&dev->struct_mutex);
2356 * i915_gem_object_sync - sync an object to a ring.
2358 * @obj: object which may be in use on another ring.
2359 * @to: ring we wish to use the object on. May be NULL.
2361 * This code is meant to abstract object synchronization with the GPU.
2362 * Calling with NULL implies synchronizing the object with the CPU
2363 * rather than a particular GPU ring.
2365 * Returns 0 if successful, else propagates up the lower layer error.
2368 i915_gem_object_sync(struct drm_i915_gem_object *obj,
2369 struct intel_ring_buffer *to)
2371 struct intel_ring_buffer *from = obj->ring;
2375 if (from == NULL || to == from)
2378 if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev))
2379 return i915_gem_object_wait_rendering(obj, false);
2381 idx = intel_ring_sync_index(from, to);
2383 seqno = obj->last_read_seqno;
2384 if (seqno <= from->sync_seqno[idx])
2387 ret = i915_gem_check_olr(obj->ring, seqno);
2391 ret = to->sync_to(to, from, seqno);
2393 /* We use last_read_seqno because sync_to()
2394 * might have just caused seqno wrap under
2397 from->sync_seqno[idx] = obj->last_read_seqno;
2402 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
2404 u32 old_write_domain, old_read_domains;
2406 /* Act a barrier for all accesses through the GTT */
2409 /* Force a pagefault for domain tracking on next user access */
2410 i915_gem_release_mmap(obj);
2412 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
2415 old_read_domains = obj->base.read_domains;
2416 old_write_domain = obj->base.write_domain;
2418 obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
2419 obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
2421 trace_i915_gem_object_change_domain(obj,
2427 * Unbinds an object from the GTT aperture.
2430 i915_gem_object_unbind(struct drm_i915_gem_object *obj)
2432 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
2435 if (obj->gtt_space == NULL)
2441 BUG_ON(obj->pages == NULL);
2443 ret = i915_gem_object_finish_gpu(obj);
2446 /* Continue on if we fail due to EIO, the GPU is hung so we
2447 * should be safe and we need to cleanup or else we might
2448 * cause memory corruption through use-after-free.
2451 i915_gem_object_finish_gtt(obj);
2453 /* release the fence reg _after_ flushing */
2454 ret = i915_gem_object_put_fence(obj);
2458 trace_i915_gem_object_unbind(obj);
2460 if (obj->has_global_gtt_mapping)
2461 i915_gem_gtt_unbind_object(obj);
2462 if (obj->has_aliasing_ppgtt_mapping) {
2463 i915_ppgtt_unbind_object(dev_priv->mm.aliasing_ppgtt, obj);
2464 obj->has_aliasing_ppgtt_mapping = 0;
2466 i915_gem_gtt_finish_object(obj);
2468 list_del(&obj->mm_list);
2469 list_move_tail(&obj->gtt_list, &dev_priv->mm.unbound_list);
2470 /* Avoid an unnecessary call to unbind on rebind. */
2471 obj->map_and_fenceable = true;
2473 drm_mm_put_block(obj->gtt_space);
2474 obj->gtt_space = NULL;
2475 obj->gtt_offset = 0;
2480 int i915_gpu_idle(struct drm_device *dev)
2482 drm_i915_private_t *dev_priv = dev->dev_private;
2483 struct intel_ring_buffer *ring;
2486 /* Flush everything onto the inactive list. */
2487 for_each_ring(ring, dev_priv, i) {
2488 ret = i915_switch_context(ring, NULL, DEFAULT_CONTEXT_ID);
2492 ret = intel_ring_idle(ring);
2500 static void sandybridge_write_fence_reg(struct drm_device *dev, int reg,
2501 struct drm_i915_gem_object *obj)
2503 drm_i915_private_t *dev_priv = dev->dev_private;
2507 u32 size = obj->gtt_space->size;
2509 val = (uint64_t)((obj->gtt_offset + size - 4096) &
2511 val |= obj->gtt_offset & 0xfffff000;
2512 val |= (uint64_t)((obj->stride / 128) - 1) <<
2513 SANDYBRIDGE_FENCE_PITCH_SHIFT;
2515 if (obj->tiling_mode == I915_TILING_Y)
2516 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2517 val |= I965_FENCE_REG_VALID;
2521 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + reg * 8, val);
2522 POSTING_READ(FENCE_REG_SANDYBRIDGE_0 + reg * 8);
2525 static void i965_write_fence_reg(struct drm_device *dev, int reg,
2526 struct drm_i915_gem_object *obj)
2528 drm_i915_private_t *dev_priv = dev->dev_private;
2532 u32 size = obj->gtt_space->size;
2534 val = (uint64_t)((obj->gtt_offset + size - 4096) &
2536 val |= obj->gtt_offset & 0xfffff000;
2537 val |= ((obj->stride / 128) - 1) << I965_FENCE_PITCH_SHIFT;
2538 if (obj->tiling_mode == I915_TILING_Y)
2539 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2540 val |= I965_FENCE_REG_VALID;
2544 I915_WRITE64(FENCE_REG_965_0 + reg * 8, val);
2545 POSTING_READ(FENCE_REG_965_0 + reg * 8);
2548 static void i915_write_fence_reg(struct drm_device *dev, int reg,
2549 struct drm_i915_gem_object *obj)
2551 drm_i915_private_t *dev_priv = dev->dev_private;
2555 u32 size = obj->gtt_space->size;
2559 WARN((obj->gtt_offset & ~I915_FENCE_START_MASK) ||
2560 (size & -size) != size ||
2561 (obj->gtt_offset & (size - 1)),
2562 "object 0x%08x [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
2563 obj->gtt_offset, obj->map_and_fenceable, size);
2565 if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
2570 /* Note: pitch better be a power of two tile widths */
2571 pitch_val = obj->stride / tile_width;
2572 pitch_val = ffs(pitch_val) - 1;
2574 val = obj->gtt_offset;
2575 if (obj->tiling_mode == I915_TILING_Y)
2576 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2577 val |= I915_FENCE_SIZE_BITS(size);
2578 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2579 val |= I830_FENCE_REG_VALID;
2584 reg = FENCE_REG_830_0 + reg * 4;
2586 reg = FENCE_REG_945_8 + (reg - 8) * 4;
2588 I915_WRITE(reg, val);
2592 static void i830_write_fence_reg(struct drm_device *dev, int reg,
2593 struct drm_i915_gem_object *obj)
2595 drm_i915_private_t *dev_priv = dev->dev_private;
2599 u32 size = obj->gtt_space->size;
2602 WARN((obj->gtt_offset & ~I830_FENCE_START_MASK) ||
2603 (size & -size) != size ||
2604 (obj->gtt_offset & (size - 1)),
2605 "object 0x%08x not 512K or pot-size 0x%08x aligned\n",
2606 obj->gtt_offset, size);
2608 pitch_val = obj->stride / 128;
2609 pitch_val = ffs(pitch_val) - 1;
2611 val = obj->gtt_offset;
2612 if (obj->tiling_mode == I915_TILING_Y)
2613 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2614 val |= I830_FENCE_SIZE_BITS(size);
2615 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2616 val |= I830_FENCE_REG_VALID;
2620 I915_WRITE(FENCE_REG_830_0 + reg * 4, val);
2621 POSTING_READ(FENCE_REG_830_0 + reg * 4);
2624 static void i915_gem_write_fence(struct drm_device *dev, int reg,
2625 struct drm_i915_gem_object *obj)
2627 switch (INTEL_INFO(dev)->gen) {
2629 case 6: sandybridge_write_fence_reg(dev, reg, obj); break;
2631 case 4: i965_write_fence_reg(dev, reg, obj); break;
2632 case 3: i915_write_fence_reg(dev, reg, obj); break;
2633 case 2: i830_write_fence_reg(dev, reg, obj); break;
2638 static inline int fence_number(struct drm_i915_private *dev_priv,
2639 struct drm_i915_fence_reg *fence)
2641 return fence - dev_priv->fence_regs;
2644 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
2645 struct drm_i915_fence_reg *fence,
2648 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2649 int reg = fence_number(dev_priv, fence);
2651 i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL);
2654 obj->fence_reg = reg;
2656 list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list);
2658 obj->fence_reg = I915_FENCE_REG_NONE;
2660 list_del_init(&fence->lru_list);
2665 i915_gem_object_flush_fence(struct drm_i915_gem_object *obj)
2667 if (obj->last_fenced_seqno) {
2668 int ret = i915_wait_seqno(obj->ring, obj->last_fenced_seqno);
2672 obj->last_fenced_seqno = 0;
2675 /* Ensure that all CPU reads are completed before installing a fence
2676 * and all writes before removing the fence.
2678 if (obj->base.read_domains & I915_GEM_DOMAIN_GTT)
2681 obj->fenced_gpu_access = false;
2686 i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
2688 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2691 ret = i915_gem_object_flush_fence(obj);
2695 if (obj->fence_reg == I915_FENCE_REG_NONE)
2698 i915_gem_object_update_fence(obj,
2699 &dev_priv->fence_regs[obj->fence_reg],
2701 i915_gem_object_fence_lost(obj);
2706 static struct drm_i915_fence_reg *
2707 i915_find_fence_reg(struct drm_device *dev)
2709 struct drm_i915_private *dev_priv = dev->dev_private;
2710 struct drm_i915_fence_reg *reg, *avail;
2713 /* First try to find a free reg */
2715 for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2716 reg = &dev_priv->fence_regs[i];
2720 if (!reg->pin_count)
2727 /* None available, try to steal one or wait for a user to finish */
2728 list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
2739 * i915_gem_object_get_fence - set up fencing for an object
2740 * @obj: object to map through a fence reg
2742 * When mapping objects through the GTT, userspace wants to be able to write
2743 * to them without having to worry about swizzling if the object is tiled.
2744 * This function walks the fence regs looking for a free one for @obj,
2745 * stealing one if it can't find any.
2747 * It then sets up the reg based on the object's properties: address, pitch
2748 * and tiling format.
2750 * For an untiled surface, this removes any existing fence.
2753 i915_gem_object_get_fence(struct drm_i915_gem_object *obj)
2755 struct drm_device *dev = obj->base.dev;
2756 struct drm_i915_private *dev_priv = dev->dev_private;
2757 bool enable = obj->tiling_mode != I915_TILING_NONE;
2758 struct drm_i915_fence_reg *reg;
2761 /* Have we updated the tiling parameters upon the object and so
2762 * will need to serialise the write to the associated fence register?
2764 if (obj->fence_dirty) {
2765 ret = i915_gem_object_flush_fence(obj);
2770 /* Just update our place in the LRU if our fence is getting reused. */
2771 if (obj->fence_reg != I915_FENCE_REG_NONE) {
2772 reg = &dev_priv->fence_regs[obj->fence_reg];
2773 if (!obj->fence_dirty) {
2774 list_move_tail(®->lru_list,
2775 &dev_priv->mm.fence_list);
2778 } else if (enable) {
2779 reg = i915_find_fence_reg(dev);
2784 struct drm_i915_gem_object *old = reg->obj;
2786 ret = i915_gem_object_flush_fence(old);
2790 i915_gem_object_fence_lost(old);
2795 i915_gem_object_update_fence(obj, reg, enable);
2796 obj->fence_dirty = false;
2801 static bool i915_gem_valid_gtt_space(struct drm_device *dev,
2802 struct drm_mm_node *gtt_space,
2803 unsigned long cache_level)
2805 struct drm_mm_node *other;
2807 /* On non-LLC machines we have to be careful when putting differing
2808 * types of snoopable memory together to avoid the prefetcher
2809 * crossing memory domains and dieing.
2814 if (gtt_space == NULL)
2817 if (list_empty(>t_space->node_list))
2820 other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list);
2821 if (other->allocated && !other->hole_follows && other->color != cache_level)
2824 other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list);
2825 if (other->allocated && !gtt_space->hole_follows && other->color != cache_level)
2831 static void i915_gem_verify_gtt(struct drm_device *dev)
2834 struct drm_i915_private *dev_priv = dev->dev_private;
2835 struct drm_i915_gem_object *obj;
2838 list_for_each_entry(obj, &dev_priv->mm.gtt_list, gtt_list) {
2839 if (obj->gtt_space == NULL) {
2840 printk(KERN_ERR "object found on GTT list with no space reserved\n");
2845 if (obj->cache_level != obj->gtt_space->color) {
2846 printk(KERN_ERR "object reserved space [%08lx, %08lx] with wrong color, cache_level=%x, color=%lx\n",
2847 obj->gtt_space->start,
2848 obj->gtt_space->start + obj->gtt_space->size,
2850 obj->gtt_space->color);
2855 if (!i915_gem_valid_gtt_space(dev,
2857 obj->cache_level)) {
2858 printk(KERN_ERR "invalid GTT space found at [%08lx, %08lx] - color=%x\n",
2859 obj->gtt_space->start,
2860 obj->gtt_space->start + obj->gtt_space->size,
2872 * Finds free space in the GTT aperture and binds the object there.
2875 i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
2877 bool map_and_fenceable,
2880 struct drm_device *dev = obj->base.dev;
2881 drm_i915_private_t *dev_priv = dev->dev_private;
2882 struct drm_mm_node *free_space;
2883 u32 size, fence_size, fence_alignment, unfenced_alignment;
2884 bool mappable, fenceable;
2887 if (obj->madv != I915_MADV_WILLNEED) {
2888 DRM_ERROR("Attempting to bind a purgeable object\n");
2892 fence_size = i915_gem_get_gtt_size(dev,
2895 fence_alignment = i915_gem_get_gtt_alignment(dev,
2898 unfenced_alignment =
2899 i915_gem_get_unfenced_gtt_alignment(dev,
2904 alignment = map_and_fenceable ? fence_alignment :
2906 if (map_and_fenceable && alignment & (fence_alignment - 1)) {
2907 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
2911 size = map_and_fenceable ? fence_size : obj->base.size;
2913 /* If the object is bigger than the entire aperture, reject it early
2914 * before evicting everything in a vain attempt to find space.
2916 if (obj->base.size >
2917 (map_and_fenceable ? dev_priv->mm.gtt_mappable_end : dev_priv->mm.gtt_total)) {
2918 DRM_ERROR("Attempting to bind an object larger than the aperture\n");
2922 ret = i915_gem_object_get_pages(obj);
2926 i915_gem_object_pin_pages(obj);
2929 if (map_and_fenceable)
2930 free_space = drm_mm_search_free_in_range_color(&dev_priv->mm.gtt_space,
2931 size, alignment, obj->cache_level,
2932 0, dev_priv->mm.gtt_mappable_end,
2935 free_space = drm_mm_search_free_color(&dev_priv->mm.gtt_space,
2936 size, alignment, obj->cache_level,
2939 if (free_space != NULL) {
2940 if (map_and_fenceable)
2942 drm_mm_get_block_range_generic(free_space,
2943 size, alignment, obj->cache_level,
2944 0, dev_priv->mm.gtt_mappable_end,
2948 drm_mm_get_block_generic(free_space,
2949 size, alignment, obj->cache_level,
2952 if (free_space == NULL) {
2953 ret = i915_gem_evict_something(dev, size, alignment,
2958 i915_gem_object_unpin_pages(obj);
2964 if (WARN_ON(!i915_gem_valid_gtt_space(dev,
2966 obj->cache_level))) {
2967 i915_gem_object_unpin_pages(obj);
2968 drm_mm_put_block(free_space);
2972 ret = i915_gem_gtt_prepare_object(obj);
2974 i915_gem_object_unpin_pages(obj);
2975 drm_mm_put_block(free_space);
2979 list_move_tail(&obj->gtt_list, &dev_priv->mm.bound_list);
2980 list_add_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
2982 obj->gtt_space = free_space;
2983 obj->gtt_offset = free_space->start;
2986 free_space->size == fence_size &&
2987 (free_space->start & (fence_alignment - 1)) == 0;
2990 obj->gtt_offset + obj->base.size <= dev_priv->mm.gtt_mappable_end;
2992 obj->map_and_fenceable = mappable && fenceable;
2994 i915_gem_object_unpin_pages(obj);
2995 trace_i915_gem_object_bind(obj, map_and_fenceable);
2996 i915_gem_verify_gtt(dev);
3001 i915_gem_clflush_object(struct drm_i915_gem_object *obj)
3003 /* If we don't have a page list set up, then we're not pinned
3004 * to GPU, and we can ignore the cache flush because it'll happen
3005 * again at bind time.
3007 if (obj->pages == NULL)
3010 /* If the GPU is snooping the contents of the CPU cache,
3011 * we do not need to manually clear the CPU cache lines. However,
3012 * the caches are only snooped when the render cache is
3013 * flushed/invalidated. As we always have to emit invalidations
3014 * and flushes when moving into and out of the RENDER domain, correct
3015 * snooping behaviour occurs naturally as the result of our domain
3018 if (obj->cache_level != I915_CACHE_NONE)
3021 trace_i915_gem_object_clflush(obj);
3023 drm_clflush_sg(obj->pages);
3026 /** Flushes the GTT write domain for the object if it's dirty. */
3028 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
3030 uint32_t old_write_domain;
3032 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
3035 /* No actual flushing is required for the GTT write domain. Writes
3036 * to it immediately go to main memory as far as we know, so there's
3037 * no chipset flush. It also doesn't land in render cache.
3039 * However, we do have to enforce the order so that all writes through
3040 * the GTT land before any writes to the device, such as updates to
3045 old_write_domain = obj->base.write_domain;
3046 obj->base.write_domain = 0;
3048 trace_i915_gem_object_change_domain(obj,
3049 obj->base.read_domains,
3053 /** Flushes the CPU write domain for the object if it's dirty. */
3055 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj)
3057 uint32_t old_write_domain;
3059 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
3062 i915_gem_clflush_object(obj);
3063 i915_gem_chipset_flush(obj->base.dev);
3064 old_write_domain = obj->base.write_domain;
3065 obj->base.write_domain = 0;
3067 trace_i915_gem_object_change_domain(obj,
3068 obj->base.read_domains,
3073 * Moves a single object to the GTT read, and possibly write domain.
3075 * This function returns when the move is complete, including waiting on
3079 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
3081 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
3082 uint32_t old_write_domain, old_read_domains;
3085 /* Not valid to be called on unbound objects. */
3086 if (obj->gtt_space == NULL)
3089 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
3092 ret = i915_gem_object_wait_rendering(obj, !write);
3096 i915_gem_object_flush_cpu_write_domain(obj);
3098 old_write_domain = obj->base.write_domain;
3099 old_read_domains = obj->base.read_domains;
3101 /* It should now be out of any other write domains, and we can update
3102 * the domain values for our changes.
3104 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3105 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3107 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
3108 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
3112 trace_i915_gem_object_change_domain(obj,
3116 /* And bump the LRU for this access */
3117 if (i915_gem_object_is_inactive(obj))
3118 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
3123 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
3124 enum i915_cache_level cache_level)
3126 struct drm_device *dev = obj->base.dev;
3127 drm_i915_private_t *dev_priv = dev->dev_private;
3130 if (obj->cache_level == cache_level)
3133 if (obj->pin_count) {
3134 DRM_DEBUG("can not change the cache level of pinned objects\n");
3138 if (!i915_gem_valid_gtt_space(dev, obj->gtt_space, cache_level)) {
3139 ret = i915_gem_object_unbind(obj);
3144 if (obj->gtt_space) {
3145 ret = i915_gem_object_finish_gpu(obj);
3149 i915_gem_object_finish_gtt(obj);
3151 /* Before SandyBridge, you could not use tiling or fence
3152 * registers with snooped memory, so relinquish any fences
3153 * currently pointing to our region in the aperture.
3155 if (INTEL_INFO(dev)->gen < 6) {
3156 ret = i915_gem_object_put_fence(obj);
3161 if (obj->has_global_gtt_mapping)
3162 i915_gem_gtt_bind_object(obj, cache_level);
3163 if (obj->has_aliasing_ppgtt_mapping)
3164 i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt,
3167 obj->gtt_space->color = cache_level;
3170 if (cache_level == I915_CACHE_NONE) {
3171 u32 old_read_domains, old_write_domain;
3173 /* If we're coming from LLC cached, then we haven't
3174 * actually been tracking whether the data is in the
3175 * CPU cache or not, since we only allow one bit set
3176 * in obj->write_domain and have been skipping the clflushes.
3177 * Just set it to the CPU cache for now.
3179 WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU);
3180 WARN_ON(obj->base.read_domains & ~I915_GEM_DOMAIN_CPU);
3182 old_read_domains = obj->base.read_domains;
3183 old_write_domain = obj->base.write_domain;
3185 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3186 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3188 trace_i915_gem_object_change_domain(obj,
3193 obj->cache_level = cache_level;
3194 i915_gem_verify_gtt(dev);
3198 int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data,
3199 struct drm_file *file)
3201 struct drm_i915_gem_caching *args = data;
3202 struct drm_i915_gem_object *obj;
3205 ret = i915_mutex_lock_interruptible(dev);
3209 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3210 if (&obj->base == NULL) {
3215 args->caching = obj->cache_level != I915_CACHE_NONE;
3217 drm_gem_object_unreference(&obj->base);
3219 mutex_unlock(&dev->struct_mutex);
3223 int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
3224 struct drm_file *file)
3226 struct drm_i915_gem_caching *args = data;
3227 struct drm_i915_gem_object *obj;
3228 enum i915_cache_level level;
3231 switch (args->caching) {
3232 case I915_CACHING_NONE:
3233 level = I915_CACHE_NONE;
3235 case I915_CACHING_CACHED:
3236 level = I915_CACHE_LLC;
3242 ret = i915_mutex_lock_interruptible(dev);
3246 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3247 if (&obj->base == NULL) {
3252 ret = i915_gem_object_set_cache_level(obj, level);
3254 drm_gem_object_unreference(&obj->base);
3256 mutex_unlock(&dev->struct_mutex);
3261 * Prepare buffer for display plane (scanout, cursors, etc).
3262 * Can be called from an uninterruptible phase (modesetting) and allows
3263 * any flushes to be pipelined (for pageflips).
3266 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
3268 struct intel_ring_buffer *pipelined)
3270 u32 old_read_domains, old_write_domain;
3273 if (pipelined != obj->ring) {
3274 ret = i915_gem_object_sync(obj, pipelined);
3279 /* The display engine is not coherent with the LLC cache on gen6. As
3280 * a result, we make sure that the pinning that is about to occur is
3281 * done with uncached PTEs. This is lowest common denominator for all
3284 * However for gen6+, we could do better by using the GFDT bit instead
3285 * of uncaching, which would allow us to flush all the LLC-cached data
3286 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3288 ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE);
3292 /* As the user may map the buffer once pinned in the display plane
3293 * (e.g. libkms for the bootup splash), we have to ensure that we
3294 * always use map_and_fenceable for all scanout buffers.
3296 ret = i915_gem_object_pin(obj, alignment, true, false);
3300 i915_gem_object_flush_cpu_write_domain(obj);
3302 old_write_domain = obj->base.write_domain;
3303 old_read_domains = obj->base.read_domains;
3305 /* It should now be out of any other write domains, and we can update
3306 * the domain values for our changes.
3308 obj->base.write_domain = 0;
3309 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3311 trace_i915_gem_object_change_domain(obj,
3319 i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
3323 if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
3326 ret = i915_gem_object_wait_rendering(obj, false);
3330 /* Ensure that we invalidate the GPU's caches and TLBs. */
3331 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
3336 * Moves a single object to the CPU read, and possibly write domain.
3338 * This function returns when the move is complete, including waiting on
3342 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
3344 uint32_t old_write_domain, old_read_domains;
3347 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
3350 ret = i915_gem_object_wait_rendering(obj, !write);
3354 i915_gem_object_flush_gtt_write_domain(obj);
3356 old_write_domain = obj->base.write_domain;
3357 old_read_domains = obj->base.read_domains;
3359 /* Flush the CPU cache if it's still invalid. */
3360 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
3361 i915_gem_clflush_object(obj);
3363 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3366 /* It should now be out of any other write domains, and we can update
3367 * the domain values for our changes.
3369 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3371 /* If we're writing through the CPU, then the GPU read domains will
3372 * need to be invalidated at next use.
3375 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3376 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3379 trace_i915_gem_object_change_domain(obj,
3386 /* Throttle our rendering by waiting until the ring has completed our requests
3387 * emitted over 20 msec ago.
3389 * Note that if we were to use the current jiffies each time around the loop,
3390 * we wouldn't escape the function with any frames outstanding if the time to
3391 * render a frame was over 20ms.
3393 * This should get us reasonable parallelism between CPU and GPU but also
3394 * relatively low latency when blocking on a particular request to finish.
3397 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
3399 struct drm_i915_private *dev_priv = dev->dev_private;
3400 struct drm_i915_file_private *file_priv = file->driver_priv;
3401 unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3402 struct drm_i915_gem_request *request;
3403 struct intel_ring_buffer *ring = NULL;
3407 if (atomic_read(&dev_priv->mm.wedged))
3410 spin_lock(&file_priv->mm.lock);
3411 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
3412 if (time_after_eq(request->emitted_jiffies, recent_enough))
3415 ring = request->ring;
3416 seqno = request->seqno;
3418 spin_unlock(&file_priv->mm.lock);
3423 ret = __wait_seqno(ring, seqno, true, NULL);
3425 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
3431 i915_gem_object_pin(struct drm_i915_gem_object *obj,
3433 bool map_and_fenceable,
3438 if (WARN_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT))
3441 if (obj->gtt_space != NULL) {
3442 if ((alignment && obj->gtt_offset & (alignment - 1)) ||
3443 (map_and_fenceable && !obj->map_and_fenceable)) {
3444 WARN(obj->pin_count,
3445 "bo is already pinned with incorrect alignment:"
3446 " offset=%x, req.alignment=%x, req.map_and_fenceable=%d,"
3447 " obj->map_and_fenceable=%d\n",
3448 obj->gtt_offset, alignment,
3450 obj->map_and_fenceable);
3451 ret = i915_gem_object_unbind(obj);
3457 if (obj->gtt_space == NULL) {
3458 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3460 ret = i915_gem_object_bind_to_gtt(obj, alignment,
3466 if (!dev_priv->mm.aliasing_ppgtt)
3467 i915_gem_gtt_bind_object(obj, obj->cache_level);
3470 if (!obj->has_global_gtt_mapping && map_and_fenceable)
3471 i915_gem_gtt_bind_object(obj, obj->cache_level);
3474 obj->pin_mappable |= map_and_fenceable;
3480 i915_gem_object_unpin(struct drm_i915_gem_object *obj)
3482 BUG_ON(obj->pin_count == 0);
3483 BUG_ON(obj->gtt_space == NULL);
3485 if (--obj->pin_count == 0)
3486 obj->pin_mappable = false;
3490 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
3491 struct drm_file *file)
3493 struct drm_i915_gem_pin *args = data;
3494 struct drm_i915_gem_object *obj;
3497 ret = i915_mutex_lock_interruptible(dev);
3501 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3502 if (&obj->base == NULL) {
3507 if (obj->madv != I915_MADV_WILLNEED) {
3508 DRM_ERROR("Attempting to pin a purgeable buffer\n");
3513 if (obj->pin_filp != NULL && obj->pin_filp != file) {
3514 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
3520 obj->user_pin_count++;
3521 obj->pin_filp = file;
3522 if (obj->user_pin_count == 1) {
3523 ret = i915_gem_object_pin(obj, args->alignment, true, false);
3528 /* XXX - flush the CPU caches for pinned objects
3529 * as the X server doesn't manage domains yet
3531 i915_gem_object_flush_cpu_write_domain(obj);
3532 args->offset = obj->gtt_offset;
3534 drm_gem_object_unreference(&obj->base);
3536 mutex_unlock(&dev->struct_mutex);
3541 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
3542 struct drm_file *file)
3544 struct drm_i915_gem_pin *args = data;
3545 struct drm_i915_gem_object *obj;
3548 ret = i915_mutex_lock_interruptible(dev);
3552 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3553 if (&obj->base == NULL) {
3558 if (obj->pin_filp != file) {
3559 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
3564 obj->user_pin_count--;
3565 if (obj->user_pin_count == 0) {
3566 obj->pin_filp = NULL;
3567 i915_gem_object_unpin(obj);
3571 drm_gem_object_unreference(&obj->base);
3573 mutex_unlock(&dev->struct_mutex);
3578 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
3579 struct drm_file *file)
3581 struct drm_i915_gem_busy *args = data;
3582 struct drm_i915_gem_object *obj;
3585 ret = i915_mutex_lock_interruptible(dev);
3589 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3590 if (&obj->base == NULL) {
3595 /* Count all active objects as busy, even if they are currently not used
3596 * by the gpu. Users of this interface expect objects to eventually
3597 * become non-busy without any further actions, therefore emit any
3598 * necessary flushes here.
3600 ret = i915_gem_object_flush_active(obj);
3602 args->busy = obj->active;
3604 BUILD_BUG_ON(I915_NUM_RINGS > 16);
3605 args->busy |= intel_ring_flag(obj->ring) << 16;
3608 drm_gem_object_unreference(&obj->base);
3610 mutex_unlock(&dev->struct_mutex);
3615 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
3616 struct drm_file *file_priv)
3618 return i915_gem_ring_throttle(dev, file_priv);
3622 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
3623 struct drm_file *file_priv)
3625 struct drm_i915_gem_madvise *args = data;
3626 struct drm_i915_gem_object *obj;
3629 switch (args->madv) {
3630 case I915_MADV_DONTNEED:
3631 case I915_MADV_WILLNEED:
3637 ret = i915_mutex_lock_interruptible(dev);
3641 obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
3642 if (&obj->base == NULL) {
3647 if (obj->pin_count) {
3652 if (obj->madv != __I915_MADV_PURGED)
3653 obj->madv = args->madv;
3655 /* if the object is no longer attached, discard its backing storage */
3656 if (i915_gem_object_is_purgeable(obj) && obj->pages == NULL)
3657 i915_gem_object_truncate(obj);
3659 args->retained = obj->madv != __I915_MADV_PURGED;
3662 drm_gem_object_unreference(&obj->base);
3664 mutex_unlock(&dev->struct_mutex);
3668 void i915_gem_object_init(struct drm_i915_gem_object *obj,
3669 const struct drm_i915_gem_object_ops *ops)
3671 INIT_LIST_HEAD(&obj->mm_list);
3672 INIT_LIST_HEAD(&obj->gtt_list);
3673 INIT_LIST_HEAD(&obj->ring_list);
3674 INIT_LIST_HEAD(&obj->exec_list);
3678 obj->fence_reg = I915_FENCE_REG_NONE;
3679 obj->madv = I915_MADV_WILLNEED;
3680 /* Avoid an unnecessary call to unbind on the first bind. */
3681 obj->map_and_fenceable = true;
3683 i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size);
3686 static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
3687 .get_pages = i915_gem_object_get_pages_gtt,
3688 .put_pages = i915_gem_object_put_pages_gtt,
3691 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
3694 struct drm_i915_gem_object *obj;
3695 struct address_space *mapping;
3698 obj = kzalloc(sizeof(*obj), GFP_KERNEL);
3702 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
3707 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
3708 if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) {
3709 /* 965gm cannot relocate objects above 4GiB. */
3710 mask &= ~__GFP_HIGHMEM;
3711 mask |= __GFP_DMA32;
3714 mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
3715 mapping_set_gfp_mask(mapping, mask);
3717 i915_gem_object_init(obj, &i915_gem_object_ops);
3719 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3720 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3723 /* On some devices, we can have the GPU use the LLC (the CPU
3724 * cache) for about a 10% performance improvement
3725 * compared to uncached. Graphics requests other than
3726 * display scanout are coherent with the CPU in
3727 * accessing this cache. This means in this mode we
3728 * don't need to clflush on the CPU side, and on the
3729 * GPU side we only need to flush internal caches to
3730 * get data visible to the CPU.
3732 * However, we maintain the display planes as UC, and so
3733 * need to rebind when first used as such.
3735 obj->cache_level = I915_CACHE_LLC;
3737 obj->cache_level = I915_CACHE_NONE;
3742 int i915_gem_init_object(struct drm_gem_object *obj)
3749 void i915_gem_free_object(struct drm_gem_object *gem_obj)
3751 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
3752 struct drm_device *dev = obj->base.dev;
3753 drm_i915_private_t *dev_priv = dev->dev_private;
3755 trace_i915_gem_object_destroy(obj);
3758 i915_gem_detach_phys_object(dev, obj);
3761 if (WARN_ON(i915_gem_object_unbind(obj) == -ERESTARTSYS)) {
3762 bool was_interruptible;
3764 was_interruptible = dev_priv->mm.interruptible;
3765 dev_priv->mm.interruptible = false;
3767 WARN_ON(i915_gem_object_unbind(obj));
3769 dev_priv->mm.interruptible = was_interruptible;
3772 obj->pages_pin_count = 0;
3773 i915_gem_object_put_pages(obj);
3774 i915_gem_object_free_mmap_offset(obj);
3775 i915_gem_object_release_stolen(obj);
3779 if (obj->base.import_attach)
3780 drm_prime_gem_destroy(&obj->base, NULL);
3782 drm_gem_object_release(&obj->base);
3783 i915_gem_info_remove_obj(dev_priv, obj->base.size);
3790 i915_gem_idle(struct drm_device *dev)
3792 drm_i915_private_t *dev_priv = dev->dev_private;
3795 mutex_lock(&dev->struct_mutex);
3797 if (dev_priv->mm.suspended) {
3798 mutex_unlock(&dev->struct_mutex);
3802 ret = i915_gpu_idle(dev);
3804 mutex_unlock(&dev->struct_mutex);
3807 i915_gem_retire_requests(dev);
3809 /* Under UMS, be paranoid and evict. */
3810 if (!drm_core_check_feature(dev, DRIVER_MODESET))
3811 i915_gem_evict_everything(dev);
3813 i915_gem_reset_fences(dev);
3815 /* Hack! Don't let anybody do execbuf while we don't control the chip.
3816 * We need to replace this with a semaphore, or something.
3817 * And not confound mm.suspended!
3819 dev_priv->mm.suspended = 1;
3820 del_timer_sync(&dev_priv->hangcheck_timer);
3822 i915_kernel_lost_context(dev);
3823 i915_gem_cleanup_ringbuffer(dev);
3825 mutex_unlock(&dev->struct_mutex);
3827 /* Cancel the retire work handler, which should be idle now. */
3828 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
3833 void i915_gem_l3_remap(struct drm_device *dev)
3835 drm_i915_private_t *dev_priv = dev->dev_private;
3839 if (!IS_IVYBRIDGE(dev))
3842 if (!dev_priv->l3_parity.remap_info)
3845 misccpctl = I915_READ(GEN7_MISCCPCTL);
3846 I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
3847 POSTING_READ(GEN7_MISCCPCTL);
3849 for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) {
3850 u32 remap = I915_READ(GEN7_L3LOG_BASE + i);
3851 if (remap && remap != dev_priv->l3_parity.remap_info[i/4])
3852 DRM_DEBUG("0x%x was already programmed to %x\n",
3853 GEN7_L3LOG_BASE + i, remap);
3854 if (remap && !dev_priv->l3_parity.remap_info[i/4])
3855 DRM_DEBUG_DRIVER("Clearing remapped register\n");
3856 I915_WRITE(GEN7_L3LOG_BASE + i, dev_priv->l3_parity.remap_info[i/4]);
3859 /* Make sure all the writes land before disabling dop clock gating */
3860 POSTING_READ(GEN7_L3LOG_BASE);
3862 I915_WRITE(GEN7_MISCCPCTL, misccpctl);
3865 void i915_gem_init_swizzling(struct drm_device *dev)
3867 drm_i915_private_t *dev_priv = dev->dev_private;
3869 if (INTEL_INFO(dev)->gen < 5 ||
3870 dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
3873 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
3874 DISP_TILE_SURFACE_SWIZZLING);
3879 I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
3881 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
3883 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
3887 intel_enable_blt(struct drm_device *dev)
3892 /* The blitter was dysfunctional on early prototypes */
3893 if (IS_GEN6(dev) && dev->pdev->revision < 8) {
3894 DRM_INFO("BLT not supported on this pre-production hardware;"
3895 " graphics performance will be degraded.\n");
3903 i915_gem_init_hw(struct drm_device *dev)
3905 drm_i915_private_t *dev_priv = dev->dev_private;
3908 if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt())
3911 if (IS_HASWELL(dev) && (I915_READ(0x120010) == 1))
3912 I915_WRITE(0x9008, I915_READ(0x9008) | 0xf0000);
3914 i915_gem_l3_remap(dev);
3916 i915_gem_init_swizzling(dev);
3918 ret = intel_init_render_ring_buffer(dev);
3923 ret = intel_init_bsd_ring_buffer(dev);
3925 goto cleanup_render_ring;
3928 if (intel_enable_blt(dev)) {
3929 ret = intel_init_blt_ring_buffer(dev);
3931 goto cleanup_bsd_ring;
3934 dev_priv->next_seqno = 1;
3937 * XXX: There was some w/a described somewhere suggesting loading
3938 * contexts before PPGTT.
3940 i915_gem_context_init(dev);
3941 i915_gem_init_ppgtt(dev);
3946 intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
3947 cleanup_render_ring:
3948 intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
3953 intel_enable_ppgtt(struct drm_device *dev)
3955 if (i915_enable_ppgtt >= 0)
3956 return i915_enable_ppgtt;
3958 #ifdef CONFIG_INTEL_IOMMU
3959 /* Disable ppgtt on SNB if VT-d is on. */
3960 if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped)
3967 int i915_gem_init(struct drm_device *dev)
3969 struct drm_i915_private *dev_priv = dev->dev_private;
3970 unsigned long gtt_size, mappable_size;
3973 gtt_size = dev_priv->mm.gtt->gtt_total_entries << PAGE_SHIFT;
3974 mappable_size = dev_priv->mm.gtt->gtt_mappable_entries << PAGE_SHIFT;
3976 mutex_lock(&dev->struct_mutex);
3977 if (intel_enable_ppgtt(dev) && HAS_ALIASING_PPGTT(dev)) {
3978 /* PPGTT pdes are stolen from global gtt ptes, so shrink the
3979 * aperture accordingly when using aliasing ppgtt. */
3980 gtt_size -= I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
3982 i915_gem_init_global_gtt(dev, 0, mappable_size, gtt_size);
3984 ret = i915_gem_init_aliasing_ppgtt(dev);
3986 mutex_unlock(&dev->struct_mutex);
3990 /* Let GEM Manage all of the aperture.
3992 * However, leave one page at the end still bound to the scratch
3993 * page. There are a number of places where the hardware
3994 * apparently prefetches past the end of the object, and we've
3995 * seen multiple hangs with the GPU head pointer stuck in a
3996 * batchbuffer bound at the last page of the aperture. One page
3997 * should be enough to keep any prefetching inside of the
4000 i915_gem_init_global_gtt(dev, 0, mappable_size,
4004 ret = i915_gem_init_hw(dev);
4005 mutex_unlock(&dev->struct_mutex);
4007 i915_gem_cleanup_aliasing_ppgtt(dev);
4011 /* Allow hardware batchbuffers unless told otherwise, but not for KMS. */
4012 if (!drm_core_check_feature(dev, DRIVER_MODESET))
4013 dev_priv->dri1.allow_batchbuffer = 1;
4018 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
4020 drm_i915_private_t *dev_priv = dev->dev_private;
4021 struct intel_ring_buffer *ring;
4024 for_each_ring(ring, dev_priv, i)
4025 intel_cleanup_ring_buffer(ring);
4029 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
4030 struct drm_file *file_priv)
4032 drm_i915_private_t *dev_priv = dev->dev_private;
4035 if (drm_core_check_feature(dev, DRIVER_MODESET))
4038 if (atomic_read(&dev_priv->mm.wedged)) {
4039 DRM_ERROR("Reenabling wedged hardware, good luck\n");
4040 atomic_set(&dev_priv->mm.wedged, 0);
4043 mutex_lock(&dev->struct_mutex);
4044 dev_priv->mm.suspended = 0;
4046 ret = i915_gem_init_hw(dev);
4048 mutex_unlock(&dev->struct_mutex);
4052 BUG_ON(!list_empty(&dev_priv->mm.active_list));
4053 mutex_unlock(&dev->struct_mutex);
4055 ret = drm_irq_install(dev);
4057 goto cleanup_ringbuffer;
4062 mutex_lock(&dev->struct_mutex);
4063 i915_gem_cleanup_ringbuffer(dev);
4064 dev_priv->mm.suspended = 1;
4065 mutex_unlock(&dev->struct_mutex);
4071 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
4072 struct drm_file *file_priv)
4074 if (drm_core_check_feature(dev, DRIVER_MODESET))
4077 drm_irq_uninstall(dev);
4078 return i915_gem_idle(dev);
4082 i915_gem_lastclose(struct drm_device *dev)
4086 if (drm_core_check_feature(dev, DRIVER_MODESET))
4089 ret = i915_gem_idle(dev);
4091 DRM_ERROR("failed to idle hardware: %d\n", ret);
4095 init_ring_lists(struct intel_ring_buffer *ring)
4097 INIT_LIST_HEAD(&ring->active_list);
4098 INIT_LIST_HEAD(&ring->request_list);
4102 i915_gem_load(struct drm_device *dev)
4105 drm_i915_private_t *dev_priv = dev->dev_private;
4107 INIT_LIST_HEAD(&dev_priv->mm.active_list);
4108 INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
4109 INIT_LIST_HEAD(&dev_priv->mm.unbound_list);
4110 INIT_LIST_HEAD(&dev_priv->mm.bound_list);
4111 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4112 for (i = 0; i < I915_NUM_RINGS; i++)
4113 init_ring_lists(&dev_priv->ring[i]);
4114 for (i = 0; i < I915_MAX_NUM_FENCES; i++)
4115 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
4116 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
4117 i915_gem_retire_work_handler);
4118 init_completion(&dev_priv->error_completion);
4120 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
4122 I915_WRITE(MI_ARB_STATE,
4123 _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
4126 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
4128 /* Old X drivers will take 0-2 for front, back, depth buffers */
4129 if (!drm_core_check_feature(dev, DRIVER_MODESET))
4130 dev_priv->fence_reg_start = 3;
4132 if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
4133 dev_priv->num_fence_regs = 16;
4135 dev_priv->num_fence_regs = 8;
4137 /* Initialize fence registers to zero */
4138 i915_gem_reset_fences(dev);
4140 i915_gem_detect_bit_6_swizzle(dev);
4141 init_waitqueue_head(&dev_priv->pending_flip_queue);
4143 dev_priv->mm.interruptible = true;
4145 dev_priv->mm.inactive_shrinker.shrink = i915_gem_inactive_shrink;
4146 dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS;
4147 register_shrinker(&dev_priv->mm.inactive_shrinker);
4151 * Create a physically contiguous memory object for this object
4152 * e.g. for cursor + overlay regs
4154 static int i915_gem_init_phys_object(struct drm_device *dev,
4155 int id, int size, int align)
4157 drm_i915_private_t *dev_priv = dev->dev_private;
4158 struct drm_i915_gem_phys_object *phys_obj;
4161 if (dev_priv->mm.phys_objs[id - 1] || !size)
4164 phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL);
4170 phys_obj->handle = drm_pci_alloc(dev, size, align);
4171 if (!phys_obj->handle) {
4176 set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4179 dev_priv->mm.phys_objs[id - 1] = phys_obj;
4187 static void i915_gem_free_phys_object(struct drm_device *dev, int id)
4189 drm_i915_private_t *dev_priv = dev->dev_private;
4190 struct drm_i915_gem_phys_object *phys_obj;
4192 if (!dev_priv->mm.phys_objs[id - 1])
4195 phys_obj = dev_priv->mm.phys_objs[id - 1];
4196 if (phys_obj->cur_obj) {
4197 i915_gem_detach_phys_object(dev, phys_obj->cur_obj);
4201 set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4203 drm_pci_free(dev, phys_obj->handle);
4205 dev_priv->mm.phys_objs[id - 1] = NULL;
4208 void i915_gem_free_all_phys_object(struct drm_device *dev)
4212 for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++)
4213 i915_gem_free_phys_object(dev, i);
4216 void i915_gem_detach_phys_object(struct drm_device *dev,
4217 struct drm_i915_gem_object *obj)
4219 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
4226 vaddr = obj->phys_obj->handle->vaddr;
4228 page_count = obj->base.size / PAGE_SIZE;
4229 for (i = 0; i < page_count; i++) {
4230 struct page *page = shmem_read_mapping_page(mapping, i);
4231 if (!IS_ERR(page)) {
4232 char *dst = kmap_atomic(page);
4233 memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE);
4236 drm_clflush_pages(&page, 1);
4238 set_page_dirty(page);
4239 mark_page_accessed(page);
4240 page_cache_release(page);
4243 i915_gem_chipset_flush(dev);
4245 obj->phys_obj->cur_obj = NULL;
4246 obj->phys_obj = NULL;
4250 i915_gem_attach_phys_object(struct drm_device *dev,
4251 struct drm_i915_gem_object *obj,
4255 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
4256 drm_i915_private_t *dev_priv = dev->dev_private;
4261 if (id > I915_MAX_PHYS_OBJECT)
4264 if (obj->phys_obj) {
4265 if (obj->phys_obj->id == id)
4267 i915_gem_detach_phys_object(dev, obj);
4270 /* create a new object */
4271 if (!dev_priv->mm.phys_objs[id - 1]) {
4272 ret = i915_gem_init_phys_object(dev, id,
4273 obj->base.size, align);
4275 DRM_ERROR("failed to init phys object %d size: %zu\n",
4276 id, obj->base.size);
4281 /* bind to the object */
4282 obj->phys_obj = dev_priv->mm.phys_objs[id - 1];
4283 obj->phys_obj->cur_obj = obj;
4285 page_count = obj->base.size / PAGE_SIZE;
4287 for (i = 0; i < page_count; i++) {
4291 page = shmem_read_mapping_page(mapping, i);
4293 return PTR_ERR(page);
4295 src = kmap_atomic(page);
4296 dst = obj->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4297 memcpy(dst, src, PAGE_SIZE);
4300 mark_page_accessed(page);
4301 page_cache_release(page);
4308 i915_gem_phys_pwrite(struct drm_device *dev,
4309 struct drm_i915_gem_object *obj,
4310 struct drm_i915_gem_pwrite *args,
4311 struct drm_file *file_priv)
4313 void *vaddr = obj->phys_obj->handle->vaddr + args->offset;
4314 char __user *user_data = (char __user *) (uintptr_t) args->data_ptr;
4316 if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
4317 unsigned long unwritten;
4319 /* The physical object once assigned is fixed for the lifetime
4320 * of the obj, so we can safely drop the lock and continue
4323 mutex_unlock(&dev->struct_mutex);
4324 unwritten = copy_from_user(vaddr, user_data, args->size);
4325 mutex_lock(&dev->struct_mutex);
4330 i915_gem_chipset_flush(dev);
4334 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
4336 struct drm_i915_file_private *file_priv = file->driver_priv;
4338 /* Clean up our request list when the client is going away, so that
4339 * later retire_requests won't dereference our soon-to-be-gone
4342 spin_lock(&file_priv->mm.lock);
4343 while (!list_empty(&file_priv->mm.request_list)) {
4344 struct drm_i915_gem_request *request;
4346 request = list_first_entry(&file_priv->mm.request_list,
4347 struct drm_i915_gem_request,
4349 list_del(&request->client_list);
4350 request->file_priv = NULL;
4352 spin_unlock(&file_priv->mm.lock);
4355 static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task)
4357 if (!mutex_is_locked(mutex))
4360 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES)
4361 return mutex->owner == task;
4363 /* Since UP may be pre-empted, we cannot assume that we own the lock */
4369 i915_gem_inactive_shrink(struct shrinker *shrinker, struct shrink_control *sc)
4371 struct drm_i915_private *dev_priv =
4372 container_of(shrinker,
4373 struct drm_i915_private,
4374 mm.inactive_shrinker);
4375 struct drm_device *dev = dev_priv->dev;
4376 struct drm_i915_gem_object *obj;
4377 int nr_to_scan = sc->nr_to_scan;
4381 if (!mutex_trylock(&dev->struct_mutex)) {
4382 if (!mutex_is_locked_by(&dev->struct_mutex, current))
4389 nr_to_scan -= i915_gem_purge(dev_priv, nr_to_scan);
4391 i915_gem_shrink_all(dev_priv);
4395 list_for_each_entry(obj, &dev_priv->mm.unbound_list, gtt_list)
4396 if (obj->pages_pin_count == 0)
4397 cnt += obj->base.size >> PAGE_SHIFT;
4398 list_for_each_entry(obj, &dev_priv->mm.bound_list, gtt_list)
4399 if (obj->pin_count == 0 && obj->pages_pin_count == 0)
4400 cnt += obj->base.size >> PAGE_SHIFT;
4403 mutex_unlock(&dev->struct_mutex);