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>
32 #include "i915_trace.h"
33 #include "intel_drv.h"
34 #include <linux/shmem_fs.h>
35 #include <linux/slab.h>
36 #include <linux/swap.h>
37 #include <linux/pci.h>
38 #include <linux/dma-buf.h>
40 static __must_check int i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object *obj);
41 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);
42 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj);
43 static __must_check int i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
45 bool map_and_fenceable);
46 static int i915_gem_phys_pwrite(struct drm_device *dev,
47 struct drm_i915_gem_object *obj,
48 struct drm_i915_gem_pwrite *args,
49 struct drm_file *file);
51 static void i915_gem_write_fence(struct drm_device *dev, int reg,
52 struct drm_i915_gem_object *obj);
53 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
54 struct drm_i915_fence_reg *fence,
57 static int i915_gem_inactive_shrink(struct shrinker *shrinker,
58 struct shrink_control *sc);
59 static void i915_gem_object_truncate(struct drm_i915_gem_object *obj);
61 static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object *obj)
64 i915_gem_release_mmap(obj);
66 /* As we do not have an associated fence register, we will force
67 * a tiling change if we ever need to acquire one.
69 obj->fence_dirty = false;
70 obj->fence_reg = I915_FENCE_REG_NONE;
73 /* some bookkeeping */
74 static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
77 dev_priv->mm.object_count++;
78 dev_priv->mm.object_memory += size;
81 static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
84 dev_priv->mm.object_count--;
85 dev_priv->mm.object_memory -= size;
89 i915_gem_wait_for_error(struct drm_device *dev)
91 struct drm_i915_private *dev_priv = dev->dev_private;
92 struct completion *x = &dev_priv->error_completion;
96 if (!atomic_read(&dev_priv->mm.wedged))
100 * Only wait 10 seconds for the gpu reset to complete to avoid hanging
101 * userspace. If it takes that long something really bad is going on and
102 * we should simply try to bail out and fail as gracefully as possible.
104 ret = wait_for_completion_interruptible_timeout(x, 10*HZ);
106 DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
108 } else if (ret < 0) {
112 if (atomic_read(&dev_priv->mm.wedged)) {
113 /* GPU is hung, bump the completion count to account for
114 * the token we just consumed so that we never hit zero and
115 * end up waiting upon a subsequent completion event that
118 spin_lock_irqsave(&x->wait.lock, flags);
120 spin_unlock_irqrestore(&x->wait.lock, flags);
125 int i915_mutex_lock_interruptible(struct drm_device *dev)
129 ret = i915_gem_wait_for_error(dev);
133 ret = mutex_lock_interruptible(&dev->struct_mutex);
137 WARN_ON(i915_verify_lists(dev));
142 i915_gem_object_is_inactive(struct drm_i915_gem_object *obj)
148 i915_gem_init_ioctl(struct drm_device *dev, void *data,
149 struct drm_file *file)
151 struct drm_i915_gem_init *args = data;
153 if (drm_core_check_feature(dev, DRIVER_MODESET))
156 if (args->gtt_start >= args->gtt_end ||
157 (args->gtt_end | args->gtt_start) & (PAGE_SIZE - 1))
160 /* GEM with user mode setting was never supported on ilk and later. */
161 if (INTEL_INFO(dev)->gen >= 5)
164 mutex_lock(&dev->struct_mutex);
165 i915_gem_init_global_gtt(dev, args->gtt_start,
166 args->gtt_end, args->gtt_end);
167 mutex_unlock(&dev->struct_mutex);
173 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
174 struct drm_file *file)
176 struct drm_i915_private *dev_priv = dev->dev_private;
177 struct drm_i915_gem_get_aperture *args = data;
178 struct drm_i915_gem_object *obj;
182 mutex_lock(&dev->struct_mutex);
183 list_for_each_entry(obj, &dev_priv->mm.gtt_list, gtt_list)
185 pinned += obj->gtt_space->size;
186 mutex_unlock(&dev->struct_mutex);
188 args->aper_size = dev_priv->mm.gtt_total;
189 args->aper_available_size = args->aper_size - pinned;
195 i915_gem_create(struct drm_file *file,
196 struct drm_device *dev,
200 struct drm_i915_gem_object *obj;
204 size = roundup(size, PAGE_SIZE);
208 /* Allocate the new object */
209 obj = i915_gem_alloc_object(dev, size);
213 ret = drm_gem_handle_create(file, &obj->base, &handle);
215 drm_gem_object_release(&obj->base);
216 i915_gem_info_remove_obj(dev->dev_private, obj->base.size);
221 /* drop reference from allocate - handle holds it now */
222 drm_gem_object_unreference(&obj->base);
223 trace_i915_gem_object_create(obj);
230 i915_gem_dumb_create(struct drm_file *file,
231 struct drm_device *dev,
232 struct drm_mode_create_dumb *args)
234 /* have to work out size/pitch and return them */
235 args->pitch = ALIGN(args->width * ((args->bpp + 7) / 8), 64);
236 args->size = args->pitch * args->height;
237 return i915_gem_create(file, dev,
238 args->size, &args->handle);
241 int i915_gem_dumb_destroy(struct drm_file *file,
242 struct drm_device *dev,
245 return drm_gem_handle_delete(file, handle);
249 * Creates a new mm object and returns a handle to it.
252 i915_gem_create_ioctl(struct drm_device *dev, void *data,
253 struct drm_file *file)
255 struct drm_i915_gem_create *args = data;
257 return i915_gem_create(file, dev,
258 args->size, &args->handle);
261 static int i915_gem_object_needs_bit17_swizzle(struct drm_i915_gem_object *obj)
263 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
265 return dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_9_10_17 &&
266 obj->tiling_mode != I915_TILING_NONE;
270 __copy_to_user_swizzled(char __user *cpu_vaddr,
271 const char *gpu_vaddr, int gpu_offset,
274 int ret, cpu_offset = 0;
277 int cacheline_end = ALIGN(gpu_offset + 1, 64);
278 int this_length = min(cacheline_end - gpu_offset, length);
279 int swizzled_gpu_offset = gpu_offset ^ 64;
281 ret = __copy_to_user(cpu_vaddr + cpu_offset,
282 gpu_vaddr + swizzled_gpu_offset,
287 cpu_offset += this_length;
288 gpu_offset += this_length;
289 length -= this_length;
296 __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset,
297 const char __user *cpu_vaddr,
300 int ret, cpu_offset = 0;
303 int cacheline_end = ALIGN(gpu_offset + 1, 64);
304 int this_length = min(cacheline_end - gpu_offset, length);
305 int swizzled_gpu_offset = gpu_offset ^ 64;
307 ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset,
308 cpu_vaddr + cpu_offset,
313 cpu_offset += this_length;
314 gpu_offset += this_length;
315 length -= this_length;
321 /* Per-page copy function for the shmem pread fastpath.
322 * Flushes invalid cachelines before reading the target if
323 * needs_clflush is set. */
325 shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length,
326 char __user *user_data,
327 bool page_do_bit17_swizzling, bool needs_clflush)
332 if (unlikely(page_do_bit17_swizzling))
335 vaddr = kmap_atomic(page);
337 drm_clflush_virt_range(vaddr + shmem_page_offset,
339 ret = __copy_to_user_inatomic(user_data,
340 vaddr + shmem_page_offset,
342 kunmap_atomic(vaddr);
348 shmem_clflush_swizzled_range(char *addr, unsigned long length,
351 if (unlikely(swizzled)) {
352 unsigned long start = (unsigned long) addr;
353 unsigned long end = (unsigned long) addr + length;
355 /* For swizzling simply ensure that we always flush both
356 * channels. Lame, but simple and it works. Swizzled
357 * pwrite/pread is far from a hotpath - current userspace
358 * doesn't use it at all. */
359 start = round_down(start, 128);
360 end = round_up(end, 128);
362 drm_clflush_virt_range((void *)start, end - start);
364 drm_clflush_virt_range(addr, length);
369 /* Only difference to the fast-path function is that this can handle bit17
370 * and uses non-atomic copy and kmap functions. */
372 shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length,
373 char __user *user_data,
374 bool page_do_bit17_swizzling, bool needs_clflush)
381 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
383 page_do_bit17_swizzling);
385 if (page_do_bit17_swizzling)
386 ret = __copy_to_user_swizzled(user_data,
387 vaddr, shmem_page_offset,
390 ret = __copy_to_user(user_data,
391 vaddr + shmem_page_offset,
399 i915_gem_shmem_pread(struct drm_device *dev,
400 struct drm_i915_gem_object *obj,
401 struct drm_i915_gem_pread *args,
402 struct drm_file *file)
404 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
405 char __user *user_data;
408 int shmem_page_offset, page_length, ret = 0;
409 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
410 int hit_slowpath = 0;
412 int needs_clflush = 0;
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 ret = i915_gem_object_set_to_gtt_domain(obj, false);
432 offset = args->offset;
437 /* Operation in this page
439 * shmem_page_offset = offset within page in shmem file
440 * page_length = bytes to copy for this page
442 shmem_page_offset = offset_in_page(offset);
443 page_length = remain;
444 if ((shmem_page_offset + page_length) > PAGE_SIZE)
445 page_length = PAGE_SIZE - shmem_page_offset;
448 page = obj->pages[offset >> PAGE_SHIFT];
451 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
459 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
460 (page_to_phys(page) & (1 << 17)) != 0;
462 ret = shmem_pread_fast(page, shmem_page_offset, page_length,
463 user_data, page_do_bit17_swizzling,
469 page_cache_get(page);
470 mutex_unlock(&dev->struct_mutex);
473 ret = fault_in_multipages_writeable(user_data, remain);
474 /* Userspace is tricking us, but we've already clobbered
475 * its pages with the prefault and promised to write the
476 * data up to the first fault. Hence ignore any errors
477 * and just continue. */
482 ret = shmem_pread_slow(page, shmem_page_offset, page_length,
483 user_data, page_do_bit17_swizzling,
486 mutex_lock(&dev->struct_mutex);
487 page_cache_release(page);
489 mark_page_accessed(page);
491 page_cache_release(page);
498 remain -= page_length;
499 user_data += page_length;
500 offset += page_length;
505 /* Fixup: Kill any reinstated backing storage pages */
506 if (obj->madv == __I915_MADV_PURGED)
507 i915_gem_object_truncate(obj);
514 * Reads data from the object referenced by handle.
516 * On error, the contents of *data are undefined.
519 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
520 struct drm_file *file)
522 struct drm_i915_gem_pread *args = data;
523 struct drm_i915_gem_object *obj;
529 if (!access_ok(VERIFY_WRITE,
530 (char __user *)(uintptr_t)args->data_ptr,
534 ret = i915_mutex_lock_interruptible(dev);
538 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
539 if (&obj->base == NULL) {
544 /* Bounds check source. */
545 if (args->offset > obj->base.size ||
546 args->size > obj->base.size - args->offset) {
551 /* prime objects have no backing filp to GEM pread/pwrite
554 if (!obj->base.filp) {
559 trace_i915_gem_object_pread(obj, args->offset, args->size);
561 ret = i915_gem_shmem_pread(dev, obj, args, file);
564 drm_gem_object_unreference(&obj->base);
566 mutex_unlock(&dev->struct_mutex);
570 /* This is the fast write path which cannot handle
571 * page faults in the source data
575 fast_user_write(struct io_mapping *mapping,
576 loff_t page_base, int page_offset,
577 char __user *user_data,
580 void __iomem *vaddr_atomic;
582 unsigned long unwritten;
584 vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
585 /* We can use the cpu mem copy function because this is X86. */
586 vaddr = (void __force*)vaddr_atomic + page_offset;
587 unwritten = __copy_from_user_inatomic_nocache(vaddr,
589 io_mapping_unmap_atomic(vaddr_atomic);
594 * This is the fast pwrite path, where we copy the data directly from the
595 * user into the GTT, uncached.
598 i915_gem_gtt_pwrite_fast(struct drm_device *dev,
599 struct drm_i915_gem_object *obj,
600 struct drm_i915_gem_pwrite *args,
601 struct drm_file *file)
603 drm_i915_private_t *dev_priv = dev->dev_private;
605 loff_t offset, page_base;
606 char __user *user_data;
607 int page_offset, page_length, ret;
609 ret = i915_gem_object_pin(obj, 0, true);
613 ret = i915_gem_object_set_to_gtt_domain(obj, true);
617 ret = i915_gem_object_put_fence(obj);
621 user_data = (char __user *) (uintptr_t) args->data_ptr;
624 offset = obj->gtt_offset + args->offset;
627 /* Operation in this page
629 * page_base = page offset within aperture
630 * page_offset = offset within page
631 * page_length = bytes to copy for this page
633 page_base = offset & PAGE_MASK;
634 page_offset = offset_in_page(offset);
635 page_length = remain;
636 if ((page_offset + remain) > PAGE_SIZE)
637 page_length = PAGE_SIZE - page_offset;
639 /* If we get a fault while copying data, then (presumably) our
640 * source page isn't available. Return the error and we'll
641 * retry in the slow path.
643 if (fast_user_write(dev_priv->mm.gtt_mapping, page_base,
644 page_offset, user_data, page_length)) {
649 remain -= page_length;
650 user_data += page_length;
651 offset += page_length;
655 i915_gem_object_unpin(obj);
660 /* Per-page copy function for the shmem pwrite fastpath.
661 * Flushes invalid cachelines before writing to the target if
662 * needs_clflush_before is set and flushes out any written cachelines after
663 * writing if needs_clflush is set. */
665 shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
666 char __user *user_data,
667 bool page_do_bit17_swizzling,
668 bool needs_clflush_before,
669 bool needs_clflush_after)
674 if (unlikely(page_do_bit17_swizzling))
677 vaddr = kmap_atomic(page);
678 if (needs_clflush_before)
679 drm_clflush_virt_range(vaddr + shmem_page_offset,
681 ret = __copy_from_user_inatomic_nocache(vaddr + shmem_page_offset,
684 if (needs_clflush_after)
685 drm_clflush_virt_range(vaddr + shmem_page_offset,
687 kunmap_atomic(vaddr);
692 /* Only difference to the fast-path function is that this can handle bit17
693 * and uses non-atomic copy and kmap functions. */
695 shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
696 char __user *user_data,
697 bool page_do_bit17_swizzling,
698 bool needs_clflush_before,
699 bool needs_clflush_after)
705 if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
706 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
708 page_do_bit17_swizzling);
709 if (page_do_bit17_swizzling)
710 ret = __copy_from_user_swizzled(vaddr, shmem_page_offset,
714 ret = __copy_from_user(vaddr + shmem_page_offset,
717 if (needs_clflush_after)
718 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
720 page_do_bit17_swizzling);
727 i915_gem_shmem_pwrite(struct drm_device *dev,
728 struct drm_i915_gem_object *obj,
729 struct drm_i915_gem_pwrite *args,
730 struct drm_file *file)
732 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
735 char __user *user_data;
736 int shmem_page_offset, page_length, ret = 0;
737 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
738 int hit_slowpath = 0;
739 int needs_clflush_after = 0;
740 int needs_clflush_before = 0;
743 user_data = (char __user *) (uintptr_t) args->data_ptr;
746 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
748 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
749 /* If we're not in the cpu write domain, set ourself into the gtt
750 * write domain and manually flush cachelines (if required). This
751 * optimizes for the case when the gpu will use the data
752 * right away and we therefore have to clflush anyway. */
753 if (obj->cache_level == I915_CACHE_NONE)
754 needs_clflush_after = 1;
755 ret = i915_gem_object_set_to_gtt_domain(obj, true);
759 /* Same trick applies for invalidate partially written cachelines before
761 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)
762 && obj->cache_level == I915_CACHE_NONE)
763 needs_clflush_before = 1;
765 offset = args->offset;
770 int partial_cacheline_write;
772 /* Operation in this page
774 * shmem_page_offset = offset within page in shmem file
775 * page_length = bytes to copy for this page
777 shmem_page_offset = offset_in_page(offset);
779 page_length = remain;
780 if ((shmem_page_offset + page_length) > PAGE_SIZE)
781 page_length = PAGE_SIZE - shmem_page_offset;
783 /* If we don't overwrite a cacheline completely we need to be
784 * careful to have up-to-date data by first clflushing. Don't
785 * overcomplicate things and flush the entire patch. */
786 partial_cacheline_write = needs_clflush_before &&
787 ((shmem_page_offset | page_length)
788 & (boot_cpu_data.x86_clflush_size - 1));
791 page = obj->pages[offset >> PAGE_SHIFT];
794 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
802 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
803 (page_to_phys(page) & (1 << 17)) != 0;
805 ret = shmem_pwrite_fast(page, shmem_page_offset, page_length,
806 user_data, page_do_bit17_swizzling,
807 partial_cacheline_write,
808 needs_clflush_after);
813 page_cache_get(page);
814 mutex_unlock(&dev->struct_mutex);
816 ret = shmem_pwrite_slow(page, shmem_page_offset, page_length,
817 user_data, page_do_bit17_swizzling,
818 partial_cacheline_write,
819 needs_clflush_after);
821 mutex_lock(&dev->struct_mutex);
822 page_cache_release(page);
824 set_page_dirty(page);
825 mark_page_accessed(page);
827 page_cache_release(page);
834 remain -= page_length;
835 user_data += page_length;
836 offset += page_length;
841 /* Fixup: Kill any reinstated backing storage pages */
842 if (obj->madv == __I915_MADV_PURGED)
843 i915_gem_object_truncate(obj);
844 /* and flush dirty cachelines in case the object isn't in the cpu write
846 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
847 i915_gem_clflush_object(obj);
848 intel_gtt_chipset_flush();
852 if (needs_clflush_after)
853 intel_gtt_chipset_flush();
859 * Writes data to the object referenced by handle.
861 * On error, the contents of the buffer that were to be modified are undefined.
864 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
865 struct drm_file *file)
867 struct drm_i915_gem_pwrite *args = data;
868 struct drm_i915_gem_object *obj;
874 if (!access_ok(VERIFY_READ,
875 (char __user *)(uintptr_t)args->data_ptr,
879 ret = fault_in_multipages_readable((char __user *)(uintptr_t)args->data_ptr,
884 ret = i915_mutex_lock_interruptible(dev);
888 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
889 if (&obj->base == NULL) {
894 /* Bounds check destination. */
895 if (args->offset > obj->base.size ||
896 args->size > obj->base.size - args->offset) {
901 /* prime objects have no backing filp to GEM pread/pwrite
904 if (!obj->base.filp) {
909 trace_i915_gem_object_pwrite(obj, args->offset, args->size);
912 /* We can only do the GTT pwrite on untiled buffers, as otherwise
913 * it would end up going through the fenced access, and we'll get
914 * different detiling behavior between reading and writing.
915 * pread/pwrite currently are reading and writing from the CPU
916 * perspective, requiring manual detiling by the client.
919 ret = i915_gem_phys_pwrite(dev, obj, args, file);
923 if (obj->gtt_space &&
924 obj->cache_level == I915_CACHE_NONE &&
925 obj->tiling_mode == I915_TILING_NONE &&
926 obj->map_and_fenceable &&
927 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
928 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
929 /* Note that the gtt paths might fail with non-page-backed user
930 * pointers (e.g. gtt mappings when moving data between
931 * textures). Fallback to the shmem path in that case. */
935 ret = i915_gem_shmem_pwrite(dev, obj, args, file);
938 drm_gem_object_unreference(&obj->base);
940 mutex_unlock(&dev->struct_mutex);
945 * Called when user space prepares to use an object with the CPU, either
946 * through the mmap ioctl's mapping or a GTT mapping.
949 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
950 struct drm_file *file)
952 struct drm_i915_gem_set_domain *args = data;
953 struct drm_i915_gem_object *obj;
954 uint32_t read_domains = args->read_domains;
955 uint32_t write_domain = args->write_domain;
958 /* Only handle setting domains to types used by the CPU. */
959 if (write_domain & I915_GEM_GPU_DOMAINS)
962 if (read_domains & I915_GEM_GPU_DOMAINS)
965 /* Having something in the write domain implies it's in the read
966 * domain, and only that read domain. Enforce that in the request.
968 if (write_domain != 0 && read_domains != write_domain)
971 ret = i915_mutex_lock_interruptible(dev);
975 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
976 if (&obj->base == NULL) {
981 if (read_domains & I915_GEM_DOMAIN_GTT) {
982 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
984 /* Silently promote "you're not bound, there was nothing to do"
985 * to success, since the client was just asking us to
986 * make sure everything was done.
991 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
994 drm_gem_object_unreference(&obj->base);
996 mutex_unlock(&dev->struct_mutex);
1001 * Called when user space has done writes to this buffer
1004 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1005 struct drm_file *file)
1007 struct drm_i915_gem_sw_finish *args = data;
1008 struct drm_i915_gem_object *obj;
1011 ret = i915_mutex_lock_interruptible(dev);
1015 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1016 if (&obj->base == NULL) {
1021 /* Pinned buffers may be scanout, so flush the cache */
1023 i915_gem_object_flush_cpu_write_domain(obj);
1025 drm_gem_object_unreference(&obj->base);
1027 mutex_unlock(&dev->struct_mutex);
1032 * Maps the contents of an object, returning the address it is mapped
1035 * While the mapping holds a reference on the contents of the object, it doesn't
1036 * imply a ref on the object itself.
1039 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1040 struct drm_file *file)
1042 struct drm_i915_gem_mmap *args = data;
1043 struct drm_gem_object *obj;
1046 obj = drm_gem_object_lookup(dev, file, args->handle);
1050 /* prime objects have no backing filp to GEM mmap
1054 drm_gem_object_unreference_unlocked(obj);
1058 addr = vm_mmap(obj->filp, 0, args->size,
1059 PROT_READ | PROT_WRITE, MAP_SHARED,
1061 drm_gem_object_unreference_unlocked(obj);
1062 if (IS_ERR((void *)addr))
1065 args->addr_ptr = (uint64_t) addr;
1071 * i915_gem_fault - fault a page into the GTT
1072 * vma: VMA in question
1075 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1076 * from userspace. The fault handler takes care of binding the object to
1077 * the GTT (if needed), allocating and programming a fence register (again,
1078 * only if needed based on whether the old reg is still valid or the object
1079 * is tiled) and inserting a new PTE into the faulting process.
1081 * Note that the faulting process may involve evicting existing objects
1082 * from the GTT and/or fence registers to make room. So performance may
1083 * suffer if the GTT working set is large or there are few fence registers
1086 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1088 struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1089 struct drm_device *dev = obj->base.dev;
1090 drm_i915_private_t *dev_priv = dev->dev_private;
1091 pgoff_t page_offset;
1094 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1096 /* We don't use vmf->pgoff since that has the fake offset */
1097 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1100 ret = i915_mutex_lock_interruptible(dev);
1104 trace_i915_gem_object_fault(obj, page_offset, true, write);
1106 /* Now bind it into the GTT if needed */
1107 if (!obj->map_and_fenceable) {
1108 ret = i915_gem_object_unbind(obj);
1112 if (!obj->gtt_space) {
1113 ret = i915_gem_object_bind_to_gtt(obj, 0, true);
1117 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1122 if (!obj->has_global_gtt_mapping)
1123 i915_gem_gtt_bind_object(obj, obj->cache_level);
1125 ret = i915_gem_object_get_fence(obj);
1129 if (i915_gem_object_is_inactive(obj))
1130 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1132 obj->fault_mappable = true;
1134 pfn = ((dev_priv->mm.gtt_base_addr + obj->gtt_offset) >> PAGE_SHIFT) +
1137 /* Finally, remap it using the new GTT offset */
1138 ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1140 mutex_unlock(&dev->struct_mutex);
1144 /* If this -EIO is due to a gpu hang, give the reset code a
1145 * chance to clean up the mess. Otherwise return the proper
1147 if (!atomic_read(&dev_priv->mm.wedged))
1148 return VM_FAULT_SIGBUS;
1150 /* Give the error handler a chance to run and move the
1151 * objects off the GPU active list. Next time we service the
1152 * fault, we should be able to transition the page into the
1153 * GTT without touching the GPU (and so avoid further
1154 * EIO/EGAIN). If the GPU is wedged, then there is no issue
1155 * with coherency, just lost writes.
1161 return VM_FAULT_NOPAGE;
1163 return VM_FAULT_OOM;
1165 return VM_FAULT_SIGBUS;
1170 * i915_gem_release_mmap - remove physical page mappings
1171 * @obj: obj in question
1173 * Preserve the reservation of the mmapping with the DRM core code, but
1174 * relinquish ownership of the pages back to the system.
1176 * It is vital that we remove the page mapping if we have mapped a tiled
1177 * object through the GTT and then lose the fence register due to
1178 * resource pressure. Similarly if the object has been moved out of the
1179 * aperture, than pages mapped into userspace must be revoked. Removing the
1180 * mapping will then trigger a page fault on the next user access, allowing
1181 * fixup by i915_gem_fault().
1184 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1186 if (!obj->fault_mappable)
1189 if (obj->base.dev->dev_mapping)
1190 unmap_mapping_range(obj->base.dev->dev_mapping,
1191 (loff_t)obj->base.map_list.hash.key<<PAGE_SHIFT,
1194 obj->fault_mappable = false;
1198 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1202 if (INTEL_INFO(dev)->gen >= 4 ||
1203 tiling_mode == I915_TILING_NONE)
1206 /* Previous chips need a power-of-two fence region when tiling */
1207 if (INTEL_INFO(dev)->gen == 3)
1208 gtt_size = 1024*1024;
1210 gtt_size = 512*1024;
1212 while (gtt_size < size)
1219 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1220 * @obj: object to check
1222 * Return the required GTT alignment for an object, taking into account
1223 * potential fence register mapping.
1226 i915_gem_get_gtt_alignment(struct drm_device *dev,
1231 * Minimum alignment is 4k (GTT page size), but might be greater
1232 * if a fence register is needed for the object.
1234 if (INTEL_INFO(dev)->gen >= 4 ||
1235 tiling_mode == I915_TILING_NONE)
1239 * Previous chips need to be aligned to the size of the smallest
1240 * fence register that can contain the object.
1242 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1246 * i915_gem_get_unfenced_gtt_alignment - return required GTT alignment for an
1249 * @size: size of the object
1250 * @tiling_mode: tiling mode of the object
1252 * Return the required GTT alignment for an object, only taking into account
1253 * unfenced tiled surface requirements.
1256 i915_gem_get_unfenced_gtt_alignment(struct drm_device *dev,
1261 * Minimum alignment is 4k (GTT page size) for sane hw.
1263 if (INTEL_INFO(dev)->gen >= 4 || IS_G33(dev) ||
1264 tiling_mode == I915_TILING_NONE)
1267 /* Previous hardware however needs to be aligned to a power-of-two
1268 * tile height. The simplest method for determining this is to reuse
1269 * the power-of-tile object size.
1271 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1275 i915_gem_mmap_gtt(struct drm_file *file,
1276 struct drm_device *dev,
1280 struct drm_i915_private *dev_priv = dev->dev_private;
1281 struct drm_i915_gem_object *obj;
1284 ret = i915_mutex_lock_interruptible(dev);
1288 obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
1289 if (&obj->base == NULL) {
1294 if (obj->base.size > dev_priv->mm.gtt_mappable_end) {
1299 if (obj->madv != I915_MADV_WILLNEED) {
1300 DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1305 if (!obj->base.map_list.map) {
1306 ret = drm_gem_create_mmap_offset(&obj->base);
1311 *offset = (u64)obj->base.map_list.hash.key << PAGE_SHIFT;
1314 drm_gem_object_unreference(&obj->base);
1316 mutex_unlock(&dev->struct_mutex);
1321 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1323 * @data: GTT mapping ioctl data
1324 * @file: GEM object info
1326 * Simply returns the fake offset to userspace so it can mmap it.
1327 * The mmap call will end up in drm_gem_mmap(), which will set things
1328 * up so we can get faults in the handler above.
1330 * The fault handler will take care of binding the object into the GTT
1331 * (since it may have been evicted to make room for something), allocating
1332 * a fence register, and mapping the appropriate aperture address into
1336 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1337 struct drm_file *file)
1339 struct drm_i915_gem_mmap_gtt *args = data;
1341 return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
1345 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj,
1349 struct address_space *mapping;
1350 struct inode *inode;
1353 if (obj->pages || obj->sg_table)
1356 /* Get the list of pages out of our struct file. They'll be pinned
1357 * at this point until we release them.
1359 page_count = obj->base.size / PAGE_SIZE;
1360 BUG_ON(obj->pages != NULL);
1361 obj->pages = drm_malloc_ab(page_count, sizeof(struct page *));
1362 if (obj->pages == NULL)
1365 inode = obj->base.filp->f_path.dentry->d_inode;
1366 mapping = inode->i_mapping;
1367 gfpmask |= mapping_gfp_mask(mapping);
1369 for (i = 0; i < page_count; i++) {
1370 page = shmem_read_mapping_page_gfp(mapping, i, gfpmask);
1374 obj->pages[i] = page;
1377 if (i915_gem_object_needs_bit17_swizzle(obj))
1378 i915_gem_object_do_bit_17_swizzle(obj);
1384 page_cache_release(obj->pages[i]);
1386 drm_free_large(obj->pages);
1388 return PTR_ERR(page);
1392 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
1394 int page_count = obj->base.size / PAGE_SIZE;
1400 BUG_ON(obj->madv == __I915_MADV_PURGED);
1402 if (i915_gem_object_needs_bit17_swizzle(obj))
1403 i915_gem_object_save_bit_17_swizzle(obj);
1405 if (obj->madv == I915_MADV_DONTNEED)
1408 for (i = 0; i < page_count; i++) {
1410 set_page_dirty(obj->pages[i]);
1412 if (obj->madv == I915_MADV_WILLNEED)
1413 mark_page_accessed(obj->pages[i]);
1415 page_cache_release(obj->pages[i]);
1419 drm_free_large(obj->pages);
1424 i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
1425 struct intel_ring_buffer *ring,
1428 struct drm_device *dev = obj->base.dev;
1429 struct drm_i915_private *dev_priv = dev->dev_private;
1431 BUG_ON(ring == NULL);
1434 /* Add a reference if we're newly entering the active list. */
1436 drm_gem_object_reference(&obj->base);
1440 /* Move from whatever list we were on to the tail of execution. */
1441 list_move_tail(&obj->mm_list, &dev_priv->mm.active_list);
1442 list_move_tail(&obj->ring_list, &ring->active_list);
1444 obj->last_read_seqno = seqno;
1446 if (obj->fenced_gpu_access) {
1447 obj->last_fenced_seqno = seqno;
1449 /* Bump MRU to take account of the delayed flush */
1450 if (obj->fence_reg != I915_FENCE_REG_NONE) {
1451 struct drm_i915_fence_reg *reg;
1453 reg = &dev_priv->fence_regs[obj->fence_reg];
1454 list_move_tail(®->lru_list,
1455 &dev_priv->mm.fence_list);
1461 i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
1463 struct drm_device *dev = obj->base.dev;
1464 struct drm_i915_private *dev_priv = dev->dev_private;
1466 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1468 BUG_ON(!list_empty(&obj->gpu_write_list));
1469 BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS);
1470 BUG_ON(!obj->active);
1472 list_del_init(&obj->ring_list);
1475 obj->last_read_seqno = 0;
1476 obj->last_write_seqno = 0;
1477 obj->base.write_domain = 0;
1479 obj->last_fenced_seqno = 0;
1480 obj->fenced_gpu_access = false;
1483 drm_gem_object_unreference(&obj->base);
1485 WARN_ON(i915_verify_lists(dev));
1488 /* Immediately discard the backing storage */
1490 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
1492 struct inode *inode;
1494 /* Our goal here is to return as much of the memory as
1495 * is possible back to the system as we are called from OOM.
1496 * To do this we must instruct the shmfs to drop all of its
1497 * backing pages, *now*.
1499 inode = obj->base.filp->f_path.dentry->d_inode;
1500 shmem_truncate_range(inode, 0, (loff_t)-1);
1502 if (obj->base.map_list.map)
1503 drm_gem_free_mmap_offset(&obj->base);
1505 obj->madv = __I915_MADV_PURGED;
1509 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj)
1511 return obj->madv == I915_MADV_DONTNEED;
1515 i915_gem_process_flushing_list(struct intel_ring_buffer *ring,
1516 uint32_t flush_domains)
1518 struct drm_i915_gem_object *obj, *next;
1520 list_for_each_entry_safe(obj, next,
1521 &ring->gpu_write_list,
1523 if (obj->base.write_domain & flush_domains) {
1524 uint32_t old_write_domain = obj->base.write_domain;
1526 obj->base.write_domain = 0;
1527 list_del_init(&obj->gpu_write_list);
1528 i915_gem_object_move_to_active(obj, ring,
1529 i915_gem_next_request_seqno(ring));
1531 trace_i915_gem_object_change_domain(obj,
1532 obj->base.read_domains,
1539 i915_gem_get_seqno(struct drm_device *dev)
1541 drm_i915_private_t *dev_priv = dev->dev_private;
1542 u32 seqno = dev_priv->next_seqno;
1544 /* reserve 0 for non-seqno */
1545 if (++dev_priv->next_seqno == 0)
1546 dev_priv->next_seqno = 1;
1552 i915_gem_next_request_seqno(struct intel_ring_buffer *ring)
1554 if (ring->outstanding_lazy_request == 0)
1555 ring->outstanding_lazy_request = i915_gem_get_seqno(ring->dev);
1557 return ring->outstanding_lazy_request;
1561 i915_add_request(struct intel_ring_buffer *ring,
1562 struct drm_file *file,
1563 struct drm_i915_gem_request *request)
1565 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1567 u32 request_ring_position;
1572 * Emit any outstanding flushes - execbuf can fail to emit the flush
1573 * after having emitted the batchbuffer command. Hence we need to fix
1574 * things up similar to emitting the lazy request. The difference here
1575 * is that the flush _must_ happen before the next request, no matter
1578 if (ring->gpu_caches_dirty) {
1579 ret = i915_gem_flush_ring(ring, 0, I915_GEM_GPU_DOMAINS);
1583 ring->gpu_caches_dirty = false;
1586 if (request == NULL) {
1587 request = kmalloc(sizeof(*request), GFP_KERNEL);
1588 if (request == NULL)
1592 seqno = i915_gem_next_request_seqno(ring);
1594 /* Record the position of the start of the request so that
1595 * should we detect the updated seqno part-way through the
1596 * GPU processing the request, we never over-estimate the
1597 * position of the head.
1599 request_ring_position = intel_ring_get_tail(ring);
1601 ret = ring->add_request(ring, &seqno);
1607 trace_i915_gem_request_add(ring, seqno);
1609 request->seqno = seqno;
1610 request->ring = ring;
1611 request->tail = request_ring_position;
1612 request->emitted_jiffies = jiffies;
1613 was_empty = list_empty(&ring->request_list);
1614 list_add_tail(&request->list, &ring->request_list);
1615 request->file_priv = NULL;
1618 struct drm_i915_file_private *file_priv = file->driver_priv;
1620 spin_lock(&file_priv->mm.lock);
1621 request->file_priv = file_priv;
1622 list_add_tail(&request->client_list,
1623 &file_priv->mm.request_list);
1624 spin_unlock(&file_priv->mm.lock);
1627 ring->outstanding_lazy_request = 0;
1629 if (!dev_priv->mm.suspended) {
1630 if (i915_enable_hangcheck) {
1631 mod_timer(&dev_priv->hangcheck_timer,
1633 msecs_to_jiffies(DRM_I915_HANGCHECK_PERIOD));
1636 queue_delayed_work(dev_priv->wq,
1637 &dev_priv->mm.retire_work, HZ);
1640 WARN_ON(!list_empty(&ring->gpu_write_list));
1646 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
1648 struct drm_i915_file_private *file_priv = request->file_priv;
1653 spin_lock(&file_priv->mm.lock);
1654 if (request->file_priv) {
1655 list_del(&request->client_list);
1656 request->file_priv = NULL;
1658 spin_unlock(&file_priv->mm.lock);
1661 static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv,
1662 struct intel_ring_buffer *ring)
1664 while (!list_empty(&ring->request_list)) {
1665 struct drm_i915_gem_request *request;
1667 request = list_first_entry(&ring->request_list,
1668 struct drm_i915_gem_request,
1671 list_del(&request->list);
1672 i915_gem_request_remove_from_client(request);
1676 while (!list_empty(&ring->active_list)) {
1677 struct drm_i915_gem_object *obj;
1679 obj = list_first_entry(&ring->active_list,
1680 struct drm_i915_gem_object,
1683 list_del_init(&obj->gpu_write_list);
1684 i915_gem_object_move_to_inactive(obj);
1688 static void i915_gem_reset_fences(struct drm_device *dev)
1690 struct drm_i915_private *dev_priv = dev->dev_private;
1693 for (i = 0; i < dev_priv->num_fence_regs; i++) {
1694 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
1696 i915_gem_write_fence(dev, i, NULL);
1699 i915_gem_object_fence_lost(reg->obj);
1703 INIT_LIST_HEAD(®->lru_list);
1706 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
1709 void i915_gem_reset(struct drm_device *dev)
1711 struct drm_i915_private *dev_priv = dev->dev_private;
1712 struct drm_i915_gem_object *obj;
1713 struct intel_ring_buffer *ring;
1716 for_each_ring(ring, dev_priv, i)
1717 i915_gem_reset_ring_lists(dev_priv, ring);
1719 /* Move everything out of the GPU domains to ensure we do any
1720 * necessary invalidation upon reuse.
1722 list_for_each_entry(obj,
1723 &dev_priv->mm.inactive_list,
1726 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
1729 /* The fence registers are invalidated so clear them out */
1730 i915_gem_reset_fences(dev);
1734 * This function clears the request list as sequence numbers are passed.
1737 i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
1742 if (list_empty(&ring->request_list))
1745 WARN_ON(i915_verify_lists(ring->dev));
1747 seqno = ring->get_seqno(ring);
1749 for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++)
1750 if (seqno >= ring->sync_seqno[i])
1751 ring->sync_seqno[i] = 0;
1753 while (!list_empty(&ring->request_list)) {
1754 struct drm_i915_gem_request *request;
1756 request = list_first_entry(&ring->request_list,
1757 struct drm_i915_gem_request,
1760 if (!i915_seqno_passed(seqno, request->seqno))
1763 trace_i915_gem_request_retire(ring, request->seqno);
1764 /* We know the GPU must have read the request to have
1765 * sent us the seqno + interrupt, so use the position
1766 * of tail of the request to update the last known position
1769 ring->last_retired_head = request->tail;
1771 list_del(&request->list);
1772 i915_gem_request_remove_from_client(request);
1776 /* Move any buffers on the active list that are no longer referenced
1777 * by the ringbuffer to the flushing/inactive lists as appropriate.
1779 while (!list_empty(&ring->active_list)) {
1780 struct drm_i915_gem_object *obj;
1782 obj = list_first_entry(&ring->active_list,
1783 struct drm_i915_gem_object,
1786 if (!i915_seqno_passed(seqno, obj->last_read_seqno))
1789 i915_gem_object_move_to_inactive(obj);
1792 if (unlikely(ring->trace_irq_seqno &&
1793 i915_seqno_passed(seqno, ring->trace_irq_seqno))) {
1794 ring->irq_put(ring);
1795 ring->trace_irq_seqno = 0;
1798 WARN_ON(i915_verify_lists(ring->dev));
1802 i915_gem_retire_requests(struct drm_device *dev)
1804 drm_i915_private_t *dev_priv = dev->dev_private;
1805 struct intel_ring_buffer *ring;
1808 for_each_ring(ring, dev_priv, i)
1809 i915_gem_retire_requests_ring(ring);
1813 i915_gem_retire_work_handler(struct work_struct *work)
1815 drm_i915_private_t *dev_priv;
1816 struct drm_device *dev;
1817 struct intel_ring_buffer *ring;
1821 dev_priv = container_of(work, drm_i915_private_t,
1822 mm.retire_work.work);
1823 dev = dev_priv->dev;
1825 /* Come back later if the device is busy... */
1826 if (!mutex_trylock(&dev->struct_mutex)) {
1827 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1831 i915_gem_retire_requests(dev);
1833 /* Send a periodic flush down the ring so we don't hold onto GEM
1834 * objects indefinitely.
1837 for_each_ring(ring, dev_priv, i) {
1838 if (ring->gpu_caches_dirty)
1839 i915_add_request(ring, NULL, NULL);
1841 idle &= list_empty(&ring->request_list);
1844 if (!dev_priv->mm.suspended && !idle)
1845 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1847 mutex_unlock(&dev->struct_mutex);
1851 i915_gem_check_wedge(struct drm_i915_private *dev_priv,
1854 if (atomic_read(&dev_priv->mm.wedged)) {
1855 struct completion *x = &dev_priv->error_completion;
1856 bool recovery_complete;
1857 unsigned long flags;
1859 /* Give the error handler a chance to run. */
1860 spin_lock_irqsave(&x->wait.lock, flags);
1861 recovery_complete = x->done > 0;
1862 spin_unlock_irqrestore(&x->wait.lock, flags);
1864 /* Non-interruptible callers can't handle -EAGAIN, hence return
1865 * -EIO unconditionally for these. */
1869 /* Recovery complete, but still wedged means reset failure. */
1870 if (recovery_complete)
1880 * Compare seqno against outstanding lazy request. Emit a request if they are
1884 i915_gem_check_olr(struct intel_ring_buffer *ring, u32 seqno)
1888 BUG_ON(!mutex_is_locked(&ring->dev->struct_mutex));
1891 if (seqno == ring->outstanding_lazy_request)
1892 ret = i915_add_request(ring, NULL, NULL);
1898 * __wait_seqno - wait until execution of seqno has finished
1899 * @ring: the ring expected to report seqno
1901 * @interruptible: do an interruptible wait (normally yes)
1902 * @timeout: in - how long to wait (NULL forever); out - how much time remaining
1904 * Returns 0 if the seqno was found within the alloted time. Else returns the
1905 * errno with remaining time filled in timeout argument.
1907 static int __wait_seqno(struct intel_ring_buffer *ring, u32 seqno,
1908 bool interruptible, struct timespec *timeout)
1910 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1911 struct timespec before, now, wait_time={1,0};
1912 unsigned long timeout_jiffies;
1914 bool wait_forever = true;
1917 if (i915_seqno_passed(ring->get_seqno(ring), seqno))
1920 trace_i915_gem_request_wait_begin(ring, seqno);
1922 if (timeout != NULL) {
1923 wait_time = *timeout;
1924 wait_forever = false;
1927 timeout_jiffies = timespec_to_jiffies(&wait_time);
1929 if (WARN_ON(!ring->irq_get(ring)))
1932 /* Record current time in case interrupted by signal, or wedged * */
1933 getrawmonotonic(&before);
1936 (i915_seqno_passed(ring->get_seqno(ring), seqno) || \
1937 atomic_read(&dev_priv->mm.wedged))
1940 end = wait_event_interruptible_timeout(ring->irq_queue,
1944 end = wait_event_timeout(ring->irq_queue, EXIT_COND,
1947 ret = i915_gem_check_wedge(dev_priv, interruptible);
1950 } while (end == 0 && wait_forever);
1952 getrawmonotonic(&now);
1954 ring->irq_put(ring);
1955 trace_i915_gem_request_wait_end(ring, seqno);
1959 struct timespec sleep_time = timespec_sub(now, before);
1960 *timeout = timespec_sub(*timeout, sleep_time);
1965 case -EAGAIN: /* Wedged */
1966 case -ERESTARTSYS: /* Signal */
1968 case 0: /* Timeout */
1970 set_normalized_timespec(timeout, 0, 0);
1972 default: /* Completed */
1973 WARN_ON(end < 0); /* We're not aware of other errors */
1979 * Waits for a sequence number to be signaled, and cleans up the
1980 * request and object lists appropriately for that event.
1983 i915_wait_seqno(struct intel_ring_buffer *ring, uint32_t seqno)
1985 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1990 ret = i915_gem_check_wedge(dev_priv, dev_priv->mm.interruptible);
1994 ret = i915_gem_check_olr(ring, seqno);
1998 ret = __wait_seqno(ring, seqno, dev_priv->mm.interruptible, NULL);
2004 * Ensures that all rendering to the object has completed and the object is
2005 * safe to unbind from the GTT or access from the CPU.
2007 static __must_check int
2008 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
2014 /* This function only exists to support waiting for existing rendering,
2015 * not for emitting required flushes.
2017 BUG_ON((obj->base.write_domain & I915_GEM_GPU_DOMAINS) != 0);
2019 /* If there is rendering queued on the buffer being evicted, wait for
2023 seqno = obj->last_write_seqno;
2025 seqno = obj->last_read_seqno;
2029 ret = i915_wait_seqno(obj->ring, seqno);
2033 /* Manually manage the write flush as we may have not yet retired
2036 if (obj->last_write_seqno &&
2037 i915_seqno_passed(seqno, obj->last_write_seqno)) {
2038 obj->last_write_seqno = 0;
2039 obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
2042 i915_gem_retire_requests_ring(obj->ring);
2047 * Ensures that an object will eventually get non-busy by flushing any required
2048 * write domains, emitting any outstanding lazy request and retiring and
2049 * completed requests.
2052 i915_gem_object_flush_active(struct drm_i915_gem_object *obj)
2057 ret = i915_gem_object_flush_gpu_write_domain(obj);
2061 ret = i915_gem_check_olr(obj->ring, obj->last_read_seqno);
2065 i915_gem_retire_requests_ring(obj->ring);
2072 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
2073 * @DRM_IOCTL_ARGS: standard ioctl arguments
2075 * Returns 0 if successful, else an error is returned with the remaining time in
2076 * the timeout parameter.
2077 * -ETIME: object is still busy after timeout
2078 * -ERESTARTSYS: signal interrupted the wait
2079 * -ENONENT: object doesn't exist
2080 * Also possible, but rare:
2081 * -EAGAIN: GPU wedged
2083 * -ENODEV: Internal IRQ fail
2084 * -E?: The add request failed
2086 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
2087 * non-zero timeout parameter the wait ioctl will wait for the given number of
2088 * nanoseconds on an object becoming unbusy. Since the wait itself does so
2089 * without holding struct_mutex the object may become re-busied before this
2090 * function completes. A similar but shorter * race condition exists in the busy
2094 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
2096 struct drm_i915_gem_wait *args = data;
2097 struct drm_i915_gem_object *obj;
2098 struct intel_ring_buffer *ring = NULL;
2099 struct timespec timeout_stack, *timeout = NULL;
2103 if (args->timeout_ns >= 0) {
2104 timeout_stack = ns_to_timespec(args->timeout_ns);
2105 timeout = &timeout_stack;
2108 ret = i915_mutex_lock_interruptible(dev);
2112 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle));
2113 if (&obj->base == NULL) {
2114 mutex_unlock(&dev->struct_mutex);
2118 /* Need to make sure the object gets inactive eventually. */
2119 ret = i915_gem_object_flush_active(obj);
2124 seqno = obj->last_read_seqno;
2131 /* Do this after OLR check to make sure we make forward progress polling
2132 * on this IOCTL with a 0 timeout (like busy ioctl)
2134 if (!args->timeout_ns) {
2139 drm_gem_object_unreference(&obj->base);
2140 mutex_unlock(&dev->struct_mutex);
2142 ret = __wait_seqno(ring, seqno, true, timeout);
2144 WARN_ON(!timespec_valid(timeout));
2145 args->timeout_ns = timespec_to_ns(timeout);
2150 drm_gem_object_unreference(&obj->base);
2151 mutex_unlock(&dev->struct_mutex);
2156 * i915_gem_object_sync - sync an object to a ring.
2158 * @obj: object which may be in use on another ring.
2159 * @to: ring we wish to use the object on. May be NULL.
2161 * This code is meant to abstract object synchronization with the GPU.
2162 * Calling with NULL implies synchronizing the object with the CPU
2163 * rather than a particular GPU ring.
2165 * Returns 0 if successful, else propagates up the lower layer error.
2168 i915_gem_object_sync(struct drm_i915_gem_object *obj,
2169 struct intel_ring_buffer *to)
2171 struct intel_ring_buffer *from = obj->ring;
2175 if (from == NULL || to == from)
2178 if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev))
2179 return i915_gem_object_wait_rendering(obj, false);
2181 idx = intel_ring_sync_index(from, to);
2183 seqno = obj->last_read_seqno;
2184 if (seqno <= from->sync_seqno[idx])
2187 ret = i915_gem_check_olr(obj->ring, seqno);
2191 ret = to->sync_to(to, from, seqno);
2193 from->sync_seqno[idx] = seqno;
2198 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
2200 u32 old_write_domain, old_read_domains;
2202 /* Act a barrier for all accesses through the GTT */
2205 /* Force a pagefault for domain tracking on next user access */
2206 i915_gem_release_mmap(obj);
2208 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
2211 old_read_domains = obj->base.read_domains;
2212 old_write_domain = obj->base.write_domain;
2214 obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
2215 obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
2217 trace_i915_gem_object_change_domain(obj,
2223 * Unbinds an object from the GTT aperture.
2226 i915_gem_object_unbind(struct drm_i915_gem_object *obj)
2228 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
2231 if (obj->gtt_space == NULL)
2237 ret = i915_gem_object_finish_gpu(obj);
2240 /* Continue on if we fail due to EIO, the GPU is hung so we
2241 * should be safe and we need to cleanup or else we might
2242 * cause memory corruption through use-after-free.
2245 i915_gem_object_finish_gtt(obj);
2247 /* Move the object to the CPU domain to ensure that
2248 * any possible CPU writes while it's not in the GTT
2249 * are flushed when we go to remap it.
2252 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
2253 if (ret == -ERESTARTSYS)
2256 /* In the event of a disaster, abandon all caches and
2257 * hope for the best.
2259 i915_gem_clflush_object(obj);
2260 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
2263 /* release the fence reg _after_ flushing */
2264 ret = i915_gem_object_put_fence(obj);
2268 trace_i915_gem_object_unbind(obj);
2270 if (obj->has_global_gtt_mapping)
2271 i915_gem_gtt_unbind_object(obj);
2272 if (obj->has_aliasing_ppgtt_mapping) {
2273 i915_ppgtt_unbind_object(dev_priv->mm.aliasing_ppgtt, obj);
2274 obj->has_aliasing_ppgtt_mapping = 0;
2276 i915_gem_gtt_finish_object(obj);
2278 i915_gem_object_put_pages_gtt(obj);
2280 list_del_init(&obj->gtt_list);
2281 list_del_init(&obj->mm_list);
2282 /* Avoid an unnecessary call to unbind on rebind. */
2283 obj->map_and_fenceable = true;
2285 drm_mm_put_block(obj->gtt_space);
2286 obj->gtt_space = NULL;
2287 obj->gtt_offset = 0;
2289 if (i915_gem_object_is_purgeable(obj))
2290 i915_gem_object_truncate(obj);
2296 i915_gem_flush_ring(struct intel_ring_buffer *ring,
2297 uint32_t invalidate_domains,
2298 uint32_t flush_domains)
2302 if (((invalidate_domains | flush_domains) & I915_GEM_GPU_DOMAINS) == 0)
2305 trace_i915_gem_ring_flush(ring, invalidate_domains, flush_domains);
2307 ret = ring->flush(ring, invalidate_domains, flush_domains);
2311 if (flush_domains & I915_GEM_GPU_DOMAINS)
2312 i915_gem_process_flushing_list(ring, flush_domains);
2317 static int i915_ring_idle(struct intel_ring_buffer *ring)
2321 if (list_empty(&ring->gpu_write_list) && list_empty(&ring->active_list))
2324 if (!list_empty(&ring->gpu_write_list)) {
2325 ret = i915_gem_flush_ring(ring,
2326 I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
2331 return i915_wait_seqno(ring, i915_gem_next_request_seqno(ring));
2334 int i915_gpu_idle(struct drm_device *dev)
2336 drm_i915_private_t *dev_priv = dev->dev_private;
2337 struct intel_ring_buffer *ring;
2340 /* Flush everything onto the inactive list. */
2341 for_each_ring(ring, dev_priv, i) {
2342 ret = i915_ring_idle(ring);
2346 /* Is the device fubar? */
2347 if (WARN_ON(!list_empty(&ring->gpu_write_list)))
2350 ret = i915_switch_context(ring, NULL, DEFAULT_CONTEXT_ID);
2358 static void sandybridge_write_fence_reg(struct drm_device *dev, int reg,
2359 struct drm_i915_gem_object *obj)
2361 drm_i915_private_t *dev_priv = dev->dev_private;
2365 u32 size = obj->gtt_space->size;
2367 val = (uint64_t)((obj->gtt_offset + size - 4096) &
2369 val |= obj->gtt_offset & 0xfffff000;
2370 val |= (uint64_t)((obj->stride / 128) - 1) <<
2371 SANDYBRIDGE_FENCE_PITCH_SHIFT;
2373 if (obj->tiling_mode == I915_TILING_Y)
2374 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2375 val |= I965_FENCE_REG_VALID;
2379 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + reg * 8, val);
2380 POSTING_READ(FENCE_REG_SANDYBRIDGE_0 + reg * 8);
2383 static void i965_write_fence_reg(struct drm_device *dev, int reg,
2384 struct drm_i915_gem_object *obj)
2386 drm_i915_private_t *dev_priv = dev->dev_private;
2390 u32 size = obj->gtt_space->size;
2392 val = (uint64_t)((obj->gtt_offset + size - 4096) &
2394 val |= obj->gtt_offset & 0xfffff000;
2395 val |= ((obj->stride / 128) - 1) << I965_FENCE_PITCH_SHIFT;
2396 if (obj->tiling_mode == I915_TILING_Y)
2397 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2398 val |= I965_FENCE_REG_VALID;
2402 I915_WRITE64(FENCE_REG_965_0 + reg * 8, val);
2403 POSTING_READ(FENCE_REG_965_0 + reg * 8);
2406 static void i915_write_fence_reg(struct drm_device *dev, int reg,
2407 struct drm_i915_gem_object *obj)
2409 drm_i915_private_t *dev_priv = dev->dev_private;
2413 u32 size = obj->gtt_space->size;
2417 WARN((obj->gtt_offset & ~I915_FENCE_START_MASK) ||
2418 (size & -size) != size ||
2419 (obj->gtt_offset & (size - 1)),
2420 "object 0x%08x [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
2421 obj->gtt_offset, obj->map_and_fenceable, size);
2423 if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
2428 /* Note: pitch better be a power of two tile widths */
2429 pitch_val = obj->stride / tile_width;
2430 pitch_val = ffs(pitch_val) - 1;
2432 val = obj->gtt_offset;
2433 if (obj->tiling_mode == I915_TILING_Y)
2434 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2435 val |= I915_FENCE_SIZE_BITS(size);
2436 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2437 val |= I830_FENCE_REG_VALID;
2442 reg = FENCE_REG_830_0 + reg * 4;
2444 reg = FENCE_REG_945_8 + (reg - 8) * 4;
2446 I915_WRITE(reg, val);
2450 static void i830_write_fence_reg(struct drm_device *dev, int reg,
2451 struct drm_i915_gem_object *obj)
2453 drm_i915_private_t *dev_priv = dev->dev_private;
2457 u32 size = obj->gtt_space->size;
2460 WARN((obj->gtt_offset & ~I830_FENCE_START_MASK) ||
2461 (size & -size) != size ||
2462 (obj->gtt_offset & (size - 1)),
2463 "object 0x%08x not 512K or pot-size 0x%08x aligned\n",
2464 obj->gtt_offset, size);
2466 pitch_val = obj->stride / 128;
2467 pitch_val = ffs(pitch_val) - 1;
2469 val = obj->gtt_offset;
2470 if (obj->tiling_mode == I915_TILING_Y)
2471 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2472 val |= I830_FENCE_SIZE_BITS(size);
2473 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2474 val |= I830_FENCE_REG_VALID;
2478 I915_WRITE(FENCE_REG_830_0 + reg * 4, val);
2479 POSTING_READ(FENCE_REG_830_0 + reg * 4);
2482 static void i915_gem_write_fence(struct drm_device *dev, int reg,
2483 struct drm_i915_gem_object *obj)
2485 switch (INTEL_INFO(dev)->gen) {
2487 case 6: sandybridge_write_fence_reg(dev, reg, obj); break;
2489 case 4: i965_write_fence_reg(dev, reg, obj); break;
2490 case 3: i915_write_fence_reg(dev, reg, obj); break;
2491 case 2: i830_write_fence_reg(dev, reg, obj); break;
2496 static inline int fence_number(struct drm_i915_private *dev_priv,
2497 struct drm_i915_fence_reg *fence)
2499 return fence - dev_priv->fence_regs;
2502 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
2503 struct drm_i915_fence_reg *fence,
2506 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2507 int reg = fence_number(dev_priv, fence);
2509 i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL);
2512 obj->fence_reg = reg;
2514 list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list);
2516 obj->fence_reg = I915_FENCE_REG_NONE;
2518 list_del_init(&fence->lru_list);
2523 i915_gem_object_flush_fence(struct drm_i915_gem_object *obj)
2527 if (obj->fenced_gpu_access) {
2528 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
2529 ret = i915_gem_flush_ring(obj->ring,
2530 0, obj->base.write_domain);
2535 obj->fenced_gpu_access = false;
2538 if (obj->last_fenced_seqno) {
2539 ret = i915_wait_seqno(obj->ring, obj->last_fenced_seqno);
2543 obj->last_fenced_seqno = 0;
2546 /* Ensure that all CPU reads are completed before installing a fence
2547 * and all writes before removing the fence.
2549 if (obj->base.read_domains & I915_GEM_DOMAIN_GTT)
2556 i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
2558 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2561 ret = i915_gem_object_flush_fence(obj);
2565 if (obj->fence_reg == I915_FENCE_REG_NONE)
2568 i915_gem_object_update_fence(obj,
2569 &dev_priv->fence_regs[obj->fence_reg],
2571 i915_gem_object_fence_lost(obj);
2576 static struct drm_i915_fence_reg *
2577 i915_find_fence_reg(struct drm_device *dev)
2579 struct drm_i915_private *dev_priv = dev->dev_private;
2580 struct drm_i915_fence_reg *reg, *avail;
2583 /* First try to find a free reg */
2585 for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2586 reg = &dev_priv->fence_regs[i];
2590 if (!reg->pin_count)
2597 /* None available, try to steal one or wait for a user to finish */
2598 list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
2609 * i915_gem_object_get_fence - set up fencing for an object
2610 * @obj: object to map through a fence reg
2612 * When mapping objects through the GTT, userspace wants to be able to write
2613 * to them without having to worry about swizzling if the object is tiled.
2614 * This function walks the fence regs looking for a free one for @obj,
2615 * stealing one if it can't find any.
2617 * It then sets up the reg based on the object's properties: address, pitch
2618 * and tiling format.
2620 * For an untiled surface, this removes any existing fence.
2623 i915_gem_object_get_fence(struct drm_i915_gem_object *obj)
2625 struct drm_device *dev = obj->base.dev;
2626 struct drm_i915_private *dev_priv = dev->dev_private;
2627 bool enable = obj->tiling_mode != I915_TILING_NONE;
2628 struct drm_i915_fence_reg *reg;
2631 /* Have we updated the tiling parameters upon the object and so
2632 * will need to serialise the write to the associated fence register?
2634 if (obj->fence_dirty) {
2635 ret = i915_gem_object_flush_fence(obj);
2640 /* Just update our place in the LRU if our fence is getting reused. */
2641 if (obj->fence_reg != I915_FENCE_REG_NONE) {
2642 reg = &dev_priv->fence_regs[obj->fence_reg];
2643 if (!obj->fence_dirty) {
2644 list_move_tail(®->lru_list,
2645 &dev_priv->mm.fence_list);
2648 } else if (enable) {
2649 reg = i915_find_fence_reg(dev);
2654 struct drm_i915_gem_object *old = reg->obj;
2656 ret = i915_gem_object_flush_fence(old);
2660 i915_gem_object_fence_lost(old);
2665 i915_gem_object_update_fence(obj, reg, enable);
2666 obj->fence_dirty = false;
2672 * Finds free space in the GTT aperture and binds the object there.
2675 i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
2677 bool map_and_fenceable)
2679 struct drm_device *dev = obj->base.dev;
2680 drm_i915_private_t *dev_priv = dev->dev_private;
2681 struct drm_mm_node *free_space;
2682 gfp_t gfpmask = __GFP_NORETRY | __GFP_NOWARN;
2683 u32 size, fence_size, fence_alignment, unfenced_alignment;
2684 bool mappable, fenceable;
2687 if (obj->madv != I915_MADV_WILLNEED) {
2688 DRM_ERROR("Attempting to bind a purgeable object\n");
2692 fence_size = i915_gem_get_gtt_size(dev,
2695 fence_alignment = i915_gem_get_gtt_alignment(dev,
2698 unfenced_alignment =
2699 i915_gem_get_unfenced_gtt_alignment(dev,
2704 alignment = map_and_fenceable ? fence_alignment :
2706 if (map_and_fenceable && alignment & (fence_alignment - 1)) {
2707 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
2711 size = map_and_fenceable ? fence_size : obj->base.size;
2713 /* If the object is bigger than the entire aperture, reject it early
2714 * before evicting everything in a vain attempt to find space.
2716 if (obj->base.size >
2717 (map_and_fenceable ? dev_priv->mm.gtt_mappable_end : dev_priv->mm.gtt_total)) {
2718 DRM_ERROR("Attempting to bind an object larger than the aperture\n");
2723 if (map_and_fenceable)
2725 drm_mm_search_free_in_range(&dev_priv->mm.gtt_space,
2727 0, dev_priv->mm.gtt_mappable_end,
2730 free_space = drm_mm_search_free(&dev_priv->mm.gtt_space,
2731 size, alignment, 0);
2733 if (free_space != NULL) {
2734 if (map_and_fenceable)
2736 drm_mm_get_block_range_generic(free_space,
2738 0, dev_priv->mm.gtt_mappable_end,
2742 drm_mm_get_block(free_space, size, alignment);
2744 if (obj->gtt_space == NULL) {
2745 /* If the gtt is empty and we're still having trouble
2746 * fitting our object in, we're out of memory.
2748 ret = i915_gem_evict_something(dev, size, alignment,
2756 ret = i915_gem_object_get_pages_gtt(obj, gfpmask);
2758 drm_mm_put_block(obj->gtt_space);
2759 obj->gtt_space = NULL;
2761 if (ret == -ENOMEM) {
2762 /* first try to reclaim some memory by clearing the GTT */
2763 ret = i915_gem_evict_everything(dev, false);
2765 /* now try to shrink everyone else */
2780 ret = i915_gem_gtt_prepare_object(obj);
2782 i915_gem_object_put_pages_gtt(obj);
2783 drm_mm_put_block(obj->gtt_space);
2784 obj->gtt_space = NULL;
2786 if (i915_gem_evict_everything(dev, false))
2792 if (!dev_priv->mm.aliasing_ppgtt)
2793 i915_gem_gtt_bind_object(obj, obj->cache_level);
2795 list_add_tail(&obj->gtt_list, &dev_priv->mm.gtt_list);
2796 list_add_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
2798 /* Assert that the object is not currently in any GPU domain. As it
2799 * wasn't in the GTT, there shouldn't be any way it could have been in
2802 BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
2803 BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
2805 obj->gtt_offset = obj->gtt_space->start;
2808 obj->gtt_space->size == fence_size &&
2809 (obj->gtt_space->start & (fence_alignment - 1)) == 0;
2812 obj->gtt_offset + obj->base.size <= dev_priv->mm.gtt_mappable_end;
2814 obj->map_and_fenceable = mappable && fenceable;
2816 trace_i915_gem_object_bind(obj, map_and_fenceable);
2821 i915_gem_clflush_object(struct drm_i915_gem_object *obj)
2823 /* If we don't have a page list set up, then we're not pinned
2824 * to GPU, and we can ignore the cache flush because it'll happen
2825 * again at bind time.
2827 if (obj->pages == NULL)
2830 /* If the GPU is snooping the contents of the CPU cache,
2831 * we do not need to manually clear the CPU cache lines. However,
2832 * the caches are only snooped when the render cache is
2833 * flushed/invalidated. As we always have to emit invalidations
2834 * and flushes when moving into and out of the RENDER domain, correct
2835 * snooping behaviour occurs naturally as the result of our domain
2838 if (obj->cache_level != I915_CACHE_NONE)
2841 trace_i915_gem_object_clflush(obj);
2843 drm_clflush_pages(obj->pages, obj->base.size / PAGE_SIZE);
2846 /** Flushes any GPU write domain for the object if it's dirty. */
2848 i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object *obj)
2850 if ((obj->base.write_domain & I915_GEM_GPU_DOMAINS) == 0)
2853 /* Queue the GPU write cache flushing we need. */
2854 return i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain);
2857 /** Flushes the GTT write domain for the object if it's dirty. */
2859 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
2861 uint32_t old_write_domain;
2863 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
2866 /* No actual flushing is required for the GTT write domain. Writes
2867 * to it immediately go to main memory as far as we know, so there's
2868 * no chipset flush. It also doesn't land in render cache.
2870 * However, we do have to enforce the order so that all writes through
2871 * the GTT land before any writes to the device, such as updates to
2876 old_write_domain = obj->base.write_domain;
2877 obj->base.write_domain = 0;
2879 trace_i915_gem_object_change_domain(obj,
2880 obj->base.read_domains,
2884 /** Flushes the CPU write domain for the object if it's dirty. */
2886 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj)
2888 uint32_t old_write_domain;
2890 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
2893 i915_gem_clflush_object(obj);
2894 intel_gtt_chipset_flush();
2895 old_write_domain = obj->base.write_domain;
2896 obj->base.write_domain = 0;
2898 trace_i915_gem_object_change_domain(obj,
2899 obj->base.read_domains,
2904 * Moves a single object to the GTT read, and possibly write domain.
2906 * This function returns when the move is complete, including waiting on
2910 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
2912 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
2913 uint32_t old_write_domain, old_read_domains;
2916 /* Not valid to be called on unbound objects. */
2917 if (obj->gtt_space == NULL)
2920 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
2923 ret = i915_gem_object_flush_gpu_write_domain(obj);
2927 ret = i915_gem_object_wait_rendering(obj, !write);
2931 i915_gem_object_flush_cpu_write_domain(obj);
2933 old_write_domain = obj->base.write_domain;
2934 old_read_domains = obj->base.read_domains;
2936 /* It should now be out of any other write domains, and we can update
2937 * the domain values for our changes.
2939 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2940 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
2942 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
2943 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
2947 trace_i915_gem_object_change_domain(obj,
2951 /* And bump the LRU for this access */
2952 if (i915_gem_object_is_inactive(obj))
2953 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
2958 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
2959 enum i915_cache_level cache_level)
2961 struct drm_device *dev = obj->base.dev;
2962 drm_i915_private_t *dev_priv = dev->dev_private;
2965 if (obj->cache_level == cache_level)
2968 if (obj->pin_count) {
2969 DRM_DEBUG("can not change the cache level of pinned objects\n");
2973 if (obj->gtt_space) {
2974 ret = i915_gem_object_finish_gpu(obj);
2978 i915_gem_object_finish_gtt(obj);
2980 /* Before SandyBridge, you could not use tiling or fence
2981 * registers with snooped memory, so relinquish any fences
2982 * currently pointing to our region in the aperture.
2984 if (INTEL_INFO(obj->base.dev)->gen < 6) {
2985 ret = i915_gem_object_put_fence(obj);
2990 if (obj->has_global_gtt_mapping)
2991 i915_gem_gtt_bind_object(obj, cache_level);
2992 if (obj->has_aliasing_ppgtt_mapping)
2993 i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt,
2997 if (cache_level == I915_CACHE_NONE) {
2998 u32 old_read_domains, old_write_domain;
3000 /* If we're coming from LLC cached, then we haven't
3001 * actually been tracking whether the data is in the
3002 * CPU cache or not, since we only allow one bit set
3003 * in obj->write_domain and have been skipping the clflushes.
3004 * Just set it to the CPU cache for now.
3006 WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU);
3007 WARN_ON(obj->base.read_domains & ~I915_GEM_DOMAIN_CPU);
3009 old_read_domains = obj->base.read_domains;
3010 old_write_domain = obj->base.write_domain;
3012 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3013 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3015 trace_i915_gem_object_change_domain(obj,
3020 obj->cache_level = cache_level;
3025 * Prepare buffer for display plane (scanout, cursors, etc).
3026 * Can be called from an uninterruptible phase (modesetting) and allows
3027 * any flushes to be pipelined (for pageflips).
3030 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
3032 struct intel_ring_buffer *pipelined)
3034 u32 old_read_domains, old_write_domain;
3037 ret = i915_gem_object_flush_gpu_write_domain(obj);
3041 if (pipelined != obj->ring) {
3042 ret = i915_gem_object_sync(obj, pipelined);
3047 /* The display engine is not coherent with the LLC cache on gen6. As
3048 * a result, we make sure that the pinning that is about to occur is
3049 * done with uncached PTEs. This is lowest common denominator for all
3052 * However for gen6+, we could do better by using the GFDT bit instead
3053 * of uncaching, which would allow us to flush all the LLC-cached data
3054 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3056 ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE);
3060 /* As the user may map the buffer once pinned in the display plane
3061 * (e.g. libkms for the bootup splash), we have to ensure that we
3062 * always use map_and_fenceable for all scanout buffers.
3064 ret = i915_gem_object_pin(obj, alignment, true);
3068 i915_gem_object_flush_cpu_write_domain(obj);
3070 old_write_domain = obj->base.write_domain;
3071 old_read_domains = obj->base.read_domains;
3073 /* It should now be out of any other write domains, and we can update
3074 * the domain values for our changes.
3076 obj->base.write_domain = 0;
3077 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3079 trace_i915_gem_object_change_domain(obj,
3087 i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
3091 if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
3094 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
3095 ret = i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain);
3100 ret = i915_gem_object_wait_rendering(obj, false);
3104 /* Ensure that we invalidate the GPU's caches and TLBs. */
3105 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
3110 * Moves a single object to the CPU read, and possibly write domain.
3112 * This function returns when the move is complete, including waiting on
3116 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
3118 uint32_t old_write_domain, old_read_domains;
3121 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
3124 ret = i915_gem_object_flush_gpu_write_domain(obj);
3128 ret = i915_gem_object_wait_rendering(obj, !write);
3132 i915_gem_object_flush_gtt_write_domain(obj);
3134 old_write_domain = obj->base.write_domain;
3135 old_read_domains = obj->base.read_domains;
3137 /* Flush the CPU cache if it's still invalid. */
3138 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
3139 i915_gem_clflush_object(obj);
3141 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3144 /* It should now be out of any other write domains, and we can update
3145 * the domain values for our changes.
3147 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3149 /* If we're writing through the CPU, then the GPU read domains will
3150 * need to be invalidated at next use.
3153 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3154 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3157 trace_i915_gem_object_change_domain(obj,
3164 /* Throttle our rendering by waiting until the ring has completed our requests
3165 * emitted over 20 msec ago.
3167 * Note that if we were to use the current jiffies each time around the loop,
3168 * we wouldn't escape the function with any frames outstanding if the time to
3169 * render a frame was over 20ms.
3171 * This should get us reasonable parallelism between CPU and GPU but also
3172 * relatively low latency when blocking on a particular request to finish.
3175 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
3177 struct drm_i915_private *dev_priv = dev->dev_private;
3178 struct drm_i915_file_private *file_priv = file->driver_priv;
3179 unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3180 struct drm_i915_gem_request *request;
3181 struct intel_ring_buffer *ring = NULL;
3185 if (atomic_read(&dev_priv->mm.wedged))
3188 spin_lock(&file_priv->mm.lock);
3189 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
3190 if (time_after_eq(request->emitted_jiffies, recent_enough))
3193 ring = request->ring;
3194 seqno = request->seqno;
3196 spin_unlock(&file_priv->mm.lock);
3201 ret = __wait_seqno(ring, seqno, true, NULL);
3203 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
3209 i915_gem_object_pin(struct drm_i915_gem_object *obj,
3211 bool map_and_fenceable)
3215 BUG_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT);
3217 if (obj->gtt_space != NULL) {
3218 if ((alignment && obj->gtt_offset & (alignment - 1)) ||
3219 (map_and_fenceable && !obj->map_and_fenceable)) {
3220 WARN(obj->pin_count,
3221 "bo is already pinned with incorrect alignment:"
3222 " offset=%x, req.alignment=%x, req.map_and_fenceable=%d,"
3223 " obj->map_and_fenceable=%d\n",
3224 obj->gtt_offset, alignment,
3226 obj->map_and_fenceable);
3227 ret = i915_gem_object_unbind(obj);
3233 if (obj->gtt_space == NULL) {
3234 ret = i915_gem_object_bind_to_gtt(obj, alignment,
3240 if (!obj->has_global_gtt_mapping && map_and_fenceable)
3241 i915_gem_gtt_bind_object(obj, obj->cache_level);
3244 obj->pin_mappable |= map_and_fenceable;
3250 i915_gem_object_unpin(struct drm_i915_gem_object *obj)
3252 BUG_ON(obj->pin_count == 0);
3253 BUG_ON(obj->gtt_space == NULL);
3255 if (--obj->pin_count == 0)
3256 obj->pin_mappable = false;
3260 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
3261 struct drm_file *file)
3263 struct drm_i915_gem_pin *args = data;
3264 struct drm_i915_gem_object *obj;
3267 ret = i915_mutex_lock_interruptible(dev);
3271 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3272 if (&obj->base == NULL) {
3277 if (obj->madv != I915_MADV_WILLNEED) {
3278 DRM_ERROR("Attempting to pin a purgeable buffer\n");
3283 if (obj->pin_filp != NULL && obj->pin_filp != file) {
3284 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
3290 obj->user_pin_count++;
3291 obj->pin_filp = file;
3292 if (obj->user_pin_count == 1) {
3293 ret = i915_gem_object_pin(obj, args->alignment, true);
3298 /* XXX - flush the CPU caches for pinned objects
3299 * as the X server doesn't manage domains yet
3301 i915_gem_object_flush_cpu_write_domain(obj);
3302 args->offset = obj->gtt_offset;
3304 drm_gem_object_unreference(&obj->base);
3306 mutex_unlock(&dev->struct_mutex);
3311 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
3312 struct drm_file *file)
3314 struct drm_i915_gem_pin *args = data;
3315 struct drm_i915_gem_object *obj;
3318 ret = i915_mutex_lock_interruptible(dev);
3322 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3323 if (&obj->base == NULL) {
3328 if (obj->pin_filp != file) {
3329 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
3334 obj->user_pin_count--;
3335 if (obj->user_pin_count == 0) {
3336 obj->pin_filp = NULL;
3337 i915_gem_object_unpin(obj);
3341 drm_gem_object_unreference(&obj->base);
3343 mutex_unlock(&dev->struct_mutex);
3348 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
3349 struct drm_file *file)
3351 struct drm_i915_gem_busy *args = data;
3352 struct drm_i915_gem_object *obj;
3355 ret = i915_mutex_lock_interruptible(dev);
3359 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3360 if (&obj->base == NULL) {
3365 /* Count all active objects as busy, even if they are currently not used
3366 * by the gpu. Users of this interface expect objects to eventually
3367 * become non-busy without any further actions, therefore emit any
3368 * necessary flushes here.
3370 ret = i915_gem_object_flush_active(obj);
3372 args->busy = obj->active;
3374 BUILD_BUG_ON(I915_NUM_RINGS > 16);
3375 args->busy |= intel_ring_flag(obj->ring) << 16;
3378 drm_gem_object_unreference(&obj->base);
3380 mutex_unlock(&dev->struct_mutex);
3385 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
3386 struct drm_file *file_priv)
3388 return i915_gem_ring_throttle(dev, file_priv);
3392 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
3393 struct drm_file *file_priv)
3395 struct drm_i915_gem_madvise *args = data;
3396 struct drm_i915_gem_object *obj;
3399 switch (args->madv) {
3400 case I915_MADV_DONTNEED:
3401 case I915_MADV_WILLNEED:
3407 ret = i915_mutex_lock_interruptible(dev);
3411 obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
3412 if (&obj->base == NULL) {
3417 if (obj->pin_count) {
3422 if (obj->madv != __I915_MADV_PURGED)
3423 obj->madv = args->madv;
3425 /* if the object is no longer bound, discard its backing storage */
3426 if (i915_gem_object_is_purgeable(obj) &&
3427 obj->gtt_space == NULL)
3428 i915_gem_object_truncate(obj);
3430 args->retained = obj->madv != __I915_MADV_PURGED;
3433 drm_gem_object_unreference(&obj->base);
3435 mutex_unlock(&dev->struct_mutex);
3439 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
3442 struct drm_i915_private *dev_priv = dev->dev_private;
3443 struct drm_i915_gem_object *obj;
3444 struct address_space *mapping;
3447 obj = kzalloc(sizeof(*obj), GFP_KERNEL);
3451 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
3456 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
3457 if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) {
3458 /* 965gm cannot relocate objects above 4GiB. */
3459 mask &= ~__GFP_HIGHMEM;
3460 mask |= __GFP_DMA32;
3463 mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
3464 mapping_set_gfp_mask(mapping, mask);
3466 i915_gem_info_add_obj(dev_priv, size);
3468 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3469 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3472 /* On some devices, we can have the GPU use the LLC (the CPU
3473 * cache) for about a 10% performance improvement
3474 * compared to uncached. Graphics requests other than
3475 * display scanout are coherent with the CPU in
3476 * accessing this cache. This means in this mode we
3477 * don't need to clflush on the CPU side, and on the
3478 * GPU side we only need to flush internal caches to
3479 * get data visible to the CPU.
3481 * However, we maintain the display planes as UC, and so
3482 * need to rebind when first used as such.
3484 obj->cache_level = I915_CACHE_LLC;
3486 obj->cache_level = I915_CACHE_NONE;
3488 obj->base.driver_private = NULL;
3489 obj->fence_reg = I915_FENCE_REG_NONE;
3490 INIT_LIST_HEAD(&obj->mm_list);
3491 INIT_LIST_HEAD(&obj->gtt_list);
3492 INIT_LIST_HEAD(&obj->ring_list);
3493 INIT_LIST_HEAD(&obj->exec_list);
3494 INIT_LIST_HEAD(&obj->gpu_write_list);
3495 obj->madv = I915_MADV_WILLNEED;
3496 /* Avoid an unnecessary call to unbind on the first bind. */
3497 obj->map_and_fenceable = true;
3502 int i915_gem_init_object(struct drm_gem_object *obj)
3509 void i915_gem_free_object(struct drm_gem_object *gem_obj)
3511 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
3512 struct drm_device *dev = obj->base.dev;
3513 drm_i915_private_t *dev_priv = dev->dev_private;
3515 trace_i915_gem_object_destroy(obj);
3517 if (gem_obj->import_attach)
3518 drm_prime_gem_destroy(gem_obj, obj->sg_table);
3521 i915_gem_detach_phys_object(dev, obj);
3524 if (WARN_ON(i915_gem_object_unbind(obj) == -ERESTARTSYS)) {
3525 bool was_interruptible;
3527 was_interruptible = dev_priv->mm.interruptible;
3528 dev_priv->mm.interruptible = false;
3530 WARN_ON(i915_gem_object_unbind(obj));
3532 dev_priv->mm.interruptible = was_interruptible;
3535 if (obj->base.map_list.map)
3536 drm_gem_free_mmap_offset(&obj->base);
3538 drm_gem_object_release(&obj->base);
3539 i915_gem_info_remove_obj(dev_priv, obj->base.size);
3546 i915_gem_idle(struct drm_device *dev)
3548 drm_i915_private_t *dev_priv = dev->dev_private;
3551 mutex_lock(&dev->struct_mutex);
3553 if (dev_priv->mm.suspended) {
3554 mutex_unlock(&dev->struct_mutex);
3558 ret = i915_gpu_idle(dev);
3560 mutex_unlock(&dev->struct_mutex);
3563 i915_gem_retire_requests(dev);
3565 /* Under UMS, be paranoid and evict. */
3566 if (!drm_core_check_feature(dev, DRIVER_MODESET))
3567 i915_gem_evict_everything(dev, false);
3569 i915_gem_reset_fences(dev);
3571 /* Hack! Don't let anybody do execbuf while we don't control the chip.
3572 * We need to replace this with a semaphore, or something.
3573 * And not confound mm.suspended!
3575 dev_priv->mm.suspended = 1;
3576 del_timer_sync(&dev_priv->hangcheck_timer);
3578 i915_kernel_lost_context(dev);
3579 i915_gem_cleanup_ringbuffer(dev);
3581 mutex_unlock(&dev->struct_mutex);
3583 /* Cancel the retire work handler, which should be idle now. */
3584 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
3589 void i915_gem_l3_remap(struct drm_device *dev)
3591 drm_i915_private_t *dev_priv = dev->dev_private;
3595 if (!IS_IVYBRIDGE(dev))
3598 if (!dev_priv->mm.l3_remap_info)
3601 misccpctl = I915_READ(GEN7_MISCCPCTL);
3602 I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
3603 POSTING_READ(GEN7_MISCCPCTL);
3605 for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) {
3606 u32 remap = I915_READ(GEN7_L3LOG_BASE + i);
3607 if (remap && remap != dev_priv->mm.l3_remap_info[i/4])
3608 DRM_DEBUG("0x%x was already programmed to %x\n",
3609 GEN7_L3LOG_BASE + i, remap);
3610 if (remap && !dev_priv->mm.l3_remap_info[i/4])
3611 DRM_DEBUG_DRIVER("Clearing remapped register\n");
3612 I915_WRITE(GEN7_L3LOG_BASE + i, dev_priv->mm.l3_remap_info[i/4]);
3615 /* Make sure all the writes land before disabling dop clock gating */
3616 POSTING_READ(GEN7_L3LOG_BASE);
3618 I915_WRITE(GEN7_MISCCPCTL, misccpctl);
3621 void i915_gem_init_swizzling(struct drm_device *dev)
3623 drm_i915_private_t *dev_priv = dev->dev_private;
3625 if (INTEL_INFO(dev)->gen < 5 ||
3626 dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
3629 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
3630 DISP_TILE_SURFACE_SWIZZLING);
3635 I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
3637 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
3639 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
3642 void i915_gem_init_ppgtt(struct drm_device *dev)
3644 drm_i915_private_t *dev_priv = dev->dev_private;
3646 struct intel_ring_buffer *ring;
3647 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
3648 uint32_t __iomem *pd_addr;
3652 if (!dev_priv->mm.aliasing_ppgtt)
3656 pd_addr = dev_priv->mm.gtt->gtt + ppgtt->pd_offset/sizeof(uint32_t);
3657 for (i = 0; i < ppgtt->num_pd_entries; i++) {
3660 if (dev_priv->mm.gtt->needs_dmar)
3661 pt_addr = ppgtt->pt_dma_addr[i];
3663 pt_addr = page_to_phys(ppgtt->pt_pages[i]);
3665 pd_entry = GEN6_PDE_ADDR_ENCODE(pt_addr);
3666 pd_entry |= GEN6_PDE_VALID;
3668 writel(pd_entry, pd_addr + i);
3672 pd_offset = ppgtt->pd_offset;
3673 pd_offset /= 64; /* in cachelines, */
3676 if (INTEL_INFO(dev)->gen == 6) {
3677 uint32_t ecochk, gab_ctl, ecobits;
3679 ecobits = I915_READ(GAC_ECO_BITS);
3680 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
3682 gab_ctl = I915_READ(GAB_CTL);
3683 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
3685 ecochk = I915_READ(GAM_ECOCHK);
3686 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT |
3687 ECOCHK_PPGTT_CACHE64B);
3688 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
3689 } else if (INTEL_INFO(dev)->gen >= 7) {
3690 I915_WRITE(GAM_ECOCHK, ECOCHK_PPGTT_CACHE64B);
3691 /* GFX_MODE is per-ring on gen7+ */
3694 for_each_ring(ring, dev_priv, i) {
3695 if (INTEL_INFO(dev)->gen >= 7)
3696 I915_WRITE(RING_MODE_GEN7(ring),
3697 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
3699 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
3700 I915_WRITE(RING_PP_DIR_BASE(ring), pd_offset);
3705 intel_enable_blt(struct drm_device *dev)
3710 /* The blitter was dysfunctional on early prototypes */
3711 if (IS_GEN6(dev) && dev->pdev->revision < 8) {
3712 DRM_INFO("BLT not supported on this pre-production hardware;"
3713 " graphics performance will be degraded.\n");
3721 i915_gem_init_hw(struct drm_device *dev)
3723 drm_i915_private_t *dev_priv = dev->dev_private;
3726 if (!intel_enable_gtt())
3729 i915_gem_l3_remap(dev);
3731 i915_gem_init_swizzling(dev);
3733 ret = intel_init_render_ring_buffer(dev);
3738 ret = intel_init_bsd_ring_buffer(dev);
3740 goto cleanup_render_ring;
3743 if (intel_enable_blt(dev)) {
3744 ret = intel_init_blt_ring_buffer(dev);
3746 goto cleanup_bsd_ring;
3749 dev_priv->next_seqno = 1;
3752 * XXX: There was some w/a described somewhere suggesting loading
3753 * contexts before PPGTT.
3755 i915_gem_context_init(dev);
3756 i915_gem_init_ppgtt(dev);
3761 intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
3762 cleanup_render_ring:
3763 intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
3768 intel_enable_ppgtt(struct drm_device *dev)
3770 if (i915_enable_ppgtt >= 0)
3771 return i915_enable_ppgtt;
3773 #ifdef CONFIG_INTEL_IOMMU
3774 /* Disable ppgtt on SNB if VT-d is on. */
3775 if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped)
3782 int i915_gem_init(struct drm_device *dev)
3784 struct drm_i915_private *dev_priv = dev->dev_private;
3785 unsigned long gtt_size, mappable_size;
3788 gtt_size = dev_priv->mm.gtt->gtt_total_entries << PAGE_SHIFT;
3789 mappable_size = dev_priv->mm.gtt->gtt_mappable_entries << PAGE_SHIFT;
3791 mutex_lock(&dev->struct_mutex);
3792 if (intel_enable_ppgtt(dev) && HAS_ALIASING_PPGTT(dev)) {
3793 /* PPGTT pdes are stolen from global gtt ptes, so shrink the
3794 * aperture accordingly when using aliasing ppgtt. */
3795 gtt_size -= I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
3797 i915_gem_init_global_gtt(dev, 0, mappable_size, gtt_size);
3799 ret = i915_gem_init_aliasing_ppgtt(dev);
3801 mutex_unlock(&dev->struct_mutex);
3805 /* Let GEM Manage all of the aperture.
3807 * However, leave one page at the end still bound to the scratch
3808 * page. There are a number of places where the hardware
3809 * apparently prefetches past the end of the object, and we've
3810 * seen multiple hangs with the GPU head pointer stuck in a
3811 * batchbuffer bound at the last page of the aperture. One page
3812 * should be enough to keep any prefetching inside of the
3815 i915_gem_init_global_gtt(dev, 0, mappable_size,
3819 ret = i915_gem_init_hw(dev);
3820 mutex_unlock(&dev->struct_mutex);
3822 i915_gem_cleanup_aliasing_ppgtt(dev);
3826 /* Allow hardware batchbuffers unless told otherwise, but not for KMS. */
3827 if (!drm_core_check_feature(dev, DRIVER_MODESET))
3828 dev_priv->dri1.allow_batchbuffer = 1;
3833 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
3835 drm_i915_private_t *dev_priv = dev->dev_private;
3836 struct intel_ring_buffer *ring;
3839 for_each_ring(ring, dev_priv, i)
3840 intel_cleanup_ring_buffer(ring);
3844 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
3845 struct drm_file *file_priv)
3847 drm_i915_private_t *dev_priv = dev->dev_private;
3850 if (drm_core_check_feature(dev, DRIVER_MODESET))
3853 if (atomic_read(&dev_priv->mm.wedged)) {
3854 DRM_ERROR("Reenabling wedged hardware, good luck\n");
3855 atomic_set(&dev_priv->mm.wedged, 0);
3858 mutex_lock(&dev->struct_mutex);
3859 dev_priv->mm.suspended = 0;
3861 ret = i915_gem_init_hw(dev);
3863 mutex_unlock(&dev->struct_mutex);
3867 BUG_ON(!list_empty(&dev_priv->mm.active_list));
3868 BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
3869 mutex_unlock(&dev->struct_mutex);
3871 ret = drm_irq_install(dev);
3873 goto cleanup_ringbuffer;
3878 mutex_lock(&dev->struct_mutex);
3879 i915_gem_cleanup_ringbuffer(dev);
3880 dev_priv->mm.suspended = 1;
3881 mutex_unlock(&dev->struct_mutex);
3887 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
3888 struct drm_file *file_priv)
3890 if (drm_core_check_feature(dev, DRIVER_MODESET))
3893 drm_irq_uninstall(dev);
3894 return i915_gem_idle(dev);
3898 i915_gem_lastclose(struct drm_device *dev)
3902 if (drm_core_check_feature(dev, DRIVER_MODESET))
3905 ret = i915_gem_idle(dev);
3907 DRM_ERROR("failed to idle hardware: %d\n", ret);
3911 init_ring_lists(struct intel_ring_buffer *ring)
3913 INIT_LIST_HEAD(&ring->active_list);
3914 INIT_LIST_HEAD(&ring->request_list);
3915 INIT_LIST_HEAD(&ring->gpu_write_list);
3919 i915_gem_load(struct drm_device *dev)
3922 drm_i915_private_t *dev_priv = dev->dev_private;
3924 INIT_LIST_HEAD(&dev_priv->mm.active_list);
3925 INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
3926 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
3927 INIT_LIST_HEAD(&dev_priv->mm.gtt_list);
3928 for (i = 0; i < I915_NUM_RINGS; i++)
3929 init_ring_lists(&dev_priv->ring[i]);
3930 for (i = 0; i < I915_MAX_NUM_FENCES; i++)
3931 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
3932 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
3933 i915_gem_retire_work_handler);
3934 init_completion(&dev_priv->error_completion);
3936 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
3938 I915_WRITE(MI_ARB_STATE,
3939 _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
3942 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
3944 /* Old X drivers will take 0-2 for front, back, depth buffers */
3945 if (!drm_core_check_feature(dev, DRIVER_MODESET))
3946 dev_priv->fence_reg_start = 3;
3948 if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
3949 dev_priv->num_fence_regs = 16;
3951 dev_priv->num_fence_regs = 8;
3953 /* Initialize fence registers to zero */
3954 i915_gem_reset_fences(dev);
3956 i915_gem_detect_bit_6_swizzle(dev);
3957 init_waitqueue_head(&dev_priv->pending_flip_queue);
3959 dev_priv->mm.interruptible = true;
3961 dev_priv->mm.inactive_shrinker.shrink = i915_gem_inactive_shrink;
3962 dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS;
3963 register_shrinker(&dev_priv->mm.inactive_shrinker);
3967 * Create a physically contiguous memory object for this object
3968 * e.g. for cursor + overlay regs
3970 static int i915_gem_init_phys_object(struct drm_device *dev,
3971 int id, int size, int align)
3973 drm_i915_private_t *dev_priv = dev->dev_private;
3974 struct drm_i915_gem_phys_object *phys_obj;
3977 if (dev_priv->mm.phys_objs[id - 1] || !size)
3980 phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL);
3986 phys_obj->handle = drm_pci_alloc(dev, size, align);
3987 if (!phys_obj->handle) {
3992 set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
3995 dev_priv->mm.phys_objs[id - 1] = phys_obj;
4003 static void i915_gem_free_phys_object(struct drm_device *dev, int id)
4005 drm_i915_private_t *dev_priv = dev->dev_private;
4006 struct drm_i915_gem_phys_object *phys_obj;
4008 if (!dev_priv->mm.phys_objs[id - 1])
4011 phys_obj = dev_priv->mm.phys_objs[id - 1];
4012 if (phys_obj->cur_obj) {
4013 i915_gem_detach_phys_object(dev, phys_obj->cur_obj);
4017 set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4019 drm_pci_free(dev, phys_obj->handle);
4021 dev_priv->mm.phys_objs[id - 1] = NULL;
4024 void i915_gem_free_all_phys_object(struct drm_device *dev)
4028 for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++)
4029 i915_gem_free_phys_object(dev, i);
4032 void i915_gem_detach_phys_object(struct drm_device *dev,
4033 struct drm_i915_gem_object *obj)
4035 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
4042 vaddr = obj->phys_obj->handle->vaddr;
4044 page_count = obj->base.size / PAGE_SIZE;
4045 for (i = 0; i < page_count; i++) {
4046 struct page *page = shmem_read_mapping_page(mapping, i);
4047 if (!IS_ERR(page)) {
4048 char *dst = kmap_atomic(page);
4049 memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE);
4052 drm_clflush_pages(&page, 1);
4054 set_page_dirty(page);
4055 mark_page_accessed(page);
4056 page_cache_release(page);
4059 intel_gtt_chipset_flush();
4061 obj->phys_obj->cur_obj = NULL;
4062 obj->phys_obj = NULL;
4066 i915_gem_attach_phys_object(struct drm_device *dev,
4067 struct drm_i915_gem_object *obj,
4071 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
4072 drm_i915_private_t *dev_priv = dev->dev_private;
4077 if (id > I915_MAX_PHYS_OBJECT)
4080 if (obj->phys_obj) {
4081 if (obj->phys_obj->id == id)
4083 i915_gem_detach_phys_object(dev, obj);
4086 /* create a new object */
4087 if (!dev_priv->mm.phys_objs[id - 1]) {
4088 ret = i915_gem_init_phys_object(dev, id,
4089 obj->base.size, align);
4091 DRM_ERROR("failed to init phys object %d size: %zu\n",
4092 id, obj->base.size);
4097 /* bind to the object */
4098 obj->phys_obj = dev_priv->mm.phys_objs[id - 1];
4099 obj->phys_obj->cur_obj = obj;
4101 page_count = obj->base.size / PAGE_SIZE;
4103 for (i = 0; i < page_count; i++) {
4107 page = shmem_read_mapping_page(mapping, i);
4109 return PTR_ERR(page);
4111 src = kmap_atomic(page);
4112 dst = obj->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4113 memcpy(dst, src, PAGE_SIZE);
4116 mark_page_accessed(page);
4117 page_cache_release(page);
4124 i915_gem_phys_pwrite(struct drm_device *dev,
4125 struct drm_i915_gem_object *obj,
4126 struct drm_i915_gem_pwrite *args,
4127 struct drm_file *file_priv)
4129 void *vaddr = obj->phys_obj->handle->vaddr + args->offset;
4130 char __user *user_data = (char __user *) (uintptr_t) args->data_ptr;
4132 if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
4133 unsigned long unwritten;
4135 /* The physical object once assigned is fixed for the lifetime
4136 * of the obj, so we can safely drop the lock and continue
4139 mutex_unlock(&dev->struct_mutex);
4140 unwritten = copy_from_user(vaddr, user_data, args->size);
4141 mutex_lock(&dev->struct_mutex);
4146 intel_gtt_chipset_flush();
4150 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
4152 struct drm_i915_file_private *file_priv = file->driver_priv;
4154 /* Clean up our request list when the client is going away, so that
4155 * later retire_requests won't dereference our soon-to-be-gone
4158 spin_lock(&file_priv->mm.lock);
4159 while (!list_empty(&file_priv->mm.request_list)) {
4160 struct drm_i915_gem_request *request;
4162 request = list_first_entry(&file_priv->mm.request_list,
4163 struct drm_i915_gem_request,
4165 list_del(&request->client_list);
4166 request->file_priv = NULL;
4168 spin_unlock(&file_priv->mm.lock);
4172 i915_gpu_is_active(struct drm_device *dev)
4174 drm_i915_private_t *dev_priv = dev->dev_private;
4175 return !list_empty(&dev_priv->mm.active_list);
4179 i915_gem_inactive_shrink(struct shrinker *shrinker, struct shrink_control *sc)
4181 struct drm_i915_private *dev_priv =
4182 container_of(shrinker,
4183 struct drm_i915_private,
4184 mm.inactive_shrinker);
4185 struct drm_device *dev = dev_priv->dev;
4186 struct drm_i915_gem_object *obj, *next;
4187 int nr_to_scan = sc->nr_to_scan;
4190 if (!mutex_trylock(&dev->struct_mutex))
4193 /* "fast-path" to count number of available objects */
4194 if (nr_to_scan == 0) {
4196 list_for_each_entry(obj,
4197 &dev_priv->mm.inactive_list,
4200 mutex_unlock(&dev->struct_mutex);
4201 return cnt / 100 * sysctl_vfs_cache_pressure;
4205 /* first scan for clean buffers */
4206 i915_gem_retire_requests(dev);
4208 list_for_each_entry_safe(obj, next,
4209 &dev_priv->mm.inactive_list,
4211 if (i915_gem_object_is_purgeable(obj)) {
4212 if (i915_gem_object_unbind(obj) == 0 &&
4218 /* second pass, evict/count anything still on the inactive list */
4220 list_for_each_entry_safe(obj, next,
4221 &dev_priv->mm.inactive_list,
4224 i915_gem_object_unbind(obj) == 0)
4230 if (nr_to_scan && i915_gpu_is_active(dev)) {
4232 * We are desperate for pages, so as a last resort, wait
4233 * for the GPU to finish and discard whatever we can.
4234 * This has a dramatic impact to reduce the number of
4235 * OOM-killer events whilst running the GPU aggressively.
4237 if (i915_gpu_idle(dev) == 0)
4240 mutex_unlock(&dev->struct_mutex);
4241 return cnt / 100 * sysctl_vfs_cache_pressure;