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>
39 static __must_check int i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object *obj);
40 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);
41 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj);
42 static __must_check int i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
44 bool map_and_fenceable);
45 static int i915_gem_phys_pwrite(struct drm_device *dev,
46 struct drm_i915_gem_object *obj,
47 struct drm_i915_gem_pwrite *args,
48 struct drm_file *file);
50 static void i915_gem_write_fence(struct drm_device *dev, int reg,
51 struct drm_i915_gem_object *obj);
52 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
53 struct drm_i915_fence_reg *fence,
56 static int i915_gem_inactive_shrink(struct shrinker *shrinker,
57 struct shrink_control *sc);
58 static void i915_gem_object_truncate(struct drm_i915_gem_object *obj);
60 static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object *obj)
63 i915_gem_release_mmap(obj);
65 /* As we do not have an associated fence register, we will force
66 * a tiling change if we ever need to acquire one.
68 obj->fence_dirty = false;
69 obj->fence_reg = I915_FENCE_REG_NONE;
72 /* some bookkeeping */
73 static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
76 dev_priv->mm.object_count++;
77 dev_priv->mm.object_memory += size;
80 static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
83 dev_priv->mm.object_count--;
84 dev_priv->mm.object_memory -= size;
88 i915_gem_wait_for_error(struct drm_device *dev)
90 struct drm_i915_private *dev_priv = dev->dev_private;
91 struct completion *x = &dev_priv->error_completion;
95 if (!atomic_read(&dev_priv->mm.wedged))
98 ret = wait_for_completion_interruptible(x);
102 if (atomic_read(&dev_priv->mm.wedged)) {
103 /* GPU is hung, bump the completion count to account for
104 * the token we just consumed so that we never hit zero and
105 * end up waiting upon a subsequent completion event that
108 spin_lock_irqsave(&x->wait.lock, flags);
110 spin_unlock_irqrestore(&x->wait.lock, flags);
115 int i915_mutex_lock_interruptible(struct drm_device *dev)
119 ret = i915_gem_wait_for_error(dev);
123 ret = mutex_lock_interruptible(&dev->struct_mutex);
127 WARN_ON(i915_verify_lists(dev));
132 i915_gem_object_is_inactive(struct drm_i915_gem_object *obj)
138 i915_gem_init_ioctl(struct drm_device *dev, void *data,
139 struct drm_file *file)
141 struct drm_i915_gem_init *args = data;
143 if (drm_core_check_feature(dev, DRIVER_MODESET))
146 if (args->gtt_start >= args->gtt_end ||
147 (args->gtt_end | args->gtt_start) & (PAGE_SIZE - 1))
150 /* GEM with user mode setting was never supported on ilk and later. */
151 if (INTEL_INFO(dev)->gen >= 5)
154 mutex_lock(&dev->struct_mutex);
155 i915_gem_init_global_gtt(dev, args->gtt_start,
156 args->gtt_end, args->gtt_end);
157 mutex_unlock(&dev->struct_mutex);
163 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
164 struct drm_file *file)
166 struct drm_i915_private *dev_priv = dev->dev_private;
167 struct drm_i915_gem_get_aperture *args = data;
168 struct drm_i915_gem_object *obj;
172 mutex_lock(&dev->struct_mutex);
173 list_for_each_entry(obj, &dev_priv->mm.gtt_list, gtt_list)
175 pinned += obj->gtt_space->size;
176 mutex_unlock(&dev->struct_mutex);
178 args->aper_size = dev_priv->mm.gtt_total;
179 args->aper_available_size = args->aper_size - pinned;
185 i915_gem_create(struct drm_file *file,
186 struct drm_device *dev,
190 struct drm_i915_gem_object *obj;
194 size = roundup(size, PAGE_SIZE);
198 /* Allocate the new object */
199 obj = i915_gem_alloc_object(dev, size);
203 ret = drm_gem_handle_create(file, &obj->base, &handle);
205 drm_gem_object_release(&obj->base);
206 i915_gem_info_remove_obj(dev->dev_private, obj->base.size);
211 /* drop reference from allocate - handle holds it now */
212 drm_gem_object_unreference(&obj->base);
213 trace_i915_gem_object_create(obj);
220 i915_gem_dumb_create(struct drm_file *file,
221 struct drm_device *dev,
222 struct drm_mode_create_dumb *args)
224 /* have to work out size/pitch and return them */
225 args->pitch = ALIGN(args->width * ((args->bpp + 7) / 8), 64);
226 args->size = args->pitch * args->height;
227 return i915_gem_create(file, dev,
228 args->size, &args->handle);
231 int i915_gem_dumb_destroy(struct drm_file *file,
232 struct drm_device *dev,
235 return drm_gem_handle_delete(file, handle);
239 * Creates a new mm object and returns a handle to it.
242 i915_gem_create_ioctl(struct drm_device *dev, void *data,
243 struct drm_file *file)
245 struct drm_i915_gem_create *args = data;
247 return i915_gem_create(file, dev,
248 args->size, &args->handle);
251 static int i915_gem_object_needs_bit17_swizzle(struct drm_i915_gem_object *obj)
253 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
255 return dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_9_10_17 &&
256 obj->tiling_mode != I915_TILING_NONE;
260 __copy_to_user_swizzled(char __user *cpu_vaddr,
261 const char *gpu_vaddr, int gpu_offset,
264 int ret, cpu_offset = 0;
267 int cacheline_end = ALIGN(gpu_offset + 1, 64);
268 int this_length = min(cacheline_end - gpu_offset, length);
269 int swizzled_gpu_offset = gpu_offset ^ 64;
271 ret = __copy_to_user(cpu_vaddr + cpu_offset,
272 gpu_vaddr + swizzled_gpu_offset,
277 cpu_offset += this_length;
278 gpu_offset += this_length;
279 length -= this_length;
286 __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset,
287 const char __user *cpu_vaddr,
290 int ret, cpu_offset = 0;
293 int cacheline_end = ALIGN(gpu_offset + 1, 64);
294 int this_length = min(cacheline_end - gpu_offset, length);
295 int swizzled_gpu_offset = gpu_offset ^ 64;
297 ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset,
298 cpu_vaddr + cpu_offset,
303 cpu_offset += this_length;
304 gpu_offset += this_length;
305 length -= this_length;
311 /* Per-page copy function for the shmem pread fastpath.
312 * Flushes invalid cachelines before reading the target if
313 * needs_clflush is set. */
315 shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length,
316 char __user *user_data,
317 bool page_do_bit17_swizzling, bool needs_clflush)
322 if (unlikely(page_do_bit17_swizzling))
325 vaddr = kmap_atomic(page);
327 drm_clflush_virt_range(vaddr + shmem_page_offset,
329 ret = __copy_to_user_inatomic(user_data,
330 vaddr + shmem_page_offset,
332 kunmap_atomic(vaddr);
338 shmem_clflush_swizzled_range(char *addr, unsigned long length,
341 if (unlikely(swizzled)) {
342 unsigned long start = (unsigned long) addr;
343 unsigned long end = (unsigned long) addr + length;
345 /* For swizzling simply ensure that we always flush both
346 * channels. Lame, but simple and it works. Swizzled
347 * pwrite/pread is far from a hotpath - current userspace
348 * doesn't use it at all. */
349 start = round_down(start, 128);
350 end = round_up(end, 128);
352 drm_clflush_virt_range((void *)start, end - start);
354 drm_clflush_virt_range(addr, length);
359 /* Only difference to the fast-path function is that this can handle bit17
360 * and uses non-atomic copy and kmap functions. */
362 shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length,
363 char __user *user_data,
364 bool page_do_bit17_swizzling, bool needs_clflush)
371 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
373 page_do_bit17_swizzling);
375 if (page_do_bit17_swizzling)
376 ret = __copy_to_user_swizzled(user_data,
377 vaddr, shmem_page_offset,
380 ret = __copy_to_user(user_data,
381 vaddr + shmem_page_offset,
389 i915_gem_shmem_pread(struct drm_device *dev,
390 struct drm_i915_gem_object *obj,
391 struct drm_i915_gem_pread *args,
392 struct drm_file *file)
394 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
395 char __user *user_data;
398 int shmem_page_offset, page_length, ret = 0;
399 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
400 int hit_slowpath = 0;
402 int needs_clflush = 0;
405 user_data = (char __user *) (uintptr_t) args->data_ptr;
408 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
410 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) {
411 /* If we're not in the cpu read domain, set ourself into the gtt
412 * read domain and manually flush cachelines (if required). This
413 * optimizes for the case when the gpu will dirty the data
414 * anyway again before the next pread happens. */
415 if (obj->cache_level == I915_CACHE_NONE)
417 ret = i915_gem_object_set_to_gtt_domain(obj, false);
422 offset = args->offset;
427 /* Operation in this page
429 * shmem_page_offset = offset within page in shmem file
430 * page_length = bytes to copy for this page
432 shmem_page_offset = offset_in_page(offset);
433 page_length = remain;
434 if ((shmem_page_offset + page_length) > PAGE_SIZE)
435 page_length = PAGE_SIZE - shmem_page_offset;
438 page = obj->pages[offset >> PAGE_SHIFT];
441 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
449 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
450 (page_to_phys(page) & (1 << 17)) != 0;
452 ret = shmem_pread_fast(page, shmem_page_offset, page_length,
453 user_data, page_do_bit17_swizzling,
459 page_cache_get(page);
460 mutex_unlock(&dev->struct_mutex);
463 ret = fault_in_multipages_writeable(user_data, remain);
464 /* Userspace is tricking us, but we've already clobbered
465 * its pages with the prefault and promised to write the
466 * data up to the first fault. Hence ignore any errors
467 * and just continue. */
472 ret = shmem_pread_slow(page, shmem_page_offset, page_length,
473 user_data, page_do_bit17_swizzling,
476 mutex_lock(&dev->struct_mutex);
477 page_cache_release(page);
479 mark_page_accessed(page);
481 page_cache_release(page);
488 remain -= page_length;
489 user_data += page_length;
490 offset += page_length;
495 /* Fixup: Kill any reinstated backing storage pages */
496 if (obj->madv == __I915_MADV_PURGED)
497 i915_gem_object_truncate(obj);
504 * Reads data from the object referenced by handle.
506 * On error, the contents of *data are undefined.
509 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
510 struct drm_file *file)
512 struct drm_i915_gem_pread *args = data;
513 struct drm_i915_gem_object *obj;
519 if (!access_ok(VERIFY_WRITE,
520 (char __user *)(uintptr_t)args->data_ptr,
524 ret = i915_mutex_lock_interruptible(dev);
528 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
529 if (&obj->base == NULL) {
534 /* Bounds check source. */
535 if (args->offset > obj->base.size ||
536 args->size > obj->base.size - args->offset) {
541 trace_i915_gem_object_pread(obj, args->offset, args->size);
543 ret = i915_gem_shmem_pread(dev, obj, args, file);
546 drm_gem_object_unreference(&obj->base);
548 mutex_unlock(&dev->struct_mutex);
552 /* This is the fast write path which cannot handle
553 * page faults in the source data
557 fast_user_write(struct io_mapping *mapping,
558 loff_t page_base, int page_offset,
559 char __user *user_data,
562 void __iomem *vaddr_atomic;
564 unsigned long unwritten;
566 vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
567 /* We can use the cpu mem copy function because this is X86. */
568 vaddr = (void __force*)vaddr_atomic + page_offset;
569 unwritten = __copy_from_user_inatomic_nocache(vaddr,
571 io_mapping_unmap_atomic(vaddr_atomic);
576 * This is the fast pwrite path, where we copy the data directly from the
577 * user into the GTT, uncached.
580 i915_gem_gtt_pwrite_fast(struct drm_device *dev,
581 struct drm_i915_gem_object *obj,
582 struct drm_i915_gem_pwrite *args,
583 struct drm_file *file)
585 drm_i915_private_t *dev_priv = dev->dev_private;
587 loff_t offset, page_base;
588 char __user *user_data;
589 int page_offset, page_length, ret;
591 ret = i915_gem_object_pin(obj, 0, true);
595 ret = i915_gem_object_set_to_gtt_domain(obj, true);
599 ret = i915_gem_object_put_fence(obj);
603 user_data = (char __user *) (uintptr_t) args->data_ptr;
606 offset = obj->gtt_offset + args->offset;
609 /* Operation in this page
611 * page_base = page offset within aperture
612 * page_offset = offset within page
613 * page_length = bytes to copy for this page
615 page_base = offset & PAGE_MASK;
616 page_offset = offset_in_page(offset);
617 page_length = remain;
618 if ((page_offset + remain) > PAGE_SIZE)
619 page_length = PAGE_SIZE - page_offset;
621 /* If we get a fault while copying data, then (presumably) our
622 * source page isn't available. Return the error and we'll
623 * retry in the slow path.
625 if (fast_user_write(dev_priv->mm.gtt_mapping, page_base,
626 page_offset, user_data, page_length)) {
631 remain -= page_length;
632 user_data += page_length;
633 offset += page_length;
637 i915_gem_object_unpin(obj);
642 /* Per-page copy function for the shmem pwrite fastpath.
643 * Flushes invalid cachelines before writing to the target if
644 * needs_clflush_before is set and flushes out any written cachelines after
645 * writing if needs_clflush is set. */
647 shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
648 char __user *user_data,
649 bool page_do_bit17_swizzling,
650 bool needs_clflush_before,
651 bool needs_clflush_after)
656 if (unlikely(page_do_bit17_swizzling))
659 vaddr = kmap_atomic(page);
660 if (needs_clflush_before)
661 drm_clflush_virt_range(vaddr + shmem_page_offset,
663 ret = __copy_from_user_inatomic_nocache(vaddr + shmem_page_offset,
666 if (needs_clflush_after)
667 drm_clflush_virt_range(vaddr + shmem_page_offset,
669 kunmap_atomic(vaddr);
674 /* Only difference to the fast-path function is that this can handle bit17
675 * and uses non-atomic copy and kmap functions. */
677 shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
678 char __user *user_data,
679 bool page_do_bit17_swizzling,
680 bool needs_clflush_before,
681 bool needs_clflush_after)
687 if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
688 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
690 page_do_bit17_swizzling);
691 if (page_do_bit17_swizzling)
692 ret = __copy_from_user_swizzled(vaddr, shmem_page_offset,
696 ret = __copy_from_user(vaddr + shmem_page_offset,
699 if (needs_clflush_after)
700 shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
702 page_do_bit17_swizzling);
709 i915_gem_shmem_pwrite(struct drm_device *dev,
710 struct drm_i915_gem_object *obj,
711 struct drm_i915_gem_pwrite *args,
712 struct drm_file *file)
714 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
717 char __user *user_data;
718 int shmem_page_offset, page_length, ret = 0;
719 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
720 int hit_slowpath = 0;
721 int needs_clflush_after = 0;
722 int needs_clflush_before = 0;
725 user_data = (char __user *) (uintptr_t) args->data_ptr;
728 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
730 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
731 /* If we're not in the cpu write domain, set ourself into the gtt
732 * write domain and manually flush cachelines (if required). This
733 * optimizes for the case when the gpu will use the data
734 * right away and we therefore have to clflush anyway. */
735 if (obj->cache_level == I915_CACHE_NONE)
736 needs_clflush_after = 1;
737 ret = i915_gem_object_set_to_gtt_domain(obj, true);
741 /* Same trick applies for invalidate partially written cachelines before
743 if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)
744 && obj->cache_level == I915_CACHE_NONE)
745 needs_clflush_before = 1;
747 offset = args->offset;
752 int partial_cacheline_write;
754 /* Operation in this page
756 * shmem_page_offset = offset within page in shmem file
757 * page_length = bytes to copy for this page
759 shmem_page_offset = offset_in_page(offset);
761 page_length = remain;
762 if ((shmem_page_offset + page_length) > PAGE_SIZE)
763 page_length = PAGE_SIZE - shmem_page_offset;
765 /* If we don't overwrite a cacheline completely we need to be
766 * careful to have up-to-date data by first clflushing. Don't
767 * overcomplicate things and flush the entire patch. */
768 partial_cacheline_write = needs_clflush_before &&
769 ((shmem_page_offset | page_length)
770 & (boot_cpu_data.x86_clflush_size - 1));
773 page = obj->pages[offset >> PAGE_SHIFT];
776 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
784 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
785 (page_to_phys(page) & (1 << 17)) != 0;
787 ret = shmem_pwrite_fast(page, shmem_page_offset, page_length,
788 user_data, page_do_bit17_swizzling,
789 partial_cacheline_write,
790 needs_clflush_after);
795 page_cache_get(page);
796 mutex_unlock(&dev->struct_mutex);
798 ret = shmem_pwrite_slow(page, shmem_page_offset, page_length,
799 user_data, page_do_bit17_swizzling,
800 partial_cacheline_write,
801 needs_clflush_after);
803 mutex_lock(&dev->struct_mutex);
804 page_cache_release(page);
806 set_page_dirty(page);
807 mark_page_accessed(page);
809 page_cache_release(page);
816 remain -= page_length;
817 user_data += page_length;
818 offset += page_length;
823 /* Fixup: Kill any reinstated backing storage pages */
824 if (obj->madv == __I915_MADV_PURGED)
825 i915_gem_object_truncate(obj);
826 /* and flush dirty cachelines in case the object isn't in the cpu write
828 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
829 i915_gem_clflush_object(obj);
830 intel_gtt_chipset_flush();
834 if (needs_clflush_after)
835 intel_gtt_chipset_flush();
841 * Writes data to the object referenced by handle.
843 * On error, the contents of the buffer that were to be modified are undefined.
846 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
847 struct drm_file *file)
849 struct drm_i915_gem_pwrite *args = data;
850 struct drm_i915_gem_object *obj;
856 if (!access_ok(VERIFY_READ,
857 (char __user *)(uintptr_t)args->data_ptr,
861 ret = fault_in_multipages_readable((char __user *)(uintptr_t)args->data_ptr,
866 ret = i915_mutex_lock_interruptible(dev);
870 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
871 if (&obj->base == NULL) {
876 /* Bounds check destination. */
877 if (args->offset > obj->base.size ||
878 args->size > obj->base.size - args->offset) {
883 trace_i915_gem_object_pwrite(obj, args->offset, args->size);
886 /* We can only do the GTT pwrite on untiled buffers, as otherwise
887 * it would end up going through the fenced access, and we'll get
888 * different detiling behavior between reading and writing.
889 * pread/pwrite currently are reading and writing from the CPU
890 * perspective, requiring manual detiling by the client.
893 ret = i915_gem_phys_pwrite(dev, obj, args, file);
897 if (obj->gtt_space &&
898 obj->cache_level == I915_CACHE_NONE &&
899 obj->tiling_mode == I915_TILING_NONE &&
900 obj->map_and_fenceable &&
901 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
902 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
903 /* Note that the gtt paths might fail with non-page-backed user
904 * pointers (e.g. gtt mappings when moving data between
905 * textures). Fallback to the shmem path in that case. */
909 ret = i915_gem_shmem_pwrite(dev, obj, args, file);
912 drm_gem_object_unreference(&obj->base);
914 mutex_unlock(&dev->struct_mutex);
919 * Called when user space prepares to use an object with the CPU, either
920 * through the mmap ioctl's mapping or a GTT mapping.
923 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
924 struct drm_file *file)
926 struct drm_i915_gem_set_domain *args = data;
927 struct drm_i915_gem_object *obj;
928 uint32_t read_domains = args->read_domains;
929 uint32_t write_domain = args->write_domain;
932 /* Only handle setting domains to types used by the CPU. */
933 if (write_domain & I915_GEM_GPU_DOMAINS)
936 if (read_domains & I915_GEM_GPU_DOMAINS)
939 /* Having something in the write domain implies it's in the read
940 * domain, and only that read domain. Enforce that in the request.
942 if (write_domain != 0 && read_domains != write_domain)
945 ret = i915_mutex_lock_interruptible(dev);
949 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
950 if (&obj->base == NULL) {
955 if (read_domains & I915_GEM_DOMAIN_GTT) {
956 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
958 /* Silently promote "you're not bound, there was nothing to do"
959 * to success, since the client was just asking us to
960 * make sure everything was done.
965 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
968 drm_gem_object_unreference(&obj->base);
970 mutex_unlock(&dev->struct_mutex);
975 * Called when user space has done writes to this buffer
978 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
979 struct drm_file *file)
981 struct drm_i915_gem_sw_finish *args = data;
982 struct drm_i915_gem_object *obj;
985 ret = i915_mutex_lock_interruptible(dev);
989 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
990 if (&obj->base == NULL) {
995 /* Pinned buffers may be scanout, so flush the cache */
997 i915_gem_object_flush_cpu_write_domain(obj);
999 drm_gem_object_unreference(&obj->base);
1001 mutex_unlock(&dev->struct_mutex);
1006 * Maps the contents of an object, returning the address it is mapped
1009 * While the mapping holds a reference on the contents of the object, it doesn't
1010 * imply a ref on the object itself.
1013 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1014 struct drm_file *file)
1016 struct drm_i915_gem_mmap *args = data;
1017 struct drm_gem_object *obj;
1020 obj = drm_gem_object_lookup(dev, file, args->handle);
1024 addr = vm_mmap(obj->filp, 0, args->size,
1025 PROT_READ | PROT_WRITE, MAP_SHARED,
1027 drm_gem_object_unreference_unlocked(obj);
1028 if (IS_ERR((void *)addr))
1031 args->addr_ptr = (uint64_t) addr;
1037 * i915_gem_fault - fault a page into the GTT
1038 * vma: VMA in question
1041 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1042 * from userspace. The fault handler takes care of binding the object to
1043 * the GTT (if needed), allocating and programming a fence register (again,
1044 * only if needed based on whether the old reg is still valid or the object
1045 * is tiled) and inserting a new PTE into the faulting process.
1047 * Note that the faulting process may involve evicting existing objects
1048 * from the GTT and/or fence registers to make room. So performance may
1049 * suffer if the GTT working set is large or there are few fence registers
1052 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1054 struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1055 struct drm_device *dev = obj->base.dev;
1056 drm_i915_private_t *dev_priv = dev->dev_private;
1057 pgoff_t page_offset;
1060 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1062 /* We don't use vmf->pgoff since that has the fake offset */
1063 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1066 ret = i915_mutex_lock_interruptible(dev);
1070 trace_i915_gem_object_fault(obj, page_offset, true, write);
1072 /* Now bind it into the GTT if needed */
1073 if (!obj->map_and_fenceable) {
1074 ret = i915_gem_object_unbind(obj);
1078 if (!obj->gtt_space) {
1079 ret = i915_gem_object_bind_to_gtt(obj, 0, true);
1083 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1088 if (!obj->has_global_gtt_mapping)
1089 i915_gem_gtt_bind_object(obj, obj->cache_level);
1091 ret = i915_gem_object_get_fence(obj);
1095 if (i915_gem_object_is_inactive(obj))
1096 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1098 obj->fault_mappable = true;
1100 pfn = ((dev->agp->base + obj->gtt_offset) >> PAGE_SHIFT) +
1103 /* Finally, remap it using the new GTT offset */
1104 ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1106 mutex_unlock(&dev->struct_mutex);
1111 /* Give the error handler a chance to run and move the
1112 * objects off the GPU active list. Next time we service the
1113 * fault, we should be able to transition the page into the
1114 * GTT without touching the GPU (and so avoid further
1115 * EIO/EGAIN). If the GPU is wedged, then there is no issue
1116 * with coherency, just lost writes.
1122 return VM_FAULT_NOPAGE;
1124 return VM_FAULT_OOM;
1126 return VM_FAULT_SIGBUS;
1131 * i915_gem_release_mmap - remove physical page mappings
1132 * @obj: obj in question
1134 * Preserve the reservation of the mmapping with the DRM core code, but
1135 * relinquish ownership of the pages back to the system.
1137 * It is vital that we remove the page mapping if we have mapped a tiled
1138 * object through the GTT and then lose the fence register due to
1139 * resource pressure. Similarly if the object has been moved out of the
1140 * aperture, than pages mapped into userspace must be revoked. Removing the
1141 * mapping will then trigger a page fault on the next user access, allowing
1142 * fixup by i915_gem_fault().
1145 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1147 if (!obj->fault_mappable)
1150 if (obj->base.dev->dev_mapping)
1151 unmap_mapping_range(obj->base.dev->dev_mapping,
1152 (loff_t)obj->base.map_list.hash.key<<PAGE_SHIFT,
1155 obj->fault_mappable = false;
1159 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1163 if (INTEL_INFO(dev)->gen >= 4 ||
1164 tiling_mode == I915_TILING_NONE)
1167 /* Previous chips need a power-of-two fence region when tiling */
1168 if (INTEL_INFO(dev)->gen == 3)
1169 gtt_size = 1024*1024;
1171 gtt_size = 512*1024;
1173 while (gtt_size < size)
1180 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1181 * @obj: object to check
1183 * Return the required GTT alignment for an object, taking into account
1184 * potential fence register mapping.
1187 i915_gem_get_gtt_alignment(struct drm_device *dev,
1192 * Minimum alignment is 4k (GTT page size), but might be greater
1193 * if a fence register is needed for the object.
1195 if (INTEL_INFO(dev)->gen >= 4 ||
1196 tiling_mode == I915_TILING_NONE)
1200 * Previous chips need to be aligned to the size of the smallest
1201 * fence register that can contain the object.
1203 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1207 * i915_gem_get_unfenced_gtt_alignment - return required GTT alignment for an
1210 * @size: size of the object
1211 * @tiling_mode: tiling mode of the object
1213 * Return the required GTT alignment for an object, only taking into account
1214 * unfenced tiled surface requirements.
1217 i915_gem_get_unfenced_gtt_alignment(struct drm_device *dev,
1222 * Minimum alignment is 4k (GTT page size) for sane hw.
1224 if (INTEL_INFO(dev)->gen >= 4 || IS_G33(dev) ||
1225 tiling_mode == I915_TILING_NONE)
1228 /* Previous hardware however needs to be aligned to a power-of-two
1229 * tile height. The simplest method for determining this is to reuse
1230 * the power-of-tile object size.
1232 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1236 i915_gem_mmap_gtt(struct drm_file *file,
1237 struct drm_device *dev,
1241 struct drm_i915_private *dev_priv = dev->dev_private;
1242 struct drm_i915_gem_object *obj;
1245 ret = i915_mutex_lock_interruptible(dev);
1249 obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
1250 if (&obj->base == NULL) {
1255 if (obj->base.size > dev_priv->mm.gtt_mappable_end) {
1260 if (obj->madv != I915_MADV_WILLNEED) {
1261 DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1266 if (!obj->base.map_list.map) {
1267 ret = drm_gem_create_mmap_offset(&obj->base);
1272 *offset = (u64)obj->base.map_list.hash.key << PAGE_SHIFT;
1275 drm_gem_object_unreference(&obj->base);
1277 mutex_unlock(&dev->struct_mutex);
1282 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1284 * @data: GTT mapping ioctl data
1285 * @file: GEM object info
1287 * Simply returns the fake offset to userspace so it can mmap it.
1288 * The mmap call will end up in drm_gem_mmap(), which will set things
1289 * up so we can get faults in the handler above.
1291 * The fault handler will take care of binding the object into the GTT
1292 * (since it may have been evicted to make room for something), allocating
1293 * a fence register, and mapping the appropriate aperture address into
1297 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1298 struct drm_file *file)
1300 struct drm_i915_gem_mmap_gtt *args = data;
1302 return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
1307 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj,
1311 struct address_space *mapping;
1312 struct inode *inode;
1315 /* Get the list of pages out of our struct file. They'll be pinned
1316 * at this point until we release them.
1318 page_count = obj->base.size / PAGE_SIZE;
1319 BUG_ON(obj->pages != NULL);
1320 obj->pages = drm_malloc_ab(page_count, sizeof(struct page *));
1321 if (obj->pages == NULL)
1324 inode = obj->base.filp->f_path.dentry->d_inode;
1325 mapping = inode->i_mapping;
1326 gfpmask |= mapping_gfp_mask(mapping);
1328 for (i = 0; i < page_count; i++) {
1329 page = shmem_read_mapping_page_gfp(mapping, i, gfpmask);
1333 obj->pages[i] = page;
1336 if (i915_gem_object_needs_bit17_swizzle(obj))
1337 i915_gem_object_do_bit_17_swizzle(obj);
1343 page_cache_release(obj->pages[i]);
1345 drm_free_large(obj->pages);
1347 return PTR_ERR(page);
1351 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
1353 int page_count = obj->base.size / PAGE_SIZE;
1356 BUG_ON(obj->madv == __I915_MADV_PURGED);
1358 if (i915_gem_object_needs_bit17_swizzle(obj))
1359 i915_gem_object_save_bit_17_swizzle(obj);
1361 if (obj->madv == I915_MADV_DONTNEED)
1364 for (i = 0; i < page_count; i++) {
1366 set_page_dirty(obj->pages[i]);
1368 if (obj->madv == I915_MADV_WILLNEED)
1369 mark_page_accessed(obj->pages[i]);
1371 page_cache_release(obj->pages[i]);
1375 drm_free_large(obj->pages);
1380 i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
1381 struct intel_ring_buffer *ring,
1384 struct drm_device *dev = obj->base.dev;
1385 struct drm_i915_private *dev_priv = dev->dev_private;
1387 BUG_ON(ring == NULL);
1390 /* Add a reference if we're newly entering the active list. */
1392 drm_gem_object_reference(&obj->base);
1396 /* Move from whatever list we were on to the tail of execution. */
1397 list_move_tail(&obj->mm_list, &dev_priv->mm.active_list);
1398 list_move_tail(&obj->ring_list, &ring->active_list);
1400 obj->last_rendering_seqno = seqno;
1402 if (obj->fenced_gpu_access) {
1403 obj->last_fenced_seqno = seqno;
1405 /* Bump MRU to take account of the delayed flush */
1406 if (obj->fence_reg != I915_FENCE_REG_NONE) {
1407 struct drm_i915_fence_reg *reg;
1409 reg = &dev_priv->fence_regs[obj->fence_reg];
1410 list_move_tail(®->lru_list,
1411 &dev_priv->mm.fence_list);
1417 i915_gem_object_move_off_active(struct drm_i915_gem_object *obj)
1419 list_del_init(&obj->ring_list);
1420 obj->last_rendering_seqno = 0;
1421 obj->last_fenced_seqno = 0;
1425 i915_gem_object_move_to_flushing(struct drm_i915_gem_object *obj)
1427 struct drm_device *dev = obj->base.dev;
1428 drm_i915_private_t *dev_priv = dev->dev_private;
1430 BUG_ON(!obj->active);
1431 list_move_tail(&obj->mm_list, &dev_priv->mm.flushing_list);
1433 i915_gem_object_move_off_active(obj);
1437 i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
1439 struct drm_device *dev = obj->base.dev;
1440 struct drm_i915_private *dev_priv = dev->dev_private;
1442 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1444 BUG_ON(!list_empty(&obj->gpu_write_list));
1445 BUG_ON(!obj->active);
1448 i915_gem_object_move_off_active(obj);
1449 obj->fenced_gpu_access = false;
1452 obj->pending_gpu_write = false;
1453 drm_gem_object_unreference(&obj->base);
1455 WARN_ON(i915_verify_lists(dev));
1458 /* Immediately discard the backing storage */
1460 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
1462 struct inode *inode;
1464 /* Our goal here is to return as much of the memory as
1465 * is possible back to the system as we are called from OOM.
1466 * To do this we must instruct the shmfs to drop all of its
1467 * backing pages, *now*.
1469 inode = obj->base.filp->f_path.dentry->d_inode;
1470 shmem_truncate_range(inode, 0, (loff_t)-1);
1472 if (obj->base.map_list.map)
1473 drm_gem_free_mmap_offset(&obj->base);
1475 obj->madv = __I915_MADV_PURGED;
1479 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj)
1481 return obj->madv == I915_MADV_DONTNEED;
1485 i915_gem_process_flushing_list(struct intel_ring_buffer *ring,
1486 uint32_t flush_domains)
1488 struct drm_i915_gem_object *obj, *next;
1490 list_for_each_entry_safe(obj, next,
1491 &ring->gpu_write_list,
1493 if (obj->base.write_domain & flush_domains) {
1494 uint32_t old_write_domain = obj->base.write_domain;
1496 obj->base.write_domain = 0;
1497 list_del_init(&obj->gpu_write_list);
1498 i915_gem_object_move_to_active(obj, ring,
1499 i915_gem_next_request_seqno(ring));
1501 trace_i915_gem_object_change_domain(obj,
1502 obj->base.read_domains,
1509 i915_gem_get_seqno(struct drm_device *dev)
1511 drm_i915_private_t *dev_priv = dev->dev_private;
1512 u32 seqno = dev_priv->next_seqno;
1514 /* reserve 0 for non-seqno */
1515 if (++dev_priv->next_seqno == 0)
1516 dev_priv->next_seqno = 1;
1522 i915_gem_next_request_seqno(struct intel_ring_buffer *ring)
1524 if (ring->outstanding_lazy_request == 0)
1525 ring->outstanding_lazy_request = i915_gem_get_seqno(ring->dev);
1527 return ring->outstanding_lazy_request;
1531 i915_add_request(struct intel_ring_buffer *ring,
1532 struct drm_file *file,
1533 struct drm_i915_gem_request *request)
1535 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1537 u32 request_ring_position;
1541 BUG_ON(request == NULL);
1542 seqno = i915_gem_next_request_seqno(ring);
1544 /* Record the position of the start of the request so that
1545 * should we detect the updated seqno part-way through the
1546 * GPU processing the request, we never over-estimate the
1547 * position of the head.
1549 request_ring_position = intel_ring_get_tail(ring);
1551 ret = ring->add_request(ring, &seqno);
1555 trace_i915_gem_request_add(ring, seqno);
1557 request->seqno = seqno;
1558 request->ring = ring;
1559 request->tail = request_ring_position;
1560 request->emitted_jiffies = jiffies;
1561 was_empty = list_empty(&ring->request_list);
1562 list_add_tail(&request->list, &ring->request_list);
1565 struct drm_i915_file_private *file_priv = file->driver_priv;
1567 spin_lock(&file_priv->mm.lock);
1568 request->file_priv = file_priv;
1569 list_add_tail(&request->client_list,
1570 &file_priv->mm.request_list);
1571 spin_unlock(&file_priv->mm.lock);
1574 ring->outstanding_lazy_request = 0;
1576 if (!dev_priv->mm.suspended) {
1577 if (i915_enable_hangcheck) {
1578 mod_timer(&dev_priv->hangcheck_timer,
1580 msecs_to_jiffies(DRM_I915_HANGCHECK_PERIOD));
1583 queue_delayed_work(dev_priv->wq,
1584 &dev_priv->mm.retire_work, HZ);
1590 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
1592 struct drm_i915_file_private *file_priv = request->file_priv;
1597 spin_lock(&file_priv->mm.lock);
1598 if (request->file_priv) {
1599 list_del(&request->client_list);
1600 request->file_priv = NULL;
1602 spin_unlock(&file_priv->mm.lock);
1605 static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv,
1606 struct intel_ring_buffer *ring)
1608 while (!list_empty(&ring->request_list)) {
1609 struct drm_i915_gem_request *request;
1611 request = list_first_entry(&ring->request_list,
1612 struct drm_i915_gem_request,
1615 list_del(&request->list);
1616 i915_gem_request_remove_from_client(request);
1620 while (!list_empty(&ring->active_list)) {
1621 struct drm_i915_gem_object *obj;
1623 obj = list_first_entry(&ring->active_list,
1624 struct drm_i915_gem_object,
1627 obj->base.write_domain = 0;
1628 list_del_init(&obj->gpu_write_list);
1629 i915_gem_object_move_to_inactive(obj);
1633 static void i915_gem_reset_fences(struct drm_device *dev)
1635 struct drm_i915_private *dev_priv = dev->dev_private;
1638 for (i = 0; i < dev_priv->num_fence_regs; i++) {
1639 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
1641 i915_gem_write_fence(dev, i, NULL);
1644 i915_gem_object_fence_lost(reg->obj);
1648 INIT_LIST_HEAD(®->lru_list);
1651 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
1654 void i915_gem_reset(struct drm_device *dev)
1656 struct drm_i915_private *dev_priv = dev->dev_private;
1657 struct drm_i915_gem_object *obj;
1658 struct intel_ring_buffer *ring;
1661 for_each_ring(ring, dev_priv, i)
1662 i915_gem_reset_ring_lists(dev_priv, ring);
1664 /* Remove anything from the flushing lists. The GPU cache is likely
1665 * to be lost on reset along with the data, so simply move the
1666 * lost bo to the inactive list.
1668 while (!list_empty(&dev_priv->mm.flushing_list)) {
1669 obj = list_first_entry(&dev_priv->mm.flushing_list,
1670 struct drm_i915_gem_object,
1673 obj->base.write_domain = 0;
1674 list_del_init(&obj->gpu_write_list);
1675 i915_gem_object_move_to_inactive(obj);
1678 /* Move everything out of the GPU domains to ensure we do any
1679 * necessary invalidation upon reuse.
1681 list_for_each_entry(obj,
1682 &dev_priv->mm.inactive_list,
1685 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
1688 /* The fence registers are invalidated so clear them out */
1689 i915_gem_reset_fences(dev);
1693 * This function clears the request list as sequence numbers are passed.
1696 i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
1701 if (list_empty(&ring->request_list))
1704 WARN_ON(i915_verify_lists(ring->dev));
1706 seqno = ring->get_seqno(ring);
1708 for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++)
1709 if (seqno >= ring->sync_seqno[i])
1710 ring->sync_seqno[i] = 0;
1712 while (!list_empty(&ring->request_list)) {
1713 struct drm_i915_gem_request *request;
1715 request = list_first_entry(&ring->request_list,
1716 struct drm_i915_gem_request,
1719 if (!i915_seqno_passed(seqno, request->seqno))
1722 trace_i915_gem_request_retire(ring, request->seqno);
1723 /* We know the GPU must have read the request to have
1724 * sent us the seqno + interrupt, so use the position
1725 * of tail of the request to update the last known position
1728 ring->last_retired_head = request->tail;
1730 list_del(&request->list);
1731 i915_gem_request_remove_from_client(request);
1735 /* Move any buffers on the active list that are no longer referenced
1736 * by the ringbuffer to the flushing/inactive lists as appropriate.
1738 while (!list_empty(&ring->active_list)) {
1739 struct drm_i915_gem_object *obj;
1741 obj = list_first_entry(&ring->active_list,
1742 struct drm_i915_gem_object,
1745 if (!i915_seqno_passed(seqno, obj->last_rendering_seqno))
1748 if (obj->base.write_domain != 0)
1749 i915_gem_object_move_to_flushing(obj);
1751 i915_gem_object_move_to_inactive(obj);
1754 if (unlikely(ring->trace_irq_seqno &&
1755 i915_seqno_passed(seqno, ring->trace_irq_seqno))) {
1756 ring->irq_put(ring);
1757 ring->trace_irq_seqno = 0;
1760 WARN_ON(i915_verify_lists(ring->dev));
1764 i915_gem_retire_requests(struct drm_device *dev)
1766 drm_i915_private_t *dev_priv = dev->dev_private;
1767 struct intel_ring_buffer *ring;
1770 for_each_ring(ring, dev_priv, i)
1771 i915_gem_retire_requests_ring(ring);
1775 i915_gem_retire_work_handler(struct work_struct *work)
1777 drm_i915_private_t *dev_priv;
1778 struct drm_device *dev;
1779 struct intel_ring_buffer *ring;
1783 dev_priv = container_of(work, drm_i915_private_t,
1784 mm.retire_work.work);
1785 dev = dev_priv->dev;
1787 /* Come back later if the device is busy... */
1788 if (!mutex_trylock(&dev->struct_mutex)) {
1789 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1793 i915_gem_retire_requests(dev);
1795 /* Send a periodic flush down the ring so we don't hold onto GEM
1796 * objects indefinitely.
1799 for_each_ring(ring, dev_priv, i) {
1800 if (!list_empty(&ring->gpu_write_list)) {
1801 struct drm_i915_gem_request *request;
1804 ret = i915_gem_flush_ring(ring,
1805 0, I915_GEM_GPU_DOMAINS);
1806 request = kzalloc(sizeof(*request), GFP_KERNEL);
1807 if (ret || request == NULL ||
1808 i915_add_request(ring, NULL, request))
1812 idle &= list_empty(&ring->request_list);
1815 if (!dev_priv->mm.suspended && !idle)
1816 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1818 mutex_unlock(&dev->struct_mutex);
1822 i915_gem_check_wedge(struct drm_i915_private *dev_priv)
1824 BUG_ON(!mutex_is_locked(&dev_priv->dev->struct_mutex));
1826 if (atomic_read(&dev_priv->mm.wedged)) {
1827 struct completion *x = &dev_priv->error_completion;
1828 bool recovery_complete;
1829 unsigned long flags;
1831 /* Give the error handler a chance to run. */
1832 spin_lock_irqsave(&x->wait.lock, flags);
1833 recovery_complete = x->done > 0;
1834 spin_unlock_irqrestore(&x->wait.lock, flags);
1836 return recovery_complete ? -EIO : -EAGAIN;
1843 * Compare seqno against outstanding lazy request. Emit a request if they are
1847 i915_gem_check_olr(struct intel_ring_buffer *ring, u32 seqno)
1851 BUG_ON(!mutex_is_locked(&ring->dev->struct_mutex));
1853 if (seqno == ring->outstanding_lazy_request) {
1854 struct drm_i915_gem_request *request;
1856 request = kzalloc(sizeof(*request), GFP_KERNEL);
1857 if (request == NULL)
1860 ret = i915_add_request(ring, NULL, request);
1866 BUG_ON(seqno != request->seqno);
1873 * __wait_seqno - wait until execution of seqno has finished
1874 * @ring: the ring expected to report seqno
1876 * @interruptible: do an interruptible wait (normally yes)
1877 * @timeout: in - how long to wait (NULL forever); out - how much time remaining
1879 * Returns 0 if the seqno was found within the alloted time. Else returns the
1880 * errno with remaining time filled in timeout argument.
1882 static int __wait_seqno(struct intel_ring_buffer *ring, u32 seqno,
1883 bool interruptible, struct timespec *timeout)
1885 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1886 struct timespec before, now, wait_time={1,0};
1887 unsigned long timeout_jiffies;
1889 bool wait_forever = true;
1891 if (i915_seqno_passed(ring->get_seqno(ring), seqno))
1894 trace_i915_gem_request_wait_begin(ring, seqno);
1896 if (timeout != NULL) {
1897 wait_time = *timeout;
1898 wait_forever = false;
1901 timeout_jiffies = timespec_to_jiffies(&wait_time);
1903 if (WARN_ON(!ring->irq_get(ring)))
1906 /* Record current time in case interrupted by signal, or wedged * */
1907 getrawmonotonic(&before);
1910 (i915_seqno_passed(ring->get_seqno(ring), seqno) || \
1911 atomic_read(&dev_priv->mm.wedged))
1914 end = wait_event_interruptible_timeout(ring->irq_queue,
1918 end = wait_event_timeout(ring->irq_queue, EXIT_COND,
1921 if (atomic_read(&dev_priv->mm.wedged))
1923 } while (end == 0 && wait_forever);
1925 getrawmonotonic(&now);
1927 ring->irq_put(ring);
1928 trace_i915_gem_request_wait_end(ring, seqno);
1932 struct timespec sleep_time = timespec_sub(now, before);
1933 *timeout = timespec_sub(*timeout, sleep_time);
1937 case -EAGAIN: /* Wedged */
1938 case -ERESTARTSYS: /* Signal */
1940 case 0: /* Timeout */
1942 set_normalized_timespec(timeout, 0, 0);
1944 default: /* Completed */
1945 WARN_ON(end < 0); /* We're not aware of other errors */
1951 * Waits for a sequence number to be signaled, and cleans up the
1952 * request and object lists appropriately for that event.
1955 i915_wait_request(struct intel_ring_buffer *ring,
1958 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1963 ret = i915_gem_check_wedge(dev_priv);
1967 ret = i915_gem_check_olr(ring, seqno);
1971 ret = __wait_seqno(ring, seqno, dev_priv->mm.interruptible, NULL);
1977 * Ensures that all rendering to the object has completed and the object is
1978 * safe to unbind from the GTT or access from the CPU.
1981 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj)
1985 /* This function only exists to support waiting for existing rendering,
1986 * not for emitting required flushes.
1988 BUG_ON((obj->base.write_domain & I915_GEM_GPU_DOMAINS) != 0);
1990 /* If there is rendering queued on the buffer being evicted, wait for
1994 ret = i915_wait_request(obj->ring, obj->last_rendering_seqno);
1997 i915_gem_retire_requests_ring(obj->ring);
2004 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
2005 * @DRM_IOCTL_ARGS: standard ioctl arguments
2007 * Returns 0 if successful, else an error is returned with the remaining time in
2008 * the timeout parameter.
2009 * -ETIME: object is still busy after timeout
2010 * -ERESTARTSYS: signal interrupted the wait
2011 * -ENONENT: object doesn't exist
2012 * Also possible, but rare:
2013 * -EAGAIN: GPU wedged
2015 * -ENODEV: Internal IRQ fail
2016 * -E?: The add request failed
2018 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
2019 * non-zero timeout parameter the wait ioctl will wait for the given number of
2020 * nanoseconds on an object becoming unbusy. Since the wait itself does so
2021 * without holding struct_mutex the object may become re-busied before this
2022 * function completes. A similar but shorter * race condition exists in the busy
2026 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
2028 struct drm_i915_gem_wait *args = data;
2029 struct drm_i915_gem_object *obj;
2030 struct intel_ring_buffer *ring = NULL;
2031 struct timespec timeout;
2035 timeout = ns_to_timespec(args->timeout_ns);
2037 ret = i915_mutex_lock_interruptible(dev);
2041 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle));
2042 if (&obj->base == NULL) {
2043 mutex_unlock(&dev->struct_mutex);
2047 /* Need to make sure the object is flushed first. This non-obvious
2048 * flush is required to enforce that (active && !olr) == no wait
2051 ret = i915_gem_object_flush_gpu_write_domain(obj);
2056 seqno = obj->last_rendering_seqno;
2063 ret = i915_gem_check_olr(ring, seqno);
2067 /* Do this after OLR check to make sure we make forward progress polling
2068 * on this IOCTL with a 0 timeout (like busy ioctl)
2070 if (!args->timeout_ns) {
2075 drm_gem_object_unreference(&obj->base);
2076 mutex_unlock(&dev->struct_mutex);
2078 ret = __wait_seqno(ring, seqno, true, &timeout);
2079 WARN_ON(!timespec_valid(&timeout));
2080 args->timeout_ns = timespec_to_ns(&timeout);
2084 drm_gem_object_unreference(&obj->base);
2085 mutex_unlock(&dev->struct_mutex);
2090 * i915_gem_object_sync - sync an object to a ring.
2092 * @obj: object which may be in use on another ring.
2093 * @to: ring we wish to use the object on. May be NULL.
2095 * This code is meant to abstract object synchronization with the GPU.
2096 * Calling with NULL implies synchronizing the object with the CPU
2097 * rather than a particular GPU ring.
2099 * Returns 0 if successful, else propagates up the lower layer error.
2102 i915_gem_object_sync(struct drm_i915_gem_object *obj,
2103 struct intel_ring_buffer *to)
2105 struct intel_ring_buffer *from = obj->ring;
2109 if (from == NULL || to == from)
2112 if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev))
2113 return i915_gem_object_wait_rendering(obj);
2115 idx = intel_ring_sync_index(from, to);
2117 seqno = obj->last_rendering_seqno;
2118 if (seqno <= from->sync_seqno[idx])
2121 ret = i915_gem_check_olr(obj->ring, seqno);
2125 ret = to->sync_to(to, from, seqno);
2127 from->sync_seqno[idx] = seqno;
2132 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
2134 u32 old_write_domain, old_read_domains;
2136 /* Act a barrier for all accesses through the GTT */
2139 /* Force a pagefault for domain tracking on next user access */
2140 i915_gem_release_mmap(obj);
2142 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
2145 old_read_domains = obj->base.read_domains;
2146 old_write_domain = obj->base.write_domain;
2148 obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
2149 obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
2151 trace_i915_gem_object_change_domain(obj,
2157 * Unbinds an object from the GTT aperture.
2160 i915_gem_object_unbind(struct drm_i915_gem_object *obj)
2162 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
2165 if (obj->gtt_space == NULL)
2168 if (obj->pin_count != 0) {
2169 DRM_ERROR("Attempting to unbind pinned buffer\n");
2173 ret = i915_gem_object_finish_gpu(obj);
2176 /* Continue on if we fail due to EIO, the GPU is hung so we
2177 * should be safe and we need to cleanup or else we might
2178 * cause memory corruption through use-after-free.
2181 i915_gem_object_finish_gtt(obj);
2183 /* Move the object to the CPU domain to ensure that
2184 * any possible CPU writes while it's not in the GTT
2185 * are flushed when we go to remap it.
2188 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
2189 if (ret == -ERESTARTSYS)
2192 /* In the event of a disaster, abandon all caches and
2193 * hope for the best.
2195 i915_gem_clflush_object(obj);
2196 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
2199 /* release the fence reg _after_ flushing */
2200 ret = i915_gem_object_put_fence(obj);
2204 trace_i915_gem_object_unbind(obj);
2206 if (obj->has_global_gtt_mapping)
2207 i915_gem_gtt_unbind_object(obj);
2208 if (obj->has_aliasing_ppgtt_mapping) {
2209 i915_ppgtt_unbind_object(dev_priv->mm.aliasing_ppgtt, obj);
2210 obj->has_aliasing_ppgtt_mapping = 0;
2212 i915_gem_gtt_finish_object(obj);
2214 i915_gem_object_put_pages_gtt(obj);
2216 list_del_init(&obj->gtt_list);
2217 list_del_init(&obj->mm_list);
2218 /* Avoid an unnecessary call to unbind on rebind. */
2219 obj->map_and_fenceable = true;
2221 drm_mm_put_block(obj->gtt_space);
2222 obj->gtt_space = NULL;
2223 obj->gtt_offset = 0;
2225 if (i915_gem_object_is_purgeable(obj))
2226 i915_gem_object_truncate(obj);
2232 i915_gem_flush_ring(struct intel_ring_buffer *ring,
2233 uint32_t invalidate_domains,
2234 uint32_t flush_domains)
2238 if (((invalidate_domains | flush_domains) & I915_GEM_GPU_DOMAINS) == 0)
2241 trace_i915_gem_ring_flush(ring, invalidate_domains, flush_domains);
2243 ret = ring->flush(ring, invalidate_domains, flush_domains);
2247 if (flush_domains & I915_GEM_GPU_DOMAINS)
2248 i915_gem_process_flushing_list(ring, flush_domains);
2253 static int i915_ring_idle(struct intel_ring_buffer *ring)
2257 if (list_empty(&ring->gpu_write_list) && list_empty(&ring->active_list))
2260 if (!list_empty(&ring->gpu_write_list)) {
2261 ret = i915_gem_flush_ring(ring,
2262 I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
2267 return i915_wait_request(ring, i915_gem_next_request_seqno(ring));
2270 int i915_gpu_idle(struct drm_device *dev)
2272 drm_i915_private_t *dev_priv = dev->dev_private;
2273 struct intel_ring_buffer *ring;
2276 /* Flush everything onto the inactive list. */
2277 for_each_ring(ring, dev_priv, i) {
2278 ret = i915_ring_idle(ring);
2282 /* Is the device fubar? */
2283 if (WARN_ON(!list_empty(&ring->gpu_write_list)))
2290 static void sandybridge_write_fence_reg(struct drm_device *dev, int reg,
2291 struct drm_i915_gem_object *obj)
2293 drm_i915_private_t *dev_priv = dev->dev_private;
2297 u32 size = obj->gtt_space->size;
2299 val = (uint64_t)((obj->gtt_offset + size - 4096) &
2301 val |= obj->gtt_offset & 0xfffff000;
2302 val |= (uint64_t)((obj->stride / 128) - 1) <<
2303 SANDYBRIDGE_FENCE_PITCH_SHIFT;
2305 if (obj->tiling_mode == I915_TILING_Y)
2306 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2307 val |= I965_FENCE_REG_VALID;
2311 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + reg * 8, val);
2312 POSTING_READ(FENCE_REG_SANDYBRIDGE_0 + reg * 8);
2315 static void i965_write_fence_reg(struct drm_device *dev, int reg,
2316 struct drm_i915_gem_object *obj)
2318 drm_i915_private_t *dev_priv = dev->dev_private;
2322 u32 size = obj->gtt_space->size;
2324 val = (uint64_t)((obj->gtt_offset + size - 4096) &
2326 val |= obj->gtt_offset & 0xfffff000;
2327 val |= ((obj->stride / 128) - 1) << I965_FENCE_PITCH_SHIFT;
2328 if (obj->tiling_mode == I915_TILING_Y)
2329 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2330 val |= I965_FENCE_REG_VALID;
2334 I915_WRITE64(FENCE_REG_965_0 + reg * 8, val);
2335 POSTING_READ(FENCE_REG_965_0 + reg * 8);
2338 static void i915_write_fence_reg(struct drm_device *dev, int reg,
2339 struct drm_i915_gem_object *obj)
2341 drm_i915_private_t *dev_priv = dev->dev_private;
2345 u32 size = obj->gtt_space->size;
2349 WARN((obj->gtt_offset & ~I915_FENCE_START_MASK) ||
2350 (size & -size) != size ||
2351 (obj->gtt_offset & (size - 1)),
2352 "object 0x%08x [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
2353 obj->gtt_offset, obj->map_and_fenceable, size);
2355 if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
2360 /* Note: pitch better be a power of two tile widths */
2361 pitch_val = obj->stride / tile_width;
2362 pitch_val = ffs(pitch_val) - 1;
2364 val = obj->gtt_offset;
2365 if (obj->tiling_mode == I915_TILING_Y)
2366 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2367 val |= I915_FENCE_SIZE_BITS(size);
2368 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2369 val |= I830_FENCE_REG_VALID;
2374 reg = FENCE_REG_830_0 + reg * 4;
2376 reg = FENCE_REG_945_8 + (reg - 8) * 4;
2378 I915_WRITE(reg, val);
2382 static void i830_write_fence_reg(struct drm_device *dev, int reg,
2383 struct drm_i915_gem_object *obj)
2385 drm_i915_private_t *dev_priv = dev->dev_private;
2389 u32 size = obj->gtt_space->size;
2392 WARN((obj->gtt_offset & ~I830_FENCE_START_MASK) ||
2393 (size & -size) != size ||
2394 (obj->gtt_offset & (size - 1)),
2395 "object 0x%08x not 512K or pot-size 0x%08x aligned\n",
2396 obj->gtt_offset, size);
2398 pitch_val = obj->stride / 128;
2399 pitch_val = ffs(pitch_val) - 1;
2401 val = obj->gtt_offset;
2402 if (obj->tiling_mode == I915_TILING_Y)
2403 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2404 val |= I830_FENCE_SIZE_BITS(size);
2405 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2406 val |= I830_FENCE_REG_VALID;
2410 I915_WRITE(FENCE_REG_830_0 + reg * 4, val);
2411 POSTING_READ(FENCE_REG_830_0 + reg * 4);
2414 static void i915_gem_write_fence(struct drm_device *dev, int reg,
2415 struct drm_i915_gem_object *obj)
2417 switch (INTEL_INFO(dev)->gen) {
2419 case 6: sandybridge_write_fence_reg(dev, reg, obj); break;
2421 case 4: i965_write_fence_reg(dev, reg, obj); break;
2422 case 3: i915_write_fence_reg(dev, reg, obj); break;
2423 case 2: i830_write_fence_reg(dev, reg, obj); break;
2428 static inline int fence_number(struct drm_i915_private *dev_priv,
2429 struct drm_i915_fence_reg *fence)
2431 return fence - dev_priv->fence_regs;
2434 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
2435 struct drm_i915_fence_reg *fence,
2438 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2439 int reg = fence_number(dev_priv, fence);
2441 i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL);
2444 obj->fence_reg = reg;
2446 list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list);
2448 obj->fence_reg = I915_FENCE_REG_NONE;
2450 list_del_init(&fence->lru_list);
2455 i915_gem_object_flush_fence(struct drm_i915_gem_object *obj)
2459 if (obj->fenced_gpu_access) {
2460 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
2461 ret = i915_gem_flush_ring(obj->ring,
2462 0, obj->base.write_domain);
2467 obj->fenced_gpu_access = false;
2470 if (obj->last_fenced_seqno) {
2471 ret = i915_wait_request(obj->ring, obj->last_fenced_seqno);
2475 obj->last_fenced_seqno = 0;
2478 /* Ensure that all CPU reads are completed before installing a fence
2479 * and all writes before removing the fence.
2481 if (obj->base.read_domains & I915_GEM_DOMAIN_GTT)
2488 i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
2490 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2493 ret = i915_gem_object_flush_fence(obj);
2497 if (obj->fence_reg == I915_FENCE_REG_NONE)
2500 i915_gem_object_update_fence(obj,
2501 &dev_priv->fence_regs[obj->fence_reg],
2503 i915_gem_object_fence_lost(obj);
2508 static struct drm_i915_fence_reg *
2509 i915_find_fence_reg(struct drm_device *dev)
2511 struct drm_i915_private *dev_priv = dev->dev_private;
2512 struct drm_i915_fence_reg *reg, *avail;
2515 /* First try to find a free reg */
2517 for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2518 reg = &dev_priv->fence_regs[i];
2522 if (!reg->pin_count)
2529 /* None available, try to steal one or wait for a user to finish */
2530 list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
2541 * i915_gem_object_get_fence - set up fencing for an object
2542 * @obj: object to map through a fence reg
2544 * When mapping objects through the GTT, userspace wants to be able to write
2545 * to them without having to worry about swizzling if the object is tiled.
2546 * This function walks the fence regs looking for a free one for @obj,
2547 * stealing one if it can't find any.
2549 * It then sets up the reg based on the object's properties: address, pitch
2550 * and tiling format.
2552 * For an untiled surface, this removes any existing fence.
2555 i915_gem_object_get_fence(struct drm_i915_gem_object *obj)
2557 struct drm_device *dev = obj->base.dev;
2558 struct drm_i915_private *dev_priv = dev->dev_private;
2559 bool enable = obj->tiling_mode != I915_TILING_NONE;
2560 struct drm_i915_fence_reg *reg;
2563 /* Have we updated the tiling parameters upon the object and so
2564 * will need to serialise the write to the associated fence register?
2566 if (obj->fence_dirty) {
2567 ret = i915_gem_object_flush_fence(obj);
2572 /* Just update our place in the LRU if our fence is getting reused. */
2573 if (obj->fence_reg != I915_FENCE_REG_NONE) {
2574 reg = &dev_priv->fence_regs[obj->fence_reg];
2575 if (!obj->fence_dirty) {
2576 list_move_tail(®->lru_list,
2577 &dev_priv->mm.fence_list);
2580 } else if (enable) {
2581 reg = i915_find_fence_reg(dev);
2586 struct drm_i915_gem_object *old = reg->obj;
2588 ret = i915_gem_object_flush_fence(old);
2592 i915_gem_object_fence_lost(old);
2597 i915_gem_object_update_fence(obj, reg, enable);
2598 obj->fence_dirty = false;
2604 * Finds free space in the GTT aperture and binds the object there.
2607 i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
2609 bool map_and_fenceable)
2611 struct drm_device *dev = obj->base.dev;
2612 drm_i915_private_t *dev_priv = dev->dev_private;
2613 struct drm_mm_node *free_space;
2614 gfp_t gfpmask = __GFP_NORETRY | __GFP_NOWARN;
2615 u32 size, fence_size, fence_alignment, unfenced_alignment;
2616 bool mappable, fenceable;
2619 if (obj->madv != I915_MADV_WILLNEED) {
2620 DRM_ERROR("Attempting to bind a purgeable object\n");
2624 fence_size = i915_gem_get_gtt_size(dev,
2627 fence_alignment = i915_gem_get_gtt_alignment(dev,
2630 unfenced_alignment =
2631 i915_gem_get_unfenced_gtt_alignment(dev,
2636 alignment = map_and_fenceable ? fence_alignment :
2638 if (map_and_fenceable && alignment & (fence_alignment - 1)) {
2639 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
2643 size = map_and_fenceable ? fence_size : obj->base.size;
2645 /* If the object is bigger than the entire aperture, reject it early
2646 * before evicting everything in a vain attempt to find space.
2648 if (obj->base.size >
2649 (map_and_fenceable ? dev_priv->mm.gtt_mappable_end : dev_priv->mm.gtt_total)) {
2650 DRM_ERROR("Attempting to bind an object larger than the aperture\n");
2655 if (map_and_fenceable)
2657 drm_mm_search_free_in_range(&dev_priv->mm.gtt_space,
2659 dev_priv->mm.gtt_mappable_end,
2662 free_space = drm_mm_search_free(&dev_priv->mm.gtt_space,
2663 size, alignment, 0);
2665 if (free_space != NULL) {
2666 if (map_and_fenceable)
2668 drm_mm_get_block_range_generic(free_space,
2670 dev_priv->mm.gtt_mappable_end,
2674 drm_mm_get_block(free_space, size, alignment);
2676 if (obj->gtt_space == NULL) {
2677 /* If the gtt is empty and we're still having trouble
2678 * fitting our object in, we're out of memory.
2680 ret = i915_gem_evict_something(dev, size, alignment,
2688 ret = i915_gem_object_get_pages_gtt(obj, gfpmask);
2690 drm_mm_put_block(obj->gtt_space);
2691 obj->gtt_space = NULL;
2693 if (ret == -ENOMEM) {
2694 /* first try to reclaim some memory by clearing the GTT */
2695 ret = i915_gem_evict_everything(dev, false);
2697 /* now try to shrink everyone else */
2712 ret = i915_gem_gtt_prepare_object(obj);
2714 i915_gem_object_put_pages_gtt(obj);
2715 drm_mm_put_block(obj->gtt_space);
2716 obj->gtt_space = NULL;
2718 if (i915_gem_evict_everything(dev, false))
2724 if (!dev_priv->mm.aliasing_ppgtt)
2725 i915_gem_gtt_bind_object(obj, obj->cache_level);
2727 list_add_tail(&obj->gtt_list, &dev_priv->mm.gtt_list);
2728 list_add_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
2730 /* Assert that the object is not currently in any GPU domain. As it
2731 * wasn't in the GTT, there shouldn't be any way it could have been in
2734 BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
2735 BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
2737 obj->gtt_offset = obj->gtt_space->start;
2740 obj->gtt_space->size == fence_size &&
2741 (obj->gtt_space->start & (fence_alignment - 1)) == 0;
2744 obj->gtt_offset + obj->base.size <= dev_priv->mm.gtt_mappable_end;
2746 obj->map_and_fenceable = mappable && fenceable;
2748 trace_i915_gem_object_bind(obj, map_and_fenceable);
2753 i915_gem_clflush_object(struct drm_i915_gem_object *obj)
2755 /* If we don't have a page list set up, then we're not pinned
2756 * to GPU, and we can ignore the cache flush because it'll happen
2757 * again at bind time.
2759 if (obj->pages == NULL)
2762 /* If the GPU is snooping the contents of the CPU cache,
2763 * we do not need to manually clear the CPU cache lines. However,
2764 * the caches are only snooped when the render cache is
2765 * flushed/invalidated. As we always have to emit invalidations
2766 * and flushes when moving into and out of the RENDER domain, correct
2767 * snooping behaviour occurs naturally as the result of our domain
2770 if (obj->cache_level != I915_CACHE_NONE)
2773 trace_i915_gem_object_clflush(obj);
2775 drm_clflush_pages(obj->pages, obj->base.size / PAGE_SIZE);
2778 /** Flushes any GPU write domain for the object if it's dirty. */
2780 i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object *obj)
2782 if ((obj->base.write_domain & I915_GEM_GPU_DOMAINS) == 0)
2785 /* Queue the GPU write cache flushing we need. */
2786 return i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain);
2789 /** Flushes the GTT write domain for the object if it's dirty. */
2791 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
2793 uint32_t old_write_domain;
2795 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
2798 /* No actual flushing is required for the GTT write domain. Writes
2799 * to it immediately go to main memory as far as we know, so there's
2800 * no chipset flush. It also doesn't land in render cache.
2802 * However, we do have to enforce the order so that all writes through
2803 * the GTT land before any writes to the device, such as updates to
2808 old_write_domain = obj->base.write_domain;
2809 obj->base.write_domain = 0;
2811 trace_i915_gem_object_change_domain(obj,
2812 obj->base.read_domains,
2816 /** Flushes the CPU write domain for the object if it's dirty. */
2818 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj)
2820 uint32_t old_write_domain;
2822 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
2825 i915_gem_clflush_object(obj);
2826 intel_gtt_chipset_flush();
2827 old_write_domain = obj->base.write_domain;
2828 obj->base.write_domain = 0;
2830 trace_i915_gem_object_change_domain(obj,
2831 obj->base.read_domains,
2836 * Moves a single object to the GTT read, and possibly write domain.
2838 * This function returns when the move is complete, including waiting on
2842 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
2844 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
2845 uint32_t old_write_domain, old_read_domains;
2848 /* Not valid to be called on unbound objects. */
2849 if (obj->gtt_space == NULL)
2852 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
2855 ret = i915_gem_object_flush_gpu_write_domain(obj);
2859 if (obj->pending_gpu_write || write) {
2860 ret = i915_gem_object_wait_rendering(obj);
2865 i915_gem_object_flush_cpu_write_domain(obj);
2867 old_write_domain = obj->base.write_domain;
2868 old_read_domains = obj->base.read_domains;
2870 /* It should now be out of any other write domains, and we can update
2871 * the domain values for our changes.
2873 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2874 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
2876 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
2877 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
2881 trace_i915_gem_object_change_domain(obj,
2885 /* And bump the LRU for this access */
2886 if (i915_gem_object_is_inactive(obj))
2887 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
2892 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
2893 enum i915_cache_level cache_level)
2895 struct drm_device *dev = obj->base.dev;
2896 drm_i915_private_t *dev_priv = dev->dev_private;
2899 if (obj->cache_level == cache_level)
2902 if (obj->pin_count) {
2903 DRM_DEBUG("can not change the cache level of pinned objects\n");
2907 if (obj->gtt_space) {
2908 ret = i915_gem_object_finish_gpu(obj);
2912 i915_gem_object_finish_gtt(obj);
2914 /* Before SandyBridge, you could not use tiling or fence
2915 * registers with snooped memory, so relinquish any fences
2916 * currently pointing to our region in the aperture.
2918 if (INTEL_INFO(obj->base.dev)->gen < 6) {
2919 ret = i915_gem_object_put_fence(obj);
2924 if (obj->has_global_gtt_mapping)
2925 i915_gem_gtt_bind_object(obj, cache_level);
2926 if (obj->has_aliasing_ppgtt_mapping)
2927 i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt,
2931 if (cache_level == I915_CACHE_NONE) {
2932 u32 old_read_domains, old_write_domain;
2934 /* If we're coming from LLC cached, then we haven't
2935 * actually been tracking whether the data is in the
2936 * CPU cache or not, since we only allow one bit set
2937 * in obj->write_domain and have been skipping the clflushes.
2938 * Just set it to the CPU cache for now.
2940 WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU);
2941 WARN_ON(obj->base.read_domains & ~I915_GEM_DOMAIN_CPU);
2943 old_read_domains = obj->base.read_domains;
2944 old_write_domain = obj->base.write_domain;
2946 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
2947 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
2949 trace_i915_gem_object_change_domain(obj,
2954 obj->cache_level = cache_level;
2959 * Prepare buffer for display plane (scanout, cursors, etc).
2960 * Can be called from an uninterruptible phase (modesetting) and allows
2961 * any flushes to be pipelined (for pageflips).
2964 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
2966 struct intel_ring_buffer *pipelined)
2968 u32 old_read_domains, old_write_domain;
2971 ret = i915_gem_object_flush_gpu_write_domain(obj);
2975 if (pipelined != obj->ring) {
2976 ret = i915_gem_object_sync(obj, pipelined);
2981 /* The display engine is not coherent with the LLC cache on gen6. As
2982 * a result, we make sure that the pinning that is about to occur is
2983 * done with uncached PTEs. This is lowest common denominator for all
2986 * However for gen6+, we could do better by using the GFDT bit instead
2987 * of uncaching, which would allow us to flush all the LLC-cached data
2988 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
2990 ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE);
2994 /* As the user may map the buffer once pinned in the display plane
2995 * (e.g. libkms for the bootup splash), we have to ensure that we
2996 * always use map_and_fenceable for all scanout buffers.
2998 ret = i915_gem_object_pin(obj, alignment, true);
3002 i915_gem_object_flush_cpu_write_domain(obj);
3004 old_write_domain = obj->base.write_domain;
3005 old_read_domains = obj->base.read_domains;
3007 /* It should now be out of any other write domains, and we can update
3008 * the domain values for our changes.
3010 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3011 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3013 trace_i915_gem_object_change_domain(obj,
3021 i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
3025 if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
3028 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
3029 ret = i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain);
3034 ret = i915_gem_object_wait_rendering(obj);
3038 /* Ensure that we invalidate the GPU's caches and TLBs. */
3039 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
3044 * Moves a single object to the CPU read, and possibly write domain.
3046 * This function returns when the move is complete, including waiting on
3050 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
3052 uint32_t old_write_domain, old_read_domains;
3055 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
3058 ret = i915_gem_object_flush_gpu_write_domain(obj);
3062 if (write || obj->pending_gpu_write) {
3063 ret = i915_gem_object_wait_rendering(obj);
3068 i915_gem_object_flush_gtt_write_domain(obj);
3070 old_write_domain = obj->base.write_domain;
3071 old_read_domains = obj->base.read_domains;
3073 /* Flush the CPU cache if it's still invalid. */
3074 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
3075 i915_gem_clflush_object(obj);
3077 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3080 /* It should now be out of any other write domains, and we can update
3081 * the domain values for our changes.
3083 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3085 /* If we're writing through the CPU, then the GPU read domains will
3086 * need to be invalidated at next use.
3089 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3090 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3093 trace_i915_gem_object_change_domain(obj,
3100 /* Throttle our rendering by waiting until the ring has completed our requests
3101 * emitted over 20 msec ago.
3103 * Note that if we were to use the current jiffies each time around the loop,
3104 * we wouldn't escape the function with any frames outstanding if the time to
3105 * render a frame was over 20ms.
3107 * This should get us reasonable parallelism between CPU and GPU but also
3108 * relatively low latency when blocking on a particular request to finish.
3111 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
3113 struct drm_i915_private *dev_priv = dev->dev_private;
3114 struct drm_i915_file_private *file_priv = file->driver_priv;
3115 unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3116 struct drm_i915_gem_request *request;
3117 struct intel_ring_buffer *ring = NULL;
3121 if (atomic_read(&dev_priv->mm.wedged))
3124 spin_lock(&file_priv->mm.lock);
3125 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
3126 if (time_after_eq(request->emitted_jiffies, recent_enough))
3129 ring = request->ring;
3130 seqno = request->seqno;
3132 spin_unlock(&file_priv->mm.lock);
3137 ret = __wait_seqno(ring, seqno, true, NULL);
3139 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
3145 i915_gem_object_pin(struct drm_i915_gem_object *obj,
3147 bool map_and_fenceable)
3151 BUG_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT);
3153 if (obj->gtt_space != NULL) {
3154 if ((alignment && obj->gtt_offset & (alignment - 1)) ||
3155 (map_and_fenceable && !obj->map_and_fenceable)) {
3156 WARN(obj->pin_count,
3157 "bo is already pinned with incorrect alignment:"
3158 " offset=%x, req.alignment=%x, req.map_and_fenceable=%d,"
3159 " obj->map_and_fenceable=%d\n",
3160 obj->gtt_offset, alignment,
3162 obj->map_and_fenceable);
3163 ret = i915_gem_object_unbind(obj);
3169 if (obj->gtt_space == NULL) {
3170 ret = i915_gem_object_bind_to_gtt(obj, alignment,
3176 if (!obj->has_global_gtt_mapping && map_and_fenceable)
3177 i915_gem_gtt_bind_object(obj, obj->cache_level);
3180 obj->pin_mappable |= map_and_fenceable;
3186 i915_gem_object_unpin(struct drm_i915_gem_object *obj)
3188 BUG_ON(obj->pin_count == 0);
3189 BUG_ON(obj->gtt_space == NULL);
3191 if (--obj->pin_count == 0)
3192 obj->pin_mappable = false;
3196 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
3197 struct drm_file *file)
3199 struct drm_i915_gem_pin *args = data;
3200 struct drm_i915_gem_object *obj;
3203 ret = i915_mutex_lock_interruptible(dev);
3207 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3208 if (&obj->base == NULL) {
3213 if (obj->madv != I915_MADV_WILLNEED) {
3214 DRM_ERROR("Attempting to pin a purgeable buffer\n");
3219 if (obj->pin_filp != NULL && obj->pin_filp != file) {
3220 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
3226 obj->user_pin_count++;
3227 obj->pin_filp = file;
3228 if (obj->user_pin_count == 1) {
3229 ret = i915_gem_object_pin(obj, args->alignment, true);
3234 /* XXX - flush the CPU caches for pinned objects
3235 * as the X server doesn't manage domains yet
3237 i915_gem_object_flush_cpu_write_domain(obj);
3238 args->offset = obj->gtt_offset;
3240 drm_gem_object_unreference(&obj->base);
3242 mutex_unlock(&dev->struct_mutex);
3247 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
3248 struct drm_file *file)
3250 struct drm_i915_gem_pin *args = data;
3251 struct drm_i915_gem_object *obj;
3254 ret = i915_mutex_lock_interruptible(dev);
3258 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3259 if (&obj->base == NULL) {
3264 if (obj->pin_filp != file) {
3265 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
3270 obj->user_pin_count--;
3271 if (obj->user_pin_count == 0) {
3272 obj->pin_filp = NULL;
3273 i915_gem_object_unpin(obj);
3277 drm_gem_object_unreference(&obj->base);
3279 mutex_unlock(&dev->struct_mutex);
3284 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
3285 struct drm_file *file)
3287 struct drm_i915_gem_busy *args = data;
3288 struct drm_i915_gem_object *obj;
3291 ret = i915_mutex_lock_interruptible(dev);
3295 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3296 if (&obj->base == NULL) {
3301 /* Count all active objects as busy, even if they are currently not used
3302 * by the gpu. Users of this interface expect objects to eventually
3303 * become non-busy without any further actions, therefore emit any
3304 * necessary flushes here.
3306 args->busy = obj->active;
3308 /* Unconditionally flush objects, even when the gpu still uses this
3309 * object. Userspace calling this function indicates that it wants to
3310 * use this buffer rather sooner than later, so issuing the required
3311 * flush earlier is beneficial.
3313 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
3314 ret = i915_gem_flush_ring(obj->ring,
3315 0, obj->base.write_domain);
3317 ret = i915_gem_check_olr(obj->ring,
3318 obj->last_rendering_seqno);
3321 /* Update the active list for the hardware's current position.
3322 * Otherwise this only updates on a delayed timer or when irqs
3323 * are actually unmasked, and our working set ends up being
3324 * larger than required.
3326 i915_gem_retire_requests_ring(obj->ring);
3328 args->busy = obj->active;
3331 drm_gem_object_unreference(&obj->base);
3333 mutex_unlock(&dev->struct_mutex);
3338 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
3339 struct drm_file *file_priv)
3341 return i915_gem_ring_throttle(dev, file_priv);
3345 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
3346 struct drm_file *file_priv)
3348 struct drm_i915_gem_madvise *args = data;
3349 struct drm_i915_gem_object *obj;
3352 switch (args->madv) {
3353 case I915_MADV_DONTNEED:
3354 case I915_MADV_WILLNEED:
3360 ret = i915_mutex_lock_interruptible(dev);
3364 obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
3365 if (&obj->base == NULL) {
3370 if (obj->pin_count) {
3375 if (obj->madv != __I915_MADV_PURGED)
3376 obj->madv = args->madv;
3378 /* if the object is no longer bound, discard its backing storage */
3379 if (i915_gem_object_is_purgeable(obj) &&
3380 obj->gtt_space == NULL)
3381 i915_gem_object_truncate(obj);
3383 args->retained = obj->madv != __I915_MADV_PURGED;
3386 drm_gem_object_unreference(&obj->base);
3388 mutex_unlock(&dev->struct_mutex);
3392 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
3395 struct drm_i915_private *dev_priv = dev->dev_private;
3396 struct drm_i915_gem_object *obj;
3397 struct address_space *mapping;
3399 obj = kzalloc(sizeof(*obj), GFP_KERNEL);
3403 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
3408 mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
3409 mapping_set_gfp_mask(mapping, GFP_HIGHUSER | __GFP_RECLAIMABLE);
3411 i915_gem_info_add_obj(dev_priv, size);
3413 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3414 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3417 /* On some devices, we can have the GPU use the LLC (the CPU
3418 * cache) for about a 10% performance improvement
3419 * compared to uncached. Graphics requests other than
3420 * display scanout are coherent with the CPU in
3421 * accessing this cache. This means in this mode we
3422 * don't need to clflush on the CPU side, and on the
3423 * GPU side we only need to flush internal caches to
3424 * get data visible to the CPU.
3426 * However, we maintain the display planes as UC, and so
3427 * need to rebind when first used as such.
3429 obj->cache_level = I915_CACHE_LLC;
3431 obj->cache_level = I915_CACHE_NONE;
3433 obj->base.driver_private = NULL;
3434 obj->fence_reg = I915_FENCE_REG_NONE;
3435 INIT_LIST_HEAD(&obj->mm_list);
3436 INIT_LIST_HEAD(&obj->gtt_list);
3437 INIT_LIST_HEAD(&obj->ring_list);
3438 INIT_LIST_HEAD(&obj->exec_list);
3439 INIT_LIST_HEAD(&obj->gpu_write_list);
3440 obj->madv = I915_MADV_WILLNEED;
3441 /* Avoid an unnecessary call to unbind on the first bind. */
3442 obj->map_and_fenceable = true;
3447 int i915_gem_init_object(struct drm_gem_object *obj)
3454 void i915_gem_free_object(struct drm_gem_object *gem_obj)
3456 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
3457 struct drm_device *dev = obj->base.dev;
3458 drm_i915_private_t *dev_priv = dev->dev_private;
3460 trace_i915_gem_object_destroy(obj);
3463 i915_gem_detach_phys_object(dev, obj);
3466 if (WARN_ON(i915_gem_object_unbind(obj) == -ERESTARTSYS)) {
3467 bool was_interruptible;
3469 was_interruptible = dev_priv->mm.interruptible;
3470 dev_priv->mm.interruptible = false;
3472 WARN_ON(i915_gem_object_unbind(obj));
3474 dev_priv->mm.interruptible = was_interruptible;
3477 if (obj->base.map_list.map)
3478 drm_gem_free_mmap_offset(&obj->base);
3480 drm_gem_object_release(&obj->base);
3481 i915_gem_info_remove_obj(dev_priv, obj->base.size);
3488 i915_gem_idle(struct drm_device *dev)
3490 drm_i915_private_t *dev_priv = dev->dev_private;
3493 mutex_lock(&dev->struct_mutex);
3495 if (dev_priv->mm.suspended) {
3496 mutex_unlock(&dev->struct_mutex);
3500 ret = i915_gpu_idle(dev);
3502 mutex_unlock(&dev->struct_mutex);
3505 i915_gem_retire_requests(dev);
3507 /* Under UMS, be paranoid and evict. */
3508 if (!drm_core_check_feature(dev, DRIVER_MODESET))
3509 i915_gem_evict_everything(dev, false);
3511 i915_gem_reset_fences(dev);
3513 /* Hack! Don't let anybody do execbuf while we don't control the chip.
3514 * We need to replace this with a semaphore, or something.
3515 * And not confound mm.suspended!
3517 dev_priv->mm.suspended = 1;
3518 del_timer_sync(&dev_priv->hangcheck_timer);
3520 i915_kernel_lost_context(dev);
3521 i915_gem_cleanup_ringbuffer(dev);
3523 mutex_unlock(&dev->struct_mutex);
3525 /* Cancel the retire work handler, which should be idle now. */
3526 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
3531 void i915_gem_init_swizzling(struct drm_device *dev)
3533 drm_i915_private_t *dev_priv = dev->dev_private;
3535 if (INTEL_INFO(dev)->gen < 5 ||
3536 dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
3539 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
3540 DISP_TILE_SURFACE_SWIZZLING);
3545 I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
3547 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
3549 I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
3552 void i915_gem_init_ppgtt(struct drm_device *dev)
3554 drm_i915_private_t *dev_priv = dev->dev_private;
3556 struct intel_ring_buffer *ring;
3557 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
3558 uint32_t __iomem *pd_addr;
3562 if (!dev_priv->mm.aliasing_ppgtt)
3566 pd_addr = dev_priv->mm.gtt->gtt + ppgtt->pd_offset/sizeof(uint32_t);
3567 for (i = 0; i < ppgtt->num_pd_entries; i++) {
3570 if (dev_priv->mm.gtt->needs_dmar)
3571 pt_addr = ppgtt->pt_dma_addr[i];
3573 pt_addr = page_to_phys(ppgtt->pt_pages[i]);
3575 pd_entry = GEN6_PDE_ADDR_ENCODE(pt_addr);
3576 pd_entry |= GEN6_PDE_VALID;
3578 writel(pd_entry, pd_addr + i);
3582 pd_offset = ppgtt->pd_offset;
3583 pd_offset /= 64; /* in cachelines, */
3586 if (INTEL_INFO(dev)->gen == 6) {
3587 uint32_t ecochk, gab_ctl, ecobits;
3589 ecobits = I915_READ(GAC_ECO_BITS);
3590 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
3592 gab_ctl = I915_READ(GAB_CTL);
3593 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
3595 ecochk = I915_READ(GAM_ECOCHK);
3596 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT |
3597 ECOCHK_PPGTT_CACHE64B);
3598 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
3599 } else if (INTEL_INFO(dev)->gen >= 7) {
3600 I915_WRITE(GAM_ECOCHK, ECOCHK_PPGTT_CACHE64B);
3601 /* GFX_MODE is per-ring on gen7+ */
3604 for_each_ring(ring, dev_priv, i) {
3605 if (INTEL_INFO(dev)->gen >= 7)
3606 I915_WRITE(RING_MODE_GEN7(ring),
3607 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
3609 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
3610 I915_WRITE(RING_PP_DIR_BASE(ring), pd_offset);
3615 i915_gem_init_hw(struct drm_device *dev)
3617 drm_i915_private_t *dev_priv = dev->dev_private;
3620 i915_gem_init_swizzling(dev);
3622 ret = intel_init_render_ring_buffer(dev);
3627 ret = intel_init_bsd_ring_buffer(dev);
3629 goto cleanup_render_ring;
3633 ret = intel_init_blt_ring_buffer(dev);
3635 goto cleanup_bsd_ring;
3638 dev_priv->next_seqno = 1;
3640 i915_gem_init_ppgtt(dev);
3645 intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
3646 cleanup_render_ring:
3647 intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
3652 intel_enable_ppgtt(struct drm_device *dev)
3654 if (i915_enable_ppgtt >= 0)
3655 return i915_enable_ppgtt;
3657 #ifdef CONFIG_INTEL_IOMMU
3658 /* Disable ppgtt on SNB if VT-d is on. */
3659 if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped)
3666 int i915_gem_init(struct drm_device *dev)
3668 struct drm_i915_private *dev_priv = dev->dev_private;
3669 unsigned long gtt_size, mappable_size;
3672 gtt_size = dev_priv->mm.gtt->gtt_total_entries << PAGE_SHIFT;
3673 mappable_size = dev_priv->mm.gtt->gtt_mappable_entries << PAGE_SHIFT;
3675 mutex_lock(&dev->struct_mutex);
3676 if (intel_enable_ppgtt(dev) && HAS_ALIASING_PPGTT(dev)) {
3677 /* PPGTT pdes are stolen from global gtt ptes, so shrink the
3678 * aperture accordingly when using aliasing ppgtt. */
3679 gtt_size -= I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
3681 i915_gem_init_global_gtt(dev, 0, mappable_size, gtt_size);
3683 ret = i915_gem_init_aliasing_ppgtt(dev);
3685 mutex_unlock(&dev->struct_mutex);
3689 /* Let GEM Manage all of the aperture.
3691 * However, leave one page at the end still bound to the scratch
3692 * page. There are a number of places where the hardware
3693 * apparently prefetches past the end of the object, and we've
3694 * seen multiple hangs with the GPU head pointer stuck in a
3695 * batchbuffer bound at the last page of the aperture. One page
3696 * should be enough to keep any prefetching inside of the
3699 i915_gem_init_global_gtt(dev, 0, mappable_size,
3703 ret = i915_gem_init_hw(dev);
3704 mutex_unlock(&dev->struct_mutex);
3706 i915_gem_cleanup_aliasing_ppgtt(dev);
3710 /* Allow hardware batchbuffers unless told otherwise, but not for KMS. */
3711 if (!drm_core_check_feature(dev, DRIVER_MODESET))
3712 dev_priv->dri1.allow_batchbuffer = 1;
3717 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
3719 drm_i915_private_t *dev_priv = dev->dev_private;
3720 struct intel_ring_buffer *ring;
3723 for_each_ring(ring, dev_priv, i)
3724 intel_cleanup_ring_buffer(ring);
3728 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
3729 struct drm_file *file_priv)
3731 drm_i915_private_t *dev_priv = dev->dev_private;
3734 if (drm_core_check_feature(dev, DRIVER_MODESET))
3737 if (atomic_read(&dev_priv->mm.wedged)) {
3738 DRM_ERROR("Reenabling wedged hardware, good luck\n");
3739 atomic_set(&dev_priv->mm.wedged, 0);
3742 mutex_lock(&dev->struct_mutex);
3743 dev_priv->mm.suspended = 0;
3745 ret = i915_gem_init_hw(dev);
3747 mutex_unlock(&dev->struct_mutex);
3751 BUG_ON(!list_empty(&dev_priv->mm.active_list));
3752 BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
3753 BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
3754 mutex_unlock(&dev->struct_mutex);
3756 ret = drm_irq_install(dev);
3758 goto cleanup_ringbuffer;
3763 mutex_lock(&dev->struct_mutex);
3764 i915_gem_cleanup_ringbuffer(dev);
3765 dev_priv->mm.suspended = 1;
3766 mutex_unlock(&dev->struct_mutex);
3772 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
3773 struct drm_file *file_priv)
3775 if (drm_core_check_feature(dev, DRIVER_MODESET))
3778 drm_irq_uninstall(dev);
3779 return i915_gem_idle(dev);
3783 i915_gem_lastclose(struct drm_device *dev)
3787 if (drm_core_check_feature(dev, DRIVER_MODESET))
3790 ret = i915_gem_idle(dev);
3792 DRM_ERROR("failed to idle hardware: %d\n", ret);
3796 init_ring_lists(struct intel_ring_buffer *ring)
3798 INIT_LIST_HEAD(&ring->active_list);
3799 INIT_LIST_HEAD(&ring->request_list);
3800 INIT_LIST_HEAD(&ring->gpu_write_list);
3804 i915_gem_load(struct drm_device *dev)
3807 drm_i915_private_t *dev_priv = dev->dev_private;
3809 INIT_LIST_HEAD(&dev_priv->mm.active_list);
3810 INIT_LIST_HEAD(&dev_priv->mm.flushing_list);
3811 INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
3812 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
3813 INIT_LIST_HEAD(&dev_priv->mm.gtt_list);
3814 for (i = 0; i < I915_NUM_RINGS; i++)
3815 init_ring_lists(&dev_priv->ring[i]);
3816 for (i = 0; i < I915_MAX_NUM_FENCES; i++)
3817 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
3818 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
3819 i915_gem_retire_work_handler);
3820 init_completion(&dev_priv->error_completion);
3822 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
3824 I915_WRITE(MI_ARB_STATE,
3825 _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
3828 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
3830 /* Old X drivers will take 0-2 for front, back, depth buffers */
3831 if (!drm_core_check_feature(dev, DRIVER_MODESET))
3832 dev_priv->fence_reg_start = 3;
3834 if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
3835 dev_priv->num_fence_regs = 16;
3837 dev_priv->num_fence_regs = 8;
3839 /* Initialize fence registers to zero */
3840 i915_gem_reset_fences(dev);
3842 i915_gem_detect_bit_6_swizzle(dev);
3843 init_waitqueue_head(&dev_priv->pending_flip_queue);
3845 dev_priv->mm.interruptible = true;
3847 dev_priv->mm.inactive_shrinker.shrink = i915_gem_inactive_shrink;
3848 dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS;
3849 register_shrinker(&dev_priv->mm.inactive_shrinker);
3853 * Create a physically contiguous memory object for this object
3854 * e.g. for cursor + overlay regs
3856 static int i915_gem_init_phys_object(struct drm_device *dev,
3857 int id, int size, int align)
3859 drm_i915_private_t *dev_priv = dev->dev_private;
3860 struct drm_i915_gem_phys_object *phys_obj;
3863 if (dev_priv->mm.phys_objs[id - 1] || !size)
3866 phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL);
3872 phys_obj->handle = drm_pci_alloc(dev, size, align);
3873 if (!phys_obj->handle) {
3878 set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
3881 dev_priv->mm.phys_objs[id - 1] = phys_obj;
3889 static void i915_gem_free_phys_object(struct drm_device *dev, int id)
3891 drm_i915_private_t *dev_priv = dev->dev_private;
3892 struct drm_i915_gem_phys_object *phys_obj;
3894 if (!dev_priv->mm.phys_objs[id - 1])
3897 phys_obj = dev_priv->mm.phys_objs[id - 1];
3898 if (phys_obj->cur_obj) {
3899 i915_gem_detach_phys_object(dev, phys_obj->cur_obj);
3903 set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
3905 drm_pci_free(dev, phys_obj->handle);
3907 dev_priv->mm.phys_objs[id - 1] = NULL;
3910 void i915_gem_free_all_phys_object(struct drm_device *dev)
3914 for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++)
3915 i915_gem_free_phys_object(dev, i);
3918 void i915_gem_detach_phys_object(struct drm_device *dev,
3919 struct drm_i915_gem_object *obj)
3921 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
3928 vaddr = obj->phys_obj->handle->vaddr;
3930 page_count = obj->base.size / PAGE_SIZE;
3931 for (i = 0; i < page_count; i++) {
3932 struct page *page = shmem_read_mapping_page(mapping, i);
3933 if (!IS_ERR(page)) {
3934 char *dst = kmap_atomic(page);
3935 memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE);
3938 drm_clflush_pages(&page, 1);
3940 set_page_dirty(page);
3941 mark_page_accessed(page);
3942 page_cache_release(page);
3945 intel_gtt_chipset_flush();
3947 obj->phys_obj->cur_obj = NULL;
3948 obj->phys_obj = NULL;
3952 i915_gem_attach_phys_object(struct drm_device *dev,
3953 struct drm_i915_gem_object *obj,
3957 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
3958 drm_i915_private_t *dev_priv = dev->dev_private;
3963 if (id > I915_MAX_PHYS_OBJECT)
3966 if (obj->phys_obj) {
3967 if (obj->phys_obj->id == id)
3969 i915_gem_detach_phys_object(dev, obj);
3972 /* create a new object */
3973 if (!dev_priv->mm.phys_objs[id - 1]) {
3974 ret = i915_gem_init_phys_object(dev, id,
3975 obj->base.size, align);
3977 DRM_ERROR("failed to init phys object %d size: %zu\n",
3978 id, obj->base.size);
3983 /* bind to the object */
3984 obj->phys_obj = dev_priv->mm.phys_objs[id - 1];
3985 obj->phys_obj->cur_obj = obj;
3987 page_count = obj->base.size / PAGE_SIZE;
3989 for (i = 0; i < page_count; i++) {
3993 page = shmem_read_mapping_page(mapping, i);
3995 return PTR_ERR(page);
3997 src = kmap_atomic(page);
3998 dst = obj->phys_obj->handle->vaddr + (i * PAGE_SIZE);
3999 memcpy(dst, src, PAGE_SIZE);
4002 mark_page_accessed(page);
4003 page_cache_release(page);
4010 i915_gem_phys_pwrite(struct drm_device *dev,
4011 struct drm_i915_gem_object *obj,
4012 struct drm_i915_gem_pwrite *args,
4013 struct drm_file *file_priv)
4015 void *vaddr = obj->phys_obj->handle->vaddr + args->offset;
4016 char __user *user_data = (char __user *) (uintptr_t) args->data_ptr;
4018 if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
4019 unsigned long unwritten;
4021 /* The physical object once assigned is fixed for the lifetime
4022 * of the obj, so we can safely drop the lock and continue
4025 mutex_unlock(&dev->struct_mutex);
4026 unwritten = copy_from_user(vaddr, user_data, args->size);
4027 mutex_lock(&dev->struct_mutex);
4032 intel_gtt_chipset_flush();
4036 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
4038 struct drm_i915_file_private *file_priv = file->driver_priv;
4040 /* Clean up our request list when the client is going away, so that
4041 * later retire_requests won't dereference our soon-to-be-gone
4044 spin_lock(&file_priv->mm.lock);
4045 while (!list_empty(&file_priv->mm.request_list)) {
4046 struct drm_i915_gem_request *request;
4048 request = list_first_entry(&file_priv->mm.request_list,
4049 struct drm_i915_gem_request,
4051 list_del(&request->client_list);
4052 request->file_priv = NULL;
4054 spin_unlock(&file_priv->mm.lock);
4058 i915_gpu_is_active(struct drm_device *dev)
4060 drm_i915_private_t *dev_priv = dev->dev_private;
4063 lists_empty = list_empty(&dev_priv->mm.flushing_list) &&
4064 list_empty(&dev_priv->mm.active_list);
4066 return !lists_empty;
4070 i915_gem_inactive_shrink(struct shrinker *shrinker, struct shrink_control *sc)
4072 struct drm_i915_private *dev_priv =
4073 container_of(shrinker,
4074 struct drm_i915_private,
4075 mm.inactive_shrinker);
4076 struct drm_device *dev = dev_priv->dev;
4077 struct drm_i915_gem_object *obj, *next;
4078 int nr_to_scan = sc->nr_to_scan;
4081 if (!mutex_trylock(&dev->struct_mutex))
4084 /* "fast-path" to count number of available objects */
4085 if (nr_to_scan == 0) {
4087 list_for_each_entry(obj,
4088 &dev_priv->mm.inactive_list,
4091 mutex_unlock(&dev->struct_mutex);
4092 return cnt / 100 * sysctl_vfs_cache_pressure;
4096 /* first scan for clean buffers */
4097 i915_gem_retire_requests(dev);
4099 list_for_each_entry_safe(obj, next,
4100 &dev_priv->mm.inactive_list,
4102 if (i915_gem_object_is_purgeable(obj)) {
4103 if (i915_gem_object_unbind(obj) == 0 &&
4109 /* second pass, evict/count anything still on the inactive list */
4111 list_for_each_entry_safe(obj, next,
4112 &dev_priv->mm.inactive_list,
4115 i915_gem_object_unbind(obj) == 0)
4121 if (nr_to_scan && i915_gpu_is_active(dev)) {
4123 * We are desperate for pages, so as a last resort, wait
4124 * for the GPU to finish and discard whatever we can.
4125 * This has a dramatic impact to reduce the number of
4126 * OOM-killer events whilst running the GPU aggressively.
4128 if (i915_gpu_idle(dev) == 0)
4131 mutex_unlock(&dev->struct_mutex);
4132 return cnt / 100 * sysctl_vfs_cache_pressure;