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_set_to_cpu_domain(struct drm_i915_gem_object *obj,
44 static __must_check int i915_gem_object_set_cpu_read_domain_range(struct drm_i915_gem_object *obj,
47 static void i915_gem_object_set_to_full_cpu_read_domain(struct drm_i915_gem_object *obj);
48 static __must_check int i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
50 bool map_and_fenceable);
51 static void i915_gem_clear_fence_reg(struct drm_device *dev,
52 struct drm_i915_fence_reg *reg);
53 static int i915_gem_phys_pwrite(struct drm_device *dev,
54 struct drm_i915_gem_object *obj,
55 struct drm_i915_gem_pwrite *args,
56 struct drm_file *file);
57 static void i915_gem_free_object_tail(struct drm_i915_gem_object *obj);
59 static int i915_gem_inactive_shrink(struct shrinker *shrinker,
60 struct shrink_control *sc);
62 /* some bookkeeping */
63 static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
66 dev_priv->mm.object_count++;
67 dev_priv->mm.object_memory += size;
70 static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
73 dev_priv->mm.object_count--;
74 dev_priv->mm.object_memory -= size;
78 i915_gem_wait_for_error(struct drm_device *dev)
80 struct drm_i915_private *dev_priv = dev->dev_private;
81 struct completion *x = &dev_priv->error_completion;
85 if (!atomic_read(&dev_priv->mm.wedged))
88 ret = wait_for_completion_interruptible(x);
92 if (atomic_read(&dev_priv->mm.wedged)) {
93 /* GPU is hung, bump the completion count to account for
94 * the token we just consumed so that we never hit zero and
95 * end up waiting upon a subsequent completion event that
98 spin_lock_irqsave(&x->wait.lock, flags);
100 spin_unlock_irqrestore(&x->wait.lock, flags);
105 int i915_mutex_lock_interruptible(struct drm_device *dev)
109 ret = i915_gem_wait_for_error(dev);
113 ret = mutex_lock_interruptible(&dev->struct_mutex);
117 WARN_ON(i915_verify_lists(dev));
122 i915_gem_object_is_inactive(struct drm_i915_gem_object *obj)
124 return obj->gtt_space && !obj->active && obj->pin_count == 0;
127 void i915_gem_do_init(struct drm_device *dev,
129 unsigned long mappable_end,
132 drm_i915_private_t *dev_priv = dev->dev_private;
134 drm_mm_init(&dev_priv->mm.gtt_space, start, end - start);
136 dev_priv->mm.gtt_start = start;
137 dev_priv->mm.gtt_mappable_end = mappable_end;
138 dev_priv->mm.gtt_end = end;
139 dev_priv->mm.gtt_total = end - start;
140 dev_priv->mm.mappable_gtt_total = min(end, mappable_end) - start;
142 /* Take over this portion of the GTT */
143 intel_gtt_clear_range(start / PAGE_SIZE, (end-start) / PAGE_SIZE);
147 i915_gem_init_ioctl(struct drm_device *dev, void *data,
148 struct drm_file *file)
150 struct drm_i915_gem_init *args = data;
152 if (args->gtt_start >= args->gtt_end ||
153 (args->gtt_end | args->gtt_start) & (PAGE_SIZE - 1))
156 mutex_lock(&dev->struct_mutex);
157 i915_gem_do_init(dev, args->gtt_start, args->gtt_end, args->gtt_end);
158 mutex_unlock(&dev->struct_mutex);
164 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
165 struct drm_file *file)
167 struct drm_i915_private *dev_priv = dev->dev_private;
168 struct drm_i915_gem_get_aperture *args = data;
169 struct drm_i915_gem_object *obj;
172 if (!(dev->driver->driver_features & DRIVER_GEM))
176 mutex_lock(&dev->struct_mutex);
177 list_for_each_entry(obj, &dev_priv->mm.pinned_list, mm_list)
178 pinned += obj->gtt_space->size;
179 mutex_unlock(&dev->struct_mutex);
181 args->aper_size = dev_priv->mm.gtt_total;
182 args->aper_available_size = args->aper_size - pinned;
188 i915_gem_create(struct drm_file *file,
189 struct drm_device *dev,
193 struct drm_i915_gem_object *obj;
197 size = roundup(size, PAGE_SIZE);
201 /* Allocate the new object */
202 obj = i915_gem_alloc_object(dev, size);
206 ret = drm_gem_handle_create(file, &obj->base, &handle);
208 drm_gem_object_release(&obj->base);
209 i915_gem_info_remove_obj(dev->dev_private, obj->base.size);
214 /* drop reference from allocate - handle holds it now */
215 drm_gem_object_unreference(&obj->base);
216 trace_i915_gem_object_create(obj);
223 i915_gem_dumb_create(struct drm_file *file,
224 struct drm_device *dev,
225 struct drm_mode_create_dumb *args)
227 /* have to work out size/pitch and return them */
228 args->pitch = ALIGN(args->width * ((args->bpp + 7) / 8), 64);
229 args->size = args->pitch * args->height;
230 return i915_gem_create(file, dev,
231 args->size, &args->handle);
234 int i915_gem_dumb_destroy(struct drm_file *file,
235 struct drm_device *dev,
238 return drm_gem_handle_delete(file, handle);
242 * Creates a new mm object and returns a handle to it.
245 i915_gem_create_ioctl(struct drm_device *dev, void *data,
246 struct drm_file *file)
248 struct drm_i915_gem_create *args = data;
249 return i915_gem_create(file, dev,
250 args->size, &args->handle);
253 static int i915_gem_object_needs_bit17_swizzle(struct drm_i915_gem_object *obj)
255 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
257 return dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_9_10_17 &&
258 obj->tiling_mode != I915_TILING_NONE;
262 slow_shmem_copy(struct page *dst_page,
264 struct page *src_page,
268 char *dst_vaddr, *src_vaddr;
270 dst_vaddr = kmap(dst_page);
271 src_vaddr = kmap(src_page);
273 memcpy(dst_vaddr + dst_offset, src_vaddr + src_offset, length);
280 slow_shmem_bit17_copy(struct page *gpu_page,
282 struct page *cpu_page,
287 char *gpu_vaddr, *cpu_vaddr;
289 /* Use the unswizzled path if this page isn't affected. */
290 if ((page_to_phys(gpu_page) & (1 << 17)) == 0) {
292 return slow_shmem_copy(cpu_page, cpu_offset,
293 gpu_page, gpu_offset, length);
295 return slow_shmem_copy(gpu_page, gpu_offset,
296 cpu_page, cpu_offset, length);
299 gpu_vaddr = kmap(gpu_page);
300 cpu_vaddr = kmap(cpu_page);
302 /* Copy the data, XORing A6 with A17 (1). The user already knows he's
303 * XORing with the other bits (A9 for Y, A9 and A10 for X)
306 int cacheline_end = ALIGN(gpu_offset + 1, 64);
307 int this_length = min(cacheline_end - gpu_offset, length);
308 int swizzled_gpu_offset = gpu_offset ^ 64;
311 memcpy(cpu_vaddr + cpu_offset,
312 gpu_vaddr + swizzled_gpu_offset,
315 memcpy(gpu_vaddr + swizzled_gpu_offset,
316 cpu_vaddr + cpu_offset,
319 cpu_offset += this_length;
320 gpu_offset += this_length;
321 length -= this_length;
329 * This is the fast shmem pread path, which attempts to copy_from_user directly
330 * from the backing pages of the object to the user's address space. On a
331 * fault, it fails so we can fall back to i915_gem_shmem_pwrite_slow().
334 i915_gem_shmem_pread_fast(struct drm_device *dev,
335 struct drm_i915_gem_object *obj,
336 struct drm_i915_gem_pread *args,
337 struct drm_file *file)
339 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
342 char __user *user_data;
343 int page_offset, page_length;
345 user_data = (char __user *) (uintptr_t) args->data_ptr;
348 offset = args->offset;
355 /* Operation in this page
357 * page_offset = offset within page
358 * page_length = bytes to copy for this page
360 page_offset = offset_in_page(offset);
361 page_length = remain;
362 if ((page_offset + remain) > PAGE_SIZE)
363 page_length = PAGE_SIZE - page_offset;
365 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
367 return PTR_ERR(page);
369 vaddr = kmap_atomic(page);
370 ret = __copy_to_user_inatomic(user_data,
373 kunmap_atomic(vaddr);
375 mark_page_accessed(page);
376 page_cache_release(page);
380 remain -= page_length;
381 user_data += page_length;
382 offset += page_length;
389 * This is the fallback shmem pread path, which allocates temporary storage
390 * in kernel space to copy_to_user into outside of the struct_mutex, so we
391 * can copy out of the object's backing pages while holding the struct mutex
392 * and not take page faults.
395 i915_gem_shmem_pread_slow(struct drm_device *dev,
396 struct drm_i915_gem_object *obj,
397 struct drm_i915_gem_pread *args,
398 struct drm_file *file)
400 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
401 struct mm_struct *mm = current->mm;
402 struct page **user_pages;
404 loff_t offset, pinned_pages, i;
405 loff_t first_data_page, last_data_page, num_pages;
406 int shmem_page_offset;
407 int data_page_index, data_page_offset;
410 uint64_t data_ptr = args->data_ptr;
411 int do_bit17_swizzling;
415 /* Pin the user pages containing the data. We can't fault while
416 * holding the struct mutex, yet we want to hold it while
417 * dereferencing the user data.
419 first_data_page = data_ptr / PAGE_SIZE;
420 last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
421 num_pages = last_data_page - first_data_page + 1;
423 user_pages = drm_malloc_ab(num_pages, sizeof(struct page *));
424 if (user_pages == NULL)
427 mutex_unlock(&dev->struct_mutex);
428 down_read(&mm->mmap_sem);
429 pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
430 num_pages, 1, 0, user_pages, NULL);
431 up_read(&mm->mmap_sem);
432 mutex_lock(&dev->struct_mutex);
433 if (pinned_pages < num_pages) {
438 ret = i915_gem_object_set_cpu_read_domain_range(obj,
444 do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
446 offset = args->offset;
451 /* Operation in this page
453 * shmem_page_offset = offset within page in shmem file
454 * data_page_index = page number in get_user_pages return
455 * data_page_offset = offset with data_page_index page.
456 * page_length = bytes to copy for this page
458 shmem_page_offset = offset_in_page(offset);
459 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
460 data_page_offset = offset_in_page(data_ptr);
462 page_length = remain;
463 if ((shmem_page_offset + page_length) > PAGE_SIZE)
464 page_length = PAGE_SIZE - shmem_page_offset;
465 if ((data_page_offset + page_length) > PAGE_SIZE)
466 page_length = PAGE_SIZE - data_page_offset;
468 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
474 if (do_bit17_swizzling) {
475 slow_shmem_bit17_copy(page,
477 user_pages[data_page_index],
482 slow_shmem_copy(user_pages[data_page_index],
489 mark_page_accessed(page);
490 page_cache_release(page);
492 remain -= page_length;
493 data_ptr += page_length;
494 offset += page_length;
498 for (i = 0; i < pinned_pages; i++) {
499 SetPageDirty(user_pages[i]);
500 mark_page_accessed(user_pages[i]);
501 page_cache_release(user_pages[i]);
503 drm_free_large(user_pages);
509 * Reads data from the object referenced by handle.
511 * On error, the contents of *data are undefined.
514 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
515 struct drm_file *file)
517 struct drm_i915_gem_pread *args = data;
518 struct drm_i915_gem_object *obj;
524 if (!access_ok(VERIFY_WRITE,
525 (char __user *)(uintptr_t)args->data_ptr,
529 ret = fault_in_pages_writeable((char __user *)(uintptr_t)args->data_ptr,
534 ret = i915_mutex_lock_interruptible(dev);
538 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
539 if (&obj->base == NULL) {
544 /* Bounds check source. */
545 if (args->offset > obj->base.size ||
546 args->size > obj->base.size - args->offset) {
551 trace_i915_gem_object_pread(obj, args->offset, args->size);
553 ret = i915_gem_object_set_cpu_read_domain_range(obj,
560 if (!i915_gem_object_needs_bit17_swizzle(obj))
561 ret = i915_gem_shmem_pread_fast(dev, obj, args, file);
563 ret = i915_gem_shmem_pread_slow(dev, obj, args, file);
566 drm_gem_object_unreference(&obj->base);
568 mutex_unlock(&dev->struct_mutex);
572 /* This is the fast write path which cannot handle
573 * page faults in the source data
577 fast_user_write(struct io_mapping *mapping,
578 loff_t page_base, int page_offset,
579 char __user *user_data,
583 unsigned long unwritten;
585 vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
586 unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + page_offset,
588 io_mapping_unmap_atomic(vaddr_atomic);
592 /* Here's the write path which can sleep for
597 slow_kernel_write(struct io_mapping *mapping,
598 loff_t gtt_base, int gtt_offset,
599 struct page *user_page, int user_offset,
602 char __iomem *dst_vaddr;
605 dst_vaddr = io_mapping_map_wc(mapping, gtt_base);
606 src_vaddr = kmap(user_page);
608 memcpy_toio(dst_vaddr + gtt_offset,
609 src_vaddr + user_offset,
613 io_mapping_unmap(dst_vaddr);
617 * This is the fast pwrite path, where we copy the data directly from the
618 * user into the GTT, uncached.
621 i915_gem_gtt_pwrite_fast(struct drm_device *dev,
622 struct drm_i915_gem_object *obj,
623 struct drm_i915_gem_pwrite *args,
624 struct drm_file *file)
626 drm_i915_private_t *dev_priv = dev->dev_private;
628 loff_t offset, page_base;
629 char __user *user_data;
630 int page_offset, page_length;
632 user_data = (char __user *) (uintptr_t) args->data_ptr;
635 offset = obj->gtt_offset + args->offset;
638 /* Operation in this page
640 * page_base = page offset within aperture
641 * page_offset = offset within page
642 * page_length = bytes to copy for this page
644 page_base = offset & PAGE_MASK;
645 page_offset = offset_in_page(offset);
646 page_length = remain;
647 if ((page_offset + remain) > PAGE_SIZE)
648 page_length = PAGE_SIZE - page_offset;
650 /* If we get a fault while copying data, then (presumably) our
651 * source page isn't available. Return the error and we'll
652 * retry in the slow path.
654 if (fast_user_write(dev_priv->mm.gtt_mapping, page_base,
655 page_offset, user_data, page_length))
658 remain -= page_length;
659 user_data += page_length;
660 offset += page_length;
667 * This is the fallback GTT pwrite path, which uses get_user_pages to pin
668 * the memory and maps it using kmap_atomic for copying.
670 * This code resulted in x11perf -rgb10text consuming about 10% more CPU
671 * than using i915_gem_gtt_pwrite_fast on a G45 (32-bit).
674 i915_gem_gtt_pwrite_slow(struct drm_device *dev,
675 struct drm_i915_gem_object *obj,
676 struct drm_i915_gem_pwrite *args,
677 struct drm_file *file)
679 drm_i915_private_t *dev_priv = dev->dev_private;
681 loff_t gtt_page_base, offset;
682 loff_t first_data_page, last_data_page, num_pages;
683 loff_t pinned_pages, i;
684 struct page **user_pages;
685 struct mm_struct *mm = current->mm;
686 int gtt_page_offset, data_page_offset, data_page_index, page_length;
688 uint64_t data_ptr = args->data_ptr;
692 /* Pin the user pages containing the data. We can't fault while
693 * holding the struct mutex, and all of the pwrite implementations
694 * want to hold it while dereferencing the user data.
696 first_data_page = data_ptr / PAGE_SIZE;
697 last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
698 num_pages = last_data_page - first_data_page + 1;
700 user_pages = drm_malloc_ab(num_pages, sizeof(struct page *));
701 if (user_pages == NULL)
704 mutex_unlock(&dev->struct_mutex);
705 down_read(&mm->mmap_sem);
706 pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
707 num_pages, 0, 0, user_pages, NULL);
708 up_read(&mm->mmap_sem);
709 mutex_lock(&dev->struct_mutex);
710 if (pinned_pages < num_pages) {
712 goto out_unpin_pages;
715 ret = i915_gem_object_set_to_gtt_domain(obj, true);
717 goto out_unpin_pages;
719 ret = i915_gem_object_put_fence(obj);
721 goto out_unpin_pages;
723 offset = obj->gtt_offset + args->offset;
726 /* Operation in this page
728 * gtt_page_base = page offset within aperture
729 * gtt_page_offset = offset within page in aperture
730 * data_page_index = page number in get_user_pages return
731 * data_page_offset = offset with data_page_index page.
732 * page_length = bytes to copy for this page
734 gtt_page_base = offset & PAGE_MASK;
735 gtt_page_offset = offset_in_page(offset);
736 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
737 data_page_offset = offset_in_page(data_ptr);
739 page_length = remain;
740 if ((gtt_page_offset + page_length) > PAGE_SIZE)
741 page_length = PAGE_SIZE - gtt_page_offset;
742 if ((data_page_offset + page_length) > PAGE_SIZE)
743 page_length = PAGE_SIZE - data_page_offset;
745 slow_kernel_write(dev_priv->mm.gtt_mapping,
746 gtt_page_base, gtt_page_offset,
747 user_pages[data_page_index],
751 remain -= page_length;
752 offset += page_length;
753 data_ptr += page_length;
757 for (i = 0; i < pinned_pages; i++)
758 page_cache_release(user_pages[i]);
759 drm_free_large(user_pages);
765 * This is the fast shmem pwrite path, which attempts to directly
766 * copy_from_user into the kmapped pages backing the object.
769 i915_gem_shmem_pwrite_fast(struct drm_device *dev,
770 struct drm_i915_gem_object *obj,
771 struct drm_i915_gem_pwrite *args,
772 struct drm_file *file)
774 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
777 char __user *user_data;
778 int page_offset, page_length;
780 user_data = (char __user *) (uintptr_t) args->data_ptr;
783 offset = args->offset;
791 /* Operation in this page
793 * page_offset = offset within page
794 * page_length = bytes to copy for this page
796 page_offset = offset_in_page(offset);
797 page_length = remain;
798 if ((page_offset + remain) > PAGE_SIZE)
799 page_length = PAGE_SIZE - page_offset;
801 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
803 return PTR_ERR(page);
805 vaddr = kmap_atomic(page);
806 ret = __copy_from_user_inatomic(vaddr + page_offset,
809 kunmap_atomic(vaddr);
811 set_page_dirty(page);
812 mark_page_accessed(page);
813 page_cache_release(page);
815 /* If we get a fault while copying data, then (presumably) our
816 * source page isn't available. Return the error and we'll
817 * retry in the slow path.
822 remain -= page_length;
823 user_data += page_length;
824 offset += page_length;
831 * This is the fallback shmem pwrite path, which uses get_user_pages to pin
832 * the memory and maps it using kmap_atomic for copying.
834 * This avoids taking mmap_sem for faulting on the user's address while the
835 * struct_mutex is held.
838 i915_gem_shmem_pwrite_slow(struct drm_device *dev,
839 struct drm_i915_gem_object *obj,
840 struct drm_i915_gem_pwrite *args,
841 struct drm_file *file)
843 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
844 struct mm_struct *mm = current->mm;
845 struct page **user_pages;
847 loff_t offset, pinned_pages, i;
848 loff_t first_data_page, last_data_page, num_pages;
849 int shmem_page_offset;
850 int data_page_index, data_page_offset;
853 uint64_t data_ptr = args->data_ptr;
854 int do_bit17_swizzling;
858 /* Pin the user pages containing the data. We can't fault while
859 * holding the struct mutex, and all of the pwrite implementations
860 * want to hold it while dereferencing the user data.
862 first_data_page = data_ptr / PAGE_SIZE;
863 last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
864 num_pages = last_data_page - first_data_page + 1;
866 user_pages = drm_malloc_ab(num_pages, sizeof(struct page *));
867 if (user_pages == NULL)
870 mutex_unlock(&dev->struct_mutex);
871 down_read(&mm->mmap_sem);
872 pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
873 num_pages, 0, 0, user_pages, NULL);
874 up_read(&mm->mmap_sem);
875 mutex_lock(&dev->struct_mutex);
876 if (pinned_pages < num_pages) {
881 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
885 do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
887 offset = args->offset;
893 /* Operation in this page
895 * shmem_page_offset = offset within page in shmem file
896 * data_page_index = page number in get_user_pages return
897 * data_page_offset = offset with data_page_index page.
898 * page_length = bytes to copy for this page
900 shmem_page_offset = offset_in_page(offset);
901 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
902 data_page_offset = offset_in_page(data_ptr);
904 page_length = remain;
905 if ((shmem_page_offset + page_length) > PAGE_SIZE)
906 page_length = PAGE_SIZE - shmem_page_offset;
907 if ((data_page_offset + page_length) > PAGE_SIZE)
908 page_length = PAGE_SIZE - data_page_offset;
910 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
916 if (do_bit17_swizzling) {
917 slow_shmem_bit17_copy(page,
919 user_pages[data_page_index],
924 slow_shmem_copy(page,
926 user_pages[data_page_index],
931 set_page_dirty(page);
932 mark_page_accessed(page);
933 page_cache_release(page);
935 remain -= page_length;
936 data_ptr += page_length;
937 offset += page_length;
941 for (i = 0; i < pinned_pages; i++)
942 page_cache_release(user_pages[i]);
943 drm_free_large(user_pages);
949 * Writes data to the object referenced by handle.
951 * On error, the contents of the buffer that were to be modified are undefined.
954 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
955 struct drm_file *file)
957 struct drm_i915_gem_pwrite *args = data;
958 struct drm_i915_gem_object *obj;
964 if (!access_ok(VERIFY_READ,
965 (char __user *)(uintptr_t)args->data_ptr,
969 ret = fault_in_pages_readable((char __user *)(uintptr_t)args->data_ptr,
974 ret = i915_mutex_lock_interruptible(dev);
978 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
979 if (&obj->base == NULL) {
984 /* Bounds check destination. */
985 if (args->offset > obj->base.size ||
986 args->size > obj->base.size - args->offset) {
991 trace_i915_gem_object_pwrite(obj, args->offset, args->size);
993 /* We can only do the GTT pwrite on untiled buffers, as otherwise
994 * it would end up going through the fenced access, and we'll get
995 * different detiling behavior between reading and writing.
996 * pread/pwrite currently are reading and writing from the CPU
997 * perspective, requiring manual detiling by the client.
1000 ret = i915_gem_phys_pwrite(dev, obj, args, file);
1001 else if (obj->gtt_space &&
1002 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
1003 ret = i915_gem_object_pin(obj, 0, true);
1007 ret = i915_gem_object_set_to_gtt_domain(obj, true);
1011 ret = i915_gem_object_put_fence(obj);
1015 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
1017 ret = i915_gem_gtt_pwrite_slow(dev, obj, args, file);
1020 i915_gem_object_unpin(obj);
1022 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
1027 if (!i915_gem_object_needs_bit17_swizzle(obj))
1028 ret = i915_gem_shmem_pwrite_fast(dev, obj, args, file);
1030 ret = i915_gem_shmem_pwrite_slow(dev, obj, args, file);
1034 drm_gem_object_unreference(&obj->base);
1036 mutex_unlock(&dev->struct_mutex);
1041 * Called when user space prepares to use an object with the CPU, either
1042 * through the mmap ioctl's mapping or a GTT mapping.
1045 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1046 struct drm_file *file)
1048 struct drm_i915_gem_set_domain *args = data;
1049 struct drm_i915_gem_object *obj;
1050 uint32_t read_domains = args->read_domains;
1051 uint32_t write_domain = args->write_domain;
1054 if (!(dev->driver->driver_features & DRIVER_GEM))
1057 /* Only handle setting domains to types used by the CPU. */
1058 if (write_domain & I915_GEM_GPU_DOMAINS)
1061 if (read_domains & I915_GEM_GPU_DOMAINS)
1064 /* Having something in the write domain implies it's in the read
1065 * domain, and only that read domain. Enforce that in the request.
1067 if (write_domain != 0 && read_domains != write_domain)
1070 ret = i915_mutex_lock_interruptible(dev);
1074 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1075 if (&obj->base == NULL) {
1080 if (read_domains & I915_GEM_DOMAIN_GTT) {
1081 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1083 /* Silently promote "you're not bound, there was nothing to do"
1084 * to success, since the client was just asking us to
1085 * make sure everything was done.
1090 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1093 drm_gem_object_unreference(&obj->base);
1095 mutex_unlock(&dev->struct_mutex);
1100 * Called when user space has done writes to this buffer
1103 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1104 struct drm_file *file)
1106 struct drm_i915_gem_sw_finish *args = data;
1107 struct drm_i915_gem_object *obj;
1110 if (!(dev->driver->driver_features & DRIVER_GEM))
1113 ret = i915_mutex_lock_interruptible(dev);
1117 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1118 if (&obj->base == NULL) {
1123 /* Pinned buffers may be scanout, so flush the cache */
1125 i915_gem_object_flush_cpu_write_domain(obj);
1127 drm_gem_object_unreference(&obj->base);
1129 mutex_unlock(&dev->struct_mutex);
1134 * Maps the contents of an object, returning the address it is mapped
1137 * While the mapping holds a reference on the contents of the object, it doesn't
1138 * imply a ref on the object itself.
1141 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1142 struct drm_file *file)
1144 struct drm_i915_private *dev_priv = dev->dev_private;
1145 struct drm_i915_gem_mmap *args = data;
1146 struct drm_gem_object *obj;
1149 if (!(dev->driver->driver_features & DRIVER_GEM))
1152 obj = drm_gem_object_lookup(dev, file, args->handle);
1156 if (obj->size > dev_priv->mm.gtt_mappable_end) {
1157 drm_gem_object_unreference_unlocked(obj);
1161 down_write(¤t->mm->mmap_sem);
1162 addr = do_mmap(obj->filp, 0, args->size,
1163 PROT_READ | PROT_WRITE, MAP_SHARED,
1165 up_write(¤t->mm->mmap_sem);
1166 drm_gem_object_unreference_unlocked(obj);
1167 if (IS_ERR((void *)addr))
1170 args->addr_ptr = (uint64_t) addr;
1176 * i915_gem_fault - fault a page into the GTT
1177 * vma: VMA in question
1180 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1181 * from userspace. The fault handler takes care of binding the object to
1182 * the GTT (if needed), allocating and programming a fence register (again,
1183 * only if needed based on whether the old reg is still valid or the object
1184 * is tiled) and inserting a new PTE into the faulting process.
1186 * Note that the faulting process may involve evicting existing objects
1187 * from the GTT and/or fence registers to make room. So performance may
1188 * suffer if the GTT working set is large or there are few fence registers
1191 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1193 struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1194 struct drm_device *dev = obj->base.dev;
1195 drm_i915_private_t *dev_priv = dev->dev_private;
1196 pgoff_t page_offset;
1199 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1201 /* We don't use vmf->pgoff since that has the fake offset */
1202 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1205 ret = i915_mutex_lock_interruptible(dev);
1209 trace_i915_gem_object_fault(obj, page_offset, true, write);
1211 /* Now bind it into the GTT if needed */
1212 if (!obj->map_and_fenceable) {
1213 ret = i915_gem_object_unbind(obj);
1217 if (!obj->gtt_space) {
1218 ret = i915_gem_object_bind_to_gtt(obj, 0, true);
1222 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1227 if (obj->tiling_mode == I915_TILING_NONE)
1228 ret = i915_gem_object_put_fence(obj);
1230 ret = i915_gem_object_get_fence(obj, NULL);
1234 if (i915_gem_object_is_inactive(obj))
1235 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1237 obj->fault_mappable = true;
1239 pfn = ((dev->agp->base + obj->gtt_offset) >> PAGE_SHIFT) +
1242 /* Finally, remap it using the new GTT offset */
1243 ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1245 mutex_unlock(&dev->struct_mutex);
1250 /* Give the error handler a chance to run and move the
1251 * objects off the GPU active list. Next time we service the
1252 * fault, we should be able to transition the page into the
1253 * GTT without touching the GPU (and so avoid further
1254 * EIO/EGAIN). If the GPU is wedged, then there is no issue
1255 * with coherency, just lost writes.
1261 return VM_FAULT_NOPAGE;
1263 return VM_FAULT_OOM;
1265 return VM_FAULT_SIGBUS;
1270 * i915_gem_release_mmap - remove physical page mappings
1271 * @obj: obj in question
1273 * Preserve the reservation of the mmapping with the DRM core code, but
1274 * relinquish ownership of the pages back to the system.
1276 * It is vital that we remove the page mapping if we have mapped a tiled
1277 * object through the GTT and then lose the fence register due to
1278 * resource pressure. Similarly if the object has been moved out of the
1279 * aperture, than pages mapped into userspace must be revoked. Removing the
1280 * mapping will then trigger a page fault on the next user access, allowing
1281 * fixup by i915_gem_fault().
1284 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1286 if (!obj->fault_mappable)
1289 if (obj->base.dev->dev_mapping)
1290 unmap_mapping_range(obj->base.dev->dev_mapping,
1291 (loff_t)obj->base.map_list.hash.key<<PAGE_SHIFT,
1294 obj->fault_mappable = false;
1298 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1302 if (INTEL_INFO(dev)->gen >= 4 ||
1303 tiling_mode == I915_TILING_NONE)
1306 /* Previous chips need a power-of-two fence region when tiling */
1307 if (INTEL_INFO(dev)->gen == 3)
1308 gtt_size = 1024*1024;
1310 gtt_size = 512*1024;
1312 while (gtt_size < size)
1319 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1320 * @obj: object to check
1322 * Return the required GTT alignment for an object, taking into account
1323 * potential fence register mapping.
1326 i915_gem_get_gtt_alignment(struct drm_device *dev,
1331 * Minimum alignment is 4k (GTT page size), but might be greater
1332 * if a fence register is needed for the object.
1334 if (INTEL_INFO(dev)->gen >= 4 ||
1335 tiling_mode == I915_TILING_NONE)
1339 * Previous chips need to be aligned to the size of the smallest
1340 * fence register that can contain the object.
1342 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1346 * i915_gem_get_unfenced_gtt_alignment - return required GTT alignment for an
1349 * @size: size of the object
1350 * @tiling_mode: tiling mode of the object
1352 * Return the required GTT alignment for an object, only taking into account
1353 * unfenced tiled surface requirements.
1356 i915_gem_get_unfenced_gtt_alignment(struct drm_device *dev,
1361 * Minimum alignment is 4k (GTT page size) for sane hw.
1363 if (INTEL_INFO(dev)->gen >= 4 || IS_G33(dev) ||
1364 tiling_mode == I915_TILING_NONE)
1367 /* Previous hardware however needs to be aligned to a power-of-two
1368 * tile height. The simplest method for determining this is to reuse
1369 * the power-of-tile object size.
1371 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1375 i915_gem_mmap_gtt(struct drm_file *file,
1376 struct drm_device *dev,
1380 struct drm_i915_private *dev_priv = dev->dev_private;
1381 struct drm_i915_gem_object *obj;
1384 if (!(dev->driver->driver_features & DRIVER_GEM))
1387 ret = i915_mutex_lock_interruptible(dev);
1391 obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
1392 if (&obj->base == NULL) {
1397 if (obj->base.size > dev_priv->mm.gtt_mappable_end) {
1402 if (obj->madv != I915_MADV_WILLNEED) {
1403 DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1408 if (!obj->base.map_list.map) {
1409 ret = drm_gem_create_mmap_offset(&obj->base);
1414 *offset = (u64)obj->base.map_list.hash.key << PAGE_SHIFT;
1417 drm_gem_object_unreference(&obj->base);
1419 mutex_unlock(&dev->struct_mutex);
1424 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1426 * @data: GTT mapping ioctl data
1427 * @file: GEM object info
1429 * Simply returns the fake offset to userspace so it can mmap it.
1430 * The mmap call will end up in drm_gem_mmap(), which will set things
1431 * up so we can get faults in the handler above.
1433 * The fault handler will take care of binding the object into the GTT
1434 * (since it may have been evicted to make room for something), allocating
1435 * a fence register, and mapping the appropriate aperture address into
1439 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1440 struct drm_file *file)
1442 struct drm_i915_gem_mmap_gtt *args = data;
1444 if (!(dev->driver->driver_features & DRIVER_GEM))
1447 return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
1452 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj,
1456 struct address_space *mapping;
1457 struct inode *inode;
1460 /* Get the list of pages out of our struct file. They'll be pinned
1461 * at this point until we release them.
1463 page_count = obj->base.size / PAGE_SIZE;
1464 BUG_ON(obj->pages != NULL);
1465 obj->pages = drm_malloc_ab(page_count, sizeof(struct page *));
1466 if (obj->pages == NULL)
1469 inode = obj->base.filp->f_path.dentry->d_inode;
1470 mapping = inode->i_mapping;
1471 gfpmask |= mapping_gfp_mask(mapping);
1473 for (i = 0; i < page_count; i++) {
1474 page = shmem_read_mapping_page_gfp(mapping, i, gfpmask);
1478 obj->pages[i] = page;
1481 if (i915_gem_object_needs_bit17_swizzle(obj))
1482 i915_gem_object_do_bit_17_swizzle(obj);
1488 page_cache_release(obj->pages[i]);
1490 drm_free_large(obj->pages);
1492 return PTR_ERR(page);
1496 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
1498 int page_count = obj->base.size / PAGE_SIZE;
1501 BUG_ON(obj->madv == __I915_MADV_PURGED);
1503 if (i915_gem_object_needs_bit17_swizzle(obj))
1504 i915_gem_object_save_bit_17_swizzle(obj);
1506 if (obj->madv == I915_MADV_DONTNEED)
1509 for (i = 0; i < page_count; i++) {
1511 set_page_dirty(obj->pages[i]);
1513 if (obj->madv == I915_MADV_WILLNEED)
1514 mark_page_accessed(obj->pages[i]);
1516 page_cache_release(obj->pages[i]);
1520 drm_free_large(obj->pages);
1525 i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
1526 struct intel_ring_buffer *ring,
1529 struct drm_device *dev = obj->base.dev;
1530 struct drm_i915_private *dev_priv = dev->dev_private;
1532 BUG_ON(ring == NULL);
1535 /* Add a reference if we're newly entering the active list. */
1537 drm_gem_object_reference(&obj->base);
1541 /* Move from whatever list we were on to the tail of execution. */
1542 list_move_tail(&obj->mm_list, &dev_priv->mm.active_list);
1543 list_move_tail(&obj->ring_list, &ring->active_list);
1545 obj->last_rendering_seqno = seqno;
1546 if (obj->fenced_gpu_access) {
1547 struct drm_i915_fence_reg *reg;
1549 BUG_ON(obj->fence_reg == I915_FENCE_REG_NONE);
1551 obj->last_fenced_seqno = seqno;
1552 obj->last_fenced_ring = ring;
1554 reg = &dev_priv->fence_regs[obj->fence_reg];
1555 list_move_tail(®->lru_list, &dev_priv->mm.fence_list);
1560 i915_gem_object_move_off_active(struct drm_i915_gem_object *obj)
1562 list_del_init(&obj->ring_list);
1563 obj->last_rendering_seqno = 0;
1567 i915_gem_object_move_to_flushing(struct drm_i915_gem_object *obj)
1569 struct drm_device *dev = obj->base.dev;
1570 drm_i915_private_t *dev_priv = dev->dev_private;
1572 BUG_ON(!obj->active);
1573 list_move_tail(&obj->mm_list, &dev_priv->mm.flushing_list);
1575 i915_gem_object_move_off_active(obj);
1579 i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
1581 struct drm_device *dev = obj->base.dev;
1582 struct drm_i915_private *dev_priv = dev->dev_private;
1584 if (obj->pin_count != 0)
1585 list_move_tail(&obj->mm_list, &dev_priv->mm.pinned_list);
1587 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1589 BUG_ON(!list_empty(&obj->gpu_write_list));
1590 BUG_ON(!obj->active);
1593 i915_gem_object_move_off_active(obj);
1594 obj->fenced_gpu_access = false;
1597 obj->pending_gpu_write = false;
1598 drm_gem_object_unreference(&obj->base);
1600 WARN_ON(i915_verify_lists(dev));
1603 /* Immediately discard the backing storage */
1605 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
1607 struct inode *inode;
1609 /* Our goal here is to return as much of the memory as
1610 * is possible back to the system as we are called from OOM.
1611 * To do this we must instruct the shmfs to drop all of its
1612 * backing pages, *now*.
1614 inode = obj->base.filp->f_path.dentry->d_inode;
1615 shmem_truncate_range(inode, 0, (loff_t)-1);
1617 obj->madv = __I915_MADV_PURGED;
1621 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj)
1623 return obj->madv == I915_MADV_DONTNEED;
1627 i915_gem_process_flushing_list(struct intel_ring_buffer *ring,
1628 uint32_t flush_domains)
1630 struct drm_i915_gem_object *obj, *next;
1632 list_for_each_entry_safe(obj, next,
1633 &ring->gpu_write_list,
1635 if (obj->base.write_domain & flush_domains) {
1636 uint32_t old_write_domain = obj->base.write_domain;
1638 obj->base.write_domain = 0;
1639 list_del_init(&obj->gpu_write_list);
1640 i915_gem_object_move_to_active(obj, ring,
1641 i915_gem_next_request_seqno(ring));
1643 trace_i915_gem_object_change_domain(obj,
1644 obj->base.read_domains,
1651 i915_add_request(struct intel_ring_buffer *ring,
1652 struct drm_file *file,
1653 struct drm_i915_gem_request *request)
1655 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1660 BUG_ON(request == NULL);
1662 ret = ring->add_request(ring, &seqno);
1666 trace_i915_gem_request_add(ring, seqno);
1668 request->seqno = seqno;
1669 request->ring = ring;
1670 request->emitted_jiffies = jiffies;
1671 was_empty = list_empty(&ring->request_list);
1672 list_add_tail(&request->list, &ring->request_list);
1675 struct drm_i915_file_private *file_priv = file->driver_priv;
1677 spin_lock(&file_priv->mm.lock);
1678 request->file_priv = file_priv;
1679 list_add_tail(&request->client_list,
1680 &file_priv->mm.request_list);
1681 spin_unlock(&file_priv->mm.lock);
1684 ring->outstanding_lazy_request = false;
1686 if (!dev_priv->mm.suspended) {
1687 if (i915_enable_hangcheck) {
1688 mod_timer(&dev_priv->hangcheck_timer,
1690 msecs_to_jiffies(DRM_I915_HANGCHECK_PERIOD));
1693 queue_delayed_work(dev_priv->wq,
1694 &dev_priv->mm.retire_work, HZ);
1700 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
1702 struct drm_i915_file_private *file_priv = request->file_priv;
1707 spin_lock(&file_priv->mm.lock);
1708 if (request->file_priv) {
1709 list_del(&request->client_list);
1710 request->file_priv = NULL;
1712 spin_unlock(&file_priv->mm.lock);
1715 static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv,
1716 struct intel_ring_buffer *ring)
1718 while (!list_empty(&ring->request_list)) {
1719 struct drm_i915_gem_request *request;
1721 request = list_first_entry(&ring->request_list,
1722 struct drm_i915_gem_request,
1725 list_del(&request->list);
1726 i915_gem_request_remove_from_client(request);
1730 while (!list_empty(&ring->active_list)) {
1731 struct drm_i915_gem_object *obj;
1733 obj = list_first_entry(&ring->active_list,
1734 struct drm_i915_gem_object,
1737 obj->base.write_domain = 0;
1738 list_del_init(&obj->gpu_write_list);
1739 i915_gem_object_move_to_inactive(obj);
1743 static void i915_gem_reset_fences(struct drm_device *dev)
1745 struct drm_i915_private *dev_priv = dev->dev_private;
1748 for (i = 0; i < dev_priv->num_fence_regs; i++) {
1749 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
1750 struct drm_i915_gem_object *obj = reg->obj;
1755 if (obj->tiling_mode)
1756 i915_gem_release_mmap(obj);
1758 reg->obj->fence_reg = I915_FENCE_REG_NONE;
1759 reg->obj->fenced_gpu_access = false;
1760 reg->obj->last_fenced_seqno = 0;
1761 reg->obj->last_fenced_ring = NULL;
1762 i915_gem_clear_fence_reg(dev, reg);
1766 void i915_gem_reset(struct drm_device *dev)
1768 struct drm_i915_private *dev_priv = dev->dev_private;
1769 struct drm_i915_gem_object *obj;
1772 for (i = 0; i < I915_NUM_RINGS; i++)
1773 i915_gem_reset_ring_lists(dev_priv, &dev_priv->ring[i]);
1775 /* Remove anything from the flushing lists. The GPU cache is likely
1776 * to be lost on reset along with the data, so simply move the
1777 * lost bo to the inactive list.
1779 while (!list_empty(&dev_priv->mm.flushing_list)) {
1780 obj = list_first_entry(&dev_priv->mm.flushing_list,
1781 struct drm_i915_gem_object,
1784 obj->base.write_domain = 0;
1785 list_del_init(&obj->gpu_write_list);
1786 i915_gem_object_move_to_inactive(obj);
1789 /* Move everything out of the GPU domains to ensure we do any
1790 * necessary invalidation upon reuse.
1792 list_for_each_entry(obj,
1793 &dev_priv->mm.inactive_list,
1796 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
1799 /* The fence registers are invalidated so clear them out */
1800 i915_gem_reset_fences(dev);
1804 * This function clears the request list as sequence numbers are passed.
1807 i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
1812 if (list_empty(&ring->request_list))
1815 WARN_ON(i915_verify_lists(ring->dev));
1817 seqno = ring->get_seqno(ring);
1819 for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++)
1820 if (seqno >= ring->sync_seqno[i])
1821 ring->sync_seqno[i] = 0;
1823 while (!list_empty(&ring->request_list)) {
1824 struct drm_i915_gem_request *request;
1826 request = list_first_entry(&ring->request_list,
1827 struct drm_i915_gem_request,
1830 if (!i915_seqno_passed(seqno, request->seqno))
1833 trace_i915_gem_request_retire(ring, request->seqno);
1835 list_del(&request->list);
1836 i915_gem_request_remove_from_client(request);
1840 /* Move any buffers on the active list that are no longer referenced
1841 * by the ringbuffer to the flushing/inactive lists as appropriate.
1843 while (!list_empty(&ring->active_list)) {
1844 struct drm_i915_gem_object *obj;
1846 obj = list_first_entry(&ring->active_list,
1847 struct drm_i915_gem_object,
1850 if (!i915_seqno_passed(seqno, obj->last_rendering_seqno))
1853 if (obj->base.write_domain != 0)
1854 i915_gem_object_move_to_flushing(obj);
1856 i915_gem_object_move_to_inactive(obj);
1859 if (unlikely(ring->trace_irq_seqno &&
1860 i915_seqno_passed(seqno, ring->trace_irq_seqno))) {
1861 ring->irq_put(ring);
1862 ring->trace_irq_seqno = 0;
1865 WARN_ON(i915_verify_lists(ring->dev));
1869 i915_gem_retire_requests(struct drm_device *dev)
1871 drm_i915_private_t *dev_priv = dev->dev_private;
1874 if (!list_empty(&dev_priv->mm.deferred_free_list)) {
1875 struct drm_i915_gem_object *obj, *next;
1877 /* We must be careful that during unbind() we do not
1878 * accidentally infinitely recurse into retire requests.
1880 * retire -> free -> unbind -> wait -> retire_ring
1882 list_for_each_entry_safe(obj, next,
1883 &dev_priv->mm.deferred_free_list,
1885 i915_gem_free_object_tail(obj);
1888 for (i = 0; i < I915_NUM_RINGS; i++)
1889 i915_gem_retire_requests_ring(&dev_priv->ring[i]);
1893 i915_gem_retire_work_handler(struct work_struct *work)
1895 drm_i915_private_t *dev_priv;
1896 struct drm_device *dev;
1900 dev_priv = container_of(work, drm_i915_private_t,
1901 mm.retire_work.work);
1902 dev = dev_priv->dev;
1904 /* Come back later if the device is busy... */
1905 if (!mutex_trylock(&dev->struct_mutex)) {
1906 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1910 i915_gem_retire_requests(dev);
1912 /* Send a periodic flush down the ring so we don't hold onto GEM
1913 * objects indefinitely.
1916 for (i = 0; i < I915_NUM_RINGS; i++) {
1917 struct intel_ring_buffer *ring = &dev_priv->ring[i];
1919 if (!list_empty(&ring->gpu_write_list)) {
1920 struct drm_i915_gem_request *request;
1923 ret = i915_gem_flush_ring(ring,
1924 0, I915_GEM_GPU_DOMAINS);
1925 request = kzalloc(sizeof(*request), GFP_KERNEL);
1926 if (ret || request == NULL ||
1927 i915_add_request(ring, NULL, request))
1931 idle &= list_empty(&ring->request_list);
1934 if (!dev_priv->mm.suspended && !idle)
1935 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1937 mutex_unlock(&dev->struct_mutex);
1941 * Waits for a sequence number to be signaled, and cleans up the
1942 * request and object lists appropriately for that event.
1945 i915_wait_request(struct intel_ring_buffer *ring,
1948 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1954 if (atomic_read(&dev_priv->mm.wedged)) {
1955 struct completion *x = &dev_priv->error_completion;
1956 bool recovery_complete;
1957 unsigned long flags;
1959 /* Give the error handler a chance to run. */
1960 spin_lock_irqsave(&x->wait.lock, flags);
1961 recovery_complete = x->done > 0;
1962 spin_unlock_irqrestore(&x->wait.lock, flags);
1964 return recovery_complete ? -EIO : -EAGAIN;
1967 if (seqno == ring->outstanding_lazy_request) {
1968 struct drm_i915_gem_request *request;
1970 request = kzalloc(sizeof(*request), GFP_KERNEL);
1971 if (request == NULL)
1974 ret = i915_add_request(ring, NULL, request);
1980 seqno = request->seqno;
1983 if (!i915_seqno_passed(ring->get_seqno(ring), seqno)) {
1984 if (HAS_PCH_SPLIT(ring->dev))
1985 ier = I915_READ(DEIER) | I915_READ(GTIER);
1987 ier = I915_READ(IER);
1989 DRM_ERROR("something (likely vbetool) disabled "
1990 "interrupts, re-enabling\n");
1991 ring->dev->driver->irq_preinstall(ring->dev);
1992 ring->dev->driver->irq_postinstall(ring->dev);
1995 trace_i915_gem_request_wait_begin(ring, seqno);
1997 ring->waiting_seqno = seqno;
1998 if (ring->irq_get(ring)) {
1999 if (dev_priv->mm.interruptible)
2000 ret = wait_event_interruptible(ring->irq_queue,
2001 i915_seqno_passed(ring->get_seqno(ring), seqno)
2002 || atomic_read(&dev_priv->mm.wedged));
2004 wait_event(ring->irq_queue,
2005 i915_seqno_passed(ring->get_seqno(ring), seqno)
2006 || atomic_read(&dev_priv->mm.wedged));
2008 ring->irq_put(ring);
2009 } else if (wait_for(i915_seqno_passed(ring->get_seqno(ring),
2011 atomic_read(&dev_priv->mm.wedged), 3000))
2013 ring->waiting_seqno = 0;
2015 trace_i915_gem_request_wait_end(ring, seqno);
2017 if (atomic_read(&dev_priv->mm.wedged))
2020 if (ret && ret != -ERESTARTSYS)
2021 DRM_ERROR("%s returns %d (awaiting %d at %d, next %d)\n",
2022 __func__, ret, seqno, ring->get_seqno(ring),
2023 dev_priv->next_seqno);
2025 /* Directly dispatch request retiring. While we have the work queue
2026 * to handle this, the waiter on a request often wants an associated
2027 * buffer to have made it to the inactive list, and we would need
2028 * a separate wait queue to handle that.
2030 * To avoid a recursion with the ilk VT-d workaround (that calls
2031 * gpu_idle when unbinding objects with interruptible==false) don't
2032 * retire requests in that case (because it might call unbind if the
2033 * active list holds the last reference to the object).
2035 if (ret == 0 && dev_priv->mm.interruptible)
2036 i915_gem_retire_requests_ring(ring);
2042 * Ensures that all rendering to the object has completed and the object is
2043 * safe to unbind from the GTT or access from the CPU.
2046 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj)
2050 /* This function only exists to support waiting for existing rendering,
2051 * not for emitting required flushes.
2053 BUG_ON((obj->base.write_domain & I915_GEM_GPU_DOMAINS) != 0);
2055 /* If there is rendering queued on the buffer being evicted, wait for
2059 ret = i915_wait_request(obj->ring, obj->last_rendering_seqno);
2067 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
2069 u32 old_write_domain, old_read_domains;
2071 /* Act a barrier for all accesses through the GTT */
2074 /* Force a pagefault for domain tracking on next user access */
2075 i915_gem_release_mmap(obj);
2077 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
2080 old_read_domains = obj->base.read_domains;
2081 old_write_domain = obj->base.write_domain;
2083 obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
2084 obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
2086 trace_i915_gem_object_change_domain(obj,
2092 * Unbinds an object from the GTT aperture.
2095 i915_gem_object_unbind(struct drm_i915_gem_object *obj)
2099 if (obj->gtt_space == NULL)
2102 if (obj->pin_count != 0) {
2103 DRM_ERROR("Attempting to unbind pinned buffer\n");
2107 ret = i915_gem_object_finish_gpu(obj);
2108 if (ret == -ERESTARTSYS)
2110 /* Continue on if we fail due to EIO, the GPU is hung so we
2111 * should be safe and we need to cleanup or else we might
2112 * cause memory corruption through use-after-free.
2115 i915_gem_object_finish_gtt(obj);
2117 /* Move the object to the CPU domain to ensure that
2118 * any possible CPU writes while it's not in the GTT
2119 * are flushed when we go to remap it.
2122 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
2123 if (ret == -ERESTARTSYS)
2126 /* In the event of a disaster, abandon all caches and
2127 * hope for the best.
2129 i915_gem_clflush_object(obj);
2130 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
2133 /* release the fence reg _after_ flushing */
2134 ret = i915_gem_object_put_fence(obj);
2135 if (ret == -ERESTARTSYS)
2138 trace_i915_gem_object_unbind(obj);
2140 i915_gem_gtt_unbind_object(obj);
2141 i915_gem_object_put_pages_gtt(obj);
2143 list_del_init(&obj->gtt_list);
2144 list_del_init(&obj->mm_list);
2145 /* Avoid an unnecessary call to unbind on rebind. */
2146 obj->map_and_fenceable = true;
2148 drm_mm_put_block(obj->gtt_space);
2149 obj->gtt_space = NULL;
2150 obj->gtt_offset = 0;
2152 if (i915_gem_object_is_purgeable(obj))
2153 i915_gem_object_truncate(obj);
2159 i915_gem_flush_ring(struct intel_ring_buffer *ring,
2160 uint32_t invalidate_domains,
2161 uint32_t flush_domains)
2165 if (((invalidate_domains | flush_domains) & I915_GEM_GPU_DOMAINS) == 0)
2168 trace_i915_gem_ring_flush(ring, invalidate_domains, flush_domains);
2170 ret = ring->flush(ring, invalidate_domains, flush_domains);
2174 if (flush_domains & I915_GEM_GPU_DOMAINS)
2175 i915_gem_process_flushing_list(ring, flush_domains);
2180 static int i915_ring_idle(struct intel_ring_buffer *ring)
2184 if (list_empty(&ring->gpu_write_list) && list_empty(&ring->active_list))
2187 if (!list_empty(&ring->gpu_write_list)) {
2188 ret = i915_gem_flush_ring(ring,
2189 I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
2194 return i915_wait_request(ring, i915_gem_next_request_seqno(ring));
2198 i915_gpu_idle(struct drm_device *dev)
2200 drm_i915_private_t *dev_priv = dev->dev_private;
2203 /* Flush everything onto the inactive list. */
2204 for (i = 0; i < I915_NUM_RINGS; i++) {
2205 ret = i915_ring_idle(&dev_priv->ring[i]);
2213 static int sandybridge_write_fence_reg(struct drm_i915_gem_object *obj,
2214 struct intel_ring_buffer *pipelined)
2216 struct drm_device *dev = obj->base.dev;
2217 drm_i915_private_t *dev_priv = dev->dev_private;
2218 u32 size = obj->gtt_space->size;
2219 int regnum = obj->fence_reg;
2222 val = (uint64_t)((obj->gtt_offset + size - 4096) &
2224 val |= obj->gtt_offset & 0xfffff000;
2225 val |= (uint64_t)((obj->stride / 128) - 1) <<
2226 SANDYBRIDGE_FENCE_PITCH_SHIFT;
2228 if (obj->tiling_mode == I915_TILING_Y)
2229 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2230 val |= I965_FENCE_REG_VALID;
2233 int ret = intel_ring_begin(pipelined, 6);
2237 intel_ring_emit(pipelined, MI_NOOP);
2238 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(2));
2239 intel_ring_emit(pipelined, FENCE_REG_SANDYBRIDGE_0 + regnum*8);
2240 intel_ring_emit(pipelined, (u32)val);
2241 intel_ring_emit(pipelined, FENCE_REG_SANDYBRIDGE_0 + regnum*8 + 4);
2242 intel_ring_emit(pipelined, (u32)(val >> 32));
2243 intel_ring_advance(pipelined);
2245 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + regnum * 8, val);
2250 static int i965_write_fence_reg(struct drm_i915_gem_object *obj,
2251 struct intel_ring_buffer *pipelined)
2253 struct drm_device *dev = obj->base.dev;
2254 drm_i915_private_t *dev_priv = dev->dev_private;
2255 u32 size = obj->gtt_space->size;
2256 int regnum = obj->fence_reg;
2259 val = (uint64_t)((obj->gtt_offset + size - 4096) &
2261 val |= obj->gtt_offset & 0xfffff000;
2262 val |= ((obj->stride / 128) - 1) << I965_FENCE_PITCH_SHIFT;
2263 if (obj->tiling_mode == I915_TILING_Y)
2264 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2265 val |= I965_FENCE_REG_VALID;
2268 int ret = intel_ring_begin(pipelined, 6);
2272 intel_ring_emit(pipelined, MI_NOOP);
2273 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(2));
2274 intel_ring_emit(pipelined, FENCE_REG_965_0 + regnum*8);
2275 intel_ring_emit(pipelined, (u32)val);
2276 intel_ring_emit(pipelined, FENCE_REG_965_0 + regnum*8 + 4);
2277 intel_ring_emit(pipelined, (u32)(val >> 32));
2278 intel_ring_advance(pipelined);
2280 I915_WRITE64(FENCE_REG_965_0 + regnum * 8, val);
2285 static int i915_write_fence_reg(struct drm_i915_gem_object *obj,
2286 struct intel_ring_buffer *pipelined)
2288 struct drm_device *dev = obj->base.dev;
2289 drm_i915_private_t *dev_priv = dev->dev_private;
2290 u32 size = obj->gtt_space->size;
2291 u32 fence_reg, val, pitch_val;
2294 if (WARN((obj->gtt_offset & ~I915_FENCE_START_MASK) ||
2295 (size & -size) != size ||
2296 (obj->gtt_offset & (size - 1)),
2297 "object 0x%08x [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
2298 obj->gtt_offset, obj->map_and_fenceable, size))
2301 if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
2306 /* Note: pitch better be a power of two tile widths */
2307 pitch_val = obj->stride / tile_width;
2308 pitch_val = ffs(pitch_val) - 1;
2310 val = obj->gtt_offset;
2311 if (obj->tiling_mode == I915_TILING_Y)
2312 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2313 val |= I915_FENCE_SIZE_BITS(size);
2314 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2315 val |= I830_FENCE_REG_VALID;
2317 fence_reg = obj->fence_reg;
2319 fence_reg = FENCE_REG_830_0 + fence_reg * 4;
2321 fence_reg = FENCE_REG_945_8 + (fence_reg - 8) * 4;
2324 int ret = intel_ring_begin(pipelined, 4);
2328 intel_ring_emit(pipelined, MI_NOOP);
2329 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(1));
2330 intel_ring_emit(pipelined, fence_reg);
2331 intel_ring_emit(pipelined, val);
2332 intel_ring_advance(pipelined);
2334 I915_WRITE(fence_reg, val);
2339 static int i830_write_fence_reg(struct drm_i915_gem_object *obj,
2340 struct intel_ring_buffer *pipelined)
2342 struct drm_device *dev = obj->base.dev;
2343 drm_i915_private_t *dev_priv = dev->dev_private;
2344 u32 size = obj->gtt_space->size;
2345 int regnum = obj->fence_reg;
2349 if (WARN((obj->gtt_offset & ~I830_FENCE_START_MASK) ||
2350 (size & -size) != size ||
2351 (obj->gtt_offset & (size - 1)),
2352 "object 0x%08x not 512K or pot-size 0x%08x aligned\n",
2353 obj->gtt_offset, size))
2356 pitch_val = obj->stride / 128;
2357 pitch_val = ffs(pitch_val) - 1;
2359 val = obj->gtt_offset;
2360 if (obj->tiling_mode == I915_TILING_Y)
2361 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2362 val |= I830_FENCE_SIZE_BITS(size);
2363 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2364 val |= I830_FENCE_REG_VALID;
2367 int ret = intel_ring_begin(pipelined, 4);
2371 intel_ring_emit(pipelined, MI_NOOP);
2372 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(1));
2373 intel_ring_emit(pipelined, FENCE_REG_830_0 + regnum*4);
2374 intel_ring_emit(pipelined, val);
2375 intel_ring_advance(pipelined);
2377 I915_WRITE(FENCE_REG_830_0 + regnum * 4, val);
2382 static bool ring_passed_seqno(struct intel_ring_buffer *ring, u32 seqno)
2384 return i915_seqno_passed(ring->get_seqno(ring), seqno);
2388 i915_gem_object_flush_fence(struct drm_i915_gem_object *obj,
2389 struct intel_ring_buffer *pipelined)
2393 if (obj->fenced_gpu_access) {
2394 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
2395 ret = i915_gem_flush_ring(obj->last_fenced_ring,
2396 0, obj->base.write_domain);
2401 obj->fenced_gpu_access = false;
2404 if (obj->last_fenced_seqno && pipelined != obj->last_fenced_ring) {
2405 if (!ring_passed_seqno(obj->last_fenced_ring,
2406 obj->last_fenced_seqno)) {
2407 ret = i915_wait_request(obj->last_fenced_ring,
2408 obj->last_fenced_seqno);
2413 obj->last_fenced_seqno = 0;
2414 obj->last_fenced_ring = NULL;
2417 /* Ensure that all CPU reads are completed before installing a fence
2418 * and all writes before removing the fence.
2420 if (obj->base.read_domains & I915_GEM_DOMAIN_GTT)
2427 i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
2431 if (obj->tiling_mode)
2432 i915_gem_release_mmap(obj);
2434 ret = i915_gem_object_flush_fence(obj, NULL);
2438 if (obj->fence_reg != I915_FENCE_REG_NONE) {
2439 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2440 i915_gem_clear_fence_reg(obj->base.dev,
2441 &dev_priv->fence_regs[obj->fence_reg]);
2443 obj->fence_reg = I915_FENCE_REG_NONE;
2449 static struct drm_i915_fence_reg *
2450 i915_find_fence_reg(struct drm_device *dev,
2451 struct intel_ring_buffer *pipelined)
2453 struct drm_i915_private *dev_priv = dev->dev_private;
2454 struct drm_i915_fence_reg *reg, *first, *avail;
2457 /* First try to find a free reg */
2459 for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2460 reg = &dev_priv->fence_regs[i];
2464 if (!reg->obj->pin_count)
2471 /* None available, try to steal one or wait for a user to finish */
2472 avail = first = NULL;
2473 list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
2474 if (reg->obj->pin_count)
2481 !reg->obj->last_fenced_ring ||
2482 reg->obj->last_fenced_ring == pipelined) {
2495 * i915_gem_object_get_fence - set up a fence reg for an object
2496 * @obj: object to map through a fence reg
2497 * @pipelined: ring on which to queue the change, or NULL for CPU access
2498 * @interruptible: must we wait uninterruptibly for the register to retire?
2500 * When mapping objects through the GTT, userspace wants to be able to write
2501 * to them without having to worry about swizzling if the object is tiled.
2503 * This function walks the fence regs looking for a free one for @obj,
2504 * stealing one if it can't find any.
2506 * It then sets up the reg based on the object's properties: address, pitch
2507 * and tiling format.
2510 i915_gem_object_get_fence(struct drm_i915_gem_object *obj,
2511 struct intel_ring_buffer *pipelined)
2513 struct drm_device *dev = obj->base.dev;
2514 struct drm_i915_private *dev_priv = dev->dev_private;
2515 struct drm_i915_fence_reg *reg;
2518 /* XXX disable pipelining. There are bugs. Shocking. */
2521 /* Just update our place in the LRU if our fence is getting reused. */
2522 if (obj->fence_reg != I915_FENCE_REG_NONE) {
2523 reg = &dev_priv->fence_regs[obj->fence_reg];
2524 list_move_tail(®->lru_list, &dev_priv->mm.fence_list);
2526 if (obj->tiling_changed) {
2527 ret = i915_gem_object_flush_fence(obj, pipelined);
2531 if (!obj->fenced_gpu_access && !obj->last_fenced_seqno)
2536 i915_gem_next_request_seqno(pipelined);
2537 obj->last_fenced_seqno = reg->setup_seqno;
2538 obj->last_fenced_ring = pipelined;
2545 if (reg->setup_seqno) {
2546 if (!ring_passed_seqno(obj->last_fenced_ring,
2547 reg->setup_seqno)) {
2548 ret = i915_wait_request(obj->last_fenced_ring,
2554 reg->setup_seqno = 0;
2556 } else if (obj->last_fenced_ring &&
2557 obj->last_fenced_ring != pipelined) {
2558 ret = i915_gem_object_flush_fence(obj, pipelined);
2566 reg = i915_find_fence_reg(dev, pipelined);
2570 ret = i915_gem_object_flush_fence(obj, pipelined);
2575 struct drm_i915_gem_object *old = reg->obj;
2577 drm_gem_object_reference(&old->base);
2579 if (old->tiling_mode)
2580 i915_gem_release_mmap(old);
2582 ret = i915_gem_object_flush_fence(old, pipelined);
2584 drm_gem_object_unreference(&old->base);
2588 if (old->last_fenced_seqno == 0 && obj->last_fenced_seqno == 0)
2591 old->fence_reg = I915_FENCE_REG_NONE;
2592 old->last_fenced_ring = pipelined;
2593 old->last_fenced_seqno =
2594 pipelined ? i915_gem_next_request_seqno(pipelined) : 0;
2596 drm_gem_object_unreference(&old->base);
2597 } else if (obj->last_fenced_seqno == 0)
2601 list_move_tail(®->lru_list, &dev_priv->mm.fence_list);
2602 obj->fence_reg = reg - dev_priv->fence_regs;
2603 obj->last_fenced_ring = pipelined;
2606 pipelined ? i915_gem_next_request_seqno(pipelined) : 0;
2607 obj->last_fenced_seqno = reg->setup_seqno;
2610 obj->tiling_changed = false;
2611 switch (INTEL_INFO(dev)->gen) {
2614 ret = sandybridge_write_fence_reg(obj, pipelined);
2618 ret = i965_write_fence_reg(obj, pipelined);
2621 ret = i915_write_fence_reg(obj, pipelined);
2624 ret = i830_write_fence_reg(obj, pipelined);
2632 * i915_gem_clear_fence_reg - clear out fence register info
2633 * @obj: object to clear
2635 * Zeroes out the fence register itself and clears out the associated
2636 * data structures in dev_priv and obj.
2639 i915_gem_clear_fence_reg(struct drm_device *dev,
2640 struct drm_i915_fence_reg *reg)
2642 drm_i915_private_t *dev_priv = dev->dev_private;
2643 uint32_t fence_reg = reg - dev_priv->fence_regs;
2645 switch (INTEL_INFO(dev)->gen) {
2648 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + fence_reg*8, 0);
2652 I915_WRITE64(FENCE_REG_965_0 + fence_reg*8, 0);
2656 fence_reg = FENCE_REG_945_8 + (fence_reg - 8) * 4;
2659 fence_reg = FENCE_REG_830_0 + fence_reg * 4;
2661 I915_WRITE(fence_reg, 0);
2665 list_del_init(®->lru_list);
2667 reg->setup_seqno = 0;
2671 * Finds free space in the GTT aperture and binds the object there.
2674 i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
2676 bool map_and_fenceable)
2678 struct drm_device *dev = obj->base.dev;
2679 drm_i915_private_t *dev_priv = dev->dev_private;
2680 struct drm_mm_node *free_space;
2681 gfp_t gfpmask = __GFP_NORETRY | __GFP_NOWARN;
2682 u32 size, fence_size, fence_alignment, unfenced_alignment;
2683 bool mappable, fenceable;
2686 if (obj->madv != I915_MADV_WILLNEED) {
2687 DRM_ERROR("Attempting to bind a purgeable object\n");
2691 fence_size = i915_gem_get_gtt_size(dev,
2694 fence_alignment = i915_gem_get_gtt_alignment(dev,
2697 unfenced_alignment =
2698 i915_gem_get_unfenced_gtt_alignment(dev,
2703 alignment = map_and_fenceable ? fence_alignment :
2705 if (map_and_fenceable && alignment & (fence_alignment - 1)) {
2706 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
2710 size = map_and_fenceable ? fence_size : obj->base.size;
2712 /* If the object is bigger than the entire aperture, reject it early
2713 * before evicting everything in a vain attempt to find space.
2715 if (obj->base.size >
2716 (map_and_fenceable ? dev_priv->mm.gtt_mappable_end : dev_priv->mm.gtt_total)) {
2717 DRM_ERROR("Attempting to bind an object larger than the aperture\n");
2722 if (map_and_fenceable)
2724 drm_mm_search_free_in_range(&dev_priv->mm.gtt_space,
2726 dev_priv->mm.gtt_mappable_end,
2729 free_space = drm_mm_search_free(&dev_priv->mm.gtt_space,
2730 size, alignment, 0);
2732 if (free_space != NULL) {
2733 if (map_and_fenceable)
2735 drm_mm_get_block_range_generic(free_space,
2737 dev_priv->mm.gtt_mappable_end,
2741 drm_mm_get_block(free_space, size, alignment);
2743 if (obj->gtt_space == NULL) {
2744 /* If the gtt is empty and we're still having trouble
2745 * fitting our object in, we're out of memory.
2747 ret = i915_gem_evict_something(dev, size, alignment,
2755 ret = i915_gem_object_get_pages_gtt(obj, gfpmask);
2757 drm_mm_put_block(obj->gtt_space);
2758 obj->gtt_space = NULL;
2760 if (ret == -ENOMEM) {
2761 /* first try to reclaim some memory by clearing the GTT */
2762 ret = i915_gem_evict_everything(dev, false);
2764 /* now try to shrink everyone else */
2779 ret = i915_gem_gtt_bind_object(obj);
2781 i915_gem_object_put_pages_gtt(obj);
2782 drm_mm_put_block(obj->gtt_space);
2783 obj->gtt_space = NULL;
2785 if (i915_gem_evict_everything(dev, false))
2791 list_add_tail(&obj->gtt_list, &dev_priv->mm.gtt_list);
2792 list_add_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
2794 /* Assert that the object is not currently in any GPU domain. As it
2795 * wasn't in the GTT, there shouldn't be any way it could have been in
2798 BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
2799 BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
2801 obj->gtt_offset = obj->gtt_space->start;
2804 obj->gtt_space->size == fence_size &&
2805 (obj->gtt_space->start & (fence_alignment - 1)) == 0;
2808 obj->gtt_offset + obj->base.size <= dev_priv->mm.gtt_mappable_end;
2810 obj->map_and_fenceable = mappable && fenceable;
2812 trace_i915_gem_object_bind(obj, map_and_fenceable);
2817 i915_gem_clflush_object(struct drm_i915_gem_object *obj)
2819 /* If we don't have a page list set up, then we're not pinned
2820 * to GPU, and we can ignore the cache flush because it'll happen
2821 * again at bind time.
2823 if (obj->pages == NULL)
2826 /* If the GPU is snooping the contents of the CPU cache,
2827 * we do not need to manually clear the CPU cache lines. However,
2828 * the caches are only snooped when the render cache is
2829 * flushed/invalidated. As we always have to emit invalidations
2830 * and flushes when moving into and out of the RENDER domain, correct
2831 * snooping behaviour occurs naturally as the result of our domain
2834 if (obj->cache_level != I915_CACHE_NONE)
2837 trace_i915_gem_object_clflush(obj);
2839 drm_clflush_pages(obj->pages, obj->base.size / PAGE_SIZE);
2842 /** Flushes any GPU write domain for the object if it's dirty. */
2844 i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object *obj)
2846 if ((obj->base.write_domain & I915_GEM_GPU_DOMAINS) == 0)
2849 /* Queue the GPU write cache flushing we need. */
2850 return i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain);
2853 /** Flushes the GTT write domain for the object if it's dirty. */
2855 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
2857 uint32_t old_write_domain;
2859 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
2862 /* No actual flushing is required for the GTT write domain. Writes
2863 * to it immediately go to main memory as far as we know, so there's
2864 * no chipset flush. It also doesn't land in render cache.
2866 * However, we do have to enforce the order so that all writes through
2867 * the GTT land before any writes to the device, such as updates to
2872 old_write_domain = obj->base.write_domain;
2873 obj->base.write_domain = 0;
2875 trace_i915_gem_object_change_domain(obj,
2876 obj->base.read_domains,
2880 /** Flushes the CPU write domain for the object if it's dirty. */
2882 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj)
2884 uint32_t old_write_domain;
2886 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
2889 i915_gem_clflush_object(obj);
2890 intel_gtt_chipset_flush();
2891 old_write_domain = obj->base.write_domain;
2892 obj->base.write_domain = 0;
2894 trace_i915_gem_object_change_domain(obj,
2895 obj->base.read_domains,
2900 * Moves a single object to the GTT read, and possibly write domain.
2902 * This function returns when the move is complete, including waiting on
2906 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
2908 uint32_t old_write_domain, old_read_domains;
2911 /* Not valid to be called on unbound objects. */
2912 if (obj->gtt_space == NULL)
2915 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
2918 ret = i915_gem_object_flush_gpu_write_domain(obj);
2922 if (obj->pending_gpu_write || write) {
2923 ret = i915_gem_object_wait_rendering(obj);
2928 i915_gem_object_flush_cpu_write_domain(obj);
2930 old_write_domain = obj->base.write_domain;
2931 old_read_domains = obj->base.read_domains;
2933 /* It should now be out of any other write domains, and we can update
2934 * the domain values for our changes.
2936 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2937 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
2939 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
2940 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
2944 trace_i915_gem_object_change_domain(obj,
2951 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
2952 enum i915_cache_level cache_level)
2956 if (obj->cache_level == cache_level)
2959 if (obj->pin_count) {
2960 DRM_DEBUG("can not change the cache level of pinned objects\n");
2964 if (obj->gtt_space) {
2965 ret = i915_gem_object_finish_gpu(obj);
2969 i915_gem_object_finish_gtt(obj);
2971 /* Before SandyBridge, you could not use tiling or fence
2972 * registers with snooped memory, so relinquish any fences
2973 * currently pointing to our region in the aperture.
2975 if (INTEL_INFO(obj->base.dev)->gen < 6) {
2976 ret = i915_gem_object_put_fence(obj);
2981 i915_gem_gtt_rebind_object(obj, cache_level);
2984 if (cache_level == I915_CACHE_NONE) {
2985 u32 old_read_domains, old_write_domain;
2987 /* If we're coming from LLC cached, then we haven't
2988 * actually been tracking whether the data is in the
2989 * CPU cache or not, since we only allow one bit set
2990 * in obj->write_domain and have been skipping the clflushes.
2991 * Just set it to the CPU cache for now.
2993 WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU);
2994 WARN_ON(obj->base.read_domains & ~I915_GEM_DOMAIN_CPU);
2996 old_read_domains = obj->base.read_domains;
2997 old_write_domain = obj->base.write_domain;
2999 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3000 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3002 trace_i915_gem_object_change_domain(obj,
3007 obj->cache_level = cache_level;
3012 * Prepare buffer for display plane (scanout, cursors, etc).
3013 * Can be called from an uninterruptible phase (modesetting) and allows
3014 * any flushes to be pipelined (for pageflips).
3016 * For the display plane, we want to be in the GTT but out of any write
3017 * domains. So in many ways this looks like set_to_gtt_domain() apart from the
3018 * ability to pipeline the waits, pinning and any additional subtleties
3019 * that may differentiate the display plane from ordinary buffers.
3022 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
3024 struct intel_ring_buffer *pipelined)
3026 u32 old_read_domains, old_write_domain;
3029 ret = i915_gem_object_flush_gpu_write_domain(obj);
3033 if (pipelined != obj->ring) {
3034 ret = i915_gem_object_wait_rendering(obj);
3035 if (ret == -ERESTARTSYS)
3039 /* The display engine is not coherent with the LLC cache on gen6. As
3040 * a result, we make sure that the pinning that is about to occur is
3041 * done with uncached PTEs. This is lowest common denominator for all
3044 * However for gen6+, we could do better by using the GFDT bit instead
3045 * of uncaching, which would allow us to flush all the LLC-cached data
3046 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3048 ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE);
3052 /* As the user may map the buffer once pinned in the display plane
3053 * (e.g. libkms for the bootup splash), we have to ensure that we
3054 * always use map_and_fenceable for all scanout buffers.
3056 ret = i915_gem_object_pin(obj, alignment, true);
3060 i915_gem_object_flush_cpu_write_domain(obj);
3062 old_write_domain = obj->base.write_domain;
3063 old_read_domains = obj->base.read_domains;
3065 /* It should now be out of any other write domains, and we can update
3066 * the domain values for our changes.
3068 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3069 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3071 trace_i915_gem_object_change_domain(obj,
3079 i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
3083 if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
3086 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
3087 ret = i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain);
3092 /* Ensure that we invalidate the GPU's caches and TLBs. */
3093 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
3095 return i915_gem_object_wait_rendering(obj);
3099 * Moves a single object to the CPU read, and possibly write domain.
3101 * This function returns when the move is complete, including waiting on
3105 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
3107 uint32_t old_write_domain, old_read_domains;
3110 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
3113 ret = i915_gem_object_flush_gpu_write_domain(obj);
3117 ret = i915_gem_object_wait_rendering(obj);
3121 i915_gem_object_flush_gtt_write_domain(obj);
3123 /* If we have a partially-valid cache of the object in the CPU,
3124 * finish invalidating it and free the per-page flags.
3126 i915_gem_object_set_to_full_cpu_read_domain(obj);
3128 old_write_domain = obj->base.write_domain;
3129 old_read_domains = obj->base.read_domains;
3131 /* Flush the CPU cache if it's still invalid. */
3132 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
3133 i915_gem_clflush_object(obj);
3135 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3138 /* It should now be out of any other write domains, and we can update
3139 * the domain values for our changes.
3141 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3143 /* If we're writing through the CPU, then the GPU read domains will
3144 * need to be invalidated at next use.
3147 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3148 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3151 trace_i915_gem_object_change_domain(obj,
3159 * Moves the object from a partially CPU read to a full one.
3161 * Note that this only resolves i915_gem_object_set_cpu_read_domain_range(),
3162 * and doesn't handle transitioning from !(read_domains & I915_GEM_DOMAIN_CPU).
3165 i915_gem_object_set_to_full_cpu_read_domain(struct drm_i915_gem_object *obj)
3167 if (!obj->page_cpu_valid)
3170 /* If we're partially in the CPU read domain, finish moving it in.
3172 if (obj->base.read_domains & I915_GEM_DOMAIN_CPU) {
3175 for (i = 0; i <= (obj->base.size - 1) / PAGE_SIZE; i++) {
3176 if (obj->page_cpu_valid[i])
3178 drm_clflush_pages(obj->pages + i, 1);
3182 /* Free the page_cpu_valid mappings which are now stale, whether
3183 * or not we've got I915_GEM_DOMAIN_CPU.
3185 kfree(obj->page_cpu_valid);
3186 obj->page_cpu_valid = NULL;
3190 * Set the CPU read domain on a range of the object.
3192 * The object ends up with I915_GEM_DOMAIN_CPU in its read flags although it's
3193 * not entirely valid. The page_cpu_valid member of the object flags which
3194 * pages have been flushed, and will be respected by
3195 * i915_gem_object_set_to_cpu_domain() if it's called on to get a valid mapping
3196 * of the whole object.
3198 * This function returns when the move is complete, including waiting on
3202 i915_gem_object_set_cpu_read_domain_range(struct drm_i915_gem_object *obj,
3203 uint64_t offset, uint64_t size)
3205 uint32_t old_read_domains;
3208 if (offset == 0 && size == obj->base.size)
3209 return i915_gem_object_set_to_cpu_domain(obj, 0);
3211 ret = i915_gem_object_flush_gpu_write_domain(obj);
3215 ret = i915_gem_object_wait_rendering(obj);
3219 i915_gem_object_flush_gtt_write_domain(obj);
3221 /* If we're already fully in the CPU read domain, we're done. */
3222 if (obj->page_cpu_valid == NULL &&
3223 (obj->base.read_domains & I915_GEM_DOMAIN_CPU) != 0)
3226 /* Otherwise, create/clear the per-page CPU read domain flag if we're
3227 * newly adding I915_GEM_DOMAIN_CPU
3229 if (obj->page_cpu_valid == NULL) {
3230 obj->page_cpu_valid = kzalloc(obj->base.size / PAGE_SIZE,
3232 if (obj->page_cpu_valid == NULL)
3234 } else if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0)
3235 memset(obj->page_cpu_valid, 0, obj->base.size / PAGE_SIZE);
3237 /* Flush the cache on any pages that are still invalid from the CPU's
3240 for (i = offset / PAGE_SIZE; i <= (offset + size - 1) / PAGE_SIZE;
3242 if (obj->page_cpu_valid[i])
3245 drm_clflush_pages(obj->pages + i, 1);
3247 obj->page_cpu_valid[i] = 1;
3250 /* It should now be out of any other write domains, and we can update
3251 * the domain values for our changes.
3253 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3255 old_read_domains = obj->base.read_domains;
3256 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3258 trace_i915_gem_object_change_domain(obj,
3260 obj->base.write_domain);
3265 /* Throttle our rendering by waiting until the ring has completed our requests
3266 * emitted over 20 msec ago.
3268 * Note that if we were to use the current jiffies each time around the loop,
3269 * we wouldn't escape the function with any frames outstanding if the time to
3270 * render a frame was over 20ms.
3272 * This should get us reasonable parallelism between CPU and GPU but also
3273 * relatively low latency when blocking on a particular request to finish.
3276 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
3278 struct drm_i915_private *dev_priv = dev->dev_private;
3279 struct drm_i915_file_private *file_priv = file->driver_priv;
3280 unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3281 struct drm_i915_gem_request *request;
3282 struct intel_ring_buffer *ring = NULL;
3286 if (atomic_read(&dev_priv->mm.wedged))
3289 spin_lock(&file_priv->mm.lock);
3290 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
3291 if (time_after_eq(request->emitted_jiffies, recent_enough))
3294 ring = request->ring;
3295 seqno = request->seqno;
3297 spin_unlock(&file_priv->mm.lock);
3303 if (!i915_seqno_passed(ring->get_seqno(ring), seqno)) {
3304 /* And wait for the seqno passing without holding any locks and
3305 * causing extra latency for others. This is safe as the irq
3306 * generation is designed to be run atomically and so is
3309 if (ring->irq_get(ring)) {
3310 ret = wait_event_interruptible(ring->irq_queue,
3311 i915_seqno_passed(ring->get_seqno(ring), seqno)
3312 || atomic_read(&dev_priv->mm.wedged));
3313 ring->irq_put(ring);
3315 if (ret == 0 && atomic_read(&dev_priv->mm.wedged))
3321 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
3327 i915_gem_object_pin(struct drm_i915_gem_object *obj,
3329 bool map_and_fenceable)
3331 struct drm_device *dev = obj->base.dev;
3332 struct drm_i915_private *dev_priv = dev->dev_private;
3335 BUG_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT);
3336 WARN_ON(i915_verify_lists(dev));
3338 if (obj->gtt_space != NULL) {
3339 if ((alignment && obj->gtt_offset & (alignment - 1)) ||
3340 (map_and_fenceable && !obj->map_and_fenceable)) {
3341 WARN(obj->pin_count,
3342 "bo is already pinned with incorrect alignment:"
3343 " offset=%x, req.alignment=%x, req.map_and_fenceable=%d,"
3344 " obj->map_and_fenceable=%d\n",
3345 obj->gtt_offset, alignment,
3347 obj->map_and_fenceable);
3348 ret = i915_gem_object_unbind(obj);
3354 if (obj->gtt_space == NULL) {
3355 ret = i915_gem_object_bind_to_gtt(obj, alignment,
3361 if (obj->pin_count++ == 0) {
3363 list_move_tail(&obj->mm_list,
3364 &dev_priv->mm.pinned_list);
3366 obj->pin_mappable |= map_and_fenceable;
3368 WARN_ON(i915_verify_lists(dev));
3373 i915_gem_object_unpin(struct drm_i915_gem_object *obj)
3375 struct drm_device *dev = obj->base.dev;
3376 drm_i915_private_t *dev_priv = dev->dev_private;
3378 WARN_ON(i915_verify_lists(dev));
3379 BUG_ON(obj->pin_count == 0);
3380 BUG_ON(obj->gtt_space == NULL);
3382 if (--obj->pin_count == 0) {
3384 list_move_tail(&obj->mm_list,
3385 &dev_priv->mm.inactive_list);
3386 obj->pin_mappable = false;
3388 WARN_ON(i915_verify_lists(dev));
3392 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
3393 struct drm_file *file)
3395 struct drm_i915_gem_pin *args = data;
3396 struct drm_i915_gem_object *obj;
3399 ret = i915_mutex_lock_interruptible(dev);
3403 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3404 if (&obj->base == NULL) {
3409 if (obj->madv != I915_MADV_WILLNEED) {
3410 DRM_ERROR("Attempting to pin a purgeable buffer\n");
3415 if (obj->pin_filp != NULL && obj->pin_filp != file) {
3416 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
3422 obj->user_pin_count++;
3423 obj->pin_filp = file;
3424 if (obj->user_pin_count == 1) {
3425 ret = i915_gem_object_pin(obj, args->alignment, true);
3430 /* XXX - flush the CPU caches for pinned objects
3431 * as the X server doesn't manage domains yet
3433 i915_gem_object_flush_cpu_write_domain(obj);
3434 args->offset = obj->gtt_offset;
3436 drm_gem_object_unreference(&obj->base);
3438 mutex_unlock(&dev->struct_mutex);
3443 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
3444 struct drm_file *file)
3446 struct drm_i915_gem_pin *args = data;
3447 struct drm_i915_gem_object *obj;
3450 ret = i915_mutex_lock_interruptible(dev);
3454 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3455 if (&obj->base == NULL) {
3460 if (obj->pin_filp != file) {
3461 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
3466 obj->user_pin_count--;
3467 if (obj->user_pin_count == 0) {
3468 obj->pin_filp = NULL;
3469 i915_gem_object_unpin(obj);
3473 drm_gem_object_unreference(&obj->base);
3475 mutex_unlock(&dev->struct_mutex);
3480 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
3481 struct drm_file *file)
3483 struct drm_i915_gem_busy *args = data;
3484 struct drm_i915_gem_object *obj;
3487 ret = i915_mutex_lock_interruptible(dev);
3491 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3492 if (&obj->base == NULL) {
3497 /* Count all active objects as busy, even if they are currently not used
3498 * by the gpu. Users of this interface expect objects to eventually
3499 * become non-busy without any further actions, therefore emit any
3500 * necessary flushes here.
3502 args->busy = obj->active;
3504 /* Unconditionally flush objects, even when the gpu still uses this
3505 * object. Userspace calling this function indicates that it wants to
3506 * use this buffer rather sooner than later, so issuing the required
3507 * flush earlier is beneficial.
3509 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
3510 ret = i915_gem_flush_ring(obj->ring,
3511 0, obj->base.write_domain);
3512 } else if (obj->ring->outstanding_lazy_request ==
3513 obj->last_rendering_seqno) {
3514 struct drm_i915_gem_request *request;
3516 /* This ring is not being cleared by active usage,
3517 * so emit a request to do so.
3519 request = kzalloc(sizeof(*request), GFP_KERNEL);
3521 ret = i915_add_request(obj->ring, NULL, request);
3528 /* Update the active list for the hardware's current position.
3529 * Otherwise this only updates on a delayed timer or when irqs
3530 * are actually unmasked, and our working set ends up being
3531 * larger than required.
3533 i915_gem_retire_requests_ring(obj->ring);
3535 args->busy = obj->active;
3538 drm_gem_object_unreference(&obj->base);
3540 mutex_unlock(&dev->struct_mutex);
3545 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
3546 struct drm_file *file_priv)
3548 return i915_gem_ring_throttle(dev, file_priv);
3552 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
3553 struct drm_file *file_priv)
3555 struct drm_i915_gem_madvise *args = data;
3556 struct drm_i915_gem_object *obj;
3559 switch (args->madv) {
3560 case I915_MADV_DONTNEED:
3561 case I915_MADV_WILLNEED:
3567 ret = i915_mutex_lock_interruptible(dev);
3571 obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
3572 if (&obj->base == NULL) {
3577 if (obj->pin_count) {
3582 if (obj->madv != __I915_MADV_PURGED)
3583 obj->madv = args->madv;
3585 /* if the object is no longer bound, discard its backing storage */
3586 if (i915_gem_object_is_purgeable(obj) &&
3587 obj->gtt_space == NULL)
3588 i915_gem_object_truncate(obj);
3590 args->retained = obj->madv != __I915_MADV_PURGED;
3593 drm_gem_object_unreference(&obj->base);
3595 mutex_unlock(&dev->struct_mutex);
3599 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
3602 struct drm_i915_private *dev_priv = dev->dev_private;
3603 struct drm_i915_gem_object *obj;
3604 struct address_space *mapping;
3606 obj = kzalloc(sizeof(*obj), GFP_KERNEL);
3610 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
3615 mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
3616 mapping_set_gfp_mask(mapping, GFP_HIGHUSER | __GFP_RECLAIMABLE);
3618 i915_gem_info_add_obj(dev_priv, size);
3620 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3621 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3623 if (IS_GEN6(dev) || IS_GEN7(dev)) {
3624 /* On Gen6, we can have the GPU use the LLC (the CPU
3625 * cache) for about a 10% performance improvement
3626 * compared to uncached. Graphics requests other than
3627 * display scanout are coherent with the CPU in
3628 * accessing this cache. This means in this mode we
3629 * don't need to clflush on the CPU side, and on the
3630 * GPU side we only need to flush internal caches to
3631 * get data visible to the CPU.
3633 * However, we maintain the display planes as UC, and so
3634 * need to rebind when first used as such.
3636 obj->cache_level = I915_CACHE_LLC;
3638 obj->cache_level = I915_CACHE_NONE;
3640 obj->base.driver_private = NULL;
3641 obj->fence_reg = I915_FENCE_REG_NONE;
3642 INIT_LIST_HEAD(&obj->mm_list);
3643 INIT_LIST_HEAD(&obj->gtt_list);
3644 INIT_LIST_HEAD(&obj->ring_list);
3645 INIT_LIST_HEAD(&obj->exec_list);
3646 INIT_LIST_HEAD(&obj->gpu_write_list);
3647 obj->madv = I915_MADV_WILLNEED;
3648 /* Avoid an unnecessary call to unbind on the first bind. */
3649 obj->map_and_fenceable = true;
3654 int i915_gem_init_object(struct drm_gem_object *obj)
3661 static void i915_gem_free_object_tail(struct drm_i915_gem_object *obj)
3663 struct drm_device *dev = obj->base.dev;
3664 drm_i915_private_t *dev_priv = dev->dev_private;
3667 ret = i915_gem_object_unbind(obj);
3668 if (ret == -ERESTARTSYS) {
3669 list_move(&obj->mm_list,
3670 &dev_priv->mm.deferred_free_list);
3674 trace_i915_gem_object_destroy(obj);
3676 if (obj->base.map_list.map)
3677 drm_gem_free_mmap_offset(&obj->base);
3679 drm_gem_object_release(&obj->base);
3680 i915_gem_info_remove_obj(dev_priv, obj->base.size);
3682 kfree(obj->page_cpu_valid);
3687 void i915_gem_free_object(struct drm_gem_object *gem_obj)
3689 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
3690 struct drm_device *dev = obj->base.dev;
3692 while (obj->pin_count > 0)
3693 i915_gem_object_unpin(obj);
3696 i915_gem_detach_phys_object(dev, obj);
3698 i915_gem_free_object_tail(obj);
3702 i915_gem_idle(struct drm_device *dev)
3704 drm_i915_private_t *dev_priv = dev->dev_private;
3707 mutex_lock(&dev->struct_mutex);
3709 if (dev_priv->mm.suspended) {
3710 mutex_unlock(&dev->struct_mutex);
3714 ret = i915_gpu_idle(dev);
3716 mutex_unlock(&dev->struct_mutex);
3720 /* Under UMS, be paranoid and evict. */
3721 if (!drm_core_check_feature(dev, DRIVER_MODESET)) {
3722 ret = i915_gem_evict_inactive(dev, false);
3724 mutex_unlock(&dev->struct_mutex);
3729 i915_gem_reset_fences(dev);
3731 /* Hack! Don't let anybody do execbuf while we don't control the chip.
3732 * We need to replace this with a semaphore, or something.
3733 * And not confound mm.suspended!
3735 dev_priv->mm.suspended = 1;
3736 del_timer_sync(&dev_priv->hangcheck_timer);
3738 i915_kernel_lost_context(dev);
3739 i915_gem_cleanup_ringbuffer(dev);
3741 mutex_unlock(&dev->struct_mutex);
3743 /* Cancel the retire work handler, which should be idle now. */
3744 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
3750 i915_gem_init_ringbuffer(struct drm_device *dev)
3752 drm_i915_private_t *dev_priv = dev->dev_private;
3755 ret = intel_init_render_ring_buffer(dev);
3760 ret = intel_init_bsd_ring_buffer(dev);
3762 goto cleanup_render_ring;
3766 ret = intel_init_blt_ring_buffer(dev);
3768 goto cleanup_bsd_ring;
3771 dev_priv->next_seqno = 1;
3776 intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
3777 cleanup_render_ring:
3778 intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
3783 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
3785 drm_i915_private_t *dev_priv = dev->dev_private;
3788 for (i = 0; i < I915_NUM_RINGS; i++)
3789 intel_cleanup_ring_buffer(&dev_priv->ring[i]);
3793 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
3794 struct drm_file *file_priv)
3796 drm_i915_private_t *dev_priv = dev->dev_private;
3799 if (drm_core_check_feature(dev, DRIVER_MODESET))
3802 if (atomic_read(&dev_priv->mm.wedged)) {
3803 DRM_ERROR("Reenabling wedged hardware, good luck\n");
3804 atomic_set(&dev_priv->mm.wedged, 0);
3807 mutex_lock(&dev->struct_mutex);
3808 dev_priv->mm.suspended = 0;
3810 ret = i915_gem_init_ringbuffer(dev);
3812 mutex_unlock(&dev->struct_mutex);
3816 BUG_ON(!list_empty(&dev_priv->mm.active_list));
3817 BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
3818 BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
3819 for (i = 0; i < I915_NUM_RINGS; i++) {
3820 BUG_ON(!list_empty(&dev_priv->ring[i].active_list));
3821 BUG_ON(!list_empty(&dev_priv->ring[i].request_list));
3823 mutex_unlock(&dev->struct_mutex);
3825 ret = drm_irq_install(dev);
3827 goto cleanup_ringbuffer;
3832 mutex_lock(&dev->struct_mutex);
3833 i915_gem_cleanup_ringbuffer(dev);
3834 dev_priv->mm.suspended = 1;
3835 mutex_unlock(&dev->struct_mutex);
3841 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
3842 struct drm_file *file_priv)
3844 if (drm_core_check_feature(dev, DRIVER_MODESET))
3847 drm_irq_uninstall(dev);
3848 return i915_gem_idle(dev);
3852 i915_gem_lastclose(struct drm_device *dev)
3856 if (drm_core_check_feature(dev, DRIVER_MODESET))
3859 ret = i915_gem_idle(dev);
3861 DRM_ERROR("failed to idle hardware: %d\n", ret);
3865 init_ring_lists(struct intel_ring_buffer *ring)
3867 INIT_LIST_HEAD(&ring->active_list);
3868 INIT_LIST_HEAD(&ring->request_list);
3869 INIT_LIST_HEAD(&ring->gpu_write_list);
3873 i915_gem_load(struct drm_device *dev)
3876 drm_i915_private_t *dev_priv = dev->dev_private;
3878 INIT_LIST_HEAD(&dev_priv->mm.active_list);
3879 INIT_LIST_HEAD(&dev_priv->mm.flushing_list);
3880 INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
3881 INIT_LIST_HEAD(&dev_priv->mm.pinned_list);
3882 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
3883 INIT_LIST_HEAD(&dev_priv->mm.deferred_free_list);
3884 INIT_LIST_HEAD(&dev_priv->mm.gtt_list);
3885 for (i = 0; i < I915_NUM_RINGS; i++)
3886 init_ring_lists(&dev_priv->ring[i]);
3887 for (i = 0; i < I915_MAX_NUM_FENCES; i++)
3888 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
3889 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
3890 i915_gem_retire_work_handler);
3891 init_completion(&dev_priv->error_completion);
3893 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
3895 u32 tmp = I915_READ(MI_ARB_STATE);
3896 if (!(tmp & MI_ARB_C3_LP_WRITE_ENABLE)) {
3897 /* arb state is a masked write, so set bit + bit in mask */
3898 tmp = MI_ARB_C3_LP_WRITE_ENABLE | (MI_ARB_C3_LP_WRITE_ENABLE << MI_ARB_MASK_SHIFT);
3899 I915_WRITE(MI_ARB_STATE, tmp);
3903 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
3905 /* Old X drivers will take 0-2 for front, back, depth buffers */
3906 if (!drm_core_check_feature(dev, DRIVER_MODESET))
3907 dev_priv->fence_reg_start = 3;
3909 if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
3910 dev_priv->num_fence_regs = 16;
3912 dev_priv->num_fence_regs = 8;
3914 /* Initialize fence registers to zero */
3915 for (i = 0; i < dev_priv->num_fence_regs; i++) {
3916 i915_gem_clear_fence_reg(dev, &dev_priv->fence_regs[i]);
3919 i915_gem_detect_bit_6_swizzle(dev);
3920 init_waitqueue_head(&dev_priv->pending_flip_queue);
3922 dev_priv->mm.interruptible = true;
3924 dev_priv->mm.inactive_shrinker.shrink = i915_gem_inactive_shrink;
3925 dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS;
3926 register_shrinker(&dev_priv->mm.inactive_shrinker);
3930 * Create a physically contiguous memory object for this object
3931 * e.g. for cursor + overlay regs
3933 static int i915_gem_init_phys_object(struct drm_device *dev,
3934 int id, int size, int align)
3936 drm_i915_private_t *dev_priv = dev->dev_private;
3937 struct drm_i915_gem_phys_object *phys_obj;
3940 if (dev_priv->mm.phys_objs[id - 1] || !size)
3943 phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL);
3949 phys_obj->handle = drm_pci_alloc(dev, size, align);
3950 if (!phys_obj->handle) {
3955 set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
3958 dev_priv->mm.phys_objs[id - 1] = phys_obj;
3966 static void i915_gem_free_phys_object(struct drm_device *dev, int id)
3968 drm_i915_private_t *dev_priv = dev->dev_private;
3969 struct drm_i915_gem_phys_object *phys_obj;
3971 if (!dev_priv->mm.phys_objs[id - 1])
3974 phys_obj = dev_priv->mm.phys_objs[id - 1];
3975 if (phys_obj->cur_obj) {
3976 i915_gem_detach_phys_object(dev, phys_obj->cur_obj);
3980 set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
3982 drm_pci_free(dev, phys_obj->handle);
3984 dev_priv->mm.phys_objs[id - 1] = NULL;
3987 void i915_gem_free_all_phys_object(struct drm_device *dev)
3991 for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++)
3992 i915_gem_free_phys_object(dev, i);
3995 void i915_gem_detach_phys_object(struct drm_device *dev,
3996 struct drm_i915_gem_object *obj)
3998 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
4005 vaddr = obj->phys_obj->handle->vaddr;
4007 page_count = obj->base.size / PAGE_SIZE;
4008 for (i = 0; i < page_count; i++) {
4009 struct page *page = shmem_read_mapping_page(mapping, i);
4010 if (!IS_ERR(page)) {
4011 char *dst = kmap_atomic(page);
4012 memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE);
4015 drm_clflush_pages(&page, 1);
4017 set_page_dirty(page);
4018 mark_page_accessed(page);
4019 page_cache_release(page);
4022 intel_gtt_chipset_flush();
4024 obj->phys_obj->cur_obj = NULL;
4025 obj->phys_obj = NULL;
4029 i915_gem_attach_phys_object(struct drm_device *dev,
4030 struct drm_i915_gem_object *obj,
4034 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
4035 drm_i915_private_t *dev_priv = dev->dev_private;
4040 if (id > I915_MAX_PHYS_OBJECT)
4043 if (obj->phys_obj) {
4044 if (obj->phys_obj->id == id)
4046 i915_gem_detach_phys_object(dev, obj);
4049 /* create a new object */
4050 if (!dev_priv->mm.phys_objs[id - 1]) {
4051 ret = i915_gem_init_phys_object(dev, id,
4052 obj->base.size, align);
4054 DRM_ERROR("failed to init phys object %d size: %zu\n",
4055 id, obj->base.size);
4060 /* bind to the object */
4061 obj->phys_obj = dev_priv->mm.phys_objs[id - 1];
4062 obj->phys_obj->cur_obj = obj;
4064 page_count = obj->base.size / PAGE_SIZE;
4066 for (i = 0; i < page_count; i++) {
4070 page = shmem_read_mapping_page(mapping, i);
4072 return PTR_ERR(page);
4074 src = kmap_atomic(page);
4075 dst = obj->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4076 memcpy(dst, src, PAGE_SIZE);
4079 mark_page_accessed(page);
4080 page_cache_release(page);
4087 i915_gem_phys_pwrite(struct drm_device *dev,
4088 struct drm_i915_gem_object *obj,
4089 struct drm_i915_gem_pwrite *args,
4090 struct drm_file *file_priv)
4092 void *vaddr = obj->phys_obj->handle->vaddr + args->offset;
4093 char __user *user_data = (char __user *) (uintptr_t) args->data_ptr;
4095 if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
4096 unsigned long unwritten;
4098 /* The physical object once assigned is fixed for the lifetime
4099 * of the obj, so we can safely drop the lock and continue
4102 mutex_unlock(&dev->struct_mutex);
4103 unwritten = copy_from_user(vaddr, user_data, args->size);
4104 mutex_lock(&dev->struct_mutex);
4109 intel_gtt_chipset_flush();
4113 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
4115 struct drm_i915_file_private *file_priv = file->driver_priv;
4117 /* Clean up our request list when the client is going away, so that
4118 * later retire_requests won't dereference our soon-to-be-gone
4121 spin_lock(&file_priv->mm.lock);
4122 while (!list_empty(&file_priv->mm.request_list)) {
4123 struct drm_i915_gem_request *request;
4125 request = list_first_entry(&file_priv->mm.request_list,
4126 struct drm_i915_gem_request,
4128 list_del(&request->client_list);
4129 request->file_priv = NULL;
4131 spin_unlock(&file_priv->mm.lock);
4135 i915_gpu_is_active(struct drm_device *dev)
4137 drm_i915_private_t *dev_priv = dev->dev_private;
4140 lists_empty = list_empty(&dev_priv->mm.flushing_list) &&
4141 list_empty(&dev_priv->mm.active_list);
4143 return !lists_empty;
4147 i915_gem_inactive_shrink(struct shrinker *shrinker, struct shrink_control *sc)
4149 struct drm_i915_private *dev_priv =
4150 container_of(shrinker,
4151 struct drm_i915_private,
4152 mm.inactive_shrinker);
4153 struct drm_device *dev = dev_priv->dev;
4154 struct drm_i915_gem_object *obj, *next;
4155 int nr_to_scan = sc->nr_to_scan;
4158 if (!mutex_trylock(&dev->struct_mutex))
4161 /* "fast-path" to count number of available objects */
4162 if (nr_to_scan == 0) {
4164 list_for_each_entry(obj,
4165 &dev_priv->mm.inactive_list,
4168 mutex_unlock(&dev->struct_mutex);
4169 return cnt / 100 * sysctl_vfs_cache_pressure;
4173 /* first scan for clean buffers */
4174 i915_gem_retire_requests(dev);
4176 list_for_each_entry_safe(obj, next,
4177 &dev_priv->mm.inactive_list,
4179 if (i915_gem_object_is_purgeable(obj)) {
4180 if (i915_gem_object_unbind(obj) == 0 &&
4186 /* second pass, evict/count anything still on the inactive list */
4188 list_for_each_entry_safe(obj, next,
4189 &dev_priv->mm.inactive_list,
4192 i915_gem_object_unbind(obj) == 0)
4198 if (nr_to_scan && i915_gpu_is_active(dev)) {
4200 * We are desperate for pages, so as a last resort, wait
4201 * for the GPU to finish and discard whatever we can.
4202 * This has a dramatic impact to reduce the number of
4203 * OOM-killer events whilst running the GPU aggressively.
4205 if (i915_gpu_idle(dev) == 0)
4208 mutex_unlock(&dev->struct_mutex);
4209 return cnt / 100 * sysctl_vfs_cache_pressure;