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 int i915_gem_mmap(struct file *filp, struct vm_area_struct *vma)
1178 int ret = drm_gem_mmap(filp, vma);
1180 pgprot_val(vma->vm_page_prot) |= _PAGE_IOMAP;
1187 * i915_gem_fault - fault a page into the GTT
1188 * vma: VMA in question
1191 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1192 * from userspace. The fault handler takes care of binding the object to
1193 * the GTT (if needed), allocating and programming a fence register (again,
1194 * only if needed based on whether the old reg is still valid or the object
1195 * is tiled) and inserting a new PTE into the faulting process.
1197 * Note that the faulting process may involve evicting existing objects
1198 * from the GTT and/or fence registers to make room. So performance may
1199 * suffer if the GTT working set is large or there are few fence registers
1202 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1204 struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1205 struct drm_device *dev = obj->base.dev;
1206 drm_i915_private_t *dev_priv = dev->dev_private;
1207 pgoff_t page_offset;
1210 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1212 /* We don't use vmf->pgoff since that has the fake offset */
1213 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1216 ret = i915_mutex_lock_interruptible(dev);
1220 trace_i915_gem_object_fault(obj, page_offset, true, write);
1222 /* Now bind it into the GTT if needed */
1223 if (!obj->map_and_fenceable) {
1224 ret = i915_gem_object_unbind(obj);
1228 if (!obj->gtt_space) {
1229 ret = i915_gem_object_bind_to_gtt(obj, 0, true);
1233 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1238 if (obj->tiling_mode == I915_TILING_NONE)
1239 ret = i915_gem_object_put_fence(obj);
1241 ret = i915_gem_object_get_fence(obj, NULL);
1245 if (i915_gem_object_is_inactive(obj))
1246 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1248 obj->fault_mappable = true;
1250 pfn = ((dev->agp->base + obj->gtt_offset) >> PAGE_SHIFT) +
1253 /* Finally, remap it using the new GTT offset */
1254 ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1256 mutex_unlock(&dev->struct_mutex);
1261 /* Give the error handler a chance to run and move the
1262 * objects off the GPU active list. Next time we service the
1263 * fault, we should be able to transition the page into the
1264 * GTT without touching the GPU (and so avoid further
1265 * EIO/EGAIN). If the GPU is wedged, then there is no issue
1266 * with coherency, just lost writes.
1272 return VM_FAULT_NOPAGE;
1274 return VM_FAULT_OOM;
1276 return VM_FAULT_SIGBUS;
1281 * i915_gem_release_mmap - remove physical page mappings
1282 * @obj: obj in question
1284 * Preserve the reservation of the mmapping with the DRM core code, but
1285 * relinquish ownership of the pages back to the system.
1287 * It is vital that we remove the page mapping if we have mapped a tiled
1288 * object through the GTT and then lose the fence register due to
1289 * resource pressure. Similarly if the object has been moved out of the
1290 * aperture, than pages mapped into userspace must be revoked. Removing the
1291 * mapping will then trigger a page fault on the next user access, allowing
1292 * fixup by i915_gem_fault().
1295 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1297 if (!obj->fault_mappable)
1300 if (obj->base.dev->dev_mapping)
1301 unmap_mapping_range(obj->base.dev->dev_mapping,
1302 (loff_t)obj->base.map_list.hash.key<<PAGE_SHIFT,
1305 obj->fault_mappable = false;
1309 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1313 if (INTEL_INFO(dev)->gen >= 4 ||
1314 tiling_mode == I915_TILING_NONE)
1317 /* Previous chips need a power-of-two fence region when tiling */
1318 if (INTEL_INFO(dev)->gen == 3)
1319 gtt_size = 1024*1024;
1321 gtt_size = 512*1024;
1323 while (gtt_size < size)
1330 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1331 * @obj: object to check
1333 * Return the required GTT alignment for an object, taking into account
1334 * potential fence register mapping.
1337 i915_gem_get_gtt_alignment(struct drm_device *dev,
1342 * Minimum alignment is 4k (GTT page size), but might be greater
1343 * if a fence register is needed for the object.
1345 if (INTEL_INFO(dev)->gen >= 4 ||
1346 tiling_mode == I915_TILING_NONE)
1350 * Previous chips need to be aligned to the size of the smallest
1351 * fence register that can contain the object.
1353 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1357 * i915_gem_get_unfenced_gtt_alignment - return required GTT alignment for an
1360 * @size: size of the object
1361 * @tiling_mode: tiling mode of the object
1363 * Return the required GTT alignment for an object, only taking into account
1364 * unfenced tiled surface requirements.
1367 i915_gem_get_unfenced_gtt_alignment(struct drm_device *dev,
1372 * Minimum alignment is 4k (GTT page size) for sane hw.
1374 if (INTEL_INFO(dev)->gen >= 4 || IS_G33(dev) ||
1375 tiling_mode == I915_TILING_NONE)
1378 /* Previous hardware however needs to be aligned to a power-of-two
1379 * tile height. The simplest method for determining this is to reuse
1380 * the power-of-tile object size.
1382 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1386 i915_gem_mmap_gtt(struct drm_file *file,
1387 struct drm_device *dev,
1391 struct drm_i915_private *dev_priv = dev->dev_private;
1392 struct drm_i915_gem_object *obj;
1395 if (!(dev->driver->driver_features & DRIVER_GEM))
1398 ret = i915_mutex_lock_interruptible(dev);
1402 obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
1403 if (&obj->base == NULL) {
1408 if (obj->base.size > dev_priv->mm.gtt_mappable_end) {
1413 if (obj->madv != I915_MADV_WILLNEED) {
1414 DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1419 if (!obj->base.map_list.map) {
1420 ret = drm_gem_create_mmap_offset(&obj->base);
1425 *offset = (u64)obj->base.map_list.hash.key << PAGE_SHIFT;
1428 drm_gem_object_unreference(&obj->base);
1430 mutex_unlock(&dev->struct_mutex);
1435 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1437 * @data: GTT mapping ioctl data
1438 * @file: GEM object info
1440 * Simply returns the fake offset to userspace so it can mmap it.
1441 * The mmap call will end up in drm_gem_mmap(), which will set things
1442 * up so we can get faults in the handler above.
1444 * The fault handler will take care of binding the object into the GTT
1445 * (since it may have been evicted to make room for something), allocating
1446 * a fence register, and mapping the appropriate aperture address into
1450 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1451 struct drm_file *file)
1453 struct drm_i915_gem_mmap_gtt *args = data;
1455 if (!(dev->driver->driver_features & DRIVER_GEM))
1458 return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
1463 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj,
1467 struct address_space *mapping;
1468 struct inode *inode;
1471 /* Get the list of pages out of our struct file. They'll be pinned
1472 * at this point until we release them.
1474 page_count = obj->base.size / PAGE_SIZE;
1475 BUG_ON(obj->pages != NULL);
1476 obj->pages = drm_malloc_ab(page_count, sizeof(struct page *));
1477 if (obj->pages == NULL)
1480 inode = obj->base.filp->f_path.dentry->d_inode;
1481 mapping = inode->i_mapping;
1482 gfpmask |= mapping_gfp_mask(mapping);
1484 for (i = 0; i < page_count; i++) {
1485 page = shmem_read_mapping_page_gfp(mapping, i, gfpmask);
1489 obj->pages[i] = page;
1492 if (i915_gem_object_needs_bit17_swizzle(obj))
1493 i915_gem_object_do_bit_17_swizzle(obj);
1499 page_cache_release(obj->pages[i]);
1501 drm_free_large(obj->pages);
1503 return PTR_ERR(page);
1507 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
1509 int page_count = obj->base.size / PAGE_SIZE;
1512 BUG_ON(obj->madv == __I915_MADV_PURGED);
1514 if (i915_gem_object_needs_bit17_swizzle(obj))
1515 i915_gem_object_save_bit_17_swizzle(obj);
1517 if (obj->madv == I915_MADV_DONTNEED)
1520 for (i = 0; i < page_count; i++) {
1522 set_page_dirty(obj->pages[i]);
1524 if (obj->madv == I915_MADV_WILLNEED)
1525 mark_page_accessed(obj->pages[i]);
1527 page_cache_release(obj->pages[i]);
1531 drm_free_large(obj->pages);
1536 i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
1537 struct intel_ring_buffer *ring,
1540 struct drm_device *dev = obj->base.dev;
1541 struct drm_i915_private *dev_priv = dev->dev_private;
1543 BUG_ON(ring == NULL);
1546 /* Add a reference if we're newly entering the active list. */
1548 drm_gem_object_reference(&obj->base);
1552 /* Move from whatever list we were on to the tail of execution. */
1553 list_move_tail(&obj->mm_list, &dev_priv->mm.active_list);
1554 list_move_tail(&obj->ring_list, &ring->active_list);
1556 obj->last_rendering_seqno = seqno;
1557 if (obj->fenced_gpu_access) {
1558 struct drm_i915_fence_reg *reg;
1560 BUG_ON(obj->fence_reg == I915_FENCE_REG_NONE);
1562 obj->last_fenced_seqno = seqno;
1563 obj->last_fenced_ring = ring;
1565 reg = &dev_priv->fence_regs[obj->fence_reg];
1566 list_move_tail(®->lru_list, &dev_priv->mm.fence_list);
1571 i915_gem_object_move_off_active(struct drm_i915_gem_object *obj)
1573 list_del_init(&obj->ring_list);
1574 obj->last_rendering_seqno = 0;
1578 i915_gem_object_move_to_flushing(struct drm_i915_gem_object *obj)
1580 struct drm_device *dev = obj->base.dev;
1581 drm_i915_private_t *dev_priv = dev->dev_private;
1583 BUG_ON(!obj->active);
1584 list_move_tail(&obj->mm_list, &dev_priv->mm.flushing_list);
1586 i915_gem_object_move_off_active(obj);
1590 i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
1592 struct drm_device *dev = obj->base.dev;
1593 struct drm_i915_private *dev_priv = dev->dev_private;
1595 if (obj->pin_count != 0)
1596 list_move_tail(&obj->mm_list, &dev_priv->mm.pinned_list);
1598 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1600 BUG_ON(!list_empty(&obj->gpu_write_list));
1601 BUG_ON(!obj->active);
1604 i915_gem_object_move_off_active(obj);
1605 obj->fenced_gpu_access = false;
1608 obj->pending_gpu_write = false;
1609 drm_gem_object_unreference(&obj->base);
1611 WARN_ON(i915_verify_lists(dev));
1614 /* Immediately discard the backing storage */
1616 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
1618 struct inode *inode;
1620 /* Our goal here is to return as much of the memory as
1621 * is possible back to the system as we are called from OOM.
1622 * To do this we must instruct the shmfs to drop all of its
1623 * backing pages, *now*.
1625 inode = obj->base.filp->f_path.dentry->d_inode;
1626 shmem_truncate_range(inode, 0, (loff_t)-1);
1628 obj->madv = __I915_MADV_PURGED;
1632 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj)
1634 return obj->madv == I915_MADV_DONTNEED;
1638 i915_gem_process_flushing_list(struct intel_ring_buffer *ring,
1639 uint32_t flush_domains)
1641 struct drm_i915_gem_object *obj, *next;
1643 list_for_each_entry_safe(obj, next,
1644 &ring->gpu_write_list,
1646 if (obj->base.write_domain & flush_domains) {
1647 uint32_t old_write_domain = obj->base.write_domain;
1649 obj->base.write_domain = 0;
1650 list_del_init(&obj->gpu_write_list);
1651 i915_gem_object_move_to_active(obj, ring,
1652 i915_gem_next_request_seqno(ring));
1654 trace_i915_gem_object_change_domain(obj,
1655 obj->base.read_domains,
1662 i915_add_request(struct intel_ring_buffer *ring,
1663 struct drm_file *file,
1664 struct drm_i915_gem_request *request)
1666 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1671 BUG_ON(request == NULL);
1673 ret = ring->add_request(ring, &seqno);
1677 trace_i915_gem_request_add(ring, seqno);
1679 request->seqno = seqno;
1680 request->ring = ring;
1681 request->emitted_jiffies = jiffies;
1682 was_empty = list_empty(&ring->request_list);
1683 list_add_tail(&request->list, &ring->request_list);
1686 struct drm_i915_file_private *file_priv = file->driver_priv;
1688 spin_lock(&file_priv->mm.lock);
1689 request->file_priv = file_priv;
1690 list_add_tail(&request->client_list,
1691 &file_priv->mm.request_list);
1692 spin_unlock(&file_priv->mm.lock);
1695 ring->outstanding_lazy_request = false;
1697 if (!dev_priv->mm.suspended) {
1698 if (i915_enable_hangcheck) {
1699 mod_timer(&dev_priv->hangcheck_timer,
1701 msecs_to_jiffies(DRM_I915_HANGCHECK_PERIOD));
1704 queue_delayed_work(dev_priv->wq,
1705 &dev_priv->mm.retire_work, HZ);
1711 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
1713 struct drm_i915_file_private *file_priv = request->file_priv;
1718 spin_lock(&file_priv->mm.lock);
1719 if (request->file_priv) {
1720 list_del(&request->client_list);
1721 request->file_priv = NULL;
1723 spin_unlock(&file_priv->mm.lock);
1726 static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv,
1727 struct intel_ring_buffer *ring)
1729 while (!list_empty(&ring->request_list)) {
1730 struct drm_i915_gem_request *request;
1732 request = list_first_entry(&ring->request_list,
1733 struct drm_i915_gem_request,
1736 list_del(&request->list);
1737 i915_gem_request_remove_from_client(request);
1741 while (!list_empty(&ring->active_list)) {
1742 struct drm_i915_gem_object *obj;
1744 obj = list_first_entry(&ring->active_list,
1745 struct drm_i915_gem_object,
1748 obj->base.write_domain = 0;
1749 list_del_init(&obj->gpu_write_list);
1750 i915_gem_object_move_to_inactive(obj);
1754 static void i915_gem_reset_fences(struct drm_device *dev)
1756 struct drm_i915_private *dev_priv = dev->dev_private;
1759 for (i = 0; i < dev_priv->num_fence_regs; i++) {
1760 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
1761 struct drm_i915_gem_object *obj = reg->obj;
1766 if (obj->tiling_mode)
1767 i915_gem_release_mmap(obj);
1769 reg->obj->fence_reg = I915_FENCE_REG_NONE;
1770 reg->obj->fenced_gpu_access = false;
1771 reg->obj->last_fenced_seqno = 0;
1772 reg->obj->last_fenced_ring = NULL;
1773 i915_gem_clear_fence_reg(dev, reg);
1777 void i915_gem_reset(struct drm_device *dev)
1779 struct drm_i915_private *dev_priv = dev->dev_private;
1780 struct drm_i915_gem_object *obj;
1783 for (i = 0; i < I915_NUM_RINGS; i++)
1784 i915_gem_reset_ring_lists(dev_priv, &dev_priv->ring[i]);
1786 /* Remove anything from the flushing lists. The GPU cache is likely
1787 * to be lost on reset along with the data, so simply move the
1788 * lost bo to the inactive list.
1790 while (!list_empty(&dev_priv->mm.flushing_list)) {
1791 obj = list_first_entry(&dev_priv->mm.flushing_list,
1792 struct drm_i915_gem_object,
1795 obj->base.write_domain = 0;
1796 list_del_init(&obj->gpu_write_list);
1797 i915_gem_object_move_to_inactive(obj);
1800 /* Move everything out of the GPU domains to ensure we do any
1801 * necessary invalidation upon reuse.
1803 list_for_each_entry(obj,
1804 &dev_priv->mm.inactive_list,
1807 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
1810 /* The fence registers are invalidated so clear them out */
1811 i915_gem_reset_fences(dev);
1815 * This function clears the request list as sequence numbers are passed.
1818 i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
1823 if (list_empty(&ring->request_list))
1826 WARN_ON(i915_verify_lists(ring->dev));
1828 seqno = ring->get_seqno(ring);
1830 for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++)
1831 if (seqno >= ring->sync_seqno[i])
1832 ring->sync_seqno[i] = 0;
1834 while (!list_empty(&ring->request_list)) {
1835 struct drm_i915_gem_request *request;
1837 request = list_first_entry(&ring->request_list,
1838 struct drm_i915_gem_request,
1841 if (!i915_seqno_passed(seqno, request->seqno))
1844 trace_i915_gem_request_retire(ring, request->seqno);
1846 list_del(&request->list);
1847 i915_gem_request_remove_from_client(request);
1851 /* Move any buffers on the active list that are no longer referenced
1852 * by the ringbuffer to the flushing/inactive lists as appropriate.
1854 while (!list_empty(&ring->active_list)) {
1855 struct drm_i915_gem_object *obj;
1857 obj = list_first_entry(&ring->active_list,
1858 struct drm_i915_gem_object,
1861 if (!i915_seqno_passed(seqno, obj->last_rendering_seqno))
1864 if (obj->base.write_domain != 0)
1865 i915_gem_object_move_to_flushing(obj);
1867 i915_gem_object_move_to_inactive(obj);
1870 if (unlikely(ring->trace_irq_seqno &&
1871 i915_seqno_passed(seqno, ring->trace_irq_seqno))) {
1872 ring->irq_put(ring);
1873 ring->trace_irq_seqno = 0;
1876 WARN_ON(i915_verify_lists(ring->dev));
1880 i915_gem_retire_requests(struct drm_device *dev)
1882 drm_i915_private_t *dev_priv = dev->dev_private;
1885 if (!list_empty(&dev_priv->mm.deferred_free_list)) {
1886 struct drm_i915_gem_object *obj, *next;
1888 /* We must be careful that during unbind() we do not
1889 * accidentally infinitely recurse into retire requests.
1891 * retire -> free -> unbind -> wait -> retire_ring
1893 list_for_each_entry_safe(obj, next,
1894 &dev_priv->mm.deferred_free_list,
1896 i915_gem_free_object_tail(obj);
1899 for (i = 0; i < I915_NUM_RINGS; i++)
1900 i915_gem_retire_requests_ring(&dev_priv->ring[i]);
1904 i915_gem_retire_work_handler(struct work_struct *work)
1906 drm_i915_private_t *dev_priv;
1907 struct drm_device *dev;
1911 dev_priv = container_of(work, drm_i915_private_t,
1912 mm.retire_work.work);
1913 dev = dev_priv->dev;
1915 /* Come back later if the device is busy... */
1916 if (!mutex_trylock(&dev->struct_mutex)) {
1917 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1921 i915_gem_retire_requests(dev);
1923 /* Send a periodic flush down the ring so we don't hold onto GEM
1924 * objects indefinitely.
1927 for (i = 0; i < I915_NUM_RINGS; i++) {
1928 struct intel_ring_buffer *ring = &dev_priv->ring[i];
1930 if (!list_empty(&ring->gpu_write_list)) {
1931 struct drm_i915_gem_request *request;
1934 ret = i915_gem_flush_ring(ring,
1935 0, I915_GEM_GPU_DOMAINS);
1936 request = kzalloc(sizeof(*request), GFP_KERNEL);
1937 if (ret || request == NULL ||
1938 i915_add_request(ring, NULL, request))
1942 idle &= list_empty(&ring->request_list);
1945 if (!dev_priv->mm.suspended && !idle)
1946 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1948 mutex_unlock(&dev->struct_mutex);
1952 * Waits for a sequence number to be signaled, and cleans up the
1953 * request and object lists appropriately for that event.
1956 i915_wait_request(struct intel_ring_buffer *ring,
1959 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1965 if (atomic_read(&dev_priv->mm.wedged)) {
1966 struct completion *x = &dev_priv->error_completion;
1967 bool recovery_complete;
1968 unsigned long flags;
1970 /* Give the error handler a chance to run. */
1971 spin_lock_irqsave(&x->wait.lock, flags);
1972 recovery_complete = x->done > 0;
1973 spin_unlock_irqrestore(&x->wait.lock, flags);
1975 return recovery_complete ? -EIO : -EAGAIN;
1978 if (seqno == ring->outstanding_lazy_request) {
1979 struct drm_i915_gem_request *request;
1981 request = kzalloc(sizeof(*request), GFP_KERNEL);
1982 if (request == NULL)
1985 ret = i915_add_request(ring, NULL, request);
1991 seqno = request->seqno;
1994 if (!i915_seqno_passed(ring->get_seqno(ring), seqno)) {
1995 if (HAS_PCH_SPLIT(ring->dev))
1996 ier = I915_READ(DEIER) | I915_READ(GTIER);
1998 ier = I915_READ(IER);
2000 DRM_ERROR("something (likely vbetool) disabled "
2001 "interrupts, re-enabling\n");
2002 ring->dev->driver->irq_preinstall(ring->dev);
2003 ring->dev->driver->irq_postinstall(ring->dev);
2006 trace_i915_gem_request_wait_begin(ring, seqno);
2008 ring->waiting_seqno = seqno;
2009 if (ring->irq_get(ring)) {
2010 if (dev_priv->mm.interruptible)
2011 ret = wait_event_interruptible(ring->irq_queue,
2012 i915_seqno_passed(ring->get_seqno(ring), seqno)
2013 || atomic_read(&dev_priv->mm.wedged));
2015 wait_event(ring->irq_queue,
2016 i915_seqno_passed(ring->get_seqno(ring), seqno)
2017 || atomic_read(&dev_priv->mm.wedged));
2019 ring->irq_put(ring);
2020 } else if (wait_for_atomic(i915_seqno_passed(ring->get_seqno(ring),
2022 atomic_read(&dev_priv->mm.wedged), 3000))
2024 ring->waiting_seqno = 0;
2026 trace_i915_gem_request_wait_end(ring, seqno);
2028 if (atomic_read(&dev_priv->mm.wedged))
2031 if (ret && ret != -ERESTARTSYS)
2032 DRM_ERROR("%s returns %d (awaiting %d at %d, next %d)\n",
2033 __func__, ret, seqno, ring->get_seqno(ring),
2034 dev_priv->next_seqno);
2036 /* Directly dispatch request retiring. While we have the work queue
2037 * to handle this, the waiter on a request often wants an associated
2038 * buffer to have made it to the inactive list, and we would need
2039 * a separate wait queue to handle that.
2042 i915_gem_retire_requests_ring(ring);
2048 * Ensures that all rendering to the object has completed and the object is
2049 * safe to unbind from the GTT or access from the CPU.
2052 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj)
2056 /* This function only exists to support waiting for existing rendering,
2057 * not for emitting required flushes.
2059 BUG_ON((obj->base.write_domain & I915_GEM_GPU_DOMAINS) != 0);
2061 /* If there is rendering queued on the buffer being evicted, wait for
2065 ret = i915_wait_request(obj->ring, obj->last_rendering_seqno);
2073 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
2075 u32 old_write_domain, old_read_domains;
2077 /* Act a barrier for all accesses through the GTT */
2080 /* Force a pagefault for domain tracking on next user access */
2081 i915_gem_release_mmap(obj);
2083 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
2086 old_read_domains = obj->base.read_domains;
2087 old_write_domain = obj->base.write_domain;
2089 obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
2090 obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
2092 trace_i915_gem_object_change_domain(obj,
2098 * Unbinds an object from the GTT aperture.
2101 i915_gem_object_unbind(struct drm_i915_gem_object *obj)
2105 if (obj->gtt_space == NULL)
2108 if (obj->pin_count != 0) {
2109 DRM_ERROR("Attempting to unbind pinned buffer\n");
2113 ret = i915_gem_object_finish_gpu(obj);
2114 if (ret == -ERESTARTSYS)
2116 /* Continue on if we fail due to EIO, the GPU is hung so we
2117 * should be safe and we need to cleanup or else we might
2118 * cause memory corruption through use-after-free.
2121 i915_gem_object_finish_gtt(obj);
2123 /* Move the object to the CPU domain to ensure that
2124 * any possible CPU writes while it's not in the GTT
2125 * are flushed when we go to remap it.
2128 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
2129 if (ret == -ERESTARTSYS)
2132 /* In the event of a disaster, abandon all caches and
2133 * hope for the best.
2135 i915_gem_clflush_object(obj);
2136 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
2139 /* release the fence reg _after_ flushing */
2140 ret = i915_gem_object_put_fence(obj);
2141 if (ret == -ERESTARTSYS)
2144 trace_i915_gem_object_unbind(obj);
2146 i915_gem_gtt_unbind_object(obj);
2147 i915_gem_object_put_pages_gtt(obj);
2149 list_del_init(&obj->gtt_list);
2150 list_del_init(&obj->mm_list);
2151 /* Avoid an unnecessary call to unbind on rebind. */
2152 obj->map_and_fenceable = true;
2154 drm_mm_put_block(obj->gtt_space);
2155 obj->gtt_space = NULL;
2156 obj->gtt_offset = 0;
2158 if (i915_gem_object_is_purgeable(obj))
2159 i915_gem_object_truncate(obj);
2165 i915_gem_flush_ring(struct intel_ring_buffer *ring,
2166 uint32_t invalidate_domains,
2167 uint32_t flush_domains)
2171 if (((invalidate_domains | flush_domains) & I915_GEM_GPU_DOMAINS) == 0)
2174 trace_i915_gem_ring_flush(ring, invalidate_domains, flush_domains);
2176 ret = ring->flush(ring, invalidate_domains, flush_domains);
2180 if (flush_domains & I915_GEM_GPU_DOMAINS)
2181 i915_gem_process_flushing_list(ring, flush_domains);
2186 static int i915_ring_idle(struct intel_ring_buffer *ring)
2190 if (list_empty(&ring->gpu_write_list) && list_empty(&ring->active_list))
2193 if (!list_empty(&ring->gpu_write_list)) {
2194 ret = i915_gem_flush_ring(ring,
2195 I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
2200 return i915_wait_request(ring, i915_gem_next_request_seqno(ring));
2204 i915_gpu_idle(struct drm_device *dev)
2206 drm_i915_private_t *dev_priv = dev->dev_private;
2209 /* Flush everything onto the inactive list. */
2210 for (i = 0; i < I915_NUM_RINGS; i++) {
2211 ret = i915_ring_idle(&dev_priv->ring[i]);
2219 static int sandybridge_write_fence_reg(struct drm_i915_gem_object *obj,
2220 struct intel_ring_buffer *pipelined)
2222 struct drm_device *dev = obj->base.dev;
2223 drm_i915_private_t *dev_priv = dev->dev_private;
2224 u32 size = obj->gtt_space->size;
2225 int regnum = obj->fence_reg;
2228 val = (uint64_t)((obj->gtt_offset + size - 4096) &
2230 val |= obj->gtt_offset & 0xfffff000;
2231 val |= (uint64_t)((obj->stride / 128) - 1) <<
2232 SANDYBRIDGE_FENCE_PITCH_SHIFT;
2234 if (obj->tiling_mode == I915_TILING_Y)
2235 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2236 val |= I965_FENCE_REG_VALID;
2239 int ret = intel_ring_begin(pipelined, 6);
2243 intel_ring_emit(pipelined, MI_NOOP);
2244 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(2));
2245 intel_ring_emit(pipelined, FENCE_REG_SANDYBRIDGE_0 + regnum*8);
2246 intel_ring_emit(pipelined, (u32)val);
2247 intel_ring_emit(pipelined, FENCE_REG_SANDYBRIDGE_0 + regnum*8 + 4);
2248 intel_ring_emit(pipelined, (u32)(val >> 32));
2249 intel_ring_advance(pipelined);
2251 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + regnum * 8, val);
2256 static int i965_write_fence_reg(struct drm_i915_gem_object *obj,
2257 struct intel_ring_buffer *pipelined)
2259 struct drm_device *dev = obj->base.dev;
2260 drm_i915_private_t *dev_priv = dev->dev_private;
2261 u32 size = obj->gtt_space->size;
2262 int regnum = obj->fence_reg;
2265 val = (uint64_t)((obj->gtt_offset + size - 4096) &
2267 val |= obj->gtt_offset & 0xfffff000;
2268 val |= ((obj->stride / 128) - 1) << I965_FENCE_PITCH_SHIFT;
2269 if (obj->tiling_mode == I915_TILING_Y)
2270 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2271 val |= I965_FENCE_REG_VALID;
2274 int ret = intel_ring_begin(pipelined, 6);
2278 intel_ring_emit(pipelined, MI_NOOP);
2279 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(2));
2280 intel_ring_emit(pipelined, FENCE_REG_965_0 + regnum*8);
2281 intel_ring_emit(pipelined, (u32)val);
2282 intel_ring_emit(pipelined, FENCE_REG_965_0 + regnum*8 + 4);
2283 intel_ring_emit(pipelined, (u32)(val >> 32));
2284 intel_ring_advance(pipelined);
2286 I915_WRITE64(FENCE_REG_965_0 + regnum * 8, val);
2291 static int i915_write_fence_reg(struct drm_i915_gem_object *obj,
2292 struct intel_ring_buffer *pipelined)
2294 struct drm_device *dev = obj->base.dev;
2295 drm_i915_private_t *dev_priv = dev->dev_private;
2296 u32 size = obj->gtt_space->size;
2297 u32 fence_reg, val, pitch_val;
2300 if (WARN((obj->gtt_offset & ~I915_FENCE_START_MASK) ||
2301 (size & -size) != size ||
2302 (obj->gtt_offset & (size - 1)),
2303 "object 0x%08x [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
2304 obj->gtt_offset, obj->map_and_fenceable, size))
2307 if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
2312 /* Note: pitch better be a power of two tile widths */
2313 pitch_val = obj->stride / tile_width;
2314 pitch_val = ffs(pitch_val) - 1;
2316 val = obj->gtt_offset;
2317 if (obj->tiling_mode == I915_TILING_Y)
2318 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2319 val |= I915_FENCE_SIZE_BITS(size);
2320 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2321 val |= I830_FENCE_REG_VALID;
2323 fence_reg = obj->fence_reg;
2325 fence_reg = FENCE_REG_830_0 + fence_reg * 4;
2327 fence_reg = FENCE_REG_945_8 + (fence_reg - 8) * 4;
2330 int ret = intel_ring_begin(pipelined, 4);
2334 intel_ring_emit(pipelined, MI_NOOP);
2335 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(1));
2336 intel_ring_emit(pipelined, fence_reg);
2337 intel_ring_emit(pipelined, val);
2338 intel_ring_advance(pipelined);
2340 I915_WRITE(fence_reg, val);
2345 static int i830_write_fence_reg(struct drm_i915_gem_object *obj,
2346 struct intel_ring_buffer *pipelined)
2348 struct drm_device *dev = obj->base.dev;
2349 drm_i915_private_t *dev_priv = dev->dev_private;
2350 u32 size = obj->gtt_space->size;
2351 int regnum = obj->fence_reg;
2355 if (WARN((obj->gtt_offset & ~I830_FENCE_START_MASK) ||
2356 (size & -size) != size ||
2357 (obj->gtt_offset & (size - 1)),
2358 "object 0x%08x not 512K or pot-size 0x%08x aligned\n",
2359 obj->gtt_offset, size))
2362 pitch_val = obj->stride / 128;
2363 pitch_val = ffs(pitch_val) - 1;
2365 val = obj->gtt_offset;
2366 if (obj->tiling_mode == I915_TILING_Y)
2367 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2368 val |= I830_FENCE_SIZE_BITS(size);
2369 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2370 val |= I830_FENCE_REG_VALID;
2373 int ret = intel_ring_begin(pipelined, 4);
2377 intel_ring_emit(pipelined, MI_NOOP);
2378 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(1));
2379 intel_ring_emit(pipelined, FENCE_REG_830_0 + regnum*4);
2380 intel_ring_emit(pipelined, val);
2381 intel_ring_advance(pipelined);
2383 I915_WRITE(FENCE_REG_830_0 + regnum * 4, val);
2388 static bool ring_passed_seqno(struct intel_ring_buffer *ring, u32 seqno)
2390 return i915_seqno_passed(ring->get_seqno(ring), seqno);
2394 i915_gem_object_flush_fence(struct drm_i915_gem_object *obj,
2395 struct intel_ring_buffer *pipelined)
2399 if (obj->fenced_gpu_access) {
2400 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
2401 ret = i915_gem_flush_ring(obj->last_fenced_ring,
2402 0, obj->base.write_domain);
2407 obj->fenced_gpu_access = false;
2410 if (obj->last_fenced_seqno && pipelined != obj->last_fenced_ring) {
2411 if (!ring_passed_seqno(obj->last_fenced_ring,
2412 obj->last_fenced_seqno)) {
2413 ret = i915_wait_request(obj->last_fenced_ring,
2414 obj->last_fenced_seqno);
2419 obj->last_fenced_seqno = 0;
2420 obj->last_fenced_ring = NULL;
2423 /* Ensure that all CPU reads are completed before installing a fence
2424 * and all writes before removing the fence.
2426 if (obj->base.read_domains & I915_GEM_DOMAIN_GTT)
2433 i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
2437 if (obj->tiling_mode)
2438 i915_gem_release_mmap(obj);
2440 ret = i915_gem_object_flush_fence(obj, NULL);
2444 if (obj->fence_reg != I915_FENCE_REG_NONE) {
2445 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2446 i915_gem_clear_fence_reg(obj->base.dev,
2447 &dev_priv->fence_regs[obj->fence_reg]);
2449 obj->fence_reg = I915_FENCE_REG_NONE;
2455 static struct drm_i915_fence_reg *
2456 i915_find_fence_reg(struct drm_device *dev,
2457 struct intel_ring_buffer *pipelined)
2459 struct drm_i915_private *dev_priv = dev->dev_private;
2460 struct drm_i915_fence_reg *reg, *first, *avail;
2463 /* First try to find a free reg */
2465 for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2466 reg = &dev_priv->fence_regs[i];
2470 if (!reg->obj->pin_count)
2477 /* None available, try to steal one or wait for a user to finish */
2478 avail = first = NULL;
2479 list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
2480 if (reg->obj->pin_count)
2487 !reg->obj->last_fenced_ring ||
2488 reg->obj->last_fenced_ring == pipelined) {
2501 * i915_gem_object_get_fence - set up a fence reg for an object
2502 * @obj: object to map through a fence reg
2503 * @pipelined: ring on which to queue the change, or NULL for CPU access
2504 * @interruptible: must we wait uninterruptibly for the register to retire?
2506 * When mapping objects through the GTT, userspace wants to be able to write
2507 * to them without having to worry about swizzling if the object is tiled.
2509 * This function walks the fence regs looking for a free one for @obj,
2510 * stealing one if it can't find any.
2512 * It then sets up the reg based on the object's properties: address, pitch
2513 * and tiling format.
2516 i915_gem_object_get_fence(struct drm_i915_gem_object *obj,
2517 struct intel_ring_buffer *pipelined)
2519 struct drm_device *dev = obj->base.dev;
2520 struct drm_i915_private *dev_priv = dev->dev_private;
2521 struct drm_i915_fence_reg *reg;
2524 /* XXX disable pipelining. There are bugs. Shocking. */
2527 /* Just update our place in the LRU if our fence is getting reused. */
2528 if (obj->fence_reg != I915_FENCE_REG_NONE) {
2529 reg = &dev_priv->fence_regs[obj->fence_reg];
2530 list_move_tail(®->lru_list, &dev_priv->mm.fence_list);
2532 if (obj->tiling_changed) {
2533 ret = i915_gem_object_flush_fence(obj, pipelined);
2537 if (!obj->fenced_gpu_access && !obj->last_fenced_seqno)
2542 i915_gem_next_request_seqno(pipelined);
2543 obj->last_fenced_seqno = reg->setup_seqno;
2544 obj->last_fenced_ring = pipelined;
2551 if (reg->setup_seqno) {
2552 if (!ring_passed_seqno(obj->last_fenced_ring,
2553 reg->setup_seqno)) {
2554 ret = i915_wait_request(obj->last_fenced_ring,
2560 reg->setup_seqno = 0;
2562 } else if (obj->last_fenced_ring &&
2563 obj->last_fenced_ring != pipelined) {
2564 ret = i915_gem_object_flush_fence(obj, pipelined);
2572 reg = i915_find_fence_reg(dev, pipelined);
2576 ret = i915_gem_object_flush_fence(obj, pipelined);
2581 struct drm_i915_gem_object *old = reg->obj;
2583 drm_gem_object_reference(&old->base);
2585 if (old->tiling_mode)
2586 i915_gem_release_mmap(old);
2588 ret = i915_gem_object_flush_fence(old, pipelined);
2590 drm_gem_object_unreference(&old->base);
2594 if (old->last_fenced_seqno == 0 && obj->last_fenced_seqno == 0)
2597 old->fence_reg = I915_FENCE_REG_NONE;
2598 old->last_fenced_ring = pipelined;
2599 old->last_fenced_seqno =
2600 pipelined ? i915_gem_next_request_seqno(pipelined) : 0;
2602 drm_gem_object_unreference(&old->base);
2603 } else if (obj->last_fenced_seqno == 0)
2607 list_move_tail(®->lru_list, &dev_priv->mm.fence_list);
2608 obj->fence_reg = reg - dev_priv->fence_regs;
2609 obj->last_fenced_ring = pipelined;
2612 pipelined ? i915_gem_next_request_seqno(pipelined) : 0;
2613 obj->last_fenced_seqno = reg->setup_seqno;
2616 obj->tiling_changed = false;
2617 switch (INTEL_INFO(dev)->gen) {
2620 ret = sandybridge_write_fence_reg(obj, pipelined);
2624 ret = i965_write_fence_reg(obj, pipelined);
2627 ret = i915_write_fence_reg(obj, pipelined);
2630 ret = i830_write_fence_reg(obj, pipelined);
2638 * i915_gem_clear_fence_reg - clear out fence register info
2639 * @obj: object to clear
2641 * Zeroes out the fence register itself and clears out the associated
2642 * data structures in dev_priv and obj.
2645 i915_gem_clear_fence_reg(struct drm_device *dev,
2646 struct drm_i915_fence_reg *reg)
2648 drm_i915_private_t *dev_priv = dev->dev_private;
2649 uint32_t fence_reg = reg - dev_priv->fence_regs;
2651 switch (INTEL_INFO(dev)->gen) {
2654 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + fence_reg*8, 0);
2658 I915_WRITE64(FENCE_REG_965_0 + fence_reg*8, 0);
2662 fence_reg = FENCE_REG_945_8 + (fence_reg - 8) * 4;
2665 fence_reg = FENCE_REG_830_0 + fence_reg * 4;
2667 I915_WRITE(fence_reg, 0);
2671 list_del_init(®->lru_list);
2673 reg->setup_seqno = 0;
2677 * Finds free space in the GTT aperture and binds the object there.
2680 i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
2682 bool map_and_fenceable)
2684 struct drm_device *dev = obj->base.dev;
2685 drm_i915_private_t *dev_priv = dev->dev_private;
2686 struct drm_mm_node *free_space;
2687 gfp_t gfpmask = __GFP_NORETRY | __GFP_NOWARN;
2688 u32 size, fence_size, fence_alignment, unfenced_alignment;
2689 bool mappable, fenceable;
2692 if (obj->madv != I915_MADV_WILLNEED) {
2693 DRM_ERROR("Attempting to bind a purgeable object\n");
2697 fence_size = i915_gem_get_gtt_size(dev,
2700 fence_alignment = i915_gem_get_gtt_alignment(dev,
2703 unfenced_alignment =
2704 i915_gem_get_unfenced_gtt_alignment(dev,
2709 alignment = map_and_fenceable ? fence_alignment :
2711 if (map_and_fenceable && alignment & (fence_alignment - 1)) {
2712 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
2716 size = map_and_fenceable ? fence_size : obj->base.size;
2718 /* If the object is bigger than the entire aperture, reject it early
2719 * before evicting everything in a vain attempt to find space.
2721 if (obj->base.size >
2722 (map_and_fenceable ? dev_priv->mm.gtt_mappable_end : dev_priv->mm.gtt_total)) {
2723 DRM_ERROR("Attempting to bind an object larger than the aperture\n");
2728 if (map_and_fenceable)
2730 drm_mm_search_free_in_range(&dev_priv->mm.gtt_space,
2732 dev_priv->mm.gtt_mappable_end,
2735 free_space = drm_mm_search_free(&dev_priv->mm.gtt_space,
2736 size, alignment, 0);
2738 if (free_space != NULL) {
2739 if (map_and_fenceable)
2741 drm_mm_get_block_range_generic(free_space,
2743 dev_priv->mm.gtt_mappable_end,
2747 drm_mm_get_block(free_space, size, alignment);
2749 if (obj->gtt_space == NULL) {
2750 /* If the gtt is empty and we're still having trouble
2751 * fitting our object in, we're out of memory.
2753 ret = i915_gem_evict_something(dev, size, alignment,
2761 ret = i915_gem_object_get_pages_gtt(obj, gfpmask);
2763 drm_mm_put_block(obj->gtt_space);
2764 obj->gtt_space = NULL;
2766 if (ret == -ENOMEM) {
2767 /* first try to reclaim some memory by clearing the GTT */
2768 ret = i915_gem_evict_everything(dev, false);
2770 /* now try to shrink everyone else */
2785 ret = i915_gem_gtt_bind_object(obj);
2787 i915_gem_object_put_pages_gtt(obj);
2788 drm_mm_put_block(obj->gtt_space);
2789 obj->gtt_space = NULL;
2791 if (i915_gem_evict_everything(dev, false))
2797 list_add_tail(&obj->gtt_list, &dev_priv->mm.gtt_list);
2798 list_add_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
2800 /* Assert that the object is not currently in any GPU domain. As it
2801 * wasn't in the GTT, there shouldn't be any way it could have been in
2804 BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
2805 BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
2807 obj->gtt_offset = obj->gtt_space->start;
2810 obj->gtt_space->size == fence_size &&
2811 (obj->gtt_space->start & (fence_alignment - 1)) == 0;
2814 obj->gtt_offset + obj->base.size <= dev_priv->mm.gtt_mappable_end;
2816 obj->map_and_fenceable = mappable && fenceable;
2818 trace_i915_gem_object_bind(obj, map_and_fenceable);
2823 i915_gem_clflush_object(struct drm_i915_gem_object *obj)
2825 /* If we don't have a page list set up, then we're not pinned
2826 * to GPU, and we can ignore the cache flush because it'll happen
2827 * again at bind time.
2829 if (obj->pages == NULL)
2832 /* If the GPU is snooping the contents of the CPU cache,
2833 * we do not need to manually clear the CPU cache lines. However,
2834 * the caches are only snooped when the render cache is
2835 * flushed/invalidated. As we always have to emit invalidations
2836 * and flushes when moving into and out of the RENDER domain, correct
2837 * snooping behaviour occurs naturally as the result of our domain
2840 if (obj->cache_level != I915_CACHE_NONE)
2843 trace_i915_gem_object_clflush(obj);
2845 drm_clflush_pages(obj->pages, obj->base.size / PAGE_SIZE);
2848 /** Flushes any GPU write domain for the object if it's dirty. */
2850 i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object *obj)
2852 if ((obj->base.write_domain & I915_GEM_GPU_DOMAINS) == 0)
2855 /* Queue the GPU write cache flushing we need. */
2856 return i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain);
2859 /** Flushes the GTT write domain for the object if it's dirty. */
2861 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
2863 uint32_t old_write_domain;
2865 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
2868 /* No actual flushing is required for the GTT write domain. Writes
2869 * to it immediately go to main memory as far as we know, so there's
2870 * no chipset flush. It also doesn't land in render cache.
2872 * However, we do have to enforce the order so that all writes through
2873 * the GTT land before any writes to the device, such as updates to
2878 old_write_domain = obj->base.write_domain;
2879 obj->base.write_domain = 0;
2881 trace_i915_gem_object_change_domain(obj,
2882 obj->base.read_domains,
2886 /** Flushes the CPU write domain for the object if it's dirty. */
2888 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj)
2890 uint32_t old_write_domain;
2892 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
2895 i915_gem_clflush_object(obj);
2896 intel_gtt_chipset_flush();
2897 old_write_domain = obj->base.write_domain;
2898 obj->base.write_domain = 0;
2900 trace_i915_gem_object_change_domain(obj,
2901 obj->base.read_domains,
2906 * Moves a single object to the GTT read, and possibly write domain.
2908 * This function returns when the move is complete, including waiting on
2912 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
2914 uint32_t old_write_domain, old_read_domains;
2917 /* Not valid to be called on unbound objects. */
2918 if (obj->gtt_space == NULL)
2921 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
2924 ret = i915_gem_object_flush_gpu_write_domain(obj);
2928 if (obj->pending_gpu_write || write) {
2929 ret = i915_gem_object_wait_rendering(obj);
2934 i915_gem_object_flush_cpu_write_domain(obj);
2936 old_write_domain = obj->base.write_domain;
2937 old_read_domains = obj->base.read_domains;
2939 /* It should now be out of any other write domains, and we can update
2940 * the domain values for our changes.
2942 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2943 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
2945 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
2946 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
2950 trace_i915_gem_object_change_domain(obj,
2957 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
2958 enum i915_cache_level cache_level)
2962 if (obj->cache_level == cache_level)
2965 if (obj->pin_count) {
2966 DRM_DEBUG("can not change the cache level of pinned objects\n");
2970 if (obj->gtt_space) {
2971 ret = i915_gem_object_finish_gpu(obj);
2975 i915_gem_object_finish_gtt(obj);
2977 /* Before SandyBridge, you could not use tiling or fence
2978 * registers with snooped memory, so relinquish any fences
2979 * currently pointing to our region in the aperture.
2981 if (INTEL_INFO(obj->base.dev)->gen < 6) {
2982 ret = i915_gem_object_put_fence(obj);
2987 i915_gem_gtt_rebind_object(obj, cache_level);
2990 if (cache_level == I915_CACHE_NONE) {
2991 u32 old_read_domains, old_write_domain;
2993 /* If we're coming from LLC cached, then we haven't
2994 * actually been tracking whether the data is in the
2995 * CPU cache or not, since we only allow one bit set
2996 * in obj->write_domain and have been skipping the clflushes.
2997 * Just set it to the CPU cache for now.
2999 WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU);
3000 WARN_ON(obj->base.read_domains & ~I915_GEM_DOMAIN_CPU);
3002 old_read_domains = obj->base.read_domains;
3003 old_write_domain = obj->base.write_domain;
3005 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3006 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3008 trace_i915_gem_object_change_domain(obj,
3013 obj->cache_level = cache_level;
3018 * Prepare buffer for display plane (scanout, cursors, etc).
3019 * Can be called from an uninterruptible phase (modesetting) and allows
3020 * any flushes to be pipelined (for pageflips).
3022 * For the display plane, we want to be in the GTT but out of any write
3023 * domains. So in many ways this looks like set_to_gtt_domain() apart from the
3024 * ability to pipeline the waits, pinning and any additional subtleties
3025 * that may differentiate the display plane from ordinary buffers.
3028 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
3030 struct intel_ring_buffer *pipelined)
3032 u32 old_read_domains, old_write_domain;
3035 ret = i915_gem_object_flush_gpu_write_domain(obj);
3039 if (pipelined != obj->ring) {
3040 ret = i915_gem_object_wait_rendering(obj);
3041 if (ret == -ERESTARTSYS)
3045 /* The display engine is not coherent with the LLC cache on gen6. As
3046 * a result, we make sure that the pinning that is about to occur is
3047 * done with uncached PTEs. This is lowest common denominator for all
3050 * However for gen6+, we could do better by using the GFDT bit instead
3051 * of uncaching, which would allow us to flush all the LLC-cached data
3052 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3054 ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE);
3058 /* As the user may map the buffer once pinned in the display plane
3059 * (e.g. libkms for the bootup splash), we have to ensure that we
3060 * always use map_and_fenceable for all scanout buffers.
3062 ret = i915_gem_object_pin(obj, alignment, true);
3066 i915_gem_object_flush_cpu_write_domain(obj);
3068 old_write_domain = obj->base.write_domain;
3069 old_read_domains = obj->base.read_domains;
3071 /* It should now be out of any other write domains, and we can update
3072 * the domain values for our changes.
3074 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3075 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3077 trace_i915_gem_object_change_domain(obj,
3085 i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
3089 if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
3092 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
3093 ret = i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain);
3098 /* Ensure that we invalidate the GPU's caches and TLBs. */
3099 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
3101 return i915_gem_object_wait_rendering(obj);
3105 * Moves a single object to the CPU read, and possibly write domain.
3107 * This function returns when the move is complete, including waiting on
3111 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
3113 uint32_t old_write_domain, old_read_domains;
3116 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
3119 ret = i915_gem_object_flush_gpu_write_domain(obj);
3123 ret = i915_gem_object_wait_rendering(obj);
3127 i915_gem_object_flush_gtt_write_domain(obj);
3129 /* If we have a partially-valid cache of the object in the CPU,
3130 * finish invalidating it and free the per-page flags.
3132 i915_gem_object_set_to_full_cpu_read_domain(obj);
3134 old_write_domain = obj->base.write_domain;
3135 old_read_domains = obj->base.read_domains;
3137 /* Flush the CPU cache if it's still invalid. */
3138 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
3139 i915_gem_clflush_object(obj);
3141 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3144 /* It should now be out of any other write domains, and we can update
3145 * the domain values for our changes.
3147 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3149 /* If we're writing through the CPU, then the GPU read domains will
3150 * need to be invalidated at next use.
3153 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3154 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3157 trace_i915_gem_object_change_domain(obj,
3165 * Moves the object from a partially CPU read to a full one.
3167 * Note that this only resolves i915_gem_object_set_cpu_read_domain_range(),
3168 * and doesn't handle transitioning from !(read_domains & I915_GEM_DOMAIN_CPU).
3171 i915_gem_object_set_to_full_cpu_read_domain(struct drm_i915_gem_object *obj)
3173 if (!obj->page_cpu_valid)
3176 /* If we're partially in the CPU read domain, finish moving it in.
3178 if (obj->base.read_domains & I915_GEM_DOMAIN_CPU) {
3181 for (i = 0; i <= (obj->base.size - 1) / PAGE_SIZE; i++) {
3182 if (obj->page_cpu_valid[i])
3184 drm_clflush_pages(obj->pages + i, 1);
3188 /* Free the page_cpu_valid mappings which are now stale, whether
3189 * or not we've got I915_GEM_DOMAIN_CPU.
3191 kfree(obj->page_cpu_valid);
3192 obj->page_cpu_valid = NULL;
3196 * Set the CPU read domain on a range of the object.
3198 * The object ends up with I915_GEM_DOMAIN_CPU in its read flags although it's
3199 * not entirely valid. The page_cpu_valid member of the object flags which
3200 * pages have been flushed, and will be respected by
3201 * i915_gem_object_set_to_cpu_domain() if it's called on to get a valid mapping
3202 * of the whole object.
3204 * This function returns when the move is complete, including waiting on
3208 i915_gem_object_set_cpu_read_domain_range(struct drm_i915_gem_object *obj,
3209 uint64_t offset, uint64_t size)
3211 uint32_t old_read_domains;
3214 if (offset == 0 && size == obj->base.size)
3215 return i915_gem_object_set_to_cpu_domain(obj, 0);
3217 ret = i915_gem_object_flush_gpu_write_domain(obj);
3221 ret = i915_gem_object_wait_rendering(obj);
3225 i915_gem_object_flush_gtt_write_domain(obj);
3227 /* If we're already fully in the CPU read domain, we're done. */
3228 if (obj->page_cpu_valid == NULL &&
3229 (obj->base.read_domains & I915_GEM_DOMAIN_CPU) != 0)
3232 /* Otherwise, create/clear the per-page CPU read domain flag if we're
3233 * newly adding I915_GEM_DOMAIN_CPU
3235 if (obj->page_cpu_valid == NULL) {
3236 obj->page_cpu_valid = kzalloc(obj->base.size / PAGE_SIZE,
3238 if (obj->page_cpu_valid == NULL)
3240 } else if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0)
3241 memset(obj->page_cpu_valid, 0, obj->base.size / PAGE_SIZE);
3243 /* Flush the cache on any pages that are still invalid from the CPU's
3246 for (i = offset / PAGE_SIZE; i <= (offset + size - 1) / PAGE_SIZE;
3248 if (obj->page_cpu_valid[i])
3251 drm_clflush_pages(obj->pages + i, 1);
3253 obj->page_cpu_valid[i] = 1;
3256 /* It should now be out of any other write domains, and we can update
3257 * the domain values for our changes.
3259 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3261 old_read_domains = obj->base.read_domains;
3262 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3264 trace_i915_gem_object_change_domain(obj,
3266 obj->base.write_domain);
3271 /* Throttle our rendering by waiting until the ring has completed our requests
3272 * emitted over 20 msec ago.
3274 * Note that if we were to use the current jiffies each time around the loop,
3275 * we wouldn't escape the function with any frames outstanding if the time to
3276 * render a frame was over 20ms.
3278 * This should get us reasonable parallelism between CPU and GPU but also
3279 * relatively low latency when blocking on a particular request to finish.
3282 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
3284 struct drm_i915_private *dev_priv = dev->dev_private;
3285 struct drm_i915_file_private *file_priv = file->driver_priv;
3286 unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3287 struct drm_i915_gem_request *request;
3288 struct intel_ring_buffer *ring = NULL;
3292 if (atomic_read(&dev_priv->mm.wedged))
3295 spin_lock(&file_priv->mm.lock);
3296 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
3297 if (time_after_eq(request->emitted_jiffies, recent_enough))
3300 ring = request->ring;
3301 seqno = request->seqno;
3303 spin_unlock(&file_priv->mm.lock);
3309 if (!i915_seqno_passed(ring->get_seqno(ring), seqno)) {
3310 /* And wait for the seqno passing without holding any locks and
3311 * causing extra latency for others. This is safe as the irq
3312 * generation is designed to be run atomically and so is
3315 if (ring->irq_get(ring)) {
3316 ret = wait_event_interruptible(ring->irq_queue,
3317 i915_seqno_passed(ring->get_seqno(ring), seqno)
3318 || atomic_read(&dev_priv->mm.wedged));
3319 ring->irq_put(ring);
3321 if (ret == 0 && atomic_read(&dev_priv->mm.wedged))
3323 } else if (wait_for_atomic(i915_seqno_passed(ring->get_seqno(ring),
3325 atomic_read(&dev_priv->mm.wedged), 3000)) {
3331 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
3337 i915_gem_object_pin(struct drm_i915_gem_object *obj,
3339 bool map_and_fenceable)
3341 struct drm_device *dev = obj->base.dev;
3342 struct drm_i915_private *dev_priv = dev->dev_private;
3345 BUG_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT);
3346 WARN_ON(i915_verify_lists(dev));
3348 if (obj->gtt_space != NULL) {
3349 if ((alignment && obj->gtt_offset & (alignment - 1)) ||
3350 (map_and_fenceable && !obj->map_and_fenceable)) {
3351 WARN(obj->pin_count,
3352 "bo is already pinned with incorrect alignment:"
3353 " offset=%x, req.alignment=%x, req.map_and_fenceable=%d,"
3354 " obj->map_and_fenceable=%d\n",
3355 obj->gtt_offset, alignment,
3357 obj->map_and_fenceable);
3358 ret = i915_gem_object_unbind(obj);
3364 if (obj->gtt_space == NULL) {
3365 ret = i915_gem_object_bind_to_gtt(obj, alignment,
3371 if (obj->pin_count++ == 0) {
3373 list_move_tail(&obj->mm_list,
3374 &dev_priv->mm.pinned_list);
3376 obj->pin_mappable |= map_and_fenceable;
3378 WARN_ON(i915_verify_lists(dev));
3383 i915_gem_object_unpin(struct drm_i915_gem_object *obj)
3385 struct drm_device *dev = obj->base.dev;
3386 drm_i915_private_t *dev_priv = dev->dev_private;
3388 WARN_ON(i915_verify_lists(dev));
3389 BUG_ON(obj->pin_count == 0);
3390 BUG_ON(obj->gtt_space == NULL);
3392 if (--obj->pin_count == 0) {
3394 list_move_tail(&obj->mm_list,
3395 &dev_priv->mm.inactive_list);
3396 obj->pin_mappable = false;
3398 WARN_ON(i915_verify_lists(dev));
3402 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
3403 struct drm_file *file)
3405 struct drm_i915_gem_pin *args = data;
3406 struct drm_i915_gem_object *obj;
3409 ret = i915_mutex_lock_interruptible(dev);
3413 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3414 if (&obj->base == NULL) {
3419 if (obj->madv != I915_MADV_WILLNEED) {
3420 DRM_ERROR("Attempting to pin a purgeable buffer\n");
3425 if (obj->pin_filp != NULL && obj->pin_filp != file) {
3426 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
3432 obj->user_pin_count++;
3433 obj->pin_filp = file;
3434 if (obj->user_pin_count == 1) {
3435 ret = i915_gem_object_pin(obj, args->alignment, true);
3440 /* XXX - flush the CPU caches for pinned objects
3441 * as the X server doesn't manage domains yet
3443 i915_gem_object_flush_cpu_write_domain(obj);
3444 args->offset = obj->gtt_offset;
3446 drm_gem_object_unreference(&obj->base);
3448 mutex_unlock(&dev->struct_mutex);
3453 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
3454 struct drm_file *file)
3456 struct drm_i915_gem_pin *args = data;
3457 struct drm_i915_gem_object *obj;
3460 ret = i915_mutex_lock_interruptible(dev);
3464 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3465 if (&obj->base == NULL) {
3470 if (obj->pin_filp != file) {
3471 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
3476 obj->user_pin_count--;
3477 if (obj->user_pin_count == 0) {
3478 obj->pin_filp = NULL;
3479 i915_gem_object_unpin(obj);
3483 drm_gem_object_unreference(&obj->base);
3485 mutex_unlock(&dev->struct_mutex);
3490 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
3491 struct drm_file *file)
3493 struct drm_i915_gem_busy *args = data;
3494 struct drm_i915_gem_object *obj;
3497 ret = i915_mutex_lock_interruptible(dev);
3501 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3502 if (&obj->base == NULL) {
3507 /* Count all active objects as busy, even if they are currently not used
3508 * by the gpu. Users of this interface expect objects to eventually
3509 * become non-busy without any further actions, therefore emit any
3510 * necessary flushes here.
3512 args->busy = obj->active;
3514 /* Unconditionally flush objects, even when the gpu still uses this
3515 * object. Userspace calling this function indicates that it wants to
3516 * use this buffer rather sooner than later, so issuing the required
3517 * flush earlier is beneficial.
3519 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
3520 ret = i915_gem_flush_ring(obj->ring,
3521 0, obj->base.write_domain);
3522 } else if (obj->ring->outstanding_lazy_request ==
3523 obj->last_rendering_seqno) {
3524 struct drm_i915_gem_request *request;
3526 /* This ring is not being cleared by active usage,
3527 * so emit a request to do so.
3529 request = kzalloc(sizeof(*request), GFP_KERNEL);
3531 ret = i915_add_request(obj->ring, NULL, request);
3538 /* Update the active list for the hardware's current position.
3539 * Otherwise this only updates on a delayed timer or when irqs
3540 * are actually unmasked, and our working set ends up being
3541 * larger than required.
3543 i915_gem_retire_requests_ring(obj->ring);
3545 args->busy = obj->active;
3548 drm_gem_object_unreference(&obj->base);
3550 mutex_unlock(&dev->struct_mutex);
3555 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
3556 struct drm_file *file_priv)
3558 return i915_gem_ring_throttle(dev, file_priv);
3562 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
3563 struct drm_file *file_priv)
3565 struct drm_i915_gem_madvise *args = data;
3566 struct drm_i915_gem_object *obj;
3569 switch (args->madv) {
3570 case I915_MADV_DONTNEED:
3571 case I915_MADV_WILLNEED:
3577 ret = i915_mutex_lock_interruptible(dev);
3581 obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
3582 if (&obj->base == NULL) {
3587 if (obj->pin_count) {
3592 if (obj->madv != __I915_MADV_PURGED)
3593 obj->madv = args->madv;
3595 /* if the object is no longer bound, discard its backing storage */
3596 if (i915_gem_object_is_purgeable(obj) &&
3597 obj->gtt_space == NULL)
3598 i915_gem_object_truncate(obj);
3600 args->retained = obj->madv != __I915_MADV_PURGED;
3603 drm_gem_object_unreference(&obj->base);
3605 mutex_unlock(&dev->struct_mutex);
3609 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
3612 struct drm_i915_private *dev_priv = dev->dev_private;
3613 struct drm_i915_gem_object *obj;
3614 struct address_space *mapping;
3616 obj = kzalloc(sizeof(*obj), GFP_KERNEL);
3620 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
3625 mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
3626 mapping_set_gfp_mask(mapping, GFP_HIGHUSER | __GFP_RECLAIMABLE);
3628 i915_gem_info_add_obj(dev_priv, size);
3630 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3631 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3633 if (IS_GEN6(dev) || IS_GEN7(dev)) {
3634 /* On Gen6, we can have the GPU use the LLC (the CPU
3635 * cache) for about a 10% performance improvement
3636 * compared to uncached. Graphics requests other than
3637 * display scanout are coherent with the CPU in
3638 * accessing this cache. This means in this mode we
3639 * don't need to clflush on the CPU side, and on the
3640 * GPU side we only need to flush internal caches to
3641 * get data visible to the CPU.
3643 * However, we maintain the display planes as UC, and so
3644 * need to rebind when first used as such.
3646 obj->cache_level = I915_CACHE_LLC;
3648 obj->cache_level = I915_CACHE_NONE;
3650 obj->base.driver_private = NULL;
3651 obj->fence_reg = I915_FENCE_REG_NONE;
3652 INIT_LIST_HEAD(&obj->mm_list);
3653 INIT_LIST_HEAD(&obj->gtt_list);
3654 INIT_LIST_HEAD(&obj->ring_list);
3655 INIT_LIST_HEAD(&obj->exec_list);
3656 INIT_LIST_HEAD(&obj->gpu_write_list);
3657 obj->madv = I915_MADV_WILLNEED;
3658 /* Avoid an unnecessary call to unbind on the first bind. */
3659 obj->map_and_fenceable = true;
3664 int i915_gem_init_object(struct drm_gem_object *obj)
3671 static void i915_gem_free_object_tail(struct drm_i915_gem_object *obj)
3673 struct drm_device *dev = obj->base.dev;
3674 drm_i915_private_t *dev_priv = dev->dev_private;
3677 ret = i915_gem_object_unbind(obj);
3678 if (ret == -ERESTARTSYS) {
3679 list_move(&obj->mm_list,
3680 &dev_priv->mm.deferred_free_list);
3684 trace_i915_gem_object_destroy(obj);
3686 if (obj->base.map_list.map)
3687 drm_gem_free_mmap_offset(&obj->base);
3689 drm_gem_object_release(&obj->base);
3690 i915_gem_info_remove_obj(dev_priv, obj->base.size);
3692 kfree(obj->page_cpu_valid);
3697 void i915_gem_free_object(struct drm_gem_object *gem_obj)
3699 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
3700 struct drm_device *dev = obj->base.dev;
3702 while (obj->pin_count > 0)
3703 i915_gem_object_unpin(obj);
3706 i915_gem_detach_phys_object(dev, obj);
3708 i915_gem_free_object_tail(obj);
3712 i915_gem_idle(struct drm_device *dev)
3714 drm_i915_private_t *dev_priv = dev->dev_private;
3717 mutex_lock(&dev->struct_mutex);
3719 if (dev_priv->mm.suspended) {
3720 mutex_unlock(&dev->struct_mutex);
3724 ret = i915_gpu_idle(dev);
3726 mutex_unlock(&dev->struct_mutex);
3730 /* Under UMS, be paranoid and evict. */
3731 if (!drm_core_check_feature(dev, DRIVER_MODESET)) {
3732 ret = i915_gem_evict_inactive(dev, false);
3734 mutex_unlock(&dev->struct_mutex);
3739 i915_gem_reset_fences(dev);
3741 /* Hack! Don't let anybody do execbuf while we don't control the chip.
3742 * We need to replace this with a semaphore, or something.
3743 * And not confound mm.suspended!
3745 dev_priv->mm.suspended = 1;
3746 del_timer_sync(&dev_priv->hangcheck_timer);
3748 i915_kernel_lost_context(dev);
3749 i915_gem_cleanup_ringbuffer(dev);
3751 mutex_unlock(&dev->struct_mutex);
3753 /* Cancel the retire work handler, which should be idle now. */
3754 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
3760 i915_gem_init_ringbuffer(struct drm_device *dev)
3762 drm_i915_private_t *dev_priv = dev->dev_private;
3765 ret = intel_init_render_ring_buffer(dev);
3770 ret = intel_init_bsd_ring_buffer(dev);
3772 goto cleanup_render_ring;
3776 ret = intel_init_blt_ring_buffer(dev);
3778 goto cleanup_bsd_ring;
3781 dev_priv->next_seqno = 1;
3786 intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
3787 cleanup_render_ring:
3788 intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
3793 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
3795 drm_i915_private_t *dev_priv = dev->dev_private;
3798 for (i = 0; i < I915_NUM_RINGS; i++)
3799 intel_cleanup_ring_buffer(&dev_priv->ring[i]);
3803 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
3804 struct drm_file *file_priv)
3806 drm_i915_private_t *dev_priv = dev->dev_private;
3809 if (drm_core_check_feature(dev, DRIVER_MODESET))
3812 if (atomic_read(&dev_priv->mm.wedged)) {
3813 DRM_ERROR("Reenabling wedged hardware, good luck\n");
3814 atomic_set(&dev_priv->mm.wedged, 0);
3817 mutex_lock(&dev->struct_mutex);
3818 dev_priv->mm.suspended = 0;
3820 ret = i915_gem_init_ringbuffer(dev);
3822 mutex_unlock(&dev->struct_mutex);
3826 BUG_ON(!list_empty(&dev_priv->mm.active_list));
3827 BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
3828 BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
3829 for (i = 0; i < I915_NUM_RINGS; i++) {
3830 BUG_ON(!list_empty(&dev_priv->ring[i].active_list));
3831 BUG_ON(!list_empty(&dev_priv->ring[i].request_list));
3833 mutex_unlock(&dev->struct_mutex);
3835 ret = drm_irq_install(dev);
3837 goto cleanup_ringbuffer;
3842 mutex_lock(&dev->struct_mutex);
3843 i915_gem_cleanup_ringbuffer(dev);
3844 dev_priv->mm.suspended = 1;
3845 mutex_unlock(&dev->struct_mutex);
3851 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
3852 struct drm_file *file_priv)
3854 if (drm_core_check_feature(dev, DRIVER_MODESET))
3857 drm_irq_uninstall(dev);
3858 return i915_gem_idle(dev);
3862 i915_gem_lastclose(struct drm_device *dev)
3866 if (drm_core_check_feature(dev, DRIVER_MODESET))
3869 ret = i915_gem_idle(dev);
3871 DRM_ERROR("failed to idle hardware: %d\n", ret);
3875 init_ring_lists(struct intel_ring_buffer *ring)
3877 INIT_LIST_HEAD(&ring->active_list);
3878 INIT_LIST_HEAD(&ring->request_list);
3879 INIT_LIST_HEAD(&ring->gpu_write_list);
3883 i915_gem_load(struct drm_device *dev)
3886 drm_i915_private_t *dev_priv = dev->dev_private;
3888 INIT_LIST_HEAD(&dev_priv->mm.active_list);
3889 INIT_LIST_HEAD(&dev_priv->mm.flushing_list);
3890 INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
3891 INIT_LIST_HEAD(&dev_priv->mm.pinned_list);
3892 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
3893 INIT_LIST_HEAD(&dev_priv->mm.deferred_free_list);
3894 INIT_LIST_HEAD(&dev_priv->mm.gtt_list);
3895 for (i = 0; i < I915_NUM_RINGS; i++)
3896 init_ring_lists(&dev_priv->ring[i]);
3897 for (i = 0; i < I915_MAX_NUM_FENCES; i++)
3898 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
3899 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
3900 i915_gem_retire_work_handler);
3901 init_completion(&dev_priv->error_completion);
3903 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
3905 u32 tmp = I915_READ(MI_ARB_STATE);
3906 if (!(tmp & MI_ARB_C3_LP_WRITE_ENABLE)) {
3907 /* arb state is a masked write, so set bit + bit in mask */
3908 tmp = MI_ARB_C3_LP_WRITE_ENABLE | (MI_ARB_C3_LP_WRITE_ENABLE << MI_ARB_MASK_SHIFT);
3909 I915_WRITE(MI_ARB_STATE, tmp);
3913 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
3915 /* Old X drivers will take 0-2 for front, back, depth buffers */
3916 if (!drm_core_check_feature(dev, DRIVER_MODESET))
3917 dev_priv->fence_reg_start = 3;
3919 if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
3920 dev_priv->num_fence_regs = 16;
3922 dev_priv->num_fence_regs = 8;
3924 /* Initialize fence registers to zero */
3925 for (i = 0; i < dev_priv->num_fence_regs; i++) {
3926 i915_gem_clear_fence_reg(dev, &dev_priv->fence_regs[i]);
3929 i915_gem_detect_bit_6_swizzle(dev);
3930 init_waitqueue_head(&dev_priv->pending_flip_queue);
3932 dev_priv->mm.interruptible = true;
3934 dev_priv->mm.inactive_shrinker.shrink = i915_gem_inactive_shrink;
3935 dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS;
3936 register_shrinker(&dev_priv->mm.inactive_shrinker);
3940 * Create a physically contiguous memory object for this object
3941 * e.g. for cursor + overlay regs
3943 static int i915_gem_init_phys_object(struct drm_device *dev,
3944 int id, int size, int align)
3946 drm_i915_private_t *dev_priv = dev->dev_private;
3947 struct drm_i915_gem_phys_object *phys_obj;
3950 if (dev_priv->mm.phys_objs[id - 1] || !size)
3953 phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL);
3959 phys_obj->handle = drm_pci_alloc(dev, size, align);
3960 if (!phys_obj->handle) {
3965 set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
3968 dev_priv->mm.phys_objs[id - 1] = phys_obj;
3976 static void i915_gem_free_phys_object(struct drm_device *dev, int id)
3978 drm_i915_private_t *dev_priv = dev->dev_private;
3979 struct drm_i915_gem_phys_object *phys_obj;
3981 if (!dev_priv->mm.phys_objs[id - 1])
3984 phys_obj = dev_priv->mm.phys_objs[id - 1];
3985 if (phys_obj->cur_obj) {
3986 i915_gem_detach_phys_object(dev, phys_obj->cur_obj);
3990 set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
3992 drm_pci_free(dev, phys_obj->handle);
3994 dev_priv->mm.phys_objs[id - 1] = NULL;
3997 void i915_gem_free_all_phys_object(struct drm_device *dev)
4001 for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++)
4002 i915_gem_free_phys_object(dev, i);
4005 void i915_gem_detach_phys_object(struct drm_device *dev,
4006 struct drm_i915_gem_object *obj)
4008 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
4015 vaddr = obj->phys_obj->handle->vaddr;
4017 page_count = obj->base.size / PAGE_SIZE;
4018 for (i = 0; i < page_count; i++) {
4019 struct page *page = shmem_read_mapping_page(mapping, i);
4020 if (!IS_ERR(page)) {
4021 char *dst = kmap_atomic(page);
4022 memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE);
4025 drm_clflush_pages(&page, 1);
4027 set_page_dirty(page);
4028 mark_page_accessed(page);
4029 page_cache_release(page);
4032 intel_gtt_chipset_flush();
4034 obj->phys_obj->cur_obj = NULL;
4035 obj->phys_obj = NULL;
4039 i915_gem_attach_phys_object(struct drm_device *dev,
4040 struct drm_i915_gem_object *obj,
4044 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
4045 drm_i915_private_t *dev_priv = dev->dev_private;
4050 if (id > I915_MAX_PHYS_OBJECT)
4053 if (obj->phys_obj) {
4054 if (obj->phys_obj->id == id)
4056 i915_gem_detach_phys_object(dev, obj);
4059 /* create a new object */
4060 if (!dev_priv->mm.phys_objs[id - 1]) {
4061 ret = i915_gem_init_phys_object(dev, id,
4062 obj->base.size, align);
4064 DRM_ERROR("failed to init phys object %d size: %zu\n",
4065 id, obj->base.size);
4070 /* bind to the object */
4071 obj->phys_obj = dev_priv->mm.phys_objs[id - 1];
4072 obj->phys_obj->cur_obj = obj;
4074 page_count = obj->base.size / PAGE_SIZE;
4076 for (i = 0; i < page_count; i++) {
4080 page = shmem_read_mapping_page(mapping, i);
4082 return PTR_ERR(page);
4084 src = kmap_atomic(page);
4085 dst = obj->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4086 memcpy(dst, src, PAGE_SIZE);
4089 mark_page_accessed(page);
4090 page_cache_release(page);
4097 i915_gem_phys_pwrite(struct drm_device *dev,
4098 struct drm_i915_gem_object *obj,
4099 struct drm_i915_gem_pwrite *args,
4100 struct drm_file *file_priv)
4102 void *vaddr = obj->phys_obj->handle->vaddr + args->offset;
4103 char __user *user_data = (char __user *) (uintptr_t) args->data_ptr;
4105 if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
4106 unsigned long unwritten;
4108 /* The physical object once assigned is fixed for the lifetime
4109 * of the obj, so we can safely drop the lock and continue
4112 mutex_unlock(&dev->struct_mutex);
4113 unwritten = copy_from_user(vaddr, user_data, args->size);
4114 mutex_lock(&dev->struct_mutex);
4119 intel_gtt_chipset_flush();
4123 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
4125 struct drm_i915_file_private *file_priv = file->driver_priv;
4127 /* Clean up our request list when the client is going away, so that
4128 * later retire_requests won't dereference our soon-to-be-gone
4131 spin_lock(&file_priv->mm.lock);
4132 while (!list_empty(&file_priv->mm.request_list)) {
4133 struct drm_i915_gem_request *request;
4135 request = list_first_entry(&file_priv->mm.request_list,
4136 struct drm_i915_gem_request,
4138 list_del(&request->client_list);
4139 request->file_priv = NULL;
4141 spin_unlock(&file_priv->mm.lock);
4145 i915_gpu_is_active(struct drm_device *dev)
4147 drm_i915_private_t *dev_priv = dev->dev_private;
4150 lists_empty = list_empty(&dev_priv->mm.flushing_list) &&
4151 list_empty(&dev_priv->mm.active_list);
4153 return !lists_empty;
4157 i915_gem_inactive_shrink(struct shrinker *shrinker, struct shrink_control *sc)
4159 struct drm_i915_private *dev_priv =
4160 container_of(shrinker,
4161 struct drm_i915_private,
4162 mm.inactive_shrinker);
4163 struct drm_device *dev = dev_priv->dev;
4164 struct drm_i915_gem_object *obj, *next;
4165 int nr_to_scan = sc->nr_to_scan;
4168 if (!mutex_trylock(&dev->struct_mutex))
4171 /* "fast-path" to count number of available objects */
4172 if (nr_to_scan == 0) {
4174 list_for_each_entry(obj,
4175 &dev_priv->mm.inactive_list,
4178 mutex_unlock(&dev->struct_mutex);
4179 return cnt / 100 * sysctl_vfs_cache_pressure;
4183 /* first scan for clean buffers */
4184 i915_gem_retire_requests(dev);
4186 list_for_each_entry_safe(obj, next,
4187 &dev_priv->mm.inactive_list,
4189 if (i915_gem_object_is_purgeable(obj)) {
4190 if (i915_gem_object_unbind(obj) == 0 &&
4196 /* second pass, evict/count anything still on the inactive list */
4198 list_for_each_entry_safe(obj, next,
4199 &dev_priv->mm.inactive_list,
4202 i915_gem_object_unbind(obj) == 0)
4208 if (nr_to_scan && i915_gpu_is_active(dev)) {
4210 * We are desperate for pages, so as a last resort, wait
4211 * for the GPU to finish and discard whatever we can.
4212 * This has a dramatic impact to reduce the number of
4213 * OOM-killer events whilst running the GPU aggressively.
4215 if (i915_gpu_idle(dev) == 0)
4218 mutex_unlock(&dev->struct_mutex);
4219 return cnt / 100 * sysctl_vfs_cache_pressure;