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/slab.h>
35 #include <linux/swap.h>
36 #include <linux/pci.h>
38 static int i915_gem_object_flush_gpu_write_domain(struct drm_gem_object *obj);
39 static void i915_gem_object_flush_gtt_write_domain(struct drm_gem_object *obj);
40 static void i915_gem_object_flush_cpu_write_domain(struct drm_gem_object *obj);
41 static int i915_gem_object_set_to_cpu_domain(struct drm_gem_object *obj,
43 static int i915_gem_object_set_cpu_read_domain_range(struct drm_gem_object *obj,
46 static void i915_gem_object_set_to_full_cpu_read_domain(struct drm_gem_object *obj);
47 static int i915_gem_object_wait_rendering(struct drm_gem_object *obj);
48 static int i915_gem_object_bind_to_gtt(struct drm_gem_object *obj,
50 static void i915_gem_clear_fence_reg(struct drm_gem_object *obj);
51 static int i915_gem_evict_something(struct drm_device *dev, int min_size);
52 static int i915_gem_evict_from_inactive_list(struct drm_device *dev);
53 static int i915_gem_phys_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
54 struct drm_i915_gem_pwrite *args,
55 struct drm_file *file_priv);
56 static void i915_gem_free_object_tail(struct drm_gem_object *obj);
58 static LIST_HEAD(shrink_list);
59 static DEFINE_SPINLOCK(shrink_list_lock);
61 int i915_gem_do_init(struct drm_device *dev, unsigned long start,
64 drm_i915_private_t *dev_priv = dev->dev_private;
67 (start & (PAGE_SIZE - 1)) != 0 ||
68 (end & (PAGE_SIZE - 1)) != 0) {
72 drm_mm_init(&dev_priv->mm.gtt_space, start,
75 dev->gtt_total = (uint32_t) (end - start);
81 i915_gem_init_ioctl(struct drm_device *dev, void *data,
82 struct drm_file *file_priv)
84 struct drm_i915_gem_init *args = data;
87 mutex_lock(&dev->struct_mutex);
88 ret = i915_gem_do_init(dev, args->gtt_start, args->gtt_end);
89 mutex_unlock(&dev->struct_mutex);
95 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
96 struct drm_file *file_priv)
98 struct drm_i915_gem_get_aperture *args = data;
100 if (!(dev->driver->driver_features & DRIVER_GEM))
103 args->aper_size = dev->gtt_total;
104 args->aper_available_size = (args->aper_size -
105 atomic_read(&dev->pin_memory));
112 * Creates a new mm object and returns a handle to it.
115 i915_gem_create_ioctl(struct drm_device *dev, void *data,
116 struct drm_file *file_priv)
118 struct drm_i915_gem_create *args = data;
119 struct drm_gem_object *obj;
123 args->size = roundup(args->size, PAGE_SIZE);
125 /* Allocate the new object */
126 obj = i915_gem_alloc_object(dev, args->size);
130 ret = drm_gem_handle_create(file_priv, obj, &handle);
131 drm_gem_object_handle_unreference_unlocked(obj);
136 args->handle = handle;
142 fast_shmem_read(struct page **pages,
143 loff_t page_base, int page_offset,
150 vaddr = kmap_atomic(pages[page_base >> PAGE_SHIFT], KM_USER0);
153 unwritten = __copy_to_user_inatomic(data, vaddr + page_offset, length);
154 kunmap_atomic(vaddr, KM_USER0);
162 static int i915_gem_object_needs_bit17_swizzle(struct drm_gem_object *obj)
164 drm_i915_private_t *dev_priv = obj->dev->dev_private;
165 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
167 return dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_9_10_17 &&
168 obj_priv->tiling_mode != I915_TILING_NONE;
172 slow_shmem_copy(struct page *dst_page,
174 struct page *src_page,
178 char *dst_vaddr, *src_vaddr;
180 dst_vaddr = kmap(dst_page);
181 src_vaddr = kmap(src_page);
183 memcpy(dst_vaddr + dst_offset, src_vaddr + src_offset, length);
190 slow_shmem_bit17_copy(struct page *gpu_page,
192 struct page *cpu_page,
197 char *gpu_vaddr, *cpu_vaddr;
199 /* Use the unswizzled path if this page isn't affected. */
200 if ((page_to_phys(gpu_page) & (1 << 17)) == 0) {
202 return slow_shmem_copy(cpu_page, cpu_offset,
203 gpu_page, gpu_offset, length);
205 return slow_shmem_copy(gpu_page, gpu_offset,
206 cpu_page, cpu_offset, length);
209 gpu_vaddr = kmap(gpu_page);
210 cpu_vaddr = kmap(cpu_page);
212 /* Copy the data, XORing A6 with A17 (1). The user already knows he's
213 * XORing with the other bits (A9 for Y, A9 and A10 for X)
216 int cacheline_end = ALIGN(gpu_offset + 1, 64);
217 int this_length = min(cacheline_end - gpu_offset, length);
218 int swizzled_gpu_offset = gpu_offset ^ 64;
221 memcpy(cpu_vaddr + cpu_offset,
222 gpu_vaddr + swizzled_gpu_offset,
225 memcpy(gpu_vaddr + swizzled_gpu_offset,
226 cpu_vaddr + cpu_offset,
229 cpu_offset += this_length;
230 gpu_offset += this_length;
231 length -= this_length;
239 * This is the fast shmem pread path, which attempts to copy_from_user directly
240 * from the backing pages of the object to the user's address space. On a
241 * fault, it fails so we can fall back to i915_gem_shmem_pwrite_slow().
244 i915_gem_shmem_pread_fast(struct drm_device *dev, struct drm_gem_object *obj,
245 struct drm_i915_gem_pread *args,
246 struct drm_file *file_priv)
248 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
250 loff_t offset, page_base;
251 char __user *user_data;
252 int page_offset, page_length;
255 user_data = (char __user *) (uintptr_t) args->data_ptr;
258 mutex_lock(&dev->struct_mutex);
260 ret = i915_gem_object_get_pages(obj, 0);
264 ret = i915_gem_object_set_cpu_read_domain_range(obj, args->offset,
269 obj_priv = to_intel_bo(obj);
270 offset = args->offset;
273 /* Operation in this page
275 * page_base = page offset within aperture
276 * page_offset = offset within page
277 * page_length = bytes to copy for this page
279 page_base = (offset & ~(PAGE_SIZE-1));
280 page_offset = offset & (PAGE_SIZE-1);
281 page_length = remain;
282 if ((page_offset + remain) > PAGE_SIZE)
283 page_length = PAGE_SIZE - page_offset;
285 ret = fast_shmem_read(obj_priv->pages,
286 page_base, page_offset,
287 user_data, page_length);
291 remain -= page_length;
292 user_data += page_length;
293 offset += page_length;
297 i915_gem_object_put_pages(obj);
299 mutex_unlock(&dev->struct_mutex);
305 i915_gem_object_get_pages_or_evict(struct drm_gem_object *obj)
309 ret = i915_gem_object_get_pages(obj, __GFP_NORETRY | __GFP_NOWARN);
311 /* If we've insufficient memory to map in the pages, attempt
312 * to make some space by throwing out some old buffers.
314 if (ret == -ENOMEM) {
315 struct drm_device *dev = obj->dev;
317 ret = i915_gem_evict_something(dev, obj->size);
321 ret = i915_gem_object_get_pages(obj, 0);
328 * This is the fallback shmem pread path, which allocates temporary storage
329 * in kernel space to copy_to_user into outside of the struct_mutex, so we
330 * can copy out of the object's backing pages while holding the struct mutex
331 * and not take page faults.
334 i915_gem_shmem_pread_slow(struct drm_device *dev, struct drm_gem_object *obj,
335 struct drm_i915_gem_pread *args,
336 struct drm_file *file_priv)
338 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
339 struct mm_struct *mm = current->mm;
340 struct page **user_pages;
342 loff_t offset, pinned_pages, i;
343 loff_t first_data_page, last_data_page, num_pages;
344 int shmem_page_index, shmem_page_offset;
345 int data_page_index, data_page_offset;
348 uint64_t data_ptr = args->data_ptr;
349 int do_bit17_swizzling;
353 /* Pin the user pages containing the data. We can't fault while
354 * holding the struct mutex, yet we want to hold it while
355 * dereferencing the user data.
357 first_data_page = data_ptr / PAGE_SIZE;
358 last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
359 num_pages = last_data_page - first_data_page + 1;
361 user_pages = drm_calloc_large(num_pages, sizeof(struct page *));
362 if (user_pages == NULL)
365 down_read(&mm->mmap_sem);
366 pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
367 num_pages, 1, 0, user_pages, NULL);
368 up_read(&mm->mmap_sem);
369 if (pinned_pages < num_pages) {
371 goto fail_put_user_pages;
374 do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
376 mutex_lock(&dev->struct_mutex);
378 ret = i915_gem_object_get_pages_or_evict(obj);
382 ret = i915_gem_object_set_cpu_read_domain_range(obj, args->offset,
387 obj_priv = to_intel_bo(obj);
388 offset = args->offset;
391 /* Operation in this page
393 * shmem_page_index = page number within shmem file
394 * shmem_page_offset = offset within page in shmem file
395 * data_page_index = page number in get_user_pages return
396 * data_page_offset = offset with data_page_index page.
397 * page_length = bytes to copy for this page
399 shmem_page_index = offset / PAGE_SIZE;
400 shmem_page_offset = offset & ~PAGE_MASK;
401 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
402 data_page_offset = data_ptr & ~PAGE_MASK;
404 page_length = remain;
405 if ((shmem_page_offset + page_length) > PAGE_SIZE)
406 page_length = PAGE_SIZE - shmem_page_offset;
407 if ((data_page_offset + page_length) > PAGE_SIZE)
408 page_length = PAGE_SIZE - data_page_offset;
410 if (do_bit17_swizzling) {
411 slow_shmem_bit17_copy(obj_priv->pages[shmem_page_index],
413 user_pages[data_page_index],
418 slow_shmem_copy(user_pages[data_page_index],
420 obj_priv->pages[shmem_page_index],
425 remain -= page_length;
426 data_ptr += page_length;
427 offset += page_length;
431 i915_gem_object_put_pages(obj);
433 mutex_unlock(&dev->struct_mutex);
435 for (i = 0; i < pinned_pages; i++) {
436 SetPageDirty(user_pages[i]);
437 page_cache_release(user_pages[i]);
439 drm_free_large(user_pages);
445 * Reads data from the object referenced by handle.
447 * On error, the contents of *data are undefined.
450 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
451 struct drm_file *file_priv)
453 struct drm_i915_gem_pread *args = data;
454 struct drm_gem_object *obj;
455 struct drm_i915_gem_object *obj_priv;
458 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
461 obj_priv = to_intel_bo(obj);
463 /* Bounds check source.
465 * XXX: This could use review for overflow issues...
467 if (args->offset > obj->size || args->size > obj->size ||
468 args->offset + args->size > obj->size) {
469 drm_gem_object_unreference_unlocked(obj);
473 if (i915_gem_object_needs_bit17_swizzle(obj)) {
474 ret = i915_gem_shmem_pread_slow(dev, obj, args, file_priv);
476 ret = i915_gem_shmem_pread_fast(dev, obj, args, file_priv);
478 ret = i915_gem_shmem_pread_slow(dev, obj, args,
482 drm_gem_object_unreference_unlocked(obj);
487 /* This is the fast write path which cannot handle
488 * page faults in the source data
492 fast_user_write(struct io_mapping *mapping,
493 loff_t page_base, int page_offset,
494 char __user *user_data,
498 unsigned long unwritten;
500 vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
501 unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + page_offset,
503 io_mapping_unmap_atomic(vaddr_atomic);
509 /* Here's the write path which can sleep for
514 slow_kernel_write(struct io_mapping *mapping,
515 loff_t gtt_base, int gtt_offset,
516 struct page *user_page, int user_offset,
519 char __iomem *dst_vaddr;
522 dst_vaddr = io_mapping_map_wc(mapping, gtt_base);
523 src_vaddr = kmap(user_page);
525 memcpy_toio(dst_vaddr + gtt_offset,
526 src_vaddr + user_offset,
530 io_mapping_unmap(dst_vaddr);
534 fast_shmem_write(struct page **pages,
535 loff_t page_base, int page_offset,
540 unsigned long unwritten;
542 vaddr = kmap_atomic(pages[page_base >> PAGE_SHIFT], KM_USER0);
545 unwritten = __copy_from_user_inatomic(vaddr + page_offset, data, length);
546 kunmap_atomic(vaddr, KM_USER0);
554 * This is the fast pwrite path, where we copy the data directly from the
555 * user into the GTT, uncached.
558 i915_gem_gtt_pwrite_fast(struct drm_device *dev, struct drm_gem_object *obj,
559 struct drm_i915_gem_pwrite *args,
560 struct drm_file *file_priv)
562 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
563 drm_i915_private_t *dev_priv = dev->dev_private;
565 loff_t offset, page_base;
566 char __user *user_data;
567 int page_offset, page_length;
570 user_data = (char __user *) (uintptr_t) args->data_ptr;
572 if (!access_ok(VERIFY_READ, user_data, remain))
576 mutex_lock(&dev->struct_mutex);
577 ret = i915_gem_object_pin(obj, 0);
579 mutex_unlock(&dev->struct_mutex);
582 ret = i915_gem_object_set_to_gtt_domain(obj, 1);
586 obj_priv = to_intel_bo(obj);
587 offset = obj_priv->gtt_offset + args->offset;
590 /* Operation in this page
592 * page_base = page offset within aperture
593 * page_offset = offset within page
594 * page_length = bytes to copy for this page
596 page_base = (offset & ~(PAGE_SIZE-1));
597 page_offset = offset & (PAGE_SIZE-1);
598 page_length = remain;
599 if ((page_offset + remain) > PAGE_SIZE)
600 page_length = PAGE_SIZE - page_offset;
602 ret = fast_user_write (dev_priv->mm.gtt_mapping, page_base,
603 page_offset, user_data, page_length);
605 /* If we get a fault while copying data, then (presumably) our
606 * source page isn't available. Return the error and we'll
607 * retry in the slow path.
612 remain -= page_length;
613 user_data += page_length;
614 offset += page_length;
618 i915_gem_object_unpin(obj);
619 mutex_unlock(&dev->struct_mutex);
625 * This is the fallback GTT pwrite path, which uses get_user_pages to pin
626 * the memory and maps it using kmap_atomic for copying.
628 * This code resulted in x11perf -rgb10text consuming about 10% more CPU
629 * than using i915_gem_gtt_pwrite_fast on a G45 (32-bit).
632 i915_gem_gtt_pwrite_slow(struct drm_device *dev, struct drm_gem_object *obj,
633 struct drm_i915_gem_pwrite *args,
634 struct drm_file *file_priv)
636 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
637 drm_i915_private_t *dev_priv = dev->dev_private;
639 loff_t gtt_page_base, offset;
640 loff_t first_data_page, last_data_page, num_pages;
641 loff_t pinned_pages, i;
642 struct page **user_pages;
643 struct mm_struct *mm = current->mm;
644 int gtt_page_offset, data_page_offset, data_page_index, page_length;
646 uint64_t data_ptr = args->data_ptr;
650 /* Pin the user pages containing the data. We can't fault while
651 * holding the struct mutex, and all of the pwrite implementations
652 * want to hold it while dereferencing the user data.
654 first_data_page = data_ptr / PAGE_SIZE;
655 last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
656 num_pages = last_data_page - first_data_page + 1;
658 user_pages = drm_calloc_large(num_pages, sizeof(struct page *));
659 if (user_pages == NULL)
662 down_read(&mm->mmap_sem);
663 pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
664 num_pages, 0, 0, user_pages, NULL);
665 up_read(&mm->mmap_sem);
666 if (pinned_pages < num_pages) {
668 goto out_unpin_pages;
671 mutex_lock(&dev->struct_mutex);
672 ret = i915_gem_object_pin(obj, 0);
676 ret = i915_gem_object_set_to_gtt_domain(obj, 1);
678 goto out_unpin_object;
680 obj_priv = to_intel_bo(obj);
681 offset = obj_priv->gtt_offset + args->offset;
684 /* Operation in this page
686 * gtt_page_base = page offset within aperture
687 * gtt_page_offset = offset within page in aperture
688 * data_page_index = page number in get_user_pages return
689 * data_page_offset = offset with data_page_index page.
690 * page_length = bytes to copy for this page
692 gtt_page_base = offset & PAGE_MASK;
693 gtt_page_offset = offset & ~PAGE_MASK;
694 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
695 data_page_offset = data_ptr & ~PAGE_MASK;
697 page_length = remain;
698 if ((gtt_page_offset + page_length) > PAGE_SIZE)
699 page_length = PAGE_SIZE - gtt_page_offset;
700 if ((data_page_offset + page_length) > PAGE_SIZE)
701 page_length = PAGE_SIZE - data_page_offset;
703 slow_kernel_write(dev_priv->mm.gtt_mapping,
704 gtt_page_base, gtt_page_offset,
705 user_pages[data_page_index],
709 remain -= page_length;
710 offset += page_length;
711 data_ptr += page_length;
715 i915_gem_object_unpin(obj);
717 mutex_unlock(&dev->struct_mutex);
719 for (i = 0; i < pinned_pages; i++)
720 page_cache_release(user_pages[i]);
721 drm_free_large(user_pages);
727 * This is the fast shmem pwrite path, which attempts to directly
728 * copy_from_user into the kmapped pages backing the object.
731 i915_gem_shmem_pwrite_fast(struct drm_device *dev, struct drm_gem_object *obj,
732 struct drm_i915_gem_pwrite *args,
733 struct drm_file *file_priv)
735 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
737 loff_t offset, page_base;
738 char __user *user_data;
739 int page_offset, page_length;
742 user_data = (char __user *) (uintptr_t) args->data_ptr;
745 mutex_lock(&dev->struct_mutex);
747 ret = i915_gem_object_get_pages(obj, 0);
751 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
755 obj_priv = to_intel_bo(obj);
756 offset = args->offset;
760 /* Operation in this page
762 * page_base = page offset within aperture
763 * page_offset = offset within page
764 * page_length = bytes to copy for this page
766 page_base = (offset & ~(PAGE_SIZE-1));
767 page_offset = offset & (PAGE_SIZE-1);
768 page_length = remain;
769 if ((page_offset + remain) > PAGE_SIZE)
770 page_length = PAGE_SIZE - page_offset;
772 ret = fast_shmem_write(obj_priv->pages,
773 page_base, page_offset,
774 user_data, page_length);
778 remain -= page_length;
779 user_data += page_length;
780 offset += page_length;
784 i915_gem_object_put_pages(obj);
786 mutex_unlock(&dev->struct_mutex);
792 * This is the fallback shmem pwrite path, which uses get_user_pages to pin
793 * the memory and maps it using kmap_atomic for copying.
795 * This avoids taking mmap_sem for faulting on the user's address while the
796 * struct_mutex is held.
799 i915_gem_shmem_pwrite_slow(struct drm_device *dev, struct drm_gem_object *obj,
800 struct drm_i915_gem_pwrite *args,
801 struct drm_file *file_priv)
803 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
804 struct mm_struct *mm = current->mm;
805 struct page **user_pages;
807 loff_t offset, pinned_pages, i;
808 loff_t first_data_page, last_data_page, num_pages;
809 int shmem_page_index, shmem_page_offset;
810 int data_page_index, data_page_offset;
813 uint64_t data_ptr = args->data_ptr;
814 int do_bit17_swizzling;
818 /* Pin the user pages containing the data. We can't fault while
819 * holding the struct mutex, and all of the pwrite implementations
820 * want to hold it while dereferencing the user data.
822 first_data_page = data_ptr / PAGE_SIZE;
823 last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
824 num_pages = last_data_page - first_data_page + 1;
826 user_pages = drm_calloc_large(num_pages, sizeof(struct page *));
827 if (user_pages == NULL)
830 down_read(&mm->mmap_sem);
831 pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
832 num_pages, 0, 0, user_pages, NULL);
833 up_read(&mm->mmap_sem);
834 if (pinned_pages < num_pages) {
836 goto fail_put_user_pages;
839 do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
841 mutex_lock(&dev->struct_mutex);
843 ret = i915_gem_object_get_pages_or_evict(obj);
847 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
851 obj_priv = to_intel_bo(obj);
852 offset = args->offset;
856 /* Operation in this page
858 * shmem_page_index = page number within shmem file
859 * shmem_page_offset = offset within page in shmem file
860 * data_page_index = page number in get_user_pages return
861 * data_page_offset = offset with data_page_index page.
862 * page_length = bytes to copy for this page
864 shmem_page_index = offset / PAGE_SIZE;
865 shmem_page_offset = offset & ~PAGE_MASK;
866 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
867 data_page_offset = data_ptr & ~PAGE_MASK;
869 page_length = remain;
870 if ((shmem_page_offset + page_length) > PAGE_SIZE)
871 page_length = PAGE_SIZE - shmem_page_offset;
872 if ((data_page_offset + page_length) > PAGE_SIZE)
873 page_length = PAGE_SIZE - data_page_offset;
875 if (do_bit17_swizzling) {
876 slow_shmem_bit17_copy(obj_priv->pages[shmem_page_index],
878 user_pages[data_page_index],
883 slow_shmem_copy(obj_priv->pages[shmem_page_index],
885 user_pages[data_page_index],
890 remain -= page_length;
891 data_ptr += page_length;
892 offset += page_length;
896 i915_gem_object_put_pages(obj);
898 mutex_unlock(&dev->struct_mutex);
900 for (i = 0; i < pinned_pages; i++)
901 page_cache_release(user_pages[i]);
902 drm_free_large(user_pages);
908 * Writes data to the object referenced by handle.
910 * On error, the contents of the buffer that were to be modified are undefined.
913 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
914 struct drm_file *file_priv)
916 struct drm_i915_gem_pwrite *args = data;
917 struct drm_gem_object *obj;
918 struct drm_i915_gem_object *obj_priv;
921 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
924 obj_priv = to_intel_bo(obj);
926 /* Bounds check destination.
928 * XXX: This could use review for overflow issues...
930 if (args->offset > obj->size || args->size > obj->size ||
931 args->offset + args->size > obj->size) {
932 drm_gem_object_unreference_unlocked(obj);
936 /* We can only do the GTT pwrite on untiled buffers, as otherwise
937 * it would end up going through the fenced access, and we'll get
938 * different detiling behavior between reading and writing.
939 * pread/pwrite currently are reading and writing from the CPU
940 * perspective, requiring manual detiling by the client.
942 if (obj_priv->phys_obj)
943 ret = i915_gem_phys_pwrite(dev, obj, args, file_priv);
944 else if (obj_priv->tiling_mode == I915_TILING_NONE &&
945 dev->gtt_total != 0 &&
946 obj->write_domain != I915_GEM_DOMAIN_CPU) {
947 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file_priv);
948 if (ret == -EFAULT) {
949 ret = i915_gem_gtt_pwrite_slow(dev, obj, args,
952 } else if (i915_gem_object_needs_bit17_swizzle(obj)) {
953 ret = i915_gem_shmem_pwrite_slow(dev, obj, args, file_priv);
955 ret = i915_gem_shmem_pwrite_fast(dev, obj, args, file_priv);
956 if (ret == -EFAULT) {
957 ret = i915_gem_shmem_pwrite_slow(dev, obj, args,
964 DRM_INFO("pwrite failed %d\n", ret);
967 drm_gem_object_unreference_unlocked(obj);
973 * Called when user space prepares to use an object with the CPU, either
974 * through the mmap ioctl's mapping or a GTT mapping.
977 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
978 struct drm_file *file_priv)
980 struct drm_i915_private *dev_priv = dev->dev_private;
981 struct drm_i915_gem_set_domain *args = data;
982 struct drm_gem_object *obj;
983 struct drm_i915_gem_object *obj_priv;
984 uint32_t read_domains = args->read_domains;
985 uint32_t write_domain = args->write_domain;
988 if (!(dev->driver->driver_features & DRIVER_GEM))
991 /* Only handle setting domains to types used by the CPU. */
992 if (write_domain & I915_GEM_GPU_DOMAINS)
995 if (read_domains & I915_GEM_GPU_DOMAINS)
998 /* Having something in the write domain implies it's in the read
999 * domain, and only that read domain. Enforce that in the request.
1001 if (write_domain != 0 && read_domains != write_domain)
1004 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1007 obj_priv = to_intel_bo(obj);
1009 mutex_lock(&dev->struct_mutex);
1011 intel_mark_busy(dev, obj);
1014 DRM_INFO("set_domain_ioctl %p(%zd), %08x %08x\n",
1015 obj, obj->size, read_domains, write_domain);
1017 if (read_domains & I915_GEM_DOMAIN_GTT) {
1018 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1020 /* Update the LRU on the fence for the CPU access that's
1023 if (obj_priv->fence_reg != I915_FENCE_REG_NONE) {
1024 struct drm_i915_fence_reg *reg =
1025 &dev_priv->fence_regs[obj_priv->fence_reg];
1026 list_move_tail(®->lru_list,
1027 &dev_priv->mm.fence_list);
1030 /* Silently promote "you're not bound, there was nothing to do"
1031 * to success, since the client was just asking us to
1032 * make sure everything was done.
1037 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1040 drm_gem_object_unreference(obj);
1041 mutex_unlock(&dev->struct_mutex);
1046 * Called when user space has done writes to this buffer
1049 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1050 struct drm_file *file_priv)
1052 struct drm_i915_gem_sw_finish *args = data;
1053 struct drm_gem_object *obj;
1054 struct drm_i915_gem_object *obj_priv;
1057 if (!(dev->driver->driver_features & DRIVER_GEM))
1060 mutex_lock(&dev->struct_mutex);
1061 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1063 mutex_unlock(&dev->struct_mutex);
1068 DRM_INFO("%s: sw_finish %d (%p %zd)\n",
1069 __func__, args->handle, obj, obj->size);
1071 obj_priv = to_intel_bo(obj);
1073 /* Pinned buffers may be scanout, so flush the cache */
1074 if (obj_priv->pin_count)
1075 i915_gem_object_flush_cpu_write_domain(obj);
1077 drm_gem_object_unreference(obj);
1078 mutex_unlock(&dev->struct_mutex);
1083 * Maps the contents of an object, returning the address it is mapped
1086 * While the mapping holds a reference on the contents of the object, it doesn't
1087 * imply a ref on the object itself.
1090 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1091 struct drm_file *file_priv)
1093 struct drm_i915_gem_mmap *args = data;
1094 struct drm_gem_object *obj;
1098 if (!(dev->driver->driver_features & DRIVER_GEM))
1101 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1105 offset = args->offset;
1107 down_write(¤t->mm->mmap_sem);
1108 addr = do_mmap(obj->filp, 0, args->size,
1109 PROT_READ | PROT_WRITE, MAP_SHARED,
1111 up_write(¤t->mm->mmap_sem);
1112 drm_gem_object_unreference_unlocked(obj);
1113 if (IS_ERR((void *)addr))
1116 args->addr_ptr = (uint64_t) addr;
1122 * i915_gem_fault - fault a page into the GTT
1123 * vma: VMA in question
1126 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1127 * from userspace. The fault handler takes care of binding the object to
1128 * the GTT (if needed), allocating and programming a fence register (again,
1129 * only if needed based on whether the old reg is still valid or the object
1130 * is tiled) and inserting a new PTE into the faulting process.
1132 * Note that the faulting process may involve evicting existing objects
1133 * from the GTT and/or fence registers to make room. So performance may
1134 * suffer if the GTT working set is large or there are few fence registers
1137 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1139 struct drm_gem_object *obj = vma->vm_private_data;
1140 struct drm_device *dev = obj->dev;
1141 struct drm_i915_private *dev_priv = dev->dev_private;
1142 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1143 pgoff_t page_offset;
1146 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1148 /* We don't use vmf->pgoff since that has the fake offset */
1149 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1152 /* Now bind it into the GTT if needed */
1153 mutex_lock(&dev->struct_mutex);
1154 if (!obj_priv->gtt_space) {
1155 ret = i915_gem_object_bind_to_gtt(obj, 0);
1159 list_add_tail(&obj_priv->list, &dev_priv->mm.inactive_list);
1161 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1166 /* Need a new fence register? */
1167 if (obj_priv->tiling_mode != I915_TILING_NONE) {
1168 ret = i915_gem_object_get_fence_reg(obj);
1173 pfn = ((dev->agp->base + obj_priv->gtt_offset) >> PAGE_SHIFT) +
1176 /* Finally, remap it using the new GTT offset */
1177 ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1179 mutex_unlock(&dev->struct_mutex);
1184 return VM_FAULT_NOPAGE;
1187 return VM_FAULT_OOM;
1189 return VM_FAULT_SIGBUS;
1194 * i915_gem_create_mmap_offset - create a fake mmap offset for an object
1195 * @obj: obj in question
1197 * GEM memory mapping works by handing back to userspace a fake mmap offset
1198 * it can use in a subsequent mmap(2) call. The DRM core code then looks
1199 * up the object based on the offset and sets up the various memory mapping
1202 * This routine allocates and attaches a fake offset for @obj.
1205 i915_gem_create_mmap_offset(struct drm_gem_object *obj)
1207 struct drm_device *dev = obj->dev;
1208 struct drm_gem_mm *mm = dev->mm_private;
1209 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1210 struct drm_map_list *list;
1211 struct drm_local_map *map;
1214 /* Set the object up for mmap'ing */
1215 list = &obj->map_list;
1216 list->map = kzalloc(sizeof(struct drm_map_list), GFP_KERNEL);
1221 map->type = _DRM_GEM;
1222 map->size = obj->size;
1225 /* Get a DRM GEM mmap offset allocated... */
1226 list->file_offset_node = drm_mm_search_free(&mm->offset_manager,
1227 obj->size / PAGE_SIZE, 0, 0);
1228 if (!list->file_offset_node) {
1229 DRM_ERROR("failed to allocate offset for bo %d\n", obj->name);
1234 list->file_offset_node = drm_mm_get_block(list->file_offset_node,
1235 obj->size / PAGE_SIZE, 0);
1236 if (!list->file_offset_node) {
1241 list->hash.key = list->file_offset_node->start;
1242 if (drm_ht_insert_item(&mm->offset_hash, &list->hash)) {
1243 DRM_ERROR("failed to add to map hash\n");
1248 /* By now we should be all set, any drm_mmap request on the offset
1249 * below will get to our mmap & fault handler */
1250 obj_priv->mmap_offset = ((uint64_t) list->hash.key) << PAGE_SHIFT;
1255 drm_mm_put_block(list->file_offset_node);
1263 * i915_gem_release_mmap - remove physical page mappings
1264 * @obj: obj in question
1266 * Preserve the reservation of the mmapping with the DRM core code, but
1267 * relinquish ownership of the pages back to the system.
1269 * It is vital that we remove the page mapping if we have mapped a tiled
1270 * object through the GTT and then lose the fence register due to
1271 * resource pressure. Similarly if the object has been moved out of the
1272 * aperture, than pages mapped into userspace must be revoked. Removing the
1273 * mapping will then trigger a page fault on the next user access, allowing
1274 * fixup by i915_gem_fault().
1277 i915_gem_release_mmap(struct drm_gem_object *obj)
1279 struct drm_device *dev = obj->dev;
1280 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1282 if (dev->dev_mapping)
1283 unmap_mapping_range(dev->dev_mapping,
1284 obj_priv->mmap_offset, obj->size, 1);
1288 i915_gem_free_mmap_offset(struct drm_gem_object *obj)
1290 struct drm_device *dev = obj->dev;
1291 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1292 struct drm_gem_mm *mm = dev->mm_private;
1293 struct drm_map_list *list;
1295 list = &obj->map_list;
1296 drm_ht_remove_item(&mm->offset_hash, &list->hash);
1298 if (list->file_offset_node) {
1299 drm_mm_put_block(list->file_offset_node);
1300 list->file_offset_node = NULL;
1308 obj_priv->mmap_offset = 0;
1312 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1313 * @obj: object to check
1315 * Return the required GTT alignment for an object, taking into account
1316 * potential fence register mapping if needed.
1319 i915_gem_get_gtt_alignment(struct drm_gem_object *obj)
1321 struct drm_device *dev = obj->dev;
1322 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1326 * Minimum alignment is 4k (GTT page size), but might be greater
1327 * if a fence register is needed for the object.
1329 if (IS_I965G(dev) || obj_priv->tiling_mode == I915_TILING_NONE)
1333 * Previous chips need to be aligned to the size of the smallest
1334 * fence register that can contain the object.
1341 for (i = start; i < obj->size; i <<= 1)
1348 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1350 * @data: GTT mapping ioctl data
1351 * @file_priv: GEM object info
1353 * Simply returns the fake offset to userspace so it can mmap it.
1354 * The mmap call will end up in drm_gem_mmap(), which will set things
1355 * up so we can get faults in the handler above.
1357 * The fault handler will take care of binding the object into the GTT
1358 * (since it may have been evicted to make room for something), allocating
1359 * a fence register, and mapping the appropriate aperture address into
1363 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1364 struct drm_file *file_priv)
1366 struct drm_i915_gem_mmap_gtt *args = data;
1367 struct drm_i915_private *dev_priv = dev->dev_private;
1368 struct drm_gem_object *obj;
1369 struct drm_i915_gem_object *obj_priv;
1372 if (!(dev->driver->driver_features & DRIVER_GEM))
1375 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1379 mutex_lock(&dev->struct_mutex);
1381 obj_priv = to_intel_bo(obj);
1383 if (obj_priv->madv != I915_MADV_WILLNEED) {
1384 DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1385 drm_gem_object_unreference(obj);
1386 mutex_unlock(&dev->struct_mutex);
1391 if (!obj_priv->mmap_offset) {
1392 ret = i915_gem_create_mmap_offset(obj);
1394 drm_gem_object_unreference(obj);
1395 mutex_unlock(&dev->struct_mutex);
1400 args->offset = obj_priv->mmap_offset;
1403 * Pull it into the GTT so that we have a page list (makes the
1404 * initial fault faster and any subsequent flushing possible).
1406 if (!obj_priv->agp_mem) {
1407 ret = i915_gem_object_bind_to_gtt(obj, 0);
1409 drm_gem_object_unreference(obj);
1410 mutex_unlock(&dev->struct_mutex);
1413 list_add_tail(&obj_priv->list, &dev_priv->mm.inactive_list);
1416 drm_gem_object_unreference(obj);
1417 mutex_unlock(&dev->struct_mutex);
1423 i915_gem_object_put_pages(struct drm_gem_object *obj)
1425 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1426 int page_count = obj->size / PAGE_SIZE;
1429 BUG_ON(obj_priv->pages_refcount == 0);
1430 BUG_ON(obj_priv->madv == __I915_MADV_PURGED);
1432 if (--obj_priv->pages_refcount != 0)
1435 if (obj_priv->tiling_mode != I915_TILING_NONE)
1436 i915_gem_object_save_bit_17_swizzle(obj);
1438 if (obj_priv->madv == I915_MADV_DONTNEED)
1439 obj_priv->dirty = 0;
1441 for (i = 0; i < page_count; i++) {
1442 if (obj_priv->dirty)
1443 set_page_dirty(obj_priv->pages[i]);
1445 if (obj_priv->madv == I915_MADV_WILLNEED)
1446 mark_page_accessed(obj_priv->pages[i]);
1448 page_cache_release(obj_priv->pages[i]);
1450 obj_priv->dirty = 0;
1452 drm_free_large(obj_priv->pages);
1453 obj_priv->pages = NULL;
1457 i915_gem_object_move_to_active(struct drm_gem_object *obj, uint32_t seqno,
1458 struct intel_ring_buffer *ring)
1460 struct drm_device *dev = obj->dev;
1461 drm_i915_private_t *dev_priv = dev->dev_private;
1462 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1463 BUG_ON(ring == NULL);
1464 obj_priv->ring = ring;
1466 /* Add a reference if we're newly entering the active list. */
1467 if (!obj_priv->active) {
1468 drm_gem_object_reference(obj);
1469 obj_priv->active = 1;
1471 /* Move from whatever list we were on to the tail of execution. */
1472 spin_lock(&dev_priv->mm.active_list_lock);
1473 list_move_tail(&obj_priv->list, &ring->active_list);
1474 spin_unlock(&dev_priv->mm.active_list_lock);
1475 obj_priv->last_rendering_seqno = seqno;
1479 i915_gem_object_move_to_flushing(struct drm_gem_object *obj)
1481 struct drm_device *dev = obj->dev;
1482 drm_i915_private_t *dev_priv = dev->dev_private;
1483 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1485 BUG_ON(!obj_priv->active);
1486 list_move_tail(&obj_priv->list, &dev_priv->mm.flushing_list);
1487 obj_priv->last_rendering_seqno = 0;
1490 /* Immediately discard the backing storage */
1492 i915_gem_object_truncate(struct drm_gem_object *obj)
1494 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1495 struct inode *inode;
1497 inode = obj->filp->f_path.dentry->d_inode;
1498 if (inode->i_op->truncate)
1499 inode->i_op->truncate (inode);
1501 obj_priv->madv = __I915_MADV_PURGED;
1505 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj_priv)
1507 return obj_priv->madv == I915_MADV_DONTNEED;
1511 i915_gem_object_move_to_inactive(struct drm_gem_object *obj)
1513 struct drm_device *dev = obj->dev;
1514 drm_i915_private_t *dev_priv = dev->dev_private;
1515 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1517 i915_verify_inactive(dev, __FILE__, __LINE__);
1518 if (obj_priv->pin_count != 0)
1519 list_del_init(&obj_priv->list);
1521 list_move_tail(&obj_priv->list, &dev_priv->mm.inactive_list);
1523 BUG_ON(!list_empty(&obj_priv->gpu_write_list));
1525 obj_priv->last_rendering_seqno = 0;
1526 obj_priv->ring = NULL;
1527 if (obj_priv->active) {
1528 obj_priv->active = 0;
1529 drm_gem_object_unreference(obj);
1531 i915_verify_inactive(dev, __FILE__, __LINE__);
1535 i915_gem_process_flushing_list(struct drm_device *dev,
1536 uint32_t flush_domains, uint32_t seqno,
1537 struct intel_ring_buffer *ring)
1539 drm_i915_private_t *dev_priv = dev->dev_private;
1540 struct drm_i915_gem_object *obj_priv, *next;
1542 list_for_each_entry_safe(obj_priv, next,
1543 &dev_priv->mm.gpu_write_list,
1545 struct drm_gem_object *obj = &obj_priv->base;
1547 if ((obj->write_domain & flush_domains) ==
1548 obj->write_domain &&
1549 obj_priv->ring->ring_flag == ring->ring_flag) {
1550 uint32_t old_write_domain = obj->write_domain;
1552 obj->write_domain = 0;
1553 list_del_init(&obj_priv->gpu_write_list);
1554 i915_gem_object_move_to_active(obj, seqno, ring);
1556 /* update the fence lru list */
1557 if (obj_priv->fence_reg != I915_FENCE_REG_NONE) {
1558 struct drm_i915_fence_reg *reg =
1559 &dev_priv->fence_regs[obj_priv->fence_reg];
1560 list_move_tail(®->lru_list,
1561 &dev_priv->mm.fence_list);
1564 trace_i915_gem_object_change_domain(obj,
1572 i915_add_request(struct drm_device *dev, struct drm_file *file_priv,
1573 uint32_t flush_domains, struct intel_ring_buffer *ring)
1575 drm_i915_private_t *dev_priv = dev->dev_private;
1576 struct drm_i915_file_private *i915_file_priv = NULL;
1577 struct drm_i915_gem_request *request;
1581 if (file_priv != NULL)
1582 i915_file_priv = file_priv->driver_priv;
1584 request = kzalloc(sizeof(*request), GFP_KERNEL);
1585 if (request == NULL)
1588 seqno = ring->add_request(dev, ring, file_priv, flush_domains);
1590 request->seqno = seqno;
1591 request->ring = ring;
1592 request->emitted_jiffies = jiffies;
1593 was_empty = list_empty(&ring->request_list);
1594 list_add_tail(&request->list, &ring->request_list);
1596 if (i915_file_priv) {
1597 list_add_tail(&request->client_list,
1598 &i915_file_priv->mm.request_list);
1600 INIT_LIST_HEAD(&request->client_list);
1603 /* Associate any objects on the flushing list matching the write
1604 * domain we're flushing with our flush.
1606 if (flush_domains != 0)
1607 i915_gem_process_flushing_list(dev, flush_domains, seqno, ring);
1609 if (!dev_priv->mm.suspended) {
1610 mod_timer(&dev_priv->hangcheck_timer, jiffies + DRM_I915_HANGCHECK_PERIOD);
1612 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1618 * Command execution barrier
1620 * Ensures that all commands in the ring are finished
1621 * before signalling the CPU
1624 i915_retire_commands(struct drm_device *dev, struct intel_ring_buffer *ring)
1626 uint32_t flush_domains = 0;
1628 /* The sampler always gets flushed on i965 (sigh) */
1630 flush_domains |= I915_GEM_DOMAIN_SAMPLER;
1632 ring->flush(dev, ring,
1633 I915_GEM_DOMAIN_COMMAND, flush_domains);
1634 return flush_domains;
1638 * Moves buffers associated only with the given active seqno from the active
1639 * to inactive list, potentially freeing them.
1642 i915_gem_retire_request(struct drm_device *dev,
1643 struct drm_i915_gem_request *request)
1645 drm_i915_private_t *dev_priv = dev->dev_private;
1647 trace_i915_gem_request_retire(dev, request->seqno);
1649 /* Move any buffers on the active list that are no longer referenced
1650 * by the ringbuffer to the flushing/inactive lists as appropriate.
1652 spin_lock(&dev_priv->mm.active_list_lock);
1653 while (!list_empty(&request->ring->active_list)) {
1654 struct drm_gem_object *obj;
1655 struct drm_i915_gem_object *obj_priv;
1657 obj_priv = list_first_entry(&request->ring->active_list,
1658 struct drm_i915_gem_object,
1660 obj = &obj_priv->base;
1662 /* If the seqno being retired doesn't match the oldest in the
1663 * list, then the oldest in the list must still be newer than
1666 if (obj_priv->last_rendering_seqno != request->seqno)
1670 DRM_INFO("%s: retire %d moves to inactive list %p\n",
1671 __func__, request->seqno, obj);
1674 if (obj->write_domain != 0)
1675 i915_gem_object_move_to_flushing(obj);
1677 /* Take a reference on the object so it won't be
1678 * freed while the spinlock is held. The list
1679 * protection for this spinlock is safe when breaking
1680 * the lock like this since the next thing we do
1681 * is just get the head of the list again.
1683 drm_gem_object_reference(obj);
1684 i915_gem_object_move_to_inactive(obj);
1685 spin_unlock(&dev_priv->mm.active_list_lock);
1686 drm_gem_object_unreference(obj);
1687 spin_lock(&dev_priv->mm.active_list_lock);
1691 spin_unlock(&dev_priv->mm.active_list_lock);
1695 * Returns true if seq1 is later than seq2.
1698 i915_seqno_passed(uint32_t seq1, uint32_t seq2)
1700 return (int32_t)(seq1 - seq2) >= 0;
1704 i915_get_gem_seqno(struct drm_device *dev,
1705 struct intel_ring_buffer *ring)
1707 return ring->get_gem_seqno(dev, ring);
1711 * This function clears the request list as sequence numbers are passed.
1714 i915_gem_retire_requests_ring(struct drm_device *dev,
1715 struct intel_ring_buffer *ring)
1717 drm_i915_private_t *dev_priv = dev->dev_private;
1720 if (!ring->status_page.page_addr
1721 || list_empty(&ring->request_list))
1724 seqno = i915_get_gem_seqno(dev, ring);
1726 while (!list_empty(&ring->request_list)) {
1727 struct drm_i915_gem_request *request;
1728 uint32_t retiring_seqno;
1730 request = list_first_entry(&ring->request_list,
1731 struct drm_i915_gem_request,
1733 retiring_seqno = request->seqno;
1735 if (i915_seqno_passed(seqno, retiring_seqno) ||
1736 atomic_read(&dev_priv->mm.wedged)) {
1737 i915_gem_retire_request(dev, request);
1739 list_del(&request->list);
1740 list_del(&request->client_list);
1746 if (unlikely (dev_priv->trace_irq_seqno &&
1747 i915_seqno_passed(dev_priv->trace_irq_seqno, seqno))) {
1749 ring->user_irq_put(dev, ring);
1750 dev_priv->trace_irq_seqno = 0;
1755 i915_gem_retire_requests(struct drm_device *dev)
1757 drm_i915_private_t *dev_priv = dev->dev_private;
1759 if (!list_empty(&dev_priv->mm.deferred_free_list)) {
1760 struct drm_i915_gem_object *obj_priv, *tmp;
1762 /* We must be careful that during unbind() we do not
1763 * accidentally infinitely recurse into retire requests.
1765 * retire -> free -> unbind -> wait -> retire_ring
1767 list_for_each_entry_safe(obj_priv, tmp,
1768 &dev_priv->mm.deferred_free_list,
1770 i915_gem_free_object_tail(&obj_priv->base);
1773 i915_gem_retire_requests_ring(dev, &dev_priv->render_ring);
1775 i915_gem_retire_requests_ring(dev, &dev_priv->bsd_ring);
1779 i915_gem_retire_work_handler(struct work_struct *work)
1781 drm_i915_private_t *dev_priv;
1782 struct drm_device *dev;
1784 dev_priv = container_of(work, drm_i915_private_t,
1785 mm.retire_work.work);
1786 dev = dev_priv->dev;
1788 mutex_lock(&dev->struct_mutex);
1789 i915_gem_retire_requests(dev);
1791 if (!dev_priv->mm.suspended &&
1792 (!list_empty(&dev_priv->render_ring.request_list) ||
1794 !list_empty(&dev_priv->bsd_ring.request_list))))
1795 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1796 mutex_unlock(&dev->struct_mutex);
1800 i915_do_wait_request(struct drm_device *dev, uint32_t seqno,
1801 int interruptible, struct intel_ring_buffer *ring)
1803 drm_i915_private_t *dev_priv = dev->dev_private;
1809 if (atomic_read(&dev_priv->mm.wedged))
1812 if (!i915_seqno_passed(ring->get_gem_seqno(dev, ring), seqno)) {
1813 if (HAS_PCH_SPLIT(dev))
1814 ier = I915_READ(DEIER) | I915_READ(GTIER);
1816 ier = I915_READ(IER);
1818 DRM_ERROR("something (likely vbetool) disabled "
1819 "interrupts, re-enabling\n");
1820 i915_driver_irq_preinstall(dev);
1821 i915_driver_irq_postinstall(dev);
1824 trace_i915_gem_request_wait_begin(dev, seqno);
1826 ring->waiting_gem_seqno = seqno;
1827 ring->user_irq_get(dev, ring);
1829 ret = wait_event_interruptible(ring->irq_queue,
1831 ring->get_gem_seqno(dev, ring), seqno)
1832 || atomic_read(&dev_priv->mm.wedged));
1834 wait_event(ring->irq_queue,
1836 ring->get_gem_seqno(dev, ring), seqno)
1837 || atomic_read(&dev_priv->mm.wedged));
1839 ring->user_irq_put(dev, ring);
1840 ring->waiting_gem_seqno = 0;
1842 trace_i915_gem_request_wait_end(dev, seqno);
1844 if (atomic_read(&dev_priv->mm.wedged))
1847 if (ret && ret != -ERESTARTSYS)
1848 DRM_ERROR("%s returns %d (awaiting %d at %d)\n",
1849 __func__, ret, seqno, ring->get_gem_seqno(dev, ring));
1851 /* Directly dispatch request retiring. While we have the work queue
1852 * to handle this, the waiter on a request often wants an associated
1853 * buffer to have made it to the inactive list, and we would need
1854 * a separate wait queue to handle that.
1857 i915_gem_retire_requests_ring(dev, ring);
1863 * Waits for a sequence number to be signaled, and cleans up the
1864 * request and object lists appropriately for that event.
1867 i915_wait_request(struct drm_device *dev, uint32_t seqno,
1868 struct intel_ring_buffer *ring)
1870 return i915_do_wait_request(dev, seqno, 1, ring);
1874 i915_gem_flush(struct drm_device *dev,
1875 uint32_t invalidate_domains,
1876 uint32_t flush_domains)
1878 drm_i915_private_t *dev_priv = dev->dev_private;
1879 if (flush_domains & I915_GEM_DOMAIN_CPU)
1880 drm_agp_chipset_flush(dev);
1881 dev_priv->render_ring.flush(dev, &dev_priv->render_ring,
1886 dev_priv->bsd_ring.flush(dev, &dev_priv->bsd_ring,
1892 i915_gem_flush_ring(struct drm_device *dev,
1893 uint32_t invalidate_domains,
1894 uint32_t flush_domains,
1895 struct intel_ring_buffer *ring)
1897 if (flush_domains & I915_GEM_DOMAIN_CPU)
1898 drm_agp_chipset_flush(dev);
1899 ring->flush(dev, ring,
1905 * Ensures that all rendering to the object has completed and the object is
1906 * safe to unbind from the GTT or access from the CPU.
1909 i915_gem_object_wait_rendering(struct drm_gem_object *obj)
1911 struct drm_device *dev = obj->dev;
1912 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1915 /* This function only exists to support waiting for existing rendering,
1916 * not for emitting required flushes.
1918 BUG_ON((obj->write_domain & I915_GEM_GPU_DOMAINS) != 0);
1920 /* If there is rendering queued on the buffer being evicted, wait for
1923 if (obj_priv->active) {
1925 DRM_INFO("%s: object %p wait for seqno %08x\n",
1926 __func__, obj, obj_priv->last_rendering_seqno);
1928 ret = i915_wait_request(dev,
1929 obj_priv->last_rendering_seqno, obj_priv->ring);
1938 * Unbinds an object from the GTT aperture.
1941 i915_gem_object_unbind(struct drm_gem_object *obj)
1943 struct drm_device *dev = obj->dev;
1944 drm_i915_private_t *dev_priv = dev->dev_private;
1945 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1949 DRM_INFO("%s:%d %p\n", __func__, __LINE__, obj);
1950 DRM_INFO("gtt_space %p\n", obj_priv->gtt_space);
1952 if (obj_priv->gtt_space == NULL)
1955 if (obj_priv->pin_count != 0) {
1956 DRM_ERROR("Attempting to unbind pinned buffer\n");
1960 /* blow away mappings if mapped through GTT */
1961 i915_gem_release_mmap(obj);
1963 /* Move the object to the CPU domain to ensure that
1964 * any possible CPU writes while it's not in the GTT
1965 * are flushed when we go to remap it. This will
1966 * also ensure that all pending GPU writes are finished
1969 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
1971 if (ret != -ERESTARTSYS)
1972 DRM_ERROR("set_domain failed: %d\n", ret);
1976 BUG_ON(obj_priv->active);
1978 /* release the fence reg _after_ flushing */
1979 if (obj_priv->fence_reg != I915_FENCE_REG_NONE)
1980 i915_gem_clear_fence_reg(obj);
1982 if (obj_priv->agp_mem != NULL) {
1983 drm_unbind_agp(obj_priv->agp_mem);
1984 drm_free_agp(obj_priv->agp_mem, obj->size / PAGE_SIZE);
1985 obj_priv->agp_mem = NULL;
1988 i915_gem_object_put_pages(obj);
1989 BUG_ON(obj_priv->pages_refcount);
1991 if (obj_priv->gtt_space) {
1992 atomic_dec(&dev->gtt_count);
1993 atomic_sub(obj->size, &dev->gtt_memory);
1995 drm_mm_put_block(obj_priv->gtt_space);
1996 obj_priv->gtt_space = NULL;
1999 /* Remove ourselves from the LRU list if present. */
2000 spin_lock(&dev_priv->mm.active_list_lock);
2001 if (!list_empty(&obj_priv->list))
2002 list_del_init(&obj_priv->list);
2003 spin_unlock(&dev_priv->mm.active_list_lock);
2005 if (i915_gem_object_is_purgeable(obj_priv))
2006 i915_gem_object_truncate(obj);
2008 trace_i915_gem_object_unbind(obj);
2013 static struct drm_gem_object *
2014 i915_gem_find_inactive_object(struct drm_device *dev, int min_size)
2016 drm_i915_private_t *dev_priv = dev->dev_private;
2017 struct drm_i915_gem_object *obj_priv;
2018 struct drm_gem_object *best = NULL;
2019 struct drm_gem_object *first = NULL;
2021 /* Try to find the smallest clean object */
2022 list_for_each_entry(obj_priv, &dev_priv->mm.inactive_list, list) {
2023 struct drm_gem_object *obj = &obj_priv->base;
2024 if (obj->size >= min_size) {
2025 if ((!obj_priv->dirty ||
2026 i915_gem_object_is_purgeable(obj_priv)) &&
2027 (!best || obj->size < best->size)) {
2029 if (best->size == min_size)
2037 return best ? best : first;
2041 i915_gpu_idle(struct drm_device *dev)
2043 drm_i915_private_t *dev_priv = dev->dev_private;
2045 uint32_t seqno1, seqno2;
2048 spin_lock(&dev_priv->mm.active_list_lock);
2049 lists_empty = (list_empty(&dev_priv->mm.flushing_list) &&
2050 list_empty(&dev_priv->render_ring.active_list) &&
2052 list_empty(&dev_priv->bsd_ring.active_list)));
2053 spin_unlock(&dev_priv->mm.active_list_lock);
2058 /* Flush everything onto the inactive list. */
2059 i915_gem_flush(dev, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
2060 seqno1 = i915_add_request(dev, NULL, I915_GEM_GPU_DOMAINS,
2061 &dev_priv->render_ring);
2064 ret = i915_wait_request(dev, seqno1, &dev_priv->render_ring);
2067 seqno2 = i915_add_request(dev, NULL, I915_GEM_GPU_DOMAINS,
2068 &dev_priv->bsd_ring);
2072 ret = i915_wait_request(dev, seqno2, &dev_priv->bsd_ring);
2082 i915_gem_evict_everything(struct drm_device *dev)
2084 drm_i915_private_t *dev_priv = dev->dev_private;
2088 spin_lock(&dev_priv->mm.active_list_lock);
2089 lists_empty = (list_empty(&dev_priv->mm.inactive_list) &&
2090 list_empty(&dev_priv->mm.flushing_list) &&
2091 list_empty(&dev_priv->render_ring.active_list) &&
2093 || list_empty(&dev_priv->bsd_ring.active_list)));
2094 spin_unlock(&dev_priv->mm.active_list_lock);
2099 /* Flush everything (on to the inactive lists) and evict */
2100 ret = i915_gpu_idle(dev);
2104 BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
2106 ret = i915_gem_evict_from_inactive_list(dev);
2110 spin_lock(&dev_priv->mm.active_list_lock);
2111 lists_empty = (list_empty(&dev_priv->mm.inactive_list) &&
2112 list_empty(&dev_priv->mm.flushing_list) &&
2113 list_empty(&dev_priv->render_ring.active_list) &&
2115 || list_empty(&dev_priv->bsd_ring.active_list)));
2116 spin_unlock(&dev_priv->mm.active_list_lock);
2117 BUG_ON(!lists_empty);
2123 i915_gem_evict_something(struct drm_device *dev, int min_size)
2125 drm_i915_private_t *dev_priv = dev->dev_private;
2126 struct drm_gem_object *obj;
2129 struct intel_ring_buffer *render_ring = &dev_priv->render_ring;
2130 struct intel_ring_buffer *bsd_ring = &dev_priv->bsd_ring;
2132 i915_gem_retire_requests(dev);
2134 /* If there's an inactive buffer available now, grab it
2137 obj = i915_gem_find_inactive_object(dev, min_size);
2139 struct drm_i915_gem_object *obj_priv;
2142 DRM_INFO("%s: evicting %p\n", __func__, obj);
2144 obj_priv = to_intel_bo(obj);
2145 BUG_ON(obj_priv->pin_count != 0);
2146 BUG_ON(obj_priv->active);
2148 /* Wait on the rendering and unbind the buffer. */
2149 return i915_gem_object_unbind(obj);
2152 /* If we didn't get anything, but the ring is still processing
2153 * things, wait for the next to finish and hopefully leave us
2154 * a buffer to evict.
2156 if (!list_empty(&render_ring->request_list)) {
2157 struct drm_i915_gem_request *request;
2159 request = list_first_entry(&render_ring->request_list,
2160 struct drm_i915_gem_request,
2163 ret = i915_wait_request(dev,
2164 request->seqno, request->ring);
2171 if (HAS_BSD(dev) && !list_empty(&bsd_ring->request_list)) {
2172 struct drm_i915_gem_request *request;
2174 request = list_first_entry(&bsd_ring->request_list,
2175 struct drm_i915_gem_request,
2178 ret = i915_wait_request(dev,
2179 request->seqno, request->ring);
2186 /* If we didn't have anything on the request list but there
2187 * are buffers awaiting a flush, emit one and try again.
2188 * When we wait on it, those buffers waiting for that flush
2189 * will get moved to inactive.
2191 if (!list_empty(&dev_priv->mm.flushing_list)) {
2192 struct drm_i915_gem_object *obj_priv;
2194 /* Find an object that we can immediately reuse */
2195 list_for_each_entry(obj_priv, &dev_priv->mm.flushing_list, list) {
2196 obj = &obj_priv->base;
2197 if (obj->size >= min_size)
2206 i915_gem_flush_ring(dev,
2210 seqno = i915_add_request(dev, NULL,
2219 /* If we didn't do any of the above, there's no single buffer
2220 * large enough to swap out for the new one, so just evict
2221 * everything and start again. (This should be rare.)
2223 if (!list_empty (&dev_priv->mm.inactive_list))
2224 return i915_gem_evict_from_inactive_list(dev);
2226 return i915_gem_evict_everything(dev);
2231 i915_gem_object_get_pages(struct drm_gem_object *obj,
2234 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2236 struct address_space *mapping;
2237 struct inode *inode;
2240 BUG_ON(obj_priv->pages_refcount
2241 == DRM_I915_GEM_OBJECT_MAX_PAGES_REFCOUNT);
2243 if (obj_priv->pages_refcount++ != 0)
2246 /* Get the list of pages out of our struct file. They'll be pinned
2247 * at this point until we release them.
2249 page_count = obj->size / PAGE_SIZE;
2250 BUG_ON(obj_priv->pages != NULL);
2251 obj_priv->pages = drm_calloc_large(page_count, sizeof(struct page *));
2252 if (obj_priv->pages == NULL) {
2253 obj_priv->pages_refcount--;
2257 inode = obj->filp->f_path.dentry->d_inode;
2258 mapping = inode->i_mapping;
2259 for (i = 0; i < page_count; i++) {
2260 page = read_cache_page_gfp(mapping, i,
2268 obj_priv->pages[i] = page;
2271 if (obj_priv->tiling_mode != I915_TILING_NONE)
2272 i915_gem_object_do_bit_17_swizzle(obj);
2278 page_cache_release(obj_priv->pages[i]);
2280 drm_free_large(obj_priv->pages);
2281 obj_priv->pages = NULL;
2282 obj_priv->pages_refcount--;
2283 return PTR_ERR(page);
2286 static void sandybridge_write_fence_reg(struct drm_i915_fence_reg *reg)
2288 struct drm_gem_object *obj = reg->obj;
2289 struct drm_device *dev = obj->dev;
2290 drm_i915_private_t *dev_priv = dev->dev_private;
2291 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2292 int regnum = obj_priv->fence_reg;
2295 val = (uint64_t)((obj_priv->gtt_offset + obj->size - 4096) &
2297 val |= obj_priv->gtt_offset & 0xfffff000;
2298 val |= (uint64_t)((obj_priv->stride / 128) - 1) <<
2299 SANDYBRIDGE_FENCE_PITCH_SHIFT;
2301 if (obj_priv->tiling_mode == I915_TILING_Y)
2302 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2303 val |= I965_FENCE_REG_VALID;
2305 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + (regnum * 8), val);
2308 static void i965_write_fence_reg(struct drm_i915_fence_reg *reg)
2310 struct drm_gem_object *obj = reg->obj;
2311 struct drm_device *dev = obj->dev;
2312 drm_i915_private_t *dev_priv = dev->dev_private;
2313 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2314 int regnum = obj_priv->fence_reg;
2317 val = (uint64_t)((obj_priv->gtt_offset + obj->size - 4096) &
2319 val |= obj_priv->gtt_offset & 0xfffff000;
2320 val |= ((obj_priv->stride / 128) - 1) << I965_FENCE_PITCH_SHIFT;
2321 if (obj_priv->tiling_mode == I915_TILING_Y)
2322 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2323 val |= I965_FENCE_REG_VALID;
2325 I915_WRITE64(FENCE_REG_965_0 + (regnum * 8), val);
2328 static void i915_write_fence_reg(struct drm_i915_fence_reg *reg)
2330 struct drm_gem_object *obj = reg->obj;
2331 struct drm_device *dev = obj->dev;
2332 drm_i915_private_t *dev_priv = dev->dev_private;
2333 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2334 int regnum = obj_priv->fence_reg;
2336 uint32_t fence_reg, val;
2339 if ((obj_priv->gtt_offset & ~I915_FENCE_START_MASK) ||
2340 (obj_priv->gtt_offset & (obj->size - 1))) {
2341 WARN(1, "%s: object 0x%08x not 1M or size (0x%zx) aligned\n",
2342 __func__, obj_priv->gtt_offset, obj->size);
2346 if (obj_priv->tiling_mode == I915_TILING_Y &&
2347 HAS_128_BYTE_Y_TILING(dev))
2352 /* Note: pitch better be a power of two tile widths */
2353 pitch_val = obj_priv->stride / tile_width;
2354 pitch_val = ffs(pitch_val) - 1;
2356 if (obj_priv->tiling_mode == I915_TILING_Y &&
2357 HAS_128_BYTE_Y_TILING(dev))
2358 WARN_ON(pitch_val > I830_FENCE_MAX_PITCH_VAL);
2360 WARN_ON(pitch_val > I915_FENCE_MAX_PITCH_VAL);
2362 val = obj_priv->gtt_offset;
2363 if (obj_priv->tiling_mode == I915_TILING_Y)
2364 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2365 val |= I915_FENCE_SIZE_BITS(obj->size);
2366 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2367 val |= I830_FENCE_REG_VALID;
2370 fence_reg = FENCE_REG_830_0 + (regnum * 4);
2372 fence_reg = FENCE_REG_945_8 + ((regnum - 8) * 4);
2373 I915_WRITE(fence_reg, val);
2376 static void i830_write_fence_reg(struct drm_i915_fence_reg *reg)
2378 struct drm_gem_object *obj = reg->obj;
2379 struct drm_device *dev = obj->dev;
2380 drm_i915_private_t *dev_priv = dev->dev_private;
2381 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2382 int regnum = obj_priv->fence_reg;
2385 uint32_t fence_size_bits;
2387 if ((obj_priv->gtt_offset & ~I830_FENCE_START_MASK) ||
2388 (obj_priv->gtt_offset & (obj->size - 1))) {
2389 WARN(1, "%s: object 0x%08x not 512K or size aligned\n",
2390 __func__, obj_priv->gtt_offset);
2394 pitch_val = obj_priv->stride / 128;
2395 pitch_val = ffs(pitch_val) - 1;
2396 WARN_ON(pitch_val > I830_FENCE_MAX_PITCH_VAL);
2398 val = obj_priv->gtt_offset;
2399 if (obj_priv->tiling_mode == I915_TILING_Y)
2400 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2401 fence_size_bits = I830_FENCE_SIZE_BITS(obj->size);
2402 WARN_ON(fence_size_bits & ~0x00000f00);
2403 val |= fence_size_bits;
2404 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2405 val |= I830_FENCE_REG_VALID;
2407 I915_WRITE(FENCE_REG_830_0 + (regnum * 4), val);
2410 static int i915_find_fence_reg(struct drm_device *dev)
2412 struct drm_i915_fence_reg *reg = NULL;
2413 struct drm_i915_gem_object *obj_priv = NULL;
2414 struct drm_i915_private *dev_priv = dev->dev_private;
2415 struct drm_gem_object *obj = NULL;
2418 /* First try to find a free reg */
2420 for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2421 reg = &dev_priv->fence_regs[i];
2425 obj_priv = to_intel_bo(reg->obj);
2426 if (!obj_priv->pin_count)
2433 /* None available, try to steal one or wait for a user to finish */
2434 i = I915_FENCE_REG_NONE;
2435 list_for_each_entry(reg, &dev_priv->mm.fence_list,
2438 obj_priv = to_intel_bo(obj);
2440 if (obj_priv->pin_count)
2444 i = obj_priv->fence_reg;
2448 BUG_ON(i == I915_FENCE_REG_NONE);
2450 /* We only have a reference on obj from the active list. put_fence_reg
2451 * might drop that one, causing a use-after-free in it. So hold a
2452 * private reference to obj like the other callers of put_fence_reg
2453 * (set_tiling ioctl) do. */
2454 drm_gem_object_reference(obj);
2455 ret = i915_gem_object_put_fence_reg(obj);
2456 drm_gem_object_unreference(obj);
2464 * i915_gem_object_get_fence_reg - set up a fence reg for an object
2465 * @obj: object to map through a fence reg
2467 * When mapping objects through the GTT, userspace wants to be able to write
2468 * to them without having to worry about swizzling if the object is tiled.
2470 * This function walks the fence regs looking for a free one for @obj,
2471 * stealing one if it can't find any.
2473 * It then sets up the reg based on the object's properties: address, pitch
2474 * and tiling format.
2477 i915_gem_object_get_fence_reg(struct drm_gem_object *obj)
2479 struct drm_device *dev = obj->dev;
2480 struct drm_i915_private *dev_priv = dev->dev_private;
2481 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2482 struct drm_i915_fence_reg *reg = NULL;
2485 /* Just update our place in the LRU if our fence is getting used. */
2486 if (obj_priv->fence_reg != I915_FENCE_REG_NONE) {
2487 reg = &dev_priv->fence_regs[obj_priv->fence_reg];
2488 list_move_tail(®->lru_list, &dev_priv->mm.fence_list);
2492 switch (obj_priv->tiling_mode) {
2493 case I915_TILING_NONE:
2494 WARN(1, "allocating a fence for non-tiled object?\n");
2497 if (!obj_priv->stride)
2499 WARN((obj_priv->stride & (512 - 1)),
2500 "object 0x%08x is X tiled but has non-512B pitch\n",
2501 obj_priv->gtt_offset);
2504 if (!obj_priv->stride)
2506 WARN((obj_priv->stride & (128 - 1)),
2507 "object 0x%08x is Y tiled but has non-128B pitch\n",
2508 obj_priv->gtt_offset);
2512 ret = i915_find_fence_reg(dev);
2516 obj_priv->fence_reg = ret;
2517 reg = &dev_priv->fence_regs[obj_priv->fence_reg];
2518 list_add_tail(®->lru_list, &dev_priv->mm.fence_list);
2523 sandybridge_write_fence_reg(reg);
2524 else if (IS_I965G(dev))
2525 i965_write_fence_reg(reg);
2526 else if (IS_I9XX(dev))
2527 i915_write_fence_reg(reg);
2529 i830_write_fence_reg(reg);
2531 trace_i915_gem_object_get_fence(obj, obj_priv->fence_reg,
2532 obj_priv->tiling_mode);
2538 * i915_gem_clear_fence_reg - clear out fence register info
2539 * @obj: object to clear
2541 * Zeroes out the fence register itself and clears out the associated
2542 * data structures in dev_priv and obj_priv.
2545 i915_gem_clear_fence_reg(struct drm_gem_object *obj)
2547 struct drm_device *dev = obj->dev;
2548 drm_i915_private_t *dev_priv = dev->dev_private;
2549 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2550 struct drm_i915_fence_reg *reg =
2551 &dev_priv->fence_regs[obj_priv->fence_reg];
2554 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 +
2555 (obj_priv->fence_reg * 8), 0);
2556 } else if (IS_I965G(dev)) {
2557 I915_WRITE64(FENCE_REG_965_0 + (obj_priv->fence_reg * 8), 0);
2561 if (obj_priv->fence_reg < 8)
2562 fence_reg = FENCE_REG_830_0 + obj_priv->fence_reg * 4;
2564 fence_reg = FENCE_REG_945_8 + (obj_priv->fence_reg -
2567 I915_WRITE(fence_reg, 0);
2571 obj_priv->fence_reg = I915_FENCE_REG_NONE;
2572 list_del_init(®->lru_list);
2576 * i915_gem_object_put_fence_reg - waits on outstanding fenced access
2577 * to the buffer to finish, and then resets the fence register.
2578 * @obj: tiled object holding a fence register.
2580 * Zeroes out the fence register itself and clears out the associated
2581 * data structures in dev_priv and obj_priv.
2584 i915_gem_object_put_fence_reg(struct drm_gem_object *obj)
2586 struct drm_device *dev = obj->dev;
2587 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2589 if (obj_priv->fence_reg == I915_FENCE_REG_NONE)
2592 /* If we've changed tiling, GTT-mappings of the object
2593 * need to re-fault to ensure that the correct fence register
2594 * setup is in place.
2596 i915_gem_release_mmap(obj);
2598 /* On the i915, GPU access to tiled buffers is via a fence,
2599 * therefore we must wait for any outstanding access to complete
2600 * before clearing the fence.
2602 if (!IS_I965G(dev)) {
2605 ret = i915_gem_object_flush_gpu_write_domain(obj);
2609 ret = i915_gem_object_wait_rendering(obj);
2614 i915_gem_object_flush_gtt_write_domain(obj);
2615 i915_gem_clear_fence_reg (obj);
2621 * Finds free space in the GTT aperture and binds the object there.
2624 i915_gem_object_bind_to_gtt(struct drm_gem_object *obj, unsigned alignment)
2626 struct drm_device *dev = obj->dev;
2627 drm_i915_private_t *dev_priv = dev->dev_private;
2628 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2629 struct drm_mm_node *free_space;
2630 gfp_t gfpmask = __GFP_NORETRY | __GFP_NOWARN;
2633 if (obj_priv->madv != I915_MADV_WILLNEED) {
2634 DRM_ERROR("Attempting to bind a purgeable object\n");
2639 alignment = i915_gem_get_gtt_alignment(obj);
2640 if (alignment & (i915_gem_get_gtt_alignment(obj) - 1)) {
2641 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
2645 /* If the object is bigger than the entire aperture, reject it early
2646 * before evicting everything in a vain attempt to find space.
2648 if (obj->size > dev->gtt_total) {
2649 DRM_ERROR("Attempting to bind an object larger than the aperture\n");
2654 free_space = drm_mm_search_free(&dev_priv->mm.gtt_space,
2655 obj->size, alignment, 0);
2656 if (free_space != NULL) {
2657 obj_priv->gtt_space = drm_mm_get_block(free_space, obj->size,
2659 if (obj_priv->gtt_space != NULL)
2660 obj_priv->gtt_offset = obj_priv->gtt_space->start;
2662 if (obj_priv->gtt_space == NULL) {
2663 /* If the gtt is empty and we're still having trouble
2664 * fitting our object in, we're out of memory.
2667 DRM_INFO("%s: GTT full, evicting something\n", __func__);
2669 ret = i915_gem_evict_something(dev, obj->size);
2677 DRM_INFO("Binding object of size %zd at 0x%08x\n",
2678 obj->size, obj_priv->gtt_offset);
2680 ret = i915_gem_object_get_pages(obj, gfpmask);
2682 drm_mm_put_block(obj_priv->gtt_space);
2683 obj_priv->gtt_space = NULL;
2685 if (ret == -ENOMEM) {
2686 /* first try to clear up some space from the GTT */
2687 ret = i915_gem_evict_something(dev, obj->size);
2689 /* now try to shrink everyone else */
2704 /* Create an AGP memory structure pointing at our pages, and bind it
2707 obj_priv->agp_mem = drm_agp_bind_pages(dev,
2709 obj->size >> PAGE_SHIFT,
2710 obj_priv->gtt_offset,
2711 obj_priv->agp_type);
2712 if (obj_priv->agp_mem == NULL) {
2713 i915_gem_object_put_pages(obj);
2714 drm_mm_put_block(obj_priv->gtt_space);
2715 obj_priv->gtt_space = NULL;
2717 ret = i915_gem_evict_something(dev, obj->size);
2723 atomic_inc(&dev->gtt_count);
2724 atomic_add(obj->size, &dev->gtt_memory);
2726 /* Assert that the object is not currently in any GPU domain. As it
2727 * wasn't in the GTT, there shouldn't be any way it could have been in
2730 BUG_ON(obj->read_domains & I915_GEM_GPU_DOMAINS);
2731 BUG_ON(obj->write_domain & I915_GEM_GPU_DOMAINS);
2733 trace_i915_gem_object_bind(obj, obj_priv->gtt_offset);
2739 i915_gem_clflush_object(struct drm_gem_object *obj)
2741 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2743 /* If we don't have a page list set up, then we're not pinned
2744 * to GPU, and we can ignore the cache flush because it'll happen
2745 * again at bind time.
2747 if (obj_priv->pages == NULL)
2750 trace_i915_gem_object_clflush(obj);
2752 drm_clflush_pages(obj_priv->pages, obj->size / PAGE_SIZE);
2755 /** Flushes any GPU write domain for the object if it's dirty. */
2757 i915_gem_object_flush_gpu_write_domain(struct drm_gem_object *obj)
2759 struct drm_device *dev = obj->dev;
2760 uint32_t old_write_domain;
2761 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2763 if ((obj->write_domain & I915_GEM_GPU_DOMAINS) == 0)
2766 /* Queue the GPU write cache flushing we need. */
2767 old_write_domain = obj->write_domain;
2768 i915_gem_flush(dev, 0, obj->write_domain);
2769 if (i915_add_request(dev, NULL, obj->write_domain, obj_priv->ring) == 0)
2772 trace_i915_gem_object_change_domain(obj,
2778 /** Flushes the GTT write domain for the object if it's dirty. */
2780 i915_gem_object_flush_gtt_write_domain(struct drm_gem_object *obj)
2782 uint32_t old_write_domain;
2784 if (obj->write_domain != I915_GEM_DOMAIN_GTT)
2787 /* No actual flushing is required for the GTT write domain. Writes
2788 * to it immediately go to main memory as far as we know, so there's
2789 * no chipset flush. It also doesn't land in render cache.
2791 old_write_domain = obj->write_domain;
2792 obj->write_domain = 0;
2794 trace_i915_gem_object_change_domain(obj,
2799 /** Flushes the CPU write domain for the object if it's dirty. */
2801 i915_gem_object_flush_cpu_write_domain(struct drm_gem_object *obj)
2803 struct drm_device *dev = obj->dev;
2804 uint32_t old_write_domain;
2806 if (obj->write_domain != I915_GEM_DOMAIN_CPU)
2809 i915_gem_clflush_object(obj);
2810 drm_agp_chipset_flush(dev);
2811 old_write_domain = obj->write_domain;
2812 obj->write_domain = 0;
2814 trace_i915_gem_object_change_domain(obj,
2820 i915_gem_object_flush_write_domain(struct drm_gem_object *obj)
2824 switch (obj->write_domain) {
2825 case I915_GEM_DOMAIN_GTT:
2826 i915_gem_object_flush_gtt_write_domain(obj);
2828 case I915_GEM_DOMAIN_CPU:
2829 i915_gem_object_flush_cpu_write_domain(obj);
2832 ret = i915_gem_object_flush_gpu_write_domain(obj);
2840 * Moves a single object to the GTT read, and possibly write domain.
2842 * This function returns when the move is complete, including waiting on
2846 i915_gem_object_set_to_gtt_domain(struct drm_gem_object *obj, int write)
2848 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2849 uint32_t old_write_domain, old_read_domains;
2852 /* Not valid to be called on unbound objects. */
2853 if (obj_priv->gtt_space == NULL)
2856 ret = i915_gem_object_flush_gpu_write_domain(obj);
2860 /* Wait on any GPU rendering and flushing to occur. */
2861 ret = i915_gem_object_wait_rendering(obj);
2865 old_write_domain = obj->write_domain;
2866 old_read_domains = obj->read_domains;
2868 /* If we're writing through the GTT domain, then CPU and GPU caches
2869 * will need to be invalidated at next use.
2872 obj->read_domains &= I915_GEM_DOMAIN_GTT;
2874 i915_gem_object_flush_cpu_write_domain(obj);
2876 /* It should now be out of any other write domains, and we can update
2877 * the domain values for our changes.
2879 BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2880 obj->read_domains |= I915_GEM_DOMAIN_GTT;
2882 obj->write_domain = I915_GEM_DOMAIN_GTT;
2883 obj_priv->dirty = 1;
2886 trace_i915_gem_object_change_domain(obj,
2894 * Prepare buffer for display plane. Use uninterruptible for possible flush
2895 * wait, as in modesetting process we're not supposed to be interrupted.
2898 i915_gem_object_set_to_display_plane(struct drm_gem_object *obj)
2900 struct drm_device *dev = obj->dev;
2901 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2902 uint32_t old_write_domain, old_read_domains;
2905 /* Not valid to be called on unbound objects. */
2906 if (obj_priv->gtt_space == NULL)
2909 ret = i915_gem_object_flush_gpu_write_domain(obj);
2913 /* Wait on any GPU rendering and flushing to occur. */
2914 if (obj_priv->active) {
2916 DRM_INFO("%s: object %p wait for seqno %08x\n",
2917 __func__, obj, obj_priv->last_rendering_seqno);
2919 ret = i915_do_wait_request(dev,
2920 obj_priv->last_rendering_seqno,
2927 i915_gem_object_flush_cpu_write_domain(obj);
2929 old_write_domain = obj->write_domain;
2930 old_read_domains = obj->read_domains;
2932 /* It should now be out of any other write domains, and we can update
2933 * the domain values for our changes.
2935 BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2936 obj->read_domains = I915_GEM_DOMAIN_GTT;
2937 obj->write_domain = I915_GEM_DOMAIN_GTT;
2938 obj_priv->dirty = 1;
2940 trace_i915_gem_object_change_domain(obj,
2948 * Moves a single object to the CPU read, and possibly write domain.
2950 * This function returns when the move is complete, including waiting on
2954 i915_gem_object_set_to_cpu_domain(struct drm_gem_object *obj, int write)
2956 uint32_t old_write_domain, old_read_domains;
2959 ret = i915_gem_object_flush_gpu_write_domain(obj);
2963 /* Wait on any GPU rendering and flushing to occur. */
2964 ret = i915_gem_object_wait_rendering(obj);
2968 i915_gem_object_flush_gtt_write_domain(obj);
2970 /* If we have a partially-valid cache of the object in the CPU,
2971 * finish invalidating it and free the per-page flags.
2973 i915_gem_object_set_to_full_cpu_read_domain(obj);
2975 old_write_domain = obj->write_domain;
2976 old_read_domains = obj->read_domains;
2978 /* Flush the CPU cache if it's still invalid. */
2979 if ((obj->read_domains & I915_GEM_DOMAIN_CPU) == 0) {
2980 i915_gem_clflush_object(obj);
2982 obj->read_domains |= I915_GEM_DOMAIN_CPU;
2985 /* It should now be out of any other write domains, and we can update
2986 * the domain values for our changes.
2988 BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
2990 /* If we're writing through the CPU, then the GPU read domains will
2991 * need to be invalidated at next use.
2994 obj->read_domains &= I915_GEM_DOMAIN_CPU;
2995 obj->write_domain = I915_GEM_DOMAIN_CPU;
2998 trace_i915_gem_object_change_domain(obj,
3006 * Set the next domain for the specified object. This
3007 * may not actually perform the necessary flushing/invaliding though,
3008 * as that may want to be batched with other set_domain operations
3010 * This is (we hope) the only really tricky part of gem. The goal
3011 * is fairly simple -- track which caches hold bits of the object
3012 * and make sure they remain coherent. A few concrete examples may
3013 * help to explain how it works. For shorthand, we use the notation
3014 * (read_domains, write_domain), e.g. (CPU, CPU) to indicate the
3015 * a pair of read and write domain masks.
3017 * Case 1: the batch buffer
3023 * 5. Unmapped from GTT
3026 * Let's take these a step at a time
3029 * Pages allocated from the kernel may still have
3030 * cache contents, so we set them to (CPU, CPU) always.
3031 * 2. Written by CPU (using pwrite)
3032 * The pwrite function calls set_domain (CPU, CPU) and
3033 * this function does nothing (as nothing changes)
3035 * This function asserts that the object is not
3036 * currently in any GPU-based read or write domains
3038 * i915_gem_execbuffer calls set_domain (COMMAND, 0).
3039 * As write_domain is zero, this function adds in the
3040 * current read domains (CPU+COMMAND, 0).
3041 * flush_domains is set to CPU.
3042 * invalidate_domains is set to COMMAND
3043 * clflush is run to get data out of the CPU caches
3044 * then i915_dev_set_domain calls i915_gem_flush to
3045 * emit an MI_FLUSH and drm_agp_chipset_flush
3046 * 5. Unmapped from GTT
3047 * i915_gem_object_unbind calls set_domain (CPU, CPU)
3048 * flush_domains and invalidate_domains end up both zero
3049 * so no flushing/invalidating happens
3053 * Case 2: The shared render buffer
3057 * 3. Read/written by GPU
3058 * 4. set_domain to (CPU,CPU)
3059 * 5. Read/written by CPU
3060 * 6. Read/written by GPU
3063 * Same as last example, (CPU, CPU)
3065 * Nothing changes (assertions find that it is not in the GPU)
3066 * 3. Read/written by GPU
3067 * execbuffer calls set_domain (RENDER, RENDER)
3068 * flush_domains gets CPU
3069 * invalidate_domains gets GPU
3071 * MI_FLUSH and drm_agp_chipset_flush
3072 * 4. set_domain (CPU, CPU)
3073 * flush_domains gets GPU
3074 * invalidate_domains gets CPU
3075 * wait_rendering (obj) to make sure all drawing is complete.
3076 * This will include an MI_FLUSH to get the data from GPU
3078 * clflush (obj) to invalidate the CPU cache
3079 * Another MI_FLUSH in i915_gem_flush (eliminate this somehow?)
3080 * 5. Read/written by CPU
3081 * cache lines are loaded and dirtied
3082 * 6. Read written by GPU
3083 * Same as last GPU access
3085 * Case 3: The constant buffer
3090 * 4. Updated (written) by CPU again
3099 * flush_domains = CPU
3100 * invalidate_domains = RENDER
3103 * drm_agp_chipset_flush
3104 * 4. Updated (written) by CPU again
3106 * flush_domains = 0 (no previous write domain)
3107 * invalidate_domains = 0 (no new read domains)
3110 * flush_domains = CPU
3111 * invalidate_domains = RENDER
3114 * drm_agp_chipset_flush
3117 i915_gem_object_set_to_gpu_domain(struct drm_gem_object *obj)
3119 struct drm_device *dev = obj->dev;
3120 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
3121 uint32_t invalidate_domains = 0;
3122 uint32_t flush_domains = 0;
3123 uint32_t old_read_domains;
3125 BUG_ON(obj->pending_read_domains & I915_GEM_DOMAIN_CPU);
3126 BUG_ON(obj->pending_write_domain == I915_GEM_DOMAIN_CPU);
3128 intel_mark_busy(dev, obj);
3131 DRM_INFO("%s: object %p read %08x -> %08x write %08x -> %08x\n",
3133 obj->read_domains, obj->pending_read_domains,
3134 obj->write_domain, obj->pending_write_domain);
3137 * If the object isn't moving to a new write domain,
3138 * let the object stay in multiple read domains
3140 if (obj->pending_write_domain == 0)
3141 obj->pending_read_domains |= obj->read_domains;
3143 obj_priv->dirty = 1;
3146 * Flush the current write domain if
3147 * the new read domains don't match. Invalidate
3148 * any read domains which differ from the old
3151 if (obj->write_domain &&
3152 obj->write_domain != obj->pending_read_domains) {
3153 flush_domains |= obj->write_domain;
3154 invalidate_domains |=
3155 obj->pending_read_domains & ~obj->write_domain;
3158 * Invalidate any read caches which may have
3159 * stale data. That is, any new read domains.
3161 invalidate_domains |= obj->pending_read_domains & ~obj->read_domains;
3162 if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_CPU) {
3164 DRM_INFO("%s: CPU domain flush %08x invalidate %08x\n",
3165 __func__, flush_domains, invalidate_domains);
3167 i915_gem_clflush_object(obj);
3170 old_read_domains = obj->read_domains;
3172 /* The actual obj->write_domain will be updated with
3173 * pending_write_domain after we emit the accumulated flush for all
3174 * of our domain changes in execbuffers (which clears objects'
3175 * write_domains). So if we have a current write domain that we
3176 * aren't changing, set pending_write_domain to that.
3178 if (flush_domains == 0 && obj->pending_write_domain == 0)
3179 obj->pending_write_domain = obj->write_domain;
3180 obj->read_domains = obj->pending_read_domains;
3182 dev->invalidate_domains |= invalidate_domains;
3183 dev->flush_domains |= flush_domains;
3185 DRM_INFO("%s: read %08x write %08x invalidate %08x flush %08x\n",
3187 obj->read_domains, obj->write_domain,
3188 dev->invalidate_domains, dev->flush_domains);
3191 trace_i915_gem_object_change_domain(obj,
3197 * Moves the object from a partially CPU read to a full one.
3199 * Note that this only resolves i915_gem_object_set_cpu_read_domain_range(),
3200 * and doesn't handle transitioning from !(read_domains & I915_GEM_DOMAIN_CPU).
3203 i915_gem_object_set_to_full_cpu_read_domain(struct drm_gem_object *obj)
3205 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
3207 if (!obj_priv->page_cpu_valid)
3210 /* If we're partially in the CPU read domain, finish moving it in.
3212 if (obj->read_domains & I915_GEM_DOMAIN_CPU) {
3215 for (i = 0; i <= (obj->size - 1) / PAGE_SIZE; i++) {
3216 if (obj_priv->page_cpu_valid[i])
3218 drm_clflush_pages(obj_priv->pages + i, 1);
3222 /* Free the page_cpu_valid mappings which are now stale, whether
3223 * or not we've got I915_GEM_DOMAIN_CPU.
3225 kfree(obj_priv->page_cpu_valid);
3226 obj_priv->page_cpu_valid = NULL;
3230 * Set the CPU read domain on a range of the object.
3232 * The object ends up with I915_GEM_DOMAIN_CPU in its read flags although it's
3233 * not entirely valid. The page_cpu_valid member of the object flags which
3234 * pages have been flushed, and will be respected by
3235 * i915_gem_object_set_to_cpu_domain() if it's called on to get a valid mapping
3236 * of the whole object.
3238 * This function returns when the move is complete, including waiting on
3242 i915_gem_object_set_cpu_read_domain_range(struct drm_gem_object *obj,
3243 uint64_t offset, uint64_t size)
3245 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
3246 uint32_t old_read_domains;
3249 if (offset == 0 && size == obj->size)
3250 return i915_gem_object_set_to_cpu_domain(obj, 0);
3252 ret = i915_gem_object_flush_gpu_write_domain(obj);
3256 /* Wait on any GPU rendering and flushing to occur. */
3257 ret = i915_gem_object_wait_rendering(obj);
3260 i915_gem_object_flush_gtt_write_domain(obj);
3262 /* If we're already fully in the CPU read domain, we're done. */
3263 if (obj_priv->page_cpu_valid == NULL &&
3264 (obj->read_domains & I915_GEM_DOMAIN_CPU) != 0)
3267 /* Otherwise, create/clear the per-page CPU read domain flag if we're
3268 * newly adding I915_GEM_DOMAIN_CPU
3270 if (obj_priv->page_cpu_valid == NULL) {
3271 obj_priv->page_cpu_valid = kzalloc(obj->size / PAGE_SIZE,
3273 if (obj_priv->page_cpu_valid == NULL)
3275 } else if ((obj->read_domains & I915_GEM_DOMAIN_CPU) == 0)
3276 memset(obj_priv->page_cpu_valid, 0, obj->size / PAGE_SIZE);
3278 /* Flush the cache on any pages that are still invalid from the CPU's
3281 for (i = offset / PAGE_SIZE; i <= (offset + size - 1) / PAGE_SIZE;
3283 if (obj_priv->page_cpu_valid[i])
3286 drm_clflush_pages(obj_priv->pages + i, 1);
3288 obj_priv->page_cpu_valid[i] = 1;
3291 /* It should now be out of any other write domains, and we can update
3292 * the domain values for our changes.
3294 BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3296 old_read_domains = obj->read_domains;
3297 obj->read_domains |= I915_GEM_DOMAIN_CPU;
3299 trace_i915_gem_object_change_domain(obj,
3307 * Pin an object to the GTT and evaluate the relocations landing in it.
3310 i915_gem_object_pin_and_relocate(struct drm_gem_object *obj,
3311 struct drm_file *file_priv,
3312 struct drm_i915_gem_exec_object2 *entry,
3313 struct drm_i915_gem_relocation_entry *relocs)
3315 struct drm_device *dev = obj->dev;
3316 drm_i915_private_t *dev_priv = dev->dev_private;
3317 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
3319 void __iomem *reloc_page;
3322 need_fence = entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
3323 obj_priv->tiling_mode != I915_TILING_NONE;
3325 /* Check fence reg constraints and rebind if necessary */
3327 !i915_gem_object_fence_offset_ok(obj,
3328 obj_priv->tiling_mode)) {
3329 ret = i915_gem_object_unbind(obj);
3334 /* Choose the GTT offset for our buffer and put it there. */
3335 ret = i915_gem_object_pin(obj, (uint32_t) entry->alignment);
3340 * Pre-965 chips need a fence register set up in order to
3341 * properly handle blits to/from tiled surfaces.
3344 ret = i915_gem_object_get_fence_reg(obj);
3346 i915_gem_object_unpin(obj);
3351 entry->offset = obj_priv->gtt_offset;
3353 /* Apply the relocations, using the GTT aperture to avoid cache
3354 * flushing requirements.
3356 for (i = 0; i < entry->relocation_count; i++) {
3357 struct drm_i915_gem_relocation_entry *reloc= &relocs[i];
3358 struct drm_gem_object *target_obj;
3359 struct drm_i915_gem_object *target_obj_priv;
3360 uint32_t reloc_val, reloc_offset;
3361 uint32_t __iomem *reloc_entry;
3363 target_obj = drm_gem_object_lookup(obj->dev, file_priv,
3364 reloc->target_handle);
3365 if (target_obj == NULL) {
3366 i915_gem_object_unpin(obj);
3369 target_obj_priv = to_intel_bo(target_obj);
3372 DRM_INFO("%s: obj %p offset %08x target %d "
3373 "read %08x write %08x gtt %08x "
3374 "presumed %08x delta %08x\n",
3377 (int) reloc->offset,
3378 (int) reloc->target_handle,
3379 (int) reloc->read_domains,
3380 (int) reloc->write_domain,
3381 (int) target_obj_priv->gtt_offset,
3382 (int) reloc->presumed_offset,
3386 /* The target buffer should have appeared before us in the
3387 * exec_object list, so it should have a GTT space bound by now.
3389 if (target_obj_priv->gtt_space == NULL) {
3390 DRM_ERROR("No GTT space found for object %d\n",
3391 reloc->target_handle);
3392 drm_gem_object_unreference(target_obj);
3393 i915_gem_object_unpin(obj);
3397 /* Validate that the target is in a valid r/w GPU domain */
3398 if (reloc->write_domain & (reloc->write_domain - 1)) {
3399 DRM_ERROR("reloc with multiple write domains: "
3400 "obj %p target %d offset %d "
3401 "read %08x write %08x",
3402 obj, reloc->target_handle,
3403 (int) reloc->offset,
3404 reloc->read_domains,
3405 reloc->write_domain);
3408 if (reloc->write_domain & I915_GEM_DOMAIN_CPU ||
3409 reloc->read_domains & I915_GEM_DOMAIN_CPU) {
3410 DRM_ERROR("reloc with read/write CPU domains: "
3411 "obj %p target %d offset %d "
3412 "read %08x write %08x",
3413 obj, reloc->target_handle,
3414 (int) reloc->offset,
3415 reloc->read_domains,
3416 reloc->write_domain);
3417 drm_gem_object_unreference(target_obj);
3418 i915_gem_object_unpin(obj);
3421 if (reloc->write_domain && target_obj->pending_write_domain &&
3422 reloc->write_domain != target_obj->pending_write_domain) {
3423 DRM_ERROR("Write domain conflict: "
3424 "obj %p target %d offset %d "
3425 "new %08x old %08x\n",
3426 obj, reloc->target_handle,
3427 (int) reloc->offset,
3428 reloc->write_domain,
3429 target_obj->pending_write_domain);
3430 drm_gem_object_unreference(target_obj);
3431 i915_gem_object_unpin(obj);
3435 target_obj->pending_read_domains |= reloc->read_domains;
3436 target_obj->pending_write_domain |= reloc->write_domain;
3438 /* If the relocation already has the right value in it, no
3439 * more work needs to be done.
3441 if (target_obj_priv->gtt_offset == reloc->presumed_offset) {
3442 drm_gem_object_unreference(target_obj);
3446 /* Check that the relocation address is valid... */
3447 if (reloc->offset > obj->size - 4) {
3448 DRM_ERROR("Relocation beyond object bounds: "
3449 "obj %p target %d offset %d size %d.\n",
3450 obj, reloc->target_handle,
3451 (int) reloc->offset, (int) obj->size);
3452 drm_gem_object_unreference(target_obj);
3453 i915_gem_object_unpin(obj);
3456 if (reloc->offset & 3) {
3457 DRM_ERROR("Relocation not 4-byte aligned: "
3458 "obj %p target %d offset %d.\n",
3459 obj, reloc->target_handle,
3460 (int) reloc->offset);
3461 drm_gem_object_unreference(target_obj);
3462 i915_gem_object_unpin(obj);
3466 /* and points to somewhere within the target object. */
3467 if (reloc->delta >= target_obj->size) {
3468 DRM_ERROR("Relocation beyond target object bounds: "
3469 "obj %p target %d delta %d size %d.\n",
3470 obj, reloc->target_handle,
3471 (int) reloc->delta, (int) target_obj->size);
3472 drm_gem_object_unreference(target_obj);
3473 i915_gem_object_unpin(obj);
3477 ret = i915_gem_object_set_to_gtt_domain(obj, 1);
3479 drm_gem_object_unreference(target_obj);
3480 i915_gem_object_unpin(obj);
3484 /* Map the page containing the relocation we're going to
3487 reloc_offset = obj_priv->gtt_offset + reloc->offset;
3488 reloc_page = io_mapping_map_atomic_wc(dev_priv->mm.gtt_mapping,
3491 reloc_entry = (uint32_t __iomem *)(reloc_page +
3492 (reloc_offset & (PAGE_SIZE - 1)));
3493 reloc_val = target_obj_priv->gtt_offset + reloc->delta;
3496 DRM_INFO("Applied relocation: %p@0x%08x %08x -> %08x\n",
3497 obj, (unsigned int) reloc->offset,
3498 readl(reloc_entry), reloc_val);
3500 writel(reloc_val, reloc_entry);
3501 io_mapping_unmap_atomic(reloc_page);
3503 /* The updated presumed offset for this entry will be
3504 * copied back out to the user.
3506 reloc->presumed_offset = target_obj_priv->gtt_offset;
3508 drm_gem_object_unreference(target_obj);
3513 i915_gem_dump_object(obj, 128, __func__, ~0);
3518 /* Throttle our rendering by waiting until the ring has completed our requests
3519 * emitted over 20 msec ago.
3521 * Note that if we were to use the current jiffies each time around the loop,
3522 * we wouldn't escape the function with any frames outstanding if the time to
3523 * render a frame was over 20ms.
3525 * This should get us reasonable parallelism between CPU and GPU but also
3526 * relatively low latency when blocking on a particular request to finish.
3529 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file_priv)
3531 struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv;
3533 unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3535 mutex_lock(&dev->struct_mutex);
3536 while (!list_empty(&i915_file_priv->mm.request_list)) {
3537 struct drm_i915_gem_request *request;
3539 request = list_first_entry(&i915_file_priv->mm.request_list,
3540 struct drm_i915_gem_request,
3543 if (time_after_eq(request->emitted_jiffies, recent_enough))
3546 ret = i915_wait_request(dev, request->seqno, request->ring);
3550 mutex_unlock(&dev->struct_mutex);
3556 i915_gem_get_relocs_from_user(struct drm_i915_gem_exec_object2 *exec_list,
3557 uint32_t buffer_count,
3558 struct drm_i915_gem_relocation_entry **relocs)
3560 uint32_t reloc_count = 0, reloc_index = 0, i;
3564 for (i = 0; i < buffer_count; i++) {
3565 if (reloc_count + exec_list[i].relocation_count < reloc_count)
3567 reloc_count += exec_list[i].relocation_count;
3570 *relocs = drm_calloc_large(reloc_count, sizeof(**relocs));
3571 if (*relocs == NULL) {
3572 DRM_ERROR("failed to alloc relocs, count %d\n", reloc_count);
3576 for (i = 0; i < buffer_count; i++) {
3577 struct drm_i915_gem_relocation_entry __user *user_relocs;
3579 user_relocs = (void __user *)(uintptr_t)exec_list[i].relocs_ptr;
3581 ret = copy_from_user(&(*relocs)[reloc_index],
3583 exec_list[i].relocation_count *
3586 drm_free_large(*relocs);
3591 reloc_index += exec_list[i].relocation_count;
3598 i915_gem_put_relocs_to_user(struct drm_i915_gem_exec_object2 *exec_list,
3599 uint32_t buffer_count,
3600 struct drm_i915_gem_relocation_entry *relocs)
3602 uint32_t reloc_count = 0, i;
3608 for (i = 0; i < buffer_count; i++) {
3609 struct drm_i915_gem_relocation_entry __user *user_relocs;
3612 user_relocs = (void __user *)(uintptr_t)exec_list[i].relocs_ptr;
3614 unwritten = copy_to_user(user_relocs,
3615 &relocs[reloc_count],
3616 exec_list[i].relocation_count *
3624 reloc_count += exec_list[i].relocation_count;
3628 drm_free_large(relocs);
3634 i915_gem_check_execbuffer (struct drm_i915_gem_execbuffer2 *exec,
3635 uint64_t exec_offset)
3637 uint32_t exec_start, exec_len;
3639 exec_start = (uint32_t) exec_offset + exec->batch_start_offset;
3640 exec_len = (uint32_t) exec->batch_len;
3642 if ((exec_start | exec_len) & 0x7)
3652 i915_gem_wait_for_pending_flip(struct drm_device *dev,
3653 struct drm_gem_object **object_list,
3656 drm_i915_private_t *dev_priv = dev->dev_private;
3657 struct drm_i915_gem_object *obj_priv;
3662 prepare_to_wait(&dev_priv->pending_flip_queue,
3663 &wait, TASK_INTERRUPTIBLE);
3664 for (i = 0; i < count; i++) {
3665 obj_priv = to_intel_bo(object_list[i]);
3666 if (atomic_read(&obj_priv->pending_flip) > 0)
3672 if (!signal_pending(current)) {
3673 mutex_unlock(&dev->struct_mutex);
3675 mutex_lock(&dev->struct_mutex);
3681 finish_wait(&dev_priv->pending_flip_queue, &wait);
3688 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
3689 struct drm_file *file_priv,
3690 struct drm_i915_gem_execbuffer2 *args,
3691 struct drm_i915_gem_exec_object2 *exec_list)
3693 drm_i915_private_t *dev_priv = dev->dev_private;
3694 struct drm_gem_object **object_list = NULL;
3695 struct drm_gem_object *batch_obj;
3696 struct drm_i915_gem_object *obj_priv;
3697 struct drm_clip_rect *cliprects = NULL;
3698 struct drm_i915_gem_relocation_entry *relocs = NULL;
3699 int ret = 0, ret2, i, pinned = 0;
3700 uint64_t exec_offset;
3701 uint32_t seqno, flush_domains, reloc_index;
3702 int pin_tries, flips;
3704 struct intel_ring_buffer *ring = NULL;
3707 DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
3708 (int) args->buffers_ptr, args->buffer_count, args->batch_len);
3710 if (args->flags & I915_EXEC_BSD) {
3711 if (!HAS_BSD(dev)) {
3712 DRM_ERROR("execbuf with wrong flag\n");
3715 ring = &dev_priv->bsd_ring;
3717 ring = &dev_priv->render_ring;
3721 if (args->buffer_count < 1) {
3722 DRM_ERROR("execbuf with %d buffers\n", args->buffer_count);
3725 object_list = drm_malloc_ab(sizeof(*object_list), args->buffer_count);
3726 if (object_list == NULL) {
3727 DRM_ERROR("Failed to allocate object list for %d buffers\n",
3728 args->buffer_count);
3733 if (args->num_cliprects != 0) {
3734 cliprects = kcalloc(args->num_cliprects, sizeof(*cliprects),
3736 if (cliprects == NULL) {
3741 ret = copy_from_user(cliprects,
3742 (struct drm_clip_rect __user *)
3743 (uintptr_t) args->cliprects_ptr,
3744 sizeof(*cliprects) * args->num_cliprects);
3746 DRM_ERROR("copy %d cliprects failed: %d\n",
3747 args->num_cliprects, ret);
3752 ret = i915_gem_get_relocs_from_user(exec_list, args->buffer_count,
3757 mutex_lock(&dev->struct_mutex);
3759 i915_verify_inactive(dev, __FILE__, __LINE__);
3761 if (atomic_read(&dev_priv->mm.wedged)) {
3762 mutex_unlock(&dev->struct_mutex);
3767 if (dev_priv->mm.suspended) {
3768 mutex_unlock(&dev->struct_mutex);
3773 /* Look up object handles */
3775 for (i = 0; i < args->buffer_count; i++) {
3776 object_list[i] = drm_gem_object_lookup(dev, file_priv,
3777 exec_list[i].handle);
3778 if (object_list[i] == NULL) {
3779 DRM_ERROR("Invalid object handle %d at index %d\n",
3780 exec_list[i].handle, i);
3781 /* prevent error path from reading uninitialized data */
3782 args->buffer_count = i + 1;
3787 obj_priv = to_intel_bo(object_list[i]);
3788 if (obj_priv->in_execbuffer) {
3789 DRM_ERROR("Object %p appears more than once in object list\n",
3791 /* prevent error path from reading uninitialized data */
3792 args->buffer_count = i + 1;
3796 obj_priv->in_execbuffer = true;
3797 flips += atomic_read(&obj_priv->pending_flip);
3801 ret = i915_gem_wait_for_pending_flip(dev, object_list,
3802 args->buffer_count);
3807 /* Pin and relocate */
3808 for (pin_tries = 0; ; pin_tries++) {
3812 for (i = 0; i < args->buffer_count; i++) {
3813 object_list[i]->pending_read_domains = 0;
3814 object_list[i]->pending_write_domain = 0;
3815 ret = i915_gem_object_pin_and_relocate(object_list[i],
3818 &relocs[reloc_index]);
3822 reloc_index += exec_list[i].relocation_count;
3828 /* error other than GTT full, or we've already tried again */
3829 if (ret != -ENOSPC || pin_tries >= 1) {
3830 if (ret != -ERESTARTSYS) {
3831 unsigned long long total_size = 0;
3833 for (i = 0; i < args->buffer_count; i++) {
3834 obj_priv = to_intel_bo(object_list[i]);
3836 total_size += object_list[i]->size;
3838 exec_list[i].flags & EXEC_OBJECT_NEEDS_FENCE &&
3839 obj_priv->tiling_mode != I915_TILING_NONE;
3841 DRM_ERROR("Failed to pin buffer %d of %d, total %llu bytes, %d fences: %d\n",
3842 pinned+1, args->buffer_count,
3843 total_size, num_fences,
3845 DRM_ERROR("%d objects [%d pinned], "
3846 "%d object bytes [%d pinned], "
3847 "%d/%d gtt bytes\n",
3848 atomic_read(&dev->object_count),
3849 atomic_read(&dev->pin_count),
3850 atomic_read(&dev->object_memory),
3851 atomic_read(&dev->pin_memory),
3852 atomic_read(&dev->gtt_memory),
3858 /* unpin all of our buffers */
3859 for (i = 0; i < pinned; i++)
3860 i915_gem_object_unpin(object_list[i]);
3863 /* evict everyone we can from the aperture */
3864 ret = i915_gem_evict_everything(dev);
3865 if (ret && ret != -ENOSPC)
3869 /* Set the pending read domains for the batch buffer to COMMAND */
3870 batch_obj = object_list[args->buffer_count-1];
3871 if (batch_obj->pending_write_domain) {
3872 DRM_ERROR("Attempting to use self-modifying batch buffer\n");
3876 batch_obj->pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
3878 /* Sanity check the batch buffer, prior to moving objects */
3879 exec_offset = exec_list[args->buffer_count - 1].offset;
3880 ret = i915_gem_check_execbuffer (args, exec_offset);
3882 DRM_ERROR("execbuf with invalid offset/length\n");
3886 i915_verify_inactive(dev, __FILE__, __LINE__);
3888 /* Zero the global flush/invalidate flags. These
3889 * will be modified as new domains are computed
3892 dev->invalidate_domains = 0;
3893 dev->flush_domains = 0;
3895 for (i = 0; i < args->buffer_count; i++) {
3896 struct drm_gem_object *obj = object_list[i];
3898 /* Compute new gpu domains and update invalidate/flush */
3899 i915_gem_object_set_to_gpu_domain(obj);
3902 i915_verify_inactive(dev, __FILE__, __LINE__);
3904 if (dev->invalidate_domains | dev->flush_domains) {
3906 DRM_INFO("%s: invalidate_domains %08x flush_domains %08x\n",
3908 dev->invalidate_domains,
3909 dev->flush_domains);
3912 dev->invalidate_domains,
3913 dev->flush_domains);
3914 if (dev->flush_domains & I915_GEM_GPU_DOMAINS) {
3915 (void)i915_add_request(dev, file_priv,
3917 &dev_priv->render_ring);
3920 (void)i915_add_request(dev, file_priv,
3922 &dev_priv->bsd_ring);
3926 for (i = 0; i < args->buffer_count; i++) {
3927 struct drm_gem_object *obj = object_list[i];
3928 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
3929 uint32_t old_write_domain = obj->write_domain;
3931 obj->write_domain = obj->pending_write_domain;
3932 if (obj->write_domain)
3933 list_move_tail(&obj_priv->gpu_write_list,
3934 &dev_priv->mm.gpu_write_list);
3936 list_del_init(&obj_priv->gpu_write_list);
3938 trace_i915_gem_object_change_domain(obj,
3943 i915_verify_inactive(dev, __FILE__, __LINE__);
3946 for (i = 0; i < args->buffer_count; i++) {
3947 i915_gem_object_check_coherency(object_list[i],
3948 exec_list[i].handle);
3953 i915_gem_dump_object(batch_obj,
3959 /* Exec the batchbuffer */
3960 ret = ring->dispatch_gem_execbuffer(dev, ring, args,
3961 cliprects, exec_offset);
3963 DRM_ERROR("dispatch failed %d\n", ret);
3968 * Ensure that the commands in the batch buffer are
3969 * finished before the interrupt fires
3971 flush_domains = i915_retire_commands(dev, ring);
3973 i915_verify_inactive(dev, __FILE__, __LINE__);
3976 * Get a seqno representing the execution of the current buffer,
3977 * which we can wait on. We would like to mitigate these interrupts,
3978 * likely by only creating seqnos occasionally (so that we have
3979 * *some* interrupts representing completion of buffers that we can
3980 * wait on when trying to clear up gtt space).
3982 seqno = i915_add_request(dev, file_priv, flush_domains, ring);
3984 for (i = 0; i < args->buffer_count; i++) {
3985 struct drm_gem_object *obj = object_list[i];
3986 obj_priv = to_intel_bo(obj);
3988 i915_gem_object_move_to_active(obj, seqno, ring);
3990 DRM_INFO("%s: move to exec list %p\n", __func__, obj);
3994 i915_dump_lru(dev, __func__);
3997 i915_verify_inactive(dev, __FILE__, __LINE__);
4000 for (i = 0; i < pinned; i++)
4001 i915_gem_object_unpin(object_list[i]);
4003 for (i = 0; i < args->buffer_count; i++) {
4004 if (object_list[i]) {
4005 obj_priv = to_intel_bo(object_list[i]);
4006 obj_priv->in_execbuffer = false;
4008 drm_gem_object_unreference(object_list[i]);
4011 mutex_unlock(&dev->struct_mutex);
4014 /* Copy the updated relocations out regardless of current error
4015 * state. Failure to update the relocs would mean that the next
4016 * time userland calls execbuf, it would do so with presumed offset
4017 * state that didn't match the actual object state.
4019 ret2 = i915_gem_put_relocs_to_user(exec_list, args->buffer_count,
4022 DRM_ERROR("Failed to copy relocations back out: %d\n", ret2);
4028 drm_free_large(object_list);
4035 * Legacy execbuffer just creates an exec2 list from the original exec object
4036 * list array and passes it to the real function.
4039 i915_gem_execbuffer(struct drm_device *dev, void *data,
4040 struct drm_file *file_priv)
4042 struct drm_i915_gem_execbuffer *args = data;
4043 struct drm_i915_gem_execbuffer2 exec2;
4044 struct drm_i915_gem_exec_object *exec_list = NULL;
4045 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
4049 DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
4050 (int) args->buffers_ptr, args->buffer_count, args->batch_len);
4053 if (args->buffer_count < 1) {
4054 DRM_ERROR("execbuf with %d buffers\n", args->buffer_count);
4058 /* Copy in the exec list from userland */
4059 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
4060 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
4061 if (exec_list == NULL || exec2_list == NULL) {
4062 DRM_ERROR("Failed to allocate exec list for %d buffers\n",
4063 args->buffer_count);
4064 drm_free_large(exec_list);
4065 drm_free_large(exec2_list);
4068 ret = copy_from_user(exec_list,
4069 (struct drm_i915_relocation_entry __user *)
4070 (uintptr_t) args->buffers_ptr,
4071 sizeof(*exec_list) * args->buffer_count);
4073 DRM_ERROR("copy %d exec entries failed %d\n",
4074 args->buffer_count, ret);
4075 drm_free_large(exec_list);
4076 drm_free_large(exec2_list);
4080 for (i = 0; i < args->buffer_count; i++) {
4081 exec2_list[i].handle = exec_list[i].handle;
4082 exec2_list[i].relocation_count = exec_list[i].relocation_count;
4083 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
4084 exec2_list[i].alignment = exec_list[i].alignment;
4085 exec2_list[i].offset = exec_list[i].offset;
4087 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
4089 exec2_list[i].flags = 0;
4092 exec2.buffers_ptr = args->buffers_ptr;
4093 exec2.buffer_count = args->buffer_count;
4094 exec2.batch_start_offset = args->batch_start_offset;
4095 exec2.batch_len = args->batch_len;
4096 exec2.DR1 = args->DR1;
4097 exec2.DR4 = args->DR4;
4098 exec2.num_cliprects = args->num_cliprects;
4099 exec2.cliprects_ptr = args->cliprects_ptr;
4100 exec2.flags = I915_EXEC_RENDER;
4102 ret = i915_gem_do_execbuffer(dev, data, file_priv, &exec2, exec2_list);
4104 /* Copy the new buffer offsets back to the user's exec list. */
4105 for (i = 0; i < args->buffer_count; i++)
4106 exec_list[i].offset = exec2_list[i].offset;
4107 /* ... and back out to userspace */
4108 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
4109 (uintptr_t) args->buffers_ptr,
4111 sizeof(*exec_list) * args->buffer_count);
4114 DRM_ERROR("failed to copy %d exec entries "
4115 "back to user (%d)\n",
4116 args->buffer_count, ret);
4120 drm_free_large(exec_list);
4121 drm_free_large(exec2_list);
4126 i915_gem_execbuffer2(struct drm_device *dev, void *data,
4127 struct drm_file *file_priv)
4129 struct drm_i915_gem_execbuffer2 *args = data;
4130 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
4134 DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
4135 (int) args->buffers_ptr, args->buffer_count, args->batch_len);
4138 if (args->buffer_count < 1) {
4139 DRM_ERROR("execbuf2 with %d buffers\n", args->buffer_count);
4143 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
4144 if (exec2_list == NULL) {
4145 DRM_ERROR("Failed to allocate exec list for %d buffers\n",
4146 args->buffer_count);
4149 ret = copy_from_user(exec2_list,
4150 (struct drm_i915_relocation_entry __user *)
4151 (uintptr_t) args->buffers_ptr,
4152 sizeof(*exec2_list) * args->buffer_count);
4154 DRM_ERROR("copy %d exec entries failed %d\n",
4155 args->buffer_count, ret);
4156 drm_free_large(exec2_list);
4160 ret = i915_gem_do_execbuffer(dev, data, file_priv, args, exec2_list);
4162 /* Copy the new buffer offsets back to the user's exec list. */
4163 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
4164 (uintptr_t) args->buffers_ptr,
4166 sizeof(*exec2_list) * args->buffer_count);
4169 DRM_ERROR("failed to copy %d exec entries "
4170 "back to user (%d)\n",
4171 args->buffer_count, ret);
4175 drm_free_large(exec2_list);
4180 i915_gem_object_pin(struct drm_gem_object *obj, uint32_t alignment)
4182 struct drm_device *dev = obj->dev;
4183 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
4186 BUG_ON(obj_priv->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT);
4188 i915_verify_inactive(dev, __FILE__, __LINE__);
4190 if (obj_priv->gtt_space != NULL) {
4192 alignment = i915_gem_get_gtt_alignment(obj);
4193 if (obj_priv->gtt_offset & (alignment - 1)) {
4194 ret = i915_gem_object_unbind(obj);
4200 if (obj_priv->gtt_space == NULL) {
4201 ret = i915_gem_object_bind_to_gtt(obj, alignment);
4206 obj_priv->pin_count++;
4208 /* If the object is not active and not pending a flush,
4209 * remove it from the inactive list
4211 if (obj_priv->pin_count == 1) {
4212 atomic_inc(&dev->pin_count);
4213 atomic_add(obj->size, &dev->pin_memory);
4214 if (!obj_priv->active &&
4215 (obj->write_domain & I915_GEM_GPU_DOMAINS) == 0 &&
4216 !list_empty(&obj_priv->list))
4217 list_del_init(&obj_priv->list);
4219 i915_verify_inactive(dev, __FILE__, __LINE__);
4225 i915_gem_object_unpin(struct drm_gem_object *obj)
4227 struct drm_device *dev = obj->dev;
4228 drm_i915_private_t *dev_priv = dev->dev_private;
4229 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
4231 i915_verify_inactive(dev, __FILE__, __LINE__);
4232 obj_priv->pin_count--;
4233 BUG_ON(obj_priv->pin_count < 0);
4234 BUG_ON(obj_priv->gtt_space == NULL);
4236 /* If the object is no longer pinned, and is
4237 * neither active nor being flushed, then stick it on
4240 if (obj_priv->pin_count == 0) {
4241 if (!obj_priv->active &&
4242 (obj->write_domain & I915_GEM_GPU_DOMAINS) == 0)
4243 list_move_tail(&obj_priv->list,
4244 &dev_priv->mm.inactive_list);
4245 atomic_dec(&dev->pin_count);
4246 atomic_sub(obj->size, &dev->pin_memory);
4248 i915_verify_inactive(dev, __FILE__, __LINE__);
4252 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
4253 struct drm_file *file_priv)
4255 struct drm_i915_gem_pin *args = data;
4256 struct drm_gem_object *obj;
4257 struct drm_i915_gem_object *obj_priv;
4260 mutex_lock(&dev->struct_mutex);
4262 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4264 DRM_ERROR("Bad handle in i915_gem_pin_ioctl(): %d\n",
4266 mutex_unlock(&dev->struct_mutex);
4269 obj_priv = to_intel_bo(obj);
4271 if (obj_priv->madv != I915_MADV_WILLNEED) {
4272 DRM_ERROR("Attempting to pin a purgeable buffer\n");
4273 drm_gem_object_unreference(obj);
4274 mutex_unlock(&dev->struct_mutex);
4278 if (obj_priv->pin_filp != NULL && obj_priv->pin_filp != file_priv) {
4279 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
4281 drm_gem_object_unreference(obj);
4282 mutex_unlock(&dev->struct_mutex);
4286 obj_priv->user_pin_count++;
4287 obj_priv->pin_filp = file_priv;
4288 if (obj_priv->user_pin_count == 1) {
4289 ret = i915_gem_object_pin(obj, args->alignment);
4291 drm_gem_object_unreference(obj);
4292 mutex_unlock(&dev->struct_mutex);
4297 /* XXX - flush the CPU caches for pinned objects
4298 * as the X server doesn't manage domains yet
4300 i915_gem_object_flush_cpu_write_domain(obj);
4301 args->offset = obj_priv->gtt_offset;
4302 drm_gem_object_unreference(obj);
4303 mutex_unlock(&dev->struct_mutex);
4309 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
4310 struct drm_file *file_priv)
4312 struct drm_i915_gem_pin *args = data;
4313 struct drm_gem_object *obj;
4314 struct drm_i915_gem_object *obj_priv;
4316 mutex_lock(&dev->struct_mutex);
4318 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4320 DRM_ERROR("Bad handle in i915_gem_unpin_ioctl(): %d\n",
4322 mutex_unlock(&dev->struct_mutex);
4326 obj_priv = to_intel_bo(obj);
4327 if (obj_priv->pin_filp != file_priv) {
4328 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
4330 drm_gem_object_unreference(obj);
4331 mutex_unlock(&dev->struct_mutex);
4334 obj_priv->user_pin_count--;
4335 if (obj_priv->user_pin_count == 0) {
4336 obj_priv->pin_filp = NULL;
4337 i915_gem_object_unpin(obj);
4340 drm_gem_object_unreference(obj);
4341 mutex_unlock(&dev->struct_mutex);
4346 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
4347 struct drm_file *file_priv)
4349 struct drm_i915_gem_busy *args = data;
4350 struct drm_gem_object *obj;
4351 struct drm_i915_gem_object *obj_priv;
4353 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4355 DRM_ERROR("Bad handle in i915_gem_busy_ioctl(): %d\n",
4360 mutex_lock(&dev->struct_mutex);
4361 /* Update the active list for the hardware's current position.
4362 * Otherwise this only updates on a delayed timer or when irqs are
4363 * actually unmasked, and our working set ends up being larger than
4366 i915_gem_retire_requests(dev);
4368 obj_priv = to_intel_bo(obj);
4369 /* Don't count being on the flushing list against the object being
4370 * done. Otherwise, a buffer left on the flushing list but not getting
4371 * flushed (because nobody's flushing that domain) won't ever return
4372 * unbusy and get reused by libdrm's bo cache. The other expected
4373 * consumer of this interface, OpenGL's occlusion queries, also specs
4374 * that the objects get unbusy "eventually" without any interference.
4376 args->busy = obj_priv->active && obj_priv->last_rendering_seqno != 0;
4378 drm_gem_object_unreference(obj);
4379 mutex_unlock(&dev->struct_mutex);
4384 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
4385 struct drm_file *file_priv)
4387 return i915_gem_ring_throttle(dev, file_priv);
4391 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
4392 struct drm_file *file_priv)
4394 struct drm_i915_gem_madvise *args = data;
4395 struct drm_gem_object *obj;
4396 struct drm_i915_gem_object *obj_priv;
4398 switch (args->madv) {
4399 case I915_MADV_DONTNEED:
4400 case I915_MADV_WILLNEED:
4406 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4408 DRM_ERROR("Bad handle in i915_gem_madvise_ioctl(): %d\n",
4413 mutex_lock(&dev->struct_mutex);
4414 obj_priv = to_intel_bo(obj);
4416 if (obj_priv->pin_count) {
4417 drm_gem_object_unreference(obj);
4418 mutex_unlock(&dev->struct_mutex);
4420 DRM_ERROR("Attempted i915_gem_madvise_ioctl() on a pinned object\n");
4424 if (obj_priv->madv != __I915_MADV_PURGED)
4425 obj_priv->madv = args->madv;
4427 /* if the object is no longer bound, discard its backing storage */
4428 if (i915_gem_object_is_purgeable(obj_priv) &&
4429 obj_priv->gtt_space == NULL)
4430 i915_gem_object_truncate(obj);
4432 args->retained = obj_priv->madv != __I915_MADV_PURGED;
4434 drm_gem_object_unreference(obj);
4435 mutex_unlock(&dev->struct_mutex);
4440 struct drm_gem_object * i915_gem_alloc_object(struct drm_device *dev,
4443 struct drm_i915_gem_object *obj;
4445 obj = kzalloc(sizeof(*obj), GFP_KERNEL);
4449 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
4454 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4455 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4457 obj->agp_type = AGP_USER_MEMORY;
4458 obj->base.driver_private = NULL;
4459 obj->fence_reg = I915_FENCE_REG_NONE;
4460 INIT_LIST_HEAD(&obj->list);
4461 INIT_LIST_HEAD(&obj->gpu_write_list);
4462 obj->madv = I915_MADV_WILLNEED;
4464 trace_i915_gem_object_create(&obj->base);
4469 int i915_gem_init_object(struct drm_gem_object *obj)
4476 static void i915_gem_free_object_tail(struct drm_gem_object *obj)
4478 struct drm_device *dev = obj->dev;
4479 drm_i915_private_t *dev_priv = dev->dev_private;
4480 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
4483 ret = i915_gem_object_unbind(obj);
4484 if (ret == -ERESTARTSYS) {
4485 list_move(&obj_priv->list,
4486 &dev_priv->mm.deferred_free_list);
4490 if (obj_priv->mmap_offset)
4491 i915_gem_free_mmap_offset(obj);
4493 drm_gem_object_release(obj);
4495 kfree(obj_priv->page_cpu_valid);
4496 kfree(obj_priv->bit_17);
4500 void i915_gem_free_object(struct drm_gem_object *obj)
4502 struct drm_device *dev = obj->dev;
4503 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
4505 trace_i915_gem_object_destroy(obj);
4507 while (obj_priv->pin_count > 0)
4508 i915_gem_object_unpin(obj);
4510 if (obj_priv->phys_obj)
4511 i915_gem_detach_phys_object(dev, obj);
4513 i915_gem_free_object_tail(obj);
4516 /** Unbinds all inactive objects. */
4518 i915_gem_evict_from_inactive_list(struct drm_device *dev)
4520 drm_i915_private_t *dev_priv = dev->dev_private;
4522 while (!list_empty(&dev_priv->mm.inactive_list)) {
4523 struct drm_gem_object *obj;
4526 obj = &list_first_entry(&dev_priv->mm.inactive_list,
4527 struct drm_i915_gem_object,
4530 ret = i915_gem_object_unbind(obj);
4532 DRM_ERROR("Error unbinding object: %d\n", ret);
4541 i915_gem_idle(struct drm_device *dev)
4543 drm_i915_private_t *dev_priv = dev->dev_private;
4546 mutex_lock(&dev->struct_mutex);
4548 if (dev_priv->mm.suspended ||
4549 (dev_priv->render_ring.gem_object == NULL) ||
4551 dev_priv->bsd_ring.gem_object == NULL)) {
4552 mutex_unlock(&dev->struct_mutex);
4556 ret = i915_gpu_idle(dev);
4558 mutex_unlock(&dev->struct_mutex);
4562 /* Under UMS, be paranoid and evict. */
4563 if (!drm_core_check_feature(dev, DRIVER_MODESET)) {
4564 ret = i915_gem_evict_from_inactive_list(dev);
4566 mutex_unlock(&dev->struct_mutex);
4571 /* Hack! Don't let anybody do execbuf while we don't control the chip.
4572 * We need to replace this with a semaphore, or something.
4573 * And not confound mm.suspended!
4575 dev_priv->mm.suspended = 1;
4576 del_timer(&dev_priv->hangcheck_timer);
4578 i915_kernel_lost_context(dev);
4579 i915_gem_cleanup_ringbuffer(dev);
4581 mutex_unlock(&dev->struct_mutex);
4583 /* Cancel the retire work handler, which should be idle now. */
4584 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
4590 * 965+ support PIPE_CONTROL commands, which provide finer grained control
4591 * over cache flushing.
4594 i915_gem_init_pipe_control(struct drm_device *dev)
4596 drm_i915_private_t *dev_priv = dev->dev_private;
4597 struct drm_gem_object *obj;
4598 struct drm_i915_gem_object *obj_priv;
4601 obj = i915_gem_alloc_object(dev, 4096);
4603 DRM_ERROR("Failed to allocate seqno page\n");
4607 obj_priv = to_intel_bo(obj);
4608 obj_priv->agp_type = AGP_USER_CACHED_MEMORY;
4610 ret = i915_gem_object_pin(obj, 4096);
4614 dev_priv->seqno_gfx_addr = obj_priv->gtt_offset;
4615 dev_priv->seqno_page = kmap(obj_priv->pages[0]);
4616 if (dev_priv->seqno_page == NULL)
4619 dev_priv->seqno_obj = obj;
4620 memset(dev_priv->seqno_page, 0, PAGE_SIZE);
4625 i915_gem_object_unpin(obj);
4627 drm_gem_object_unreference(obj);
4634 i915_gem_cleanup_pipe_control(struct drm_device *dev)
4636 drm_i915_private_t *dev_priv = dev->dev_private;
4637 struct drm_gem_object *obj;
4638 struct drm_i915_gem_object *obj_priv;
4640 obj = dev_priv->seqno_obj;
4641 obj_priv = to_intel_bo(obj);
4642 kunmap(obj_priv->pages[0]);
4643 i915_gem_object_unpin(obj);
4644 drm_gem_object_unreference(obj);
4645 dev_priv->seqno_obj = NULL;
4647 dev_priv->seqno_page = NULL;
4651 i915_gem_init_ringbuffer(struct drm_device *dev)
4653 drm_i915_private_t *dev_priv = dev->dev_private;
4656 dev_priv->render_ring = render_ring;
4658 if (!I915_NEED_GFX_HWS(dev)) {
4659 dev_priv->render_ring.status_page.page_addr
4660 = dev_priv->status_page_dmah->vaddr;
4661 memset(dev_priv->render_ring.status_page.page_addr,
4665 if (HAS_PIPE_CONTROL(dev)) {
4666 ret = i915_gem_init_pipe_control(dev);
4671 ret = intel_init_ring_buffer(dev, &dev_priv->render_ring);
4673 goto cleanup_pipe_control;
4676 dev_priv->bsd_ring = bsd_ring;
4677 ret = intel_init_ring_buffer(dev, &dev_priv->bsd_ring);
4679 goto cleanup_render_ring;
4684 cleanup_render_ring:
4685 intel_cleanup_ring_buffer(dev, &dev_priv->render_ring);
4686 cleanup_pipe_control:
4687 if (HAS_PIPE_CONTROL(dev))
4688 i915_gem_cleanup_pipe_control(dev);
4693 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
4695 drm_i915_private_t *dev_priv = dev->dev_private;
4697 intel_cleanup_ring_buffer(dev, &dev_priv->render_ring);
4699 intel_cleanup_ring_buffer(dev, &dev_priv->bsd_ring);
4700 if (HAS_PIPE_CONTROL(dev))
4701 i915_gem_cleanup_pipe_control(dev);
4705 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
4706 struct drm_file *file_priv)
4708 drm_i915_private_t *dev_priv = dev->dev_private;
4711 if (drm_core_check_feature(dev, DRIVER_MODESET))
4714 if (atomic_read(&dev_priv->mm.wedged)) {
4715 DRM_ERROR("Reenabling wedged hardware, good luck\n");
4716 atomic_set(&dev_priv->mm.wedged, 0);
4719 mutex_lock(&dev->struct_mutex);
4720 dev_priv->mm.suspended = 0;
4722 ret = i915_gem_init_ringbuffer(dev);
4724 mutex_unlock(&dev->struct_mutex);
4728 spin_lock(&dev_priv->mm.active_list_lock);
4729 BUG_ON(!list_empty(&dev_priv->render_ring.active_list));
4730 BUG_ON(HAS_BSD(dev) && !list_empty(&dev_priv->bsd_ring.active_list));
4731 spin_unlock(&dev_priv->mm.active_list_lock);
4733 BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
4734 BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
4735 BUG_ON(!list_empty(&dev_priv->render_ring.request_list));
4736 BUG_ON(HAS_BSD(dev) && !list_empty(&dev_priv->bsd_ring.request_list));
4737 mutex_unlock(&dev->struct_mutex);
4739 ret = drm_irq_install(dev);
4741 goto cleanup_ringbuffer;
4746 mutex_lock(&dev->struct_mutex);
4747 i915_gem_cleanup_ringbuffer(dev);
4748 dev_priv->mm.suspended = 1;
4749 mutex_unlock(&dev->struct_mutex);
4755 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
4756 struct drm_file *file_priv)
4758 if (drm_core_check_feature(dev, DRIVER_MODESET))
4761 drm_irq_uninstall(dev);
4762 return i915_gem_idle(dev);
4766 i915_gem_lastclose(struct drm_device *dev)
4770 if (drm_core_check_feature(dev, DRIVER_MODESET))
4773 ret = i915_gem_idle(dev);
4775 DRM_ERROR("failed to idle hardware: %d\n", ret);
4779 i915_gem_load(struct drm_device *dev)
4782 drm_i915_private_t *dev_priv = dev->dev_private;
4784 spin_lock_init(&dev_priv->mm.active_list_lock);
4785 INIT_LIST_HEAD(&dev_priv->mm.flushing_list);
4786 INIT_LIST_HEAD(&dev_priv->mm.gpu_write_list);
4787 INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
4788 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4789 INIT_LIST_HEAD(&dev_priv->mm.deferred_free_list);
4790 INIT_LIST_HEAD(&dev_priv->render_ring.active_list);
4791 INIT_LIST_HEAD(&dev_priv->render_ring.request_list);
4793 INIT_LIST_HEAD(&dev_priv->bsd_ring.active_list);
4794 INIT_LIST_HEAD(&dev_priv->bsd_ring.request_list);
4796 for (i = 0; i < 16; i++)
4797 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
4798 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
4799 i915_gem_retire_work_handler);
4800 spin_lock(&shrink_list_lock);
4801 list_add(&dev_priv->mm.shrink_list, &shrink_list);
4802 spin_unlock(&shrink_list_lock);
4804 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
4806 u32 tmp = I915_READ(MI_ARB_STATE);
4807 if (!(tmp & MI_ARB_C3_LP_WRITE_ENABLE)) {
4808 /* arb state is a masked write, so set bit + bit in mask */
4809 tmp = MI_ARB_C3_LP_WRITE_ENABLE | (MI_ARB_C3_LP_WRITE_ENABLE << MI_ARB_MASK_SHIFT);
4810 I915_WRITE(MI_ARB_STATE, tmp);
4814 /* Old X drivers will take 0-2 for front, back, depth buffers */
4815 if (!drm_core_check_feature(dev, DRIVER_MODESET))
4816 dev_priv->fence_reg_start = 3;
4818 if (IS_I965G(dev) || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
4819 dev_priv->num_fence_regs = 16;
4821 dev_priv->num_fence_regs = 8;
4823 /* Initialize fence registers to zero */
4824 if (IS_I965G(dev)) {
4825 for (i = 0; i < 16; i++)
4826 I915_WRITE64(FENCE_REG_965_0 + (i * 8), 0);
4828 for (i = 0; i < 8; i++)
4829 I915_WRITE(FENCE_REG_830_0 + (i * 4), 0);
4830 if (IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
4831 for (i = 0; i < 8; i++)
4832 I915_WRITE(FENCE_REG_945_8 + (i * 4), 0);
4834 i915_gem_detect_bit_6_swizzle(dev);
4835 init_waitqueue_head(&dev_priv->pending_flip_queue);
4839 * Create a physically contiguous memory object for this object
4840 * e.g. for cursor + overlay regs
4842 int i915_gem_init_phys_object(struct drm_device *dev,
4845 drm_i915_private_t *dev_priv = dev->dev_private;
4846 struct drm_i915_gem_phys_object *phys_obj;
4849 if (dev_priv->mm.phys_objs[id - 1] || !size)
4852 phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL);
4858 phys_obj->handle = drm_pci_alloc(dev, size, 0);
4859 if (!phys_obj->handle) {
4864 set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4867 dev_priv->mm.phys_objs[id - 1] = phys_obj;
4875 void i915_gem_free_phys_object(struct drm_device *dev, int id)
4877 drm_i915_private_t *dev_priv = dev->dev_private;
4878 struct drm_i915_gem_phys_object *phys_obj;
4880 if (!dev_priv->mm.phys_objs[id - 1])
4883 phys_obj = dev_priv->mm.phys_objs[id - 1];
4884 if (phys_obj->cur_obj) {
4885 i915_gem_detach_phys_object(dev, phys_obj->cur_obj);
4889 set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4891 drm_pci_free(dev, phys_obj->handle);
4893 dev_priv->mm.phys_objs[id - 1] = NULL;
4896 void i915_gem_free_all_phys_object(struct drm_device *dev)
4900 for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++)
4901 i915_gem_free_phys_object(dev, i);
4904 void i915_gem_detach_phys_object(struct drm_device *dev,
4905 struct drm_gem_object *obj)
4907 struct drm_i915_gem_object *obj_priv;
4912 obj_priv = to_intel_bo(obj);
4913 if (!obj_priv->phys_obj)
4916 ret = i915_gem_object_get_pages(obj, 0);
4920 page_count = obj->size / PAGE_SIZE;
4922 for (i = 0; i < page_count; i++) {
4923 char *dst = kmap_atomic(obj_priv->pages[i], KM_USER0);
4924 char *src = obj_priv->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4926 memcpy(dst, src, PAGE_SIZE);
4927 kunmap_atomic(dst, KM_USER0);
4929 drm_clflush_pages(obj_priv->pages, page_count);
4930 drm_agp_chipset_flush(dev);
4932 i915_gem_object_put_pages(obj);
4934 obj_priv->phys_obj->cur_obj = NULL;
4935 obj_priv->phys_obj = NULL;
4939 i915_gem_attach_phys_object(struct drm_device *dev,
4940 struct drm_gem_object *obj, int id)
4942 drm_i915_private_t *dev_priv = dev->dev_private;
4943 struct drm_i915_gem_object *obj_priv;
4948 if (id > I915_MAX_PHYS_OBJECT)
4951 obj_priv = to_intel_bo(obj);
4953 if (obj_priv->phys_obj) {
4954 if (obj_priv->phys_obj->id == id)
4956 i915_gem_detach_phys_object(dev, obj);
4960 /* create a new object */
4961 if (!dev_priv->mm.phys_objs[id - 1]) {
4962 ret = i915_gem_init_phys_object(dev, id,
4965 DRM_ERROR("failed to init phys object %d size: %zu\n", id, obj->size);
4970 /* bind to the object */
4971 obj_priv->phys_obj = dev_priv->mm.phys_objs[id - 1];
4972 obj_priv->phys_obj->cur_obj = obj;
4974 ret = i915_gem_object_get_pages(obj, 0);
4976 DRM_ERROR("failed to get page list\n");
4980 page_count = obj->size / PAGE_SIZE;
4982 for (i = 0; i < page_count; i++) {
4983 char *src = kmap_atomic(obj_priv->pages[i], KM_USER0);
4984 char *dst = obj_priv->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4986 memcpy(dst, src, PAGE_SIZE);
4987 kunmap_atomic(src, KM_USER0);
4990 i915_gem_object_put_pages(obj);
4998 i915_gem_phys_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
4999 struct drm_i915_gem_pwrite *args,
5000 struct drm_file *file_priv)
5002 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
5005 char __user *user_data;
5007 user_data = (char __user *) (uintptr_t) args->data_ptr;
5008 obj_addr = obj_priv->phys_obj->handle->vaddr + args->offset;
5010 DRM_DEBUG_DRIVER("obj_addr %p, %lld\n", obj_addr, args->size);
5011 ret = copy_from_user(obj_addr, user_data, args->size);
5015 drm_agp_chipset_flush(dev);
5019 void i915_gem_release(struct drm_device * dev, struct drm_file *file_priv)
5021 struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv;
5023 /* Clean up our request list when the client is going away, so that
5024 * later retire_requests won't dereference our soon-to-be-gone
5027 mutex_lock(&dev->struct_mutex);
5028 while (!list_empty(&i915_file_priv->mm.request_list))
5029 list_del_init(i915_file_priv->mm.request_list.next);
5030 mutex_unlock(&dev->struct_mutex);
5034 i915_gpu_is_active(struct drm_device *dev)
5036 drm_i915_private_t *dev_priv = dev->dev_private;
5039 spin_lock(&dev_priv->mm.active_list_lock);
5040 lists_empty = list_empty(&dev_priv->mm.flushing_list) &&
5041 list_empty(&dev_priv->render_ring.active_list);
5043 lists_empty &= list_empty(&dev_priv->bsd_ring.active_list);
5044 spin_unlock(&dev_priv->mm.active_list_lock);
5046 return !lists_empty;
5050 i915_gem_shrink(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask)
5052 drm_i915_private_t *dev_priv, *next_dev;
5053 struct drm_i915_gem_object *obj_priv, *next_obj;
5055 int would_deadlock = 1;
5057 /* "fast-path" to count number of available objects */
5058 if (nr_to_scan == 0) {
5059 spin_lock(&shrink_list_lock);
5060 list_for_each_entry(dev_priv, &shrink_list, mm.shrink_list) {
5061 struct drm_device *dev = dev_priv->dev;
5063 if (mutex_trylock(&dev->struct_mutex)) {
5064 list_for_each_entry(obj_priv,
5065 &dev_priv->mm.inactive_list,
5068 mutex_unlock(&dev->struct_mutex);
5071 spin_unlock(&shrink_list_lock);
5073 return (cnt / 100) * sysctl_vfs_cache_pressure;
5076 spin_lock(&shrink_list_lock);
5079 /* first scan for clean buffers */
5080 list_for_each_entry_safe(dev_priv, next_dev,
5081 &shrink_list, mm.shrink_list) {
5082 struct drm_device *dev = dev_priv->dev;
5084 if (! mutex_trylock(&dev->struct_mutex))
5087 spin_unlock(&shrink_list_lock);
5088 i915_gem_retire_requests(dev);
5090 list_for_each_entry_safe(obj_priv, next_obj,
5091 &dev_priv->mm.inactive_list,
5093 if (i915_gem_object_is_purgeable(obj_priv)) {
5094 i915_gem_object_unbind(&obj_priv->base);
5095 if (--nr_to_scan <= 0)
5100 spin_lock(&shrink_list_lock);
5101 mutex_unlock(&dev->struct_mutex);
5105 if (nr_to_scan <= 0)
5109 /* second pass, evict/count anything still on the inactive list */
5110 list_for_each_entry_safe(dev_priv, next_dev,
5111 &shrink_list, mm.shrink_list) {
5112 struct drm_device *dev = dev_priv->dev;
5114 if (! mutex_trylock(&dev->struct_mutex))
5117 spin_unlock(&shrink_list_lock);
5119 list_for_each_entry_safe(obj_priv, next_obj,
5120 &dev_priv->mm.inactive_list,
5122 if (nr_to_scan > 0) {
5123 i915_gem_object_unbind(&obj_priv->base);
5129 spin_lock(&shrink_list_lock);
5130 mutex_unlock(&dev->struct_mutex);
5139 * We are desperate for pages, so as a last resort, wait
5140 * for the GPU to finish and discard whatever we can.
5141 * This has a dramatic impact to reduce the number of
5142 * OOM-killer events whilst running the GPU aggressively.
5144 list_for_each_entry(dev_priv, &shrink_list, mm.shrink_list) {
5145 struct drm_device *dev = dev_priv->dev;
5147 if (!mutex_trylock(&dev->struct_mutex))
5150 spin_unlock(&shrink_list_lock);
5152 if (i915_gpu_is_active(dev)) {
5157 spin_lock(&shrink_list_lock);
5158 mutex_unlock(&dev->struct_mutex);
5165 spin_unlock(&shrink_list_lock);
5170 return (cnt / 100) * sysctl_vfs_cache_pressure;
5175 static struct shrinker shrinker = {
5176 .shrink = i915_gem_shrink,
5177 .seeks = DEFAULT_SEEKS,
5181 i915_gem_shrinker_init(void)
5183 register_shrinker(&shrinker);
5187 i915_gem_shrinker_exit(void)
5189 unregister_shrinker(&shrinker);