2 * Copyright © 2008 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eric Anholt <eric@anholt.net>
32 #include "i915_trace.h"
33 #include "intel_drv.h"
34 #include <linux/shmem_fs.h>
35 #include <linux/slab.h>
36 #include <linux/swap.h>
37 #include <linux/pci.h>
39 static __must_check int i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object *obj);
40 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);
41 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj);
42 static __must_check int i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj,
44 static __must_check int i915_gem_object_set_cpu_read_domain_range(struct drm_i915_gem_object *obj,
47 static void i915_gem_object_set_to_full_cpu_read_domain(struct drm_i915_gem_object *obj);
48 static __must_check int i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
50 bool map_and_fenceable);
51 static void i915_gem_clear_fence_reg(struct drm_device *dev,
52 struct drm_i915_fence_reg *reg);
53 static int i915_gem_phys_pwrite(struct drm_device *dev,
54 struct drm_i915_gem_object *obj,
55 struct drm_i915_gem_pwrite *args,
56 struct drm_file *file);
57 static void i915_gem_free_object_tail(struct drm_i915_gem_object *obj);
59 static int i915_gem_inactive_shrink(struct shrinker *shrinker,
60 struct shrink_control *sc);
61 static void i915_gem_object_truncate(struct drm_i915_gem_object *obj);
63 /* some bookkeeping */
64 static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
67 dev_priv->mm.object_count++;
68 dev_priv->mm.object_memory += size;
71 static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
74 dev_priv->mm.object_count--;
75 dev_priv->mm.object_memory -= size;
79 i915_gem_wait_for_error(struct drm_device *dev)
81 struct drm_i915_private *dev_priv = dev->dev_private;
82 struct completion *x = &dev_priv->error_completion;
86 if (!atomic_read(&dev_priv->mm.wedged))
89 ret = wait_for_completion_interruptible(x);
93 if (atomic_read(&dev_priv->mm.wedged)) {
94 /* GPU is hung, bump the completion count to account for
95 * the token we just consumed so that we never hit zero and
96 * end up waiting upon a subsequent completion event that
99 spin_lock_irqsave(&x->wait.lock, flags);
101 spin_unlock_irqrestore(&x->wait.lock, flags);
106 int i915_mutex_lock_interruptible(struct drm_device *dev)
110 ret = i915_gem_wait_for_error(dev);
114 ret = mutex_lock_interruptible(&dev->struct_mutex);
118 WARN_ON(i915_verify_lists(dev));
123 i915_gem_object_is_inactive(struct drm_i915_gem_object *obj)
125 return obj->gtt_space && !obj->active && obj->pin_count == 0;
128 void i915_gem_do_init(struct drm_device *dev,
130 unsigned long mappable_end,
133 drm_i915_private_t *dev_priv = dev->dev_private;
135 drm_mm_init(&dev_priv->mm.gtt_space, start, end - start);
137 dev_priv->mm.gtt_start = start;
138 dev_priv->mm.gtt_mappable_end = mappable_end;
139 dev_priv->mm.gtt_end = end;
140 dev_priv->mm.gtt_total = end - start;
141 dev_priv->mm.mappable_gtt_total = min(end, mappable_end) - start;
143 /* Take over this portion of the GTT */
144 intel_gtt_clear_range(start / PAGE_SIZE, (end-start) / PAGE_SIZE);
148 i915_gem_init_ioctl(struct drm_device *dev, void *data,
149 struct drm_file *file)
151 struct drm_i915_gem_init *args = data;
153 if (args->gtt_start >= args->gtt_end ||
154 (args->gtt_end | args->gtt_start) & (PAGE_SIZE - 1))
157 mutex_lock(&dev->struct_mutex);
158 i915_gem_do_init(dev, args->gtt_start, args->gtt_end, args->gtt_end);
159 mutex_unlock(&dev->struct_mutex);
165 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
166 struct drm_file *file)
168 struct drm_i915_private *dev_priv = dev->dev_private;
169 struct drm_i915_gem_get_aperture *args = data;
170 struct drm_i915_gem_object *obj;
173 if (!(dev->driver->driver_features & DRIVER_GEM))
177 mutex_lock(&dev->struct_mutex);
178 list_for_each_entry(obj, &dev_priv->mm.pinned_list, mm_list)
179 pinned += obj->gtt_space->size;
180 mutex_unlock(&dev->struct_mutex);
182 args->aper_size = dev_priv->mm.gtt_total;
183 args->aper_available_size = args->aper_size - pinned;
189 i915_gem_create(struct drm_file *file,
190 struct drm_device *dev,
194 struct drm_i915_gem_object *obj;
198 size = roundup(size, PAGE_SIZE);
202 /* Allocate the new object */
203 obj = i915_gem_alloc_object(dev, size);
207 ret = drm_gem_handle_create(file, &obj->base, &handle);
209 drm_gem_object_release(&obj->base);
210 i915_gem_info_remove_obj(dev->dev_private, obj->base.size);
215 /* drop reference from allocate - handle holds it now */
216 drm_gem_object_unreference(&obj->base);
217 trace_i915_gem_object_create(obj);
224 i915_gem_dumb_create(struct drm_file *file,
225 struct drm_device *dev,
226 struct drm_mode_create_dumb *args)
228 /* have to work out size/pitch and return them */
229 args->pitch = ALIGN(args->width * ((args->bpp + 7) / 8), 64);
230 args->size = args->pitch * args->height;
231 return i915_gem_create(file, dev,
232 args->size, &args->handle);
235 int i915_gem_dumb_destroy(struct drm_file *file,
236 struct drm_device *dev,
239 return drm_gem_handle_delete(file, handle);
243 * Creates a new mm object and returns a handle to it.
246 i915_gem_create_ioctl(struct drm_device *dev, void *data,
247 struct drm_file *file)
249 struct drm_i915_gem_create *args = data;
250 return i915_gem_create(file, dev,
251 args->size, &args->handle);
254 static int i915_gem_object_needs_bit17_swizzle(struct drm_i915_gem_object *obj)
256 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
258 return dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_9_10_17 &&
259 obj->tiling_mode != I915_TILING_NONE;
263 * This is the fast shmem pread path, which attempts to copy_from_user directly
264 * from the backing pages of the object to the user's address space. On a
265 * fault, it fails so we can fall back to i915_gem_shmem_pwrite_slow().
268 i915_gem_shmem_pread_fast(struct drm_device *dev,
269 struct drm_i915_gem_object *obj,
270 struct drm_i915_gem_pread *args,
271 struct drm_file *file)
273 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
276 char __user *user_data;
277 int page_offset, page_length;
279 user_data = (char __user *) (uintptr_t) args->data_ptr;
282 offset = args->offset;
289 /* Operation in this page
291 * page_offset = offset within page
292 * page_length = bytes to copy for this page
294 page_offset = offset_in_page(offset);
295 page_length = remain;
296 if ((page_offset + remain) > PAGE_SIZE)
297 page_length = PAGE_SIZE - page_offset;
299 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
301 return PTR_ERR(page);
303 vaddr = kmap_atomic(page);
304 ret = __copy_to_user_inatomic(user_data,
307 kunmap_atomic(vaddr);
309 mark_page_accessed(page);
310 page_cache_release(page);
314 remain -= page_length;
315 user_data += page_length;
316 offset += page_length;
323 __copy_to_user_swizzled(char __user *cpu_vaddr,
324 const char *gpu_vaddr, int gpu_offset,
327 int ret, cpu_offset = 0;
330 int cacheline_end = ALIGN(gpu_offset + 1, 64);
331 int this_length = min(cacheline_end - gpu_offset, length);
332 int swizzled_gpu_offset = gpu_offset ^ 64;
334 ret = __copy_to_user(cpu_vaddr + cpu_offset,
335 gpu_vaddr + swizzled_gpu_offset,
340 cpu_offset += this_length;
341 gpu_offset += this_length;
342 length -= this_length;
349 __copy_from_user_swizzled(char __user *gpu_vaddr, int gpu_offset,
350 const char *cpu_vaddr,
353 int ret, cpu_offset = 0;
356 int cacheline_end = ALIGN(gpu_offset + 1, 64);
357 int this_length = min(cacheline_end - gpu_offset, length);
358 int swizzled_gpu_offset = gpu_offset ^ 64;
360 ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset,
361 cpu_vaddr + cpu_offset,
366 cpu_offset += this_length;
367 gpu_offset += this_length;
368 length -= this_length;
375 * This is the fallback shmem pread path, which allocates temporary storage
376 * in kernel space to copy_to_user into outside of the struct_mutex, so we
377 * can copy out of the object's backing pages while holding the struct mutex
378 * and not take page faults.
381 i915_gem_shmem_pread_slow(struct drm_device *dev,
382 struct drm_i915_gem_object *obj,
383 struct drm_i915_gem_pread *args,
384 struct drm_file *file)
386 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
387 char __user *user_data;
390 int shmem_page_offset, page_length, ret;
391 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
393 user_data = (char __user *) (uintptr_t) args->data_ptr;
396 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
398 offset = args->offset;
400 mutex_unlock(&dev->struct_mutex);
406 /* Operation in this page
408 * shmem_page_offset = offset within page in shmem file
409 * page_length = bytes to copy for this page
411 shmem_page_offset = offset_in_page(offset);
412 page_length = remain;
413 if ((shmem_page_offset + page_length) > PAGE_SIZE)
414 page_length = PAGE_SIZE - shmem_page_offset;
416 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
422 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
423 (page_to_phys(page) & (1 << 17)) != 0;
426 if (page_do_bit17_swizzling)
427 ret = __copy_to_user_swizzled(user_data,
428 vaddr, shmem_page_offset,
431 ret = __copy_to_user(user_data,
432 vaddr + shmem_page_offset,
436 mark_page_accessed(page);
437 page_cache_release(page);
444 remain -= page_length;
445 user_data += page_length;
446 offset += page_length;
450 mutex_lock(&dev->struct_mutex);
451 /* Fixup: Kill any reinstated backing storage pages */
452 if (obj->madv == __I915_MADV_PURGED)
453 i915_gem_object_truncate(obj);
459 * Reads data from the object referenced by handle.
461 * On error, the contents of *data are undefined.
464 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
465 struct drm_file *file)
467 struct drm_i915_gem_pread *args = data;
468 struct drm_i915_gem_object *obj;
474 if (!access_ok(VERIFY_WRITE,
475 (char __user *)(uintptr_t)args->data_ptr,
479 ret = fault_in_pages_writeable((char __user *)(uintptr_t)args->data_ptr,
484 ret = i915_mutex_lock_interruptible(dev);
488 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
489 if (&obj->base == NULL) {
494 /* Bounds check source. */
495 if (args->offset > obj->base.size ||
496 args->size > obj->base.size - args->offset) {
501 trace_i915_gem_object_pread(obj, args->offset, args->size);
503 ret = i915_gem_object_set_cpu_read_domain_range(obj,
510 if (!i915_gem_object_needs_bit17_swizzle(obj))
511 ret = i915_gem_shmem_pread_fast(dev, obj, args, file);
513 ret = i915_gem_shmem_pread_slow(dev, obj, args, file);
516 drm_gem_object_unreference(&obj->base);
518 mutex_unlock(&dev->struct_mutex);
522 /* This is the fast write path which cannot handle
523 * page faults in the source data
527 fast_user_write(struct io_mapping *mapping,
528 loff_t page_base, int page_offset,
529 char __user *user_data,
533 unsigned long unwritten;
535 vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
536 unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + page_offset,
538 io_mapping_unmap_atomic(vaddr_atomic);
542 /* Here's the write path which can sleep for
547 slow_kernel_write(struct io_mapping *mapping,
548 loff_t gtt_base, int gtt_offset,
549 struct page *user_page, int user_offset,
552 char __iomem *dst_vaddr;
555 dst_vaddr = io_mapping_map_wc(mapping, gtt_base);
556 src_vaddr = kmap(user_page);
558 memcpy_toio(dst_vaddr + gtt_offset,
559 src_vaddr + user_offset,
563 io_mapping_unmap(dst_vaddr);
567 * This is the fast pwrite path, where we copy the data directly from the
568 * user into the GTT, uncached.
571 i915_gem_gtt_pwrite_fast(struct drm_device *dev,
572 struct drm_i915_gem_object *obj,
573 struct drm_i915_gem_pwrite *args,
574 struct drm_file *file)
576 drm_i915_private_t *dev_priv = dev->dev_private;
578 loff_t offset, page_base;
579 char __user *user_data;
580 int page_offset, page_length;
582 user_data = (char __user *) (uintptr_t) args->data_ptr;
585 offset = obj->gtt_offset + args->offset;
588 /* Operation in this page
590 * page_base = page offset within aperture
591 * page_offset = offset within page
592 * page_length = bytes to copy for this page
594 page_base = offset & PAGE_MASK;
595 page_offset = offset_in_page(offset);
596 page_length = remain;
597 if ((page_offset + remain) > PAGE_SIZE)
598 page_length = PAGE_SIZE - page_offset;
600 /* If we get a fault while copying data, then (presumably) our
601 * source page isn't available. Return the error and we'll
602 * retry in the slow path.
604 if (fast_user_write(dev_priv->mm.gtt_mapping, page_base,
605 page_offset, user_data, page_length))
608 remain -= page_length;
609 user_data += page_length;
610 offset += page_length;
617 * This is the fallback GTT pwrite path, which uses get_user_pages to pin
618 * the memory and maps it using kmap_atomic for copying.
620 * This code resulted in x11perf -rgb10text consuming about 10% more CPU
621 * than using i915_gem_gtt_pwrite_fast on a G45 (32-bit).
624 i915_gem_gtt_pwrite_slow(struct drm_device *dev,
625 struct drm_i915_gem_object *obj,
626 struct drm_i915_gem_pwrite *args,
627 struct drm_file *file)
629 drm_i915_private_t *dev_priv = dev->dev_private;
631 loff_t gtt_page_base, offset;
632 loff_t first_data_page, last_data_page, num_pages;
633 loff_t pinned_pages, i;
634 struct page **user_pages;
635 struct mm_struct *mm = current->mm;
636 int gtt_page_offset, data_page_offset, data_page_index, page_length;
638 uint64_t data_ptr = args->data_ptr;
642 /* Pin the user pages containing the data. We can't fault while
643 * holding the struct mutex, and all of the pwrite implementations
644 * want to hold it while dereferencing the user data.
646 first_data_page = data_ptr / PAGE_SIZE;
647 last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
648 num_pages = last_data_page - first_data_page + 1;
650 user_pages = drm_malloc_ab(num_pages, sizeof(struct page *));
651 if (user_pages == NULL)
654 mutex_unlock(&dev->struct_mutex);
655 down_read(&mm->mmap_sem);
656 pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
657 num_pages, 0, 0, user_pages, NULL);
658 up_read(&mm->mmap_sem);
659 mutex_lock(&dev->struct_mutex);
660 if (pinned_pages < num_pages) {
662 goto out_unpin_pages;
665 ret = i915_gem_object_set_to_gtt_domain(obj, true);
667 goto out_unpin_pages;
669 ret = i915_gem_object_put_fence(obj);
671 goto out_unpin_pages;
673 offset = obj->gtt_offset + args->offset;
676 /* Operation in this page
678 * gtt_page_base = page offset within aperture
679 * gtt_page_offset = offset within page in aperture
680 * data_page_index = page number in get_user_pages return
681 * data_page_offset = offset with data_page_index page.
682 * page_length = bytes to copy for this page
684 gtt_page_base = offset & PAGE_MASK;
685 gtt_page_offset = offset_in_page(offset);
686 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
687 data_page_offset = offset_in_page(data_ptr);
689 page_length = remain;
690 if ((gtt_page_offset + page_length) > PAGE_SIZE)
691 page_length = PAGE_SIZE - gtt_page_offset;
692 if ((data_page_offset + page_length) > PAGE_SIZE)
693 page_length = PAGE_SIZE - data_page_offset;
695 slow_kernel_write(dev_priv->mm.gtt_mapping,
696 gtt_page_base, gtt_page_offset,
697 user_pages[data_page_index],
701 remain -= page_length;
702 offset += page_length;
703 data_ptr += page_length;
707 for (i = 0; i < pinned_pages; i++)
708 page_cache_release(user_pages[i]);
709 drm_free_large(user_pages);
715 * This is the fast shmem pwrite path, which attempts to directly
716 * copy_from_user into the kmapped pages backing the object.
719 i915_gem_shmem_pwrite_fast(struct drm_device *dev,
720 struct drm_i915_gem_object *obj,
721 struct drm_i915_gem_pwrite *args,
722 struct drm_file *file)
724 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
727 char __user *user_data;
728 int page_offset, page_length;
730 user_data = (char __user *) (uintptr_t) args->data_ptr;
733 offset = args->offset;
741 /* Operation in this page
743 * page_offset = offset within page
744 * page_length = bytes to copy for this page
746 page_offset = offset_in_page(offset);
747 page_length = remain;
748 if ((page_offset + remain) > PAGE_SIZE)
749 page_length = PAGE_SIZE - page_offset;
751 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
753 return PTR_ERR(page);
755 vaddr = kmap_atomic(page);
756 ret = __copy_from_user_inatomic(vaddr + page_offset,
759 kunmap_atomic(vaddr);
761 set_page_dirty(page);
762 mark_page_accessed(page);
763 page_cache_release(page);
765 /* If we get a fault while copying data, then (presumably) our
766 * source page isn't available. Return the error and we'll
767 * retry in the slow path.
772 remain -= page_length;
773 user_data += page_length;
774 offset += page_length;
781 * This is the fallback shmem pwrite path, which uses get_user_pages to pin
782 * the memory and maps it using kmap_atomic for copying.
784 * This avoids taking mmap_sem for faulting on the user's address while the
785 * struct_mutex is held.
788 i915_gem_shmem_pwrite_slow(struct drm_device *dev,
789 struct drm_i915_gem_object *obj,
790 struct drm_i915_gem_pwrite *args,
791 struct drm_file *file)
793 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
796 char __user *user_data;
797 int shmem_page_offset, page_length, ret;
798 int obj_do_bit17_swizzling, page_do_bit17_swizzling;
800 user_data = (char __user *) (uintptr_t) args->data_ptr;
803 obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
805 offset = args->offset;
808 mutex_unlock(&dev->struct_mutex);
814 /* Operation in this page
816 * shmem_page_offset = offset within page in shmem file
817 * page_length = bytes to copy for this page
819 shmem_page_offset = offset_in_page(offset);
821 page_length = remain;
822 if ((shmem_page_offset + page_length) > PAGE_SIZE)
823 page_length = PAGE_SIZE - shmem_page_offset;
825 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
831 page_do_bit17_swizzling = obj_do_bit17_swizzling &&
832 (page_to_phys(page) & (1 << 17)) != 0;
835 if (page_do_bit17_swizzling)
836 ret = __copy_from_user_swizzled(vaddr, shmem_page_offset,
840 ret = __copy_from_user(vaddr + shmem_page_offset,
845 set_page_dirty(page);
846 mark_page_accessed(page);
847 page_cache_release(page);
854 remain -= page_length;
855 user_data += page_length;
856 offset += page_length;
860 mutex_lock(&dev->struct_mutex);
861 /* Fixup: Kill any reinstated backing storage pages */
862 if (obj->madv == __I915_MADV_PURGED)
863 i915_gem_object_truncate(obj);
864 /* and flush dirty cachelines in case the object isn't in the cpu write
866 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
867 i915_gem_clflush_object(obj);
868 intel_gtt_chipset_flush();
875 * Writes data to the object referenced by handle.
877 * On error, the contents of the buffer that were to be modified are undefined.
880 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
881 struct drm_file *file)
883 struct drm_i915_gem_pwrite *args = data;
884 struct drm_i915_gem_object *obj;
890 if (!access_ok(VERIFY_READ,
891 (char __user *)(uintptr_t)args->data_ptr,
895 ret = fault_in_pages_readable((char __user *)(uintptr_t)args->data_ptr,
900 ret = i915_mutex_lock_interruptible(dev);
904 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
905 if (&obj->base == NULL) {
910 /* Bounds check destination. */
911 if (args->offset > obj->base.size ||
912 args->size > obj->base.size - args->offset) {
917 trace_i915_gem_object_pwrite(obj, args->offset, args->size);
919 /* We can only do the GTT pwrite on untiled buffers, as otherwise
920 * it would end up going through the fenced access, and we'll get
921 * different detiling behavior between reading and writing.
922 * pread/pwrite currently are reading and writing from the CPU
923 * perspective, requiring manual detiling by the client.
926 ret = i915_gem_phys_pwrite(dev, obj, args, file);
930 if (obj->gtt_space &&
931 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
932 ret = i915_gem_object_pin(obj, 0, true);
936 ret = i915_gem_object_set_to_gtt_domain(obj, true);
940 ret = i915_gem_object_put_fence(obj);
944 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
946 ret = i915_gem_gtt_pwrite_slow(dev, obj, args, file);
949 i915_gem_object_unpin(obj);
953 /* Fall through to the shmfs paths because the gtt paths might
954 * fail with non-page-backed user pointers (e.g. gtt mappings
955 * when moving data between textures). */
958 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
963 if (!i915_gem_object_needs_bit17_swizzle(obj))
964 ret = i915_gem_shmem_pwrite_fast(dev, obj, args, file);
966 ret = i915_gem_shmem_pwrite_slow(dev, obj, args, file);
969 drm_gem_object_unreference(&obj->base);
971 mutex_unlock(&dev->struct_mutex);
976 * Called when user space prepares to use an object with the CPU, either
977 * through the mmap ioctl's mapping or a GTT mapping.
980 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
981 struct drm_file *file)
983 struct drm_i915_gem_set_domain *args = data;
984 struct drm_i915_gem_object *obj;
985 uint32_t read_domains = args->read_domains;
986 uint32_t write_domain = args->write_domain;
989 if (!(dev->driver->driver_features & DRIVER_GEM))
992 /* Only handle setting domains to types used by the CPU. */
993 if (write_domain & I915_GEM_GPU_DOMAINS)
996 if (read_domains & I915_GEM_GPU_DOMAINS)
999 /* Having something in the write domain implies it's in the read
1000 * domain, and only that read domain. Enforce that in the request.
1002 if (write_domain != 0 && read_domains != write_domain)
1005 ret = i915_mutex_lock_interruptible(dev);
1009 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1010 if (&obj->base == NULL) {
1015 if (read_domains & I915_GEM_DOMAIN_GTT) {
1016 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1018 /* Silently promote "you're not bound, there was nothing to do"
1019 * to success, since the client was just asking us to
1020 * make sure everything was done.
1025 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1028 drm_gem_object_unreference(&obj->base);
1030 mutex_unlock(&dev->struct_mutex);
1035 * Called when user space has done writes to this buffer
1038 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1039 struct drm_file *file)
1041 struct drm_i915_gem_sw_finish *args = data;
1042 struct drm_i915_gem_object *obj;
1045 if (!(dev->driver->driver_features & DRIVER_GEM))
1048 ret = i915_mutex_lock_interruptible(dev);
1052 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1053 if (&obj->base == NULL) {
1058 /* Pinned buffers may be scanout, so flush the cache */
1060 i915_gem_object_flush_cpu_write_domain(obj);
1062 drm_gem_object_unreference(&obj->base);
1064 mutex_unlock(&dev->struct_mutex);
1069 * Maps the contents of an object, returning the address it is mapped
1072 * While the mapping holds a reference on the contents of the object, it doesn't
1073 * imply a ref on the object itself.
1076 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1077 struct drm_file *file)
1079 struct drm_i915_gem_mmap *args = data;
1080 struct drm_gem_object *obj;
1083 if (!(dev->driver->driver_features & DRIVER_GEM))
1086 obj = drm_gem_object_lookup(dev, file, args->handle);
1090 addr = vm_mmap(obj->filp, 0, args->size,
1091 PROT_READ | PROT_WRITE, MAP_SHARED,
1093 drm_gem_object_unreference_unlocked(obj);
1094 if (IS_ERR((void *)addr))
1097 args->addr_ptr = (uint64_t) addr;
1103 * i915_gem_fault - fault a page into the GTT
1104 * vma: VMA in question
1107 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1108 * from userspace. The fault handler takes care of binding the object to
1109 * the GTT (if needed), allocating and programming a fence register (again,
1110 * only if needed based on whether the old reg is still valid or the object
1111 * is tiled) and inserting a new PTE into the faulting process.
1113 * Note that the faulting process may involve evicting existing objects
1114 * from the GTT and/or fence registers to make room. So performance may
1115 * suffer if the GTT working set is large or there are few fence registers
1118 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1120 struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1121 struct drm_device *dev = obj->base.dev;
1122 drm_i915_private_t *dev_priv = dev->dev_private;
1123 pgoff_t page_offset;
1126 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1128 /* We don't use vmf->pgoff since that has the fake offset */
1129 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1132 ret = i915_mutex_lock_interruptible(dev);
1136 trace_i915_gem_object_fault(obj, page_offset, true, write);
1138 /* Now bind it into the GTT if needed */
1139 if (!obj->map_and_fenceable) {
1140 ret = i915_gem_object_unbind(obj);
1144 if (!obj->gtt_space) {
1145 ret = i915_gem_object_bind_to_gtt(obj, 0, true);
1149 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1154 if (obj->tiling_mode == I915_TILING_NONE)
1155 ret = i915_gem_object_put_fence(obj);
1157 ret = i915_gem_object_get_fence(obj, NULL);
1161 if (i915_gem_object_is_inactive(obj))
1162 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1164 obj->fault_mappable = true;
1166 pfn = ((dev->agp->base + obj->gtt_offset) >> PAGE_SHIFT) +
1169 /* Finally, remap it using the new GTT offset */
1170 ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1172 mutex_unlock(&dev->struct_mutex);
1177 /* Give the error handler a chance to run and move the
1178 * objects off the GPU active list. Next time we service the
1179 * fault, we should be able to transition the page into the
1180 * GTT without touching the GPU (and so avoid further
1181 * EIO/EGAIN). If the GPU is wedged, then there is no issue
1182 * with coherency, just lost writes.
1188 return VM_FAULT_NOPAGE;
1190 return VM_FAULT_OOM;
1192 return VM_FAULT_SIGBUS;
1197 * i915_gem_release_mmap - remove physical page mappings
1198 * @obj: obj in question
1200 * Preserve the reservation of the mmapping with the DRM core code, but
1201 * relinquish ownership of the pages back to the system.
1203 * It is vital that we remove the page mapping if we have mapped a tiled
1204 * object through the GTT and then lose the fence register due to
1205 * resource pressure. Similarly if the object has been moved out of the
1206 * aperture, than pages mapped into userspace must be revoked. Removing the
1207 * mapping will then trigger a page fault on the next user access, allowing
1208 * fixup by i915_gem_fault().
1211 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1213 if (!obj->fault_mappable)
1216 if (obj->base.dev->dev_mapping)
1217 unmap_mapping_range(obj->base.dev->dev_mapping,
1218 (loff_t)obj->base.map_list.hash.key<<PAGE_SHIFT,
1221 obj->fault_mappable = false;
1225 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1229 if (INTEL_INFO(dev)->gen >= 4 ||
1230 tiling_mode == I915_TILING_NONE)
1233 /* Previous chips need a power-of-two fence region when tiling */
1234 if (INTEL_INFO(dev)->gen == 3)
1235 gtt_size = 1024*1024;
1237 gtt_size = 512*1024;
1239 while (gtt_size < size)
1246 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1247 * @obj: object to check
1249 * Return the required GTT alignment for an object, taking into account
1250 * potential fence register mapping.
1253 i915_gem_get_gtt_alignment(struct drm_device *dev,
1258 * Minimum alignment is 4k (GTT page size), but might be greater
1259 * if a fence register is needed for the object.
1261 if (INTEL_INFO(dev)->gen >= 4 ||
1262 tiling_mode == I915_TILING_NONE)
1266 * Previous chips need to be aligned to the size of the smallest
1267 * fence register that can contain the object.
1269 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1273 * i915_gem_get_unfenced_gtt_alignment - return required GTT alignment for an
1276 * @size: size of the object
1277 * @tiling_mode: tiling mode of the object
1279 * Return the required GTT alignment for an object, only taking into account
1280 * unfenced tiled surface requirements.
1283 i915_gem_get_unfenced_gtt_alignment(struct drm_device *dev,
1288 * Minimum alignment is 4k (GTT page size) for sane hw.
1290 if (INTEL_INFO(dev)->gen >= 4 || IS_G33(dev) ||
1291 tiling_mode == I915_TILING_NONE)
1294 /* Previous hardware however needs to be aligned to a power-of-two
1295 * tile height. The simplest method for determining this is to reuse
1296 * the power-of-tile object size.
1298 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1302 i915_gem_mmap_gtt(struct drm_file *file,
1303 struct drm_device *dev,
1307 struct drm_i915_private *dev_priv = dev->dev_private;
1308 struct drm_i915_gem_object *obj;
1311 if (!(dev->driver->driver_features & DRIVER_GEM))
1314 ret = i915_mutex_lock_interruptible(dev);
1318 obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
1319 if (&obj->base == NULL) {
1324 if (obj->base.size > dev_priv->mm.gtt_mappable_end) {
1329 if (obj->madv != I915_MADV_WILLNEED) {
1330 DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1335 if (!obj->base.map_list.map) {
1336 ret = drm_gem_create_mmap_offset(&obj->base);
1341 *offset = (u64)obj->base.map_list.hash.key << PAGE_SHIFT;
1344 drm_gem_object_unreference(&obj->base);
1346 mutex_unlock(&dev->struct_mutex);
1351 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1353 * @data: GTT mapping ioctl data
1354 * @file: GEM object info
1356 * Simply returns the fake offset to userspace so it can mmap it.
1357 * The mmap call will end up in drm_gem_mmap(), which will set things
1358 * up so we can get faults in the handler above.
1360 * The fault handler will take care of binding the object into the GTT
1361 * (since it may have been evicted to make room for something), allocating
1362 * a fence register, and mapping the appropriate aperture address into
1366 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1367 struct drm_file *file)
1369 struct drm_i915_gem_mmap_gtt *args = data;
1371 if (!(dev->driver->driver_features & DRIVER_GEM))
1374 return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
1379 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj,
1383 struct address_space *mapping;
1384 struct inode *inode;
1387 /* Get the list of pages out of our struct file. They'll be pinned
1388 * at this point until we release them.
1390 page_count = obj->base.size / PAGE_SIZE;
1391 BUG_ON(obj->pages != NULL);
1392 obj->pages = drm_malloc_ab(page_count, sizeof(struct page *));
1393 if (obj->pages == NULL)
1396 inode = obj->base.filp->f_path.dentry->d_inode;
1397 mapping = inode->i_mapping;
1398 gfpmask |= mapping_gfp_mask(mapping);
1400 for (i = 0; i < page_count; i++) {
1401 page = shmem_read_mapping_page_gfp(mapping, i, gfpmask);
1405 obj->pages[i] = page;
1408 if (i915_gem_object_needs_bit17_swizzle(obj))
1409 i915_gem_object_do_bit_17_swizzle(obj);
1415 page_cache_release(obj->pages[i]);
1417 drm_free_large(obj->pages);
1419 return PTR_ERR(page);
1423 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
1425 int page_count = obj->base.size / PAGE_SIZE;
1428 BUG_ON(obj->madv == __I915_MADV_PURGED);
1430 if (i915_gem_object_needs_bit17_swizzle(obj))
1431 i915_gem_object_save_bit_17_swizzle(obj);
1433 if (obj->madv == I915_MADV_DONTNEED)
1436 for (i = 0; i < page_count; i++) {
1438 set_page_dirty(obj->pages[i]);
1440 if (obj->madv == I915_MADV_WILLNEED)
1441 mark_page_accessed(obj->pages[i]);
1443 page_cache_release(obj->pages[i]);
1447 drm_free_large(obj->pages);
1452 i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
1453 struct intel_ring_buffer *ring,
1456 struct drm_device *dev = obj->base.dev;
1457 struct drm_i915_private *dev_priv = dev->dev_private;
1459 BUG_ON(ring == NULL);
1462 /* Add a reference if we're newly entering the active list. */
1464 drm_gem_object_reference(&obj->base);
1468 /* Move from whatever list we were on to the tail of execution. */
1469 list_move_tail(&obj->mm_list, &dev_priv->mm.active_list);
1470 list_move_tail(&obj->ring_list, &ring->active_list);
1472 obj->last_rendering_seqno = seqno;
1474 if (obj->fenced_gpu_access) {
1475 obj->last_fenced_seqno = seqno;
1476 obj->last_fenced_ring = ring;
1478 /* Bump MRU to take account of the delayed flush */
1479 if (obj->fence_reg != I915_FENCE_REG_NONE) {
1480 struct drm_i915_fence_reg *reg;
1482 reg = &dev_priv->fence_regs[obj->fence_reg];
1483 list_move_tail(®->lru_list,
1484 &dev_priv->mm.fence_list);
1490 i915_gem_object_move_off_active(struct drm_i915_gem_object *obj)
1492 list_del_init(&obj->ring_list);
1493 obj->last_rendering_seqno = 0;
1494 obj->last_fenced_seqno = 0;
1498 i915_gem_object_move_to_flushing(struct drm_i915_gem_object *obj)
1500 struct drm_device *dev = obj->base.dev;
1501 drm_i915_private_t *dev_priv = dev->dev_private;
1503 BUG_ON(!obj->active);
1504 list_move_tail(&obj->mm_list, &dev_priv->mm.flushing_list);
1506 i915_gem_object_move_off_active(obj);
1510 i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
1512 struct drm_device *dev = obj->base.dev;
1513 struct drm_i915_private *dev_priv = dev->dev_private;
1515 if (obj->pin_count != 0)
1516 list_move_tail(&obj->mm_list, &dev_priv->mm.pinned_list);
1518 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1520 BUG_ON(!list_empty(&obj->gpu_write_list));
1521 BUG_ON(!obj->active);
1523 obj->last_fenced_ring = NULL;
1525 i915_gem_object_move_off_active(obj);
1526 obj->fenced_gpu_access = false;
1529 obj->pending_gpu_write = false;
1530 drm_gem_object_unreference(&obj->base);
1532 WARN_ON(i915_verify_lists(dev));
1535 /* Immediately discard the backing storage */
1537 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
1539 struct inode *inode;
1541 /* Our goal here is to return as much of the memory as
1542 * is possible back to the system as we are called from OOM.
1543 * To do this we must instruct the shmfs to drop all of its
1544 * backing pages, *now*.
1546 inode = obj->base.filp->f_path.dentry->d_inode;
1547 shmem_truncate_range(inode, 0, (loff_t)-1);
1549 obj->madv = __I915_MADV_PURGED;
1553 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj)
1555 return obj->madv == I915_MADV_DONTNEED;
1559 i915_gem_process_flushing_list(struct intel_ring_buffer *ring,
1560 uint32_t flush_domains)
1562 struct drm_i915_gem_object *obj, *next;
1564 list_for_each_entry_safe(obj, next,
1565 &ring->gpu_write_list,
1567 if (obj->base.write_domain & flush_domains) {
1568 uint32_t old_write_domain = obj->base.write_domain;
1570 obj->base.write_domain = 0;
1571 list_del_init(&obj->gpu_write_list);
1572 i915_gem_object_move_to_active(obj, ring,
1573 i915_gem_next_request_seqno(ring));
1575 trace_i915_gem_object_change_domain(obj,
1576 obj->base.read_domains,
1583 i915_gem_get_seqno(struct drm_device *dev)
1585 drm_i915_private_t *dev_priv = dev->dev_private;
1586 u32 seqno = dev_priv->next_seqno;
1588 /* reserve 0 for non-seqno */
1589 if (++dev_priv->next_seqno == 0)
1590 dev_priv->next_seqno = 1;
1596 i915_gem_next_request_seqno(struct intel_ring_buffer *ring)
1598 if (ring->outstanding_lazy_request == 0)
1599 ring->outstanding_lazy_request = i915_gem_get_seqno(ring->dev);
1601 return ring->outstanding_lazy_request;
1605 i915_add_request(struct intel_ring_buffer *ring,
1606 struct drm_file *file,
1607 struct drm_i915_gem_request *request)
1609 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1611 u32 request_ring_position;
1615 BUG_ON(request == NULL);
1616 seqno = i915_gem_next_request_seqno(ring);
1618 /* Record the position of the start of the request so that
1619 * should we detect the updated seqno part-way through the
1620 * GPU processing the request, we never over-estimate the
1621 * position of the head.
1623 request_ring_position = intel_ring_get_tail(ring);
1625 ret = ring->add_request(ring, &seqno);
1629 trace_i915_gem_request_add(ring, seqno);
1631 request->seqno = seqno;
1632 request->ring = ring;
1633 request->tail = request_ring_position;
1634 request->emitted_jiffies = jiffies;
1635 was_empty = list_empty(&ring->request_list);
1636 list_add_tail(&request->list, &ring->request_list);
1639 struct drm_i915_file_private *file_priv = file->driver_priv;
1641 spin_lock(&file_priv->mm.lock);
1642 request->file_priv = file_priv;
1643 list_add_tail(&request->client_list,
1644 &file_priv->mm.request_list);
1645 spin_unlock(&file_priv->mm.lock);
1648 ring->outstanding_lazy_request = 0;
1650 if (!dev_priv->mm.suspended) {
1651 if (i915_enable_hangcheck) {
1652 mod_timer(&dev_priv->hangcheck_timer,
1654 msecs_to_jiffies(DRM_I915_HANGCHECK_PERIOD));
1657 queue_delayed_work(dev_priv->wq,
1658 &dev_priv->mm.retire_work, HZ);
1664 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
1666 struct drm_i915_file_private *file_priv = request->file_priv;
1671 spin_lock(&file_priv->mm.lock);
1672 if (request->file_priv) {
1673 list_del(&request->client_list);
1674 request->file_priv = NULL;
1676 spin_unlock(&file_priv->mm.lock);
1679 static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv,
1680 struct intel_ring_buffer *ring)
1682 while (!list_empty(&ring->request_list)) {
1683 struct drm_i915_gem_request *request;
1685 request = list_first_entry(&ring->request_list,
1686 struct drm_i915_gem_request,
1689 list_del(&request->list);
1690 i915_gem_request_remove_from_client(request);
1694 while (!list_empty(&ring->active_list)) {
1695 struct drm_i915_gem_object *obj;
1697 obj = list_first_entry(&ring->active_list,
1698 struct drm_i915_gem_object,
1701 obj->base.write_domain = 0;
1702 list_del_init(&obj->gpu_write_list);
1703 i915_gem_object_move_to_inactive(obj);
1707 static void i915_gem_reset_fences(struct drm_device *dev)
1709 struct drm_i915_private *dev_priv = dev->dev_private;
1712 for (i = 0; i < dev_priv->num_fence_regs; i++) {
1713 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
1714 struct drm_i915_gem_object *obj = reg->obj;
1719 if (obj->tiling_mode)
1720 i915_gem_release_mmap(obj);
1722 reg->obj->fence_reg = I915_FENCE_REG_NONE;
1723 reg->obj->fenced_gpu_access = false;
1724 reg->obj->last_fenced_seqno = 0;
1725 reg->obj->last_fenced_ring = NULL;
1726 i915_gem_clear_fence_reg(dev, reg);
1730 void i915_gem_reset(struct drm_device *dev)
1732 struct drm_i915_private *dev_priv = dev->dev_private;
1733 struct drm_i915_gem_object *obj;
1736 for (i = 0; i < I915_NUM_RINGS; i++)
1737 i915_gem_reset_ring_lists(dev_priv, &dev_priv->ring[i]);
1739 /* Remove anything from the flushing lists. The GPU cache is likely
1740 * to be lost on reset along with the data, so simply move the
1741 * lost bo to the inactive list.
1743 while (!list_empty(&dev_priv->mm.flushing_list)) {
1744 obj = list_first_entry(&dev_priv->mm.flushing_list,
1745 struct drm_i915_gem_object,
1748 obj->base.write_domain = 0;
1749 list_del_init(&obj->gpu_write_list);
1750 i915_gem_object_move_to_inactive(obj);
1753 /* Move everything out of the GPU domains to ensure we do any
1754 * necessary invalidation upon reuse.
1756 list_for_each_entry(obj,
1757 &dev_priv->mm.inactive_list,
1760 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
1763 /* The fence registers are invalidated so clear them out */
1764 i915_gem_reset_fences(dev);
1768 * This function clears the request list as sequence numbers are passed.
1771 i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
1776 if (list_empty(&ring->request_list))
1779 WARN_ON(i915_verify_lists(ring->dev));
1781 seqno = ring->get_seqno(ring);
1783 for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++)
1784 if (seqno >= ring->sync_seqno[i])
1785 ring->sync_seqno[i] = 0;
1787 while (!list_empty(&ring->request_list)) {
1788 struct drm_i915_gem_request *request;
1790 request = list_first_entry(&ring->request_list,
1791 struct drm_i915_gem_request,
1794 if (!i915_seqno_passed(seqno, request->seqno))
1797 trace_i915_gem_request_retire(ring, request->seqno);
1798 /* We know the GPU must have read the request to have
1799 * sent us the seqno + interrupt, so use the position
1800 * of tail of the request to update the last known position
1803 ring->last_retired_head = request->tail;
1805 list_del(&request->list);
1806 i915_gem_request_remove_from_client(request);
1810 /* Move any buffers on the active list that are no longer referenced
1811 * by the ringbuffer to the flushing/inactive lists as appropriate.
1813 while (!list_empty(&ring->active_list)) {
1814 struct drm_i915_gem_object *obj;
1816 obj = list_first_entry(&ring->active_list,
1817 struct drm_i915_gem_object,
1820 if (!i915_seqno_passed(seqno, obj->last_rendering_seqno))
1823 if (obj->base.write_domain != 0)
1824 i915_gem_object_move_to_flushing(obj);
1826 i915_gem_object_move_to_inactive(obj);
1829 if (unlikely(ring->trace_irq_seqno &&
1830 i915_seqno_passed(seqno, ring->trace_irq_seqno))) {
1831 ring->irq_put(ring);
1832 ring->trace_irq_seqno = 0;
1835 WARN_ON(i915_verify_lists(ring->dev));
1839 i915_gem_retire_requests(struct drm_device *dev)
1841 drm_i915_private_t *dev_priv = dev->dev_private;
1844 if (!list_empty(&dev_priv->mm.deferred_free_list)) {
1845 struct drm_i915_gem_object *obj, *next;
1847 /* We must be careful that during unbind() we do not
1848 * accidentally infinitely recurse into retire requests.
1850 * retire -> free -> unbind -> wait -> retire_ring
1852 list_for_each_entry_safe(obj, next,
1853 &dev_priv->mm.deferred_free_list,
1855 i915_gem_free_object_tail(obj);
1858 for (i = 0; i < I915_NUM_RINGS; i++)
1859 i915_gem_retire_requests_ring(&dev_priv->ring[i]);
1863 i915_gem_retire_work_handler(struct work_struct *work)
1865 drm_i915_private_t *dev_priv;
1866 struct drm_device *dev;
1870 dev_priv = container_of(work, drm_i915_private_t,
1871 mm.retire_work.work);
1872 dev = dev_priv->dev;
1874 /* Come back later if the device is busy... */
1875 if (!mutex_trylock(&dev->struct_mutex)) {
1876 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1880 i915_gem_retire_requests(dev);
1882 /* Send a periodic flush down the ring so we don't hold onto GEM
1883 * objects indefinitely.
1886 for (i = 0; i < I915_NUM_RINGS; i++) {
1887 struct intel_ring_buffer *ring = &dev_priv->ring[i];
1889 if (!list_empty(&ring->gpu_write_list)) {
1890 struct drm_i915_gem_request *request;
1893 ret = i915_gem_flush_ring(ring,
1894 0, I915_GEM_GPU_DOMAINS);
1895 request = kzalloc(sizeof(*request), GFP_KERNEL);
1896 if (ret || request == NULL ||
1897 i915_add_request(ring, NULL, request))
1901 idle &= list_empty(&ring->request_list);
1904 if (!dev_priv->mm.suspended && !idle)
1905 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1907 mutex_unlock(&dev->struct_mutex);
1911 * Waits for a sequence number to be signaled, and cleans up the
1912 * request and object lists appropriately for that event.
1915 i915_wait_request(struct intel_ring_buffer *ring,
1919 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1925 if (atomic_read(&dev_priv->mm.wedged)) {
1926 struct completion *x = &dev_priv->error_completion;
1927 bool recovery_complete;
1928 unsigned long flags;
1930 /* Give the error handler a chance to run. */
1931 spin_lock_irqsave(&x->wait.lock, flags);
1932 recovery_complete = x->done > 0;
1933 spin_unlock_irqrestore(&x->wait.lock, flags);
1935 return recovery_complete ? -EIO : -EAGAIN;
1938 if (seqno == ring->outstanding_lazy_request) {
1939 struct drm_i915_gem_request *request;
1941 request = kzalloc(sizeof(*request), GFP_KERNEL);
1942 if (request == NULL)
1945 ret = i915_add_request(ring, NULL, request);
1951 seqno = request->seqno;
1954 if (!i915_seqno_passed(ring->get_seqno(ring), seqno)) {
1955 if (HAS_PCH_SPLIT(ring->dev))
1956 ier = I915_READ(DEIER) | I915_READ(GTIER);
1958 ier = I915_READ(IER);
1960 DRM_ERROR("something (likely vbetool) disabled "
1961 "interrupts, re-enabling\n");
1962 ring->dev->driver->irq_preinstall(ring->dev);
1963 ring->dev->driver->irq_postinstall(ring->dev);
1966 trace_i915_gem_request_wait_begin(ring, seqno);
1968 ring->waiting_seqno = seqno;
1969 if (ring->irq_get(ring)) {
1970 if (dev_priv->mm.interruptible)
1971 ret = wait_event_interruptible(ring->irq_queue,
1972 i915_seqno_passed(ring->get_seqno(ring), seqno)
1973 || atomic_read(&dev_priv->mm.wedged));
1975 wait_event(ring->irq_queue,
1976 i915_seqno_passed(ring->get_seqno(ring), seqno)
1977 || atomic_read(&dev_priv->mm.wedged));
1979 ring->irq_put(ring);
1980 } else if (wait_for_atomic(i915_seqno_passed(ring->get_seqno(ring),
1982 atomic_read(&dev_priv->mm.wedged), 3000))
1984 ring->waiting_seqno = 0;
1986 trace_i915_gem_request_wait_end(ring, seqno);
1988 if (atomic_read(&dev_priv->mm.wedged))
1991 /* Directly dispatch request retiring. While we have the work queue
1992 * to handle this, the waiter on a request often wants an associated
1993 * buffer to have made it to the inactive list, and we would need
1994 * a separate wait queue to handle that.
1996 if (ret == 0 && do_retire)
1997 i915_gem_retire_requests_ring(ring);
2003 * Ensures that all rendering to the object has completed and the object is
2004 * safe to unbind from the GTT or access from the CPU.
2007 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj)
2011 /* This function only exists to support waiting for existing rendering,
2012 * not for emitting required flushes.
2014 BUG_ON((obj->base.write_domain & I915_GEM_GPU_DOMAINS) != 0);
2016 /* If there is rendering queued on the buffer being evicted, wait for
2020 ret = i915_wait_request(obj->ring, obj->last_rendering_seqno,
2029 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
2031 u32 old_write_domain, old_read_domains;
2033 /* Act a barrier for all accesses through the GTT */
2036 /* Force a pagefault for domain tracking on next user access */
2037 i915_gem_release_mmap(obj);
2039 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
2042 old_read_domains = obj->base.read_domains;
2043 old_write_domain = obj->base.write_domain;
2045 obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
2046 obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
2048 trace_i915_gem_object_change_domain(obj,
2054 * Unbinds an object from the GTT aperture.
2057 i915_gem_object_unbind(struct drm_i915_gem_object *obj)
2059 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
2062 if (obj->gtt_space == NULL)
2065 if (obj->pin_count != 0) {
2066 DRM_ERROR("Attempting to unbind pinned buffer\n");
2070 ret = i915_gem_object_finish_gpu(obj);
2071 if (ret == -ERESTARTSYS)
2073 /* Continue on if we fail due to EIO, the GPU is hung so we
2074 * should be safe and we need to cleanup or else we might
2075 * cause memory corruption through use-after-free.
2078 i915_gem_object_finish_gtt(obj);
2080 /* Move the object to the CPU domain to ensure that
2081 * any possible CPU writes while it's not in the GTT
2082 * are flushed when we go to remap it.
2085 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
2086 if (ret == -ERESTARTSYS)
2089 /* In the event of a disaster, abandon all caches and
2090 * hope for the best.
2092 i915_gem_clflush_object(obj);
2093 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
2096 /* release the fence reg _after_ flushing */
2097 ret = i915_gem_object_put_fence(obj);
2098 if (ret == -ERESTARTSYS)
2101 trace_i915_gem_object_unbind(obj);
2103 i915_gem_gtt_unbind_object(obj);
2104 if (obj->has_aliasing_ppgtt_mapping) {
2105 i915_ppgtt_unbind_object(dev_priv->mm.aliasing_ppgtt, obj);
2106 obj->has_aliasing_ppgtt_mapping = 0;
2109 i915_gem_object_put_pages_gtt(obj);
2111 list_del_init(&obj->gtt_list);
2112 list_del_init(&obj->mm_list);
2113 /* Avoid an unnecessary call to unbind on rebind. */
2114 obj->map_and_fenceable = true;
2116 drm_mm_put_block(obj->gtt_space);
2117 obj->gtt_space = NULL;
2118 obj->gtt_offset = 0;
2120 if (i915_gem_object_is_purgeable(obj))
2121 i915_gem_object_truncate(obj);
2127 i915_gem_flush_ring(struct intel_ring_buffer *ring,
2128 uint32_t invalidate_domains,
2129 uint32_t flush_domains)
2133 if (((invalidate_domains | flush_domains) & I915_GEM_GPU_DOMAINS) == 0)
2136 trace_i915_gem_ring_flush(ring, invalidate_domains, flush_domains);
2138 ret = ring->flush(ring, invalidate_domains, flush_domains);
2142 if (flush_domains & I915_GEM_GPU_DOMAINS)
2143 i915_gem_process_flushing_list(ring, flush_domains);
2148 static int i915_ring_idle(struct intel_ring_buffer *ring, bool do_retire)
2152 if (list_empty(&ring->gpu_write_list) && list_empty(&ring->active_list))
2155 if (!list_empty(&ring->gpu_write_list)) {
2156 ret = i915_gem_flush_ring(ring,
2157 I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
2162 return i915_wait_request(ring, i915_gem_next_request_seqno(ring),
2166 int i915_gpu_idle(struct drm_device *dev, bool do_retire)
2168 drm_i915_private_t *dev_priv = dev->dev_private;
2171 /* Flush everything onto the inactive list. */
2172 for (i = 0; i < I915_NUM_RINGS; i++) {
2173 ret = i915_ring_idle(&dev_priv->ring[i], do_retire);
2181 static int sandybridge_write_fence_reg(struct drm_i915_gem_object *obj,
2182 struct intel_ring_buffer *pipelined)
2184 struct drm_device *dev = obj->base.dev;
2185 drm_i915_private_t *dev_priv = dev->dev_private;
2186 u32 size = obj->gtt_space->size;
2187 int regnum = obj->fence_reg;
2190 val = (uint64_t)((obj->gtt_offset + size - 4096) &
2192 val |= obj->gtt_offset & 0xfffff000;
2193 val |= (uint64_t)((obj->stride / 128) - 1) <<
2194 SANDYBRIDGE_FENCE_PITCH_SHIFT;
2196 if (obj->tiling_mode == I915_TILING_Y)
2197 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2198 val |= I965_FENCE_REG_VALID;
2201 int ret = intel_ring_begin(pipelined, 6);
2205 intel_ring_emit(pipelined, MI_NOOP);
2206 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(2));
2207 intel_ring_emit(pipelined, FENCE_REG_SANDYBRIDGE_0 + regnum*8);
2208 intel_ring_emit(pipelined, (u32)val);
2209 intel_ring_emit(pipelined, FENCE_REG_SANDYBRIDGE_0 + regnum*8 + 4);
2210 intel_ring_emit(pipelined, (u32)(val >> 32));
2211 intel_ring_advance(pipelined);
2213 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + regnum * 8, val);
2218 static int i965_write_fence_reg(struct drm_i915_gem_object *obj,
2219 struct intel_ring_buffer *pipelined)
2221 struct drm_device *dev = obj->base.dev;
2222 drm_i915_private_t *dev_priv = dev->dev_private;
2223 u32 size = obj->gtt_space->size;
2224 int regnum = obj->fence_reg;
2227 val = (uint64_t)((obj->gtt_offset + size - 4096) &
2229 val |= obj->gtt_offset & 0xfffff000;
2230 val |= ((obj->stride / 128) - 1) << I965_FENCE_PITCH_SHIFT;
2231 if (obj->tiling_mode == I915_TILING_Y)
2232 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2233 val |= I965_FENCE_REG_VALID;
2236 int ret = intel_ring_begin(pipelined, 6);
2240 intel_ring_emit(pipelined, MI_NOOP);
2241 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(2));
2242 intel_ring_emit(pipelined, FENCE_REG_965_0 + regnum*8);
2243 intel_ring_emit(pipelined, (u32)val);
2244 intel_ring_emit(pipelined, FENCE_REG_965_0 + regnum*8 + 4);
2245 intel_ring_emit(pipelined, (u32)(val >> 32));
2246 intel_ring_advance(pipelined);
2248 I915_WRITE64(FENCE_REG_965_0 + regnum * 8, val);
2253 static int i915_write_fence_reg(struct drm_i915_gem_object *obj,
2254 struct intel_ring_buffer *pipelined)
2256 struct drm_device *dev = obj->base.dev;
2257 drm_i915_private_t *dev_priv = dev->dev_private;
2258 u32 size = obj->gtt_space->size;
2259 u32 fence_reg, val, pitch_val;
2262 if (WARN((obj->gtt_offset & ~I915_FENCE_START_MASK) ||
2263 (size & -size) != size ||
2264 (obj->gtt_offset & (size - 1)),
2265 "object 0x%08x [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
2266 obj->gtt_offset, obj->map_and_fenceable, size))
2269 if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
2274 /* Note: pitch better be a power of two tile widths */
2275 pitch_val = obj->stride / tile_width;
2276 pitch_val = ffs(pitch_val) - 1;
2278 val = obj->gtt_offset;
2279 if (obj->tiling_mode == I915_TILING_Y)
2280 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2281 val |= I915_FENCE_SIZE_BITS(size);
2282 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2283 val |= I830_FENCE_REG_VALID;
2285 fence_reg = obj->fence_reg;
2287 fence_reg = FENCE_REG_830_0 + fence_reg * 4;
2289 fence_reg = FENCE_REG_945_8 + (fence_reg - 8) * 4;
2292 int ret = intel_ring_begin(pipelined, 4);
2296 intel_ring_emit(pipelined, MI_NOOP);
2297 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(1));
2298 intel_ring_emit(pipelined, fence_reg);
2299 intel_ring_emit(pipelined, val);
2300 intel_ring_advance(pipelined);
2302 I915_WRITE(fence_reg, val);
2307 static int i830_write_fence_reg(struct drm_i915_gem_object *obj,
2308 struct intel_ring_buffer *pipelined)
2310 struct drm_device *dev = obj->base.dev;
2311 drm_i915_private_t *dev_priv = dev->dev_private;
2312 u32 size = obj->gtt_space->size;
2313 int regnum = obj->fence_reg;
2317 if (WARN((obj->gtt_offset & ~I830_FENCE_START_MASK) ||
2318 (size & -size) != size ||
2319 (obj->gtt_offset & (size - 1)),
2320 "object 0x%08x not 512K or pot-size 0x%08x aligned\n",
2321 obj->gtt_offset, size))
2324 pitch_val = obj->stride / 128;
2325 pitch_val = ffs(pitch_val) - 1;
2327 val = obj->gtt_offset;
2328 if (obj->tiling_mode == I915_TILING_Y)
2329 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2330 val |= I830_FENCE_SIZE_BITS(size);
2331 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2332 val |= I830_FENCE_REG_VALID;
2335 int ret = intel_ring_begin(pipelined, 4);
2339 intel_ring_emit(pipelined, MI_NOOP);
2340 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(1));
2341 intel_ring_emit(pipelined, FENCE_REG_830_0 + regnum*4);
2342 intel_ring_emit(pipelined, val);
2343 intel_ring_advance(pipelined);
2345 I915_WRITE(FENCE_REG_830_0 + regnum * 4, val);
2350 static bool ring_passed_seqno(struct intel_ring_buffer *ring, u32 seqno)
2352 return i915_seqno_passed(ring->get_seqno(ring), seqno);
2356 i915_gem_object_flush_fence(struct drm_i915_gem_object *obj,
2357 struct intel_ring_buffer *pipelined)
2361 if (obj->fenced_gpu_access) {
2362 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
2363 ret = i915_gem_flush_ring(obj->last_fenced_ring,
2364 0, obj->base.write_domain);
2369 obj->fenced_gpu_access = false;
2372 if (obj->last_fenced_seqno && pipelined != obj->last_fenced_ring) {
2373 if (!ring_passed_seqno(obj->last_fenced_ring,
2374 obj->last_fenced_seqno)) {
2375 ret = i915_wait_request(obj->last_fenced_ring,
2376 obj->last_fenced_seqno,
2382 obj->last_fenced_seqno = 0;
2383 obj->last_fenced_ring = NULL;
2386 /* Ensure that all CPU reads are completed before installing a fence
2387 * and all writes before removing the fence.
2389 if (obj->base.read_domains & I915_GEM_DOMAIN_GTT)
2396 i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
2400 if (obj->tiling_mode)
2401 i915_gem_release_mmap(obj);
2403 ret = i915_gem_object_flush_fence(obj, NULL);
2407 if (obj->fence_reg != I915_FENCE_REG_NONE) {
2408 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2410 WARN_ON(dev_priv->fence_regs[obj->fence_reg].pin_count);
2411 i915_gem_clear_fence_reg(obj->base.dev,
2412 &dev_priv->fence_regs[obj->fence_reg]);
2414 obj->fence_reg = I915_FENCE_REG_NONE;
2420 static struct drm_i915_fence_reg *
2421 i915_find_fence_reg(struct drm_device *dev,
2422 struct intel_ring_buffer *pipelined)
2424 struct drm_i915_private *dev_priv = dev->dev_private;
2425 struct drm_i915_fence_reg *reg, *first, *avail;
2428 /* First try to find a free reg */
2430 for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2431 reg = &dev_priv->fence_regs[i];
2435 if (!reg->pin_count)
2442 /* None available, try to steal one or wait for a user to finish */
2443 avail = first = NULL;
2444 list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
2452 !reg->obj->last_fenced_ring ||
2453 reg->obj->last_fenced_ring == pipelined) {
2466 * i915_gem_object_get_fence - set up a fence reg for an object
2467 * @obj: object to map through a fence reg
2468 * @pipelined: ring on which to queue the change, or NULL for CPU access
2469 * @interruptible: must we wait uninterruptibly for the register to retire?
2471 * When mapping objects through the GTT, userspace wants to be able to write
2472 * to them without having to worry about swizzling if the object is tiled.
2474 * This function walks the fence regs looking for a free one for @obj,
2475 * stealing one if it can't find any.
2477 * It then sets up the reg based on the object's properties: address, pitch
2478 * and tiling format.
2481 i915_gem_object_get_fence(struct drm_i915_gem_object *obj,
2482 struct intel_ring_buffer *pipelined)
2484 struct drm_device *dev = obj->base.dev;
2485 struct drm_i915_private *dev_priv = dev->dev_private;
2486 struct drm_i915_fence_reg *reg;
2489 /* XXX disable pipelining. There are bugs. Shocking. */
2492 /* Just update our place in the LRU if our fence is getting reused. */
2493 if (obj->fence_reg != I915_FENCE_REG_NONE) {
2494 reg = &dev_priv->fence_regs[obj->fence_reg];
2495 list_move_tail(®->lru_list, &dev_priv->mm.fence_list);
2497 if (obj->tiling_changed) {
2498 ret = i915_gem_object_flush_fence(obj, pipelined);
2502 if (!obj->fenced_gpu_access && !obj->last_fenced_seqno)
2507 i915_gem_next_request_seqno(pipelined);
2508 obj->last_fenced_seqno = reg->setup_seqno;
2509 obj->last_fenced_ring = pipelined;
2516 if (reg->setup_seqno) {
2517 if (!ring_passed_seqno(obj->last_fenced_ring,
2518 reg->setup_seqno)) {
2519 ret = i915_wait_request(obj->last_fenced_ring,
2526 reg->setup_seqno = 0;
2528 } else if (obj->last_fenced_ring &&
2529 obj->last_fenced_ring != pipelined) {
2530 ret = i915_gem_object_flush_fence(obj, pipelined);
2538 reg = i915_find_fence_reg(dev, pipelined);
2542 ret = i915_gem_object_flush_fence(obj, pipelined);
2547 struct drm_i915_gem_object *old = reg->obj;
2549 drm_gem_object_reference(&old->base);
2551 if (old->tiling_mode)
2552 i915_gem_release_mmap(old);
2554 ret = i915_gem_object_flush_fence(old, pipelined);
2556 drm_gem_object_unreference(&old->base);
2560 if (old->last_fenced_seqno == 0 && obj->last_fenced_seqno == 0)
2563 old->fence_reg = I915_FENCE_REG_NONE;
2564 old->last_fenced_ring = pipelined;
2565 old->last_fenced_seqno =
2566 pipelined ? i915_gem_next_request_seqno(pipelined) : 0;
2568 drm_gem_object_unreference(&old->base);
2569 } else if (obj->last_fenced_seqno == 0)
2573 list_move_tail(®->lru_list, &dev_priv->mm.fence_list);
2574 obj->fence_reg = reg - dev_priv->fence_regs;
2575 obj->last_fenced_ring = pipelined;
2578 pipelined ? i915_gem_next_request_seqno(pipelined) : 0;
2579 obj->last_fenced_seqno = reg->setup_seqno;
2582 obj->tiling_changed = false;
2583 switch (INTEL_INFO(dev)->gen) {
2586 ret = sandybridge_write_fence_reg(obj, pipelined);
2590 ret = i965_write_fence_reg(obj, pipelined);
2593 ret = i915_write_fence_reg(obj, pipelined);
2596 ret = i830_write_fence_reg(obj, pipelined);
2604 * i915_gem_clear_fence_reg - clear out fence register info
2605 * @obj: object to clear
2607 * Zeroes out the fence register itself and clears out the associated
2608 * data structures in dev_priv and obj.
2611 i915_gem_clear_fence_reg(struct drm_device *dev,
2612 struct drm_i915_fence_reg *reg)
2614 drm_i915_private_t *dev_priv = dev->dev_private;
2615 uint32_t fence_reg = reg - dev_priv->fence_regs;
2617 switch (INTEL_INFO(dev)->gen) {
2620 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + fence_reg*8, 0);
2624 I915_WRITE64(FENCE_REG_965_0 + fence_reg*8, 0);
2628 fence_reg = FENCE_REG_945_8 + (fence_reg - 8) * 4;
2631 fence_reg = FENCE_REG_830_0 + fence_reg * 4;
2633 I915_WRITE(fence_reg, 0);
2637 list_del_init(®->lru_list);
2639 reg->setup_seqno = 0;
2644 * Finds free space in the GTT aperture and binds the object there.
2647 i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
2649 bool map_and_fenceable)
2651 struct drm_device *dev = obj->base.dev;
2652 drm_i915_private_t *dev_priv = dev->dev_private;
2653 struct drm_mm_node *free_space;
2654 gfp_t gfpmask = __GFP_NORETRY | __GFP_NOWARN;
2655 u32 size, fence_size, fence_alignment, unfenced_alignment;
2656 bool mappable, fenceable;
2659 if (obj->madv != I915_MADV_WILLNEED) {
2660 DRM_ERROR("Attempting to bind a purgeable object\n");
2664 fence_size = i915_gem_get_gtt_size(dev,
2667 fence_alignment = i915_gem_get_gtt_alignment(dev,
2670 unfenced_alignment =
2671 i915_gem_get_unfenced_gtt_alignment(dev,
2676 alignment = map_and_fenceable ? fence_alignment :
2678 if (map_and_fenceable && alignment & (fence_alignment - 1)) {
2679 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
2683 size = map_and_fenceable ? fence_size : obj->base.size;
2685 /* If the object is bigger than the entire aperture, reject it early
2686 * before evicting everything in a vain attempt to find space.
2688 if (obj->base.size >
2689 (map_and_fenceable ? dev_priv->mm.gtt_mappable_end : dev_priv->mm.gtt_total)) {
2690 DRM_ERROR("Attempting to bind an object larger than the aperture\n");
2695 if (map_and_fenceable)
2697 drm_mm_search_free_in_range(&dev_priv->mm.gtt_space,
2699 dev_priv->mm.gtt_mappable_end,
2702 free_space = drm_mm_search_free(&dev_priv->mm.gtt_space,
2703 size, alignment, 0);
2705 if (free_space != NULL) {
2706 if (map_and_fenceable)
2708 drm_mm_get_block_range_generic(free_space,
2710 dev_priv->mm.gtt_mappable_end,
2714 drm_mm_get_block(free_space, size, alignment);
2716 if (obj->gtt_space == NULL) {
2717 /* If the gtt is empty and we're still having trouble
2718 * fitting our object in, we're out of memory.
2720 ret = i915_gem_evict_something(dev, size, alignment,
2728 ret = i915_gem_object_get_pages_gtt(obj, gfpmask);
2730 drm_mm_put_block(obj->gtt_space);
2731 obj->gtt_space = NULL;
2733 if (ret == -ENOMEM) {
2734 /* first try to reclaim some memory by clearing the GTT */
2735 ret = i915_gem_evict_everything(dev, false);
2737 /* now try to shrink everyone else */
2752 ret = i915_gem_gtt_bind_object(obj);
2754 i915_gem_object_put_pages_gtt(obj);
2755 drm_mm_put_block(obj->gtt_space);
2756 obj->gtt_space = NULL;
2758 if (i915_gem_evict_everything(dev, false))
2764 list_add_tail(&obj->gtt_list, &dev_priv->mm.gtt_list);
2765 list_add_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
2767 /* Assert that the object is not currently in any GPU domain. As it
2768 * wasn't in the GTT, there shouldn't be any way it could have been in
2771 BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
2772 BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
2774 obj->gtt_offset = obj->gtt_space->start;
2777 obj->gtt_space->size == fence_size &&
2778 (obj->gtt_space->start & (fence_alignment - 1)) == 0;
2781 obj->gtt_offset + obj->base.size <= dev_priv->mm.gtt_mappable_end;
2783 obj->map_and_fenceable = mappable && fenceable;
2785 trace_i915_gem_object_bind(obj, map_and_fenceable);
2790 i915_gem_clflush_object(struct drm_i915_gem_object *obj)
2792 /* If we don't have a page list set up, then we're not pinned
2793 * to GPU, and we can ignore the cache flush because it'll happen
2794 * again at bind time.
2796 if (obj->pages == NULL)
2799 /* If the GPU is snooping the contents of the CPU cache,
2800 * we do not need to manually clear the CPU cache lines. However,
2801 * the caches are only snooped when the render cache is
2802 * flushed/invalidated. As we always have to emit invalidations
2803 * and flushes when moving into and out of the RENDER domain, correct
2804 * snooping behaviour occurs naturally as the result of our domain
2807 if (obj->cache_level != I915_CACHE_NONE)
2810 trace_i915_gem_object_clflush(obj);
2812 drm_clflush_pages(obj->pages, obj->base.size / PAGE_SIZE);
2815 /** Flushes any GPU write domain for the object if it's dirty. */
2817 i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object *obj)
2819 if ((obj->base.write_domain & I915_GEM_GPU_DOMAINS) == 0)
2822 /* Queue the GPU write cache flushing we need. */
2823 return i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain);
2826 /** Flushes the GTT write domain for the object if it's dirty. */
2828 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
2830 uint32_t old_write_domain;
2832 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
2835 /* No actual flushing is required for the GTT write domain. Writes
2836 * to it immediately go to main memory as far as we know, so there's
2837 * no chipset flush. It also doesn't land in render cache.
2839 * However, we do have to enforce the order so that all writes through
2840 * the GTT land before any writes to the device, such as updates to
2845 old_write_domain = obj->base.write_domain;
2846 obj->base.write_domain = 0;
2848 trace_i915_gem_object_change_domain(obj,
2849 obj->base.read_domains,
2853 /** Flushes the CPU write domain for the object if it's dirty. */
2855 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj)
2857 uint32_t old_write_domain;
2859 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
2862 i915_gem_clflush_object(obj);
2863 intel_gtt_chipset_flush();
2864 old_write_domain = obj->base.write_domain;
2865 obj->base.write_domain = 0;
2867 trace_i915_gem_object_change_domain(obj,
2868 obj->base.read_domains,
2873 * Moves a single object to the GTT read, and possibly write domain.
2875 * This function returns when the move is complete, including waiting on
2879 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
2881 uint32_t old_write_domain, old_read_domains;
2884 /* Not valid to be called on unbound objects. */
2885 if (obj->gtt_space == NULL)
2888 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
2891 ret = i915_gem_object_flush_gpu_write_domain(obj);
2895 if (obj->pending_gpu_write || write) {
2896 ret = i915_gem_object_wait_rendering(obj);
2901 i915_gem_object_flush_cpu_write_domain(obj);
2903 old_write_domain = obj->base.write_domain;
2904 old_read_domains = obj->base.read_domains;
2906 /* It should now be out of any other write domains, and we can update
2907 * the domain values for our changes.
2909 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2910 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
2912 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
2913 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
2917 trace_i915_gem_object_change_domain(obj,
2924 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
2925 enum i915_cache_level cache_level)
2927 struct drm_device *dev = obj->base.dev;
2928 drm_i915_private_t *dev_priv = dev->dev_private;
2931 if (obj->cache_level == cache_level)
2934 if (obj->pin_count) {
2935 DRM_DEBUG("can not change the cache level of pinned objects\n");
2939 if (obj->gtt_space) {
2940 ret = i915_gem_object_finish_gpu(obj);
2944 i915_gem_object_finish_gtt(obj);
2946 /* Before SandyBridge, you could not use tiling or fence
2947 * registers with snooped memory, so relinquish any fences
2948 * currently pointing to our region in the aperture.
2950 if (INTEL_INFO(obj->base.dev)->gen < 6) {
2951 ret = i915_gem_object_put_fence(obj);
2956 i915_gem_gtt_rebind_object(obj, cache_level);
2957 if (obj->has_aliasing_ppgtt_mapping)
2958 i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt,
2962 if (cache_level == I915_CACHE_NONE) {
2963 u32 old_read_domains, old_write_domain;
2965 /* If we're coming from LLC cached, then we haven't
2966 * actually been tracking whether the data is in the
2967 * CPU cache or not, since we only allow one bit set
2968 * in obj->write_domain and have been skipping the clflushes.
2969 * Just set it to the CPU cache for now.
2971 WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU);
2972 WARN_ON(obj->base.read_domains & ~I915_GEM_DOMAIN_CPU);
2974 old_read_domains = obj->base.read_domains;
2975 old_write_domain = obj->base.write_domain;
2977 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
2978 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
2980 trace_i915_gem_object_change_domain(obj,
2985 obj->cache_level = cache_level;
2990 * Prepare buffer for display plane (scanout, cursors, etc).
2991 * Can be called from an uninterruptible phase (modesetting) and allows
2992 * any flushes to be pipelined (for pageflips).
2994 * For the display plane, we want to be in the GTT but out of any write
2995 * domains. So in many ways this looks like set_to_gtt_domain() apart from the
2996 * ability to pipeline the waits, pinning and any additional subtleties
2997 * that may differentiate the display plane from ordinary buffers.
3000 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
3002 struct intel_ring_buffer *pipelined)
3004 u32 old_read_domains, old_write_domain;
3007 ret = i915_gem_object_flush_gpu_write_domain(obj);
3011 if (pipelined != obj->ring) {
3012 ret = i915_gem_object_wait_rendering(obj);
3013 if (ret == -ERESTARTSYS)
3017 /* The display engine is not coherent with the LLC cache on gen6. As
3018 * a result, we make sure that the pinning that is about to occur is
3019 * done with uncached PTEs. This is lowest common denominator for all
3022 * However for gen6+, we could do better by using the GFDT bit instead
3023 * of uncaching, which would allow us to flush all the LLC-cached data
3024 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3026 ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE);
3030 /* As the user may map the buffer once pinned in the display plane
3031 * (e.g. libkms for the bootup splash), we have to ensure that we
3032 * always use map_and_fenceable for all scanout buffers.
3034 ret = i915_gem_object_pin(obj, alignment, true);
3038 i915_gem_object_flush_cpu_write_domain(obj);
3040 old_write_domain = obj->base.write_domain;
3041 old_read_domains = obj->base.read_domains;
3043 /* It should now be out of any other write domains, and we can update
3044 * the domain values for our changes.
3046 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3047 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3049 trace_i915_gem_object_change_domain(obj,
3057 i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
3061 if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
3064 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
3065 ret = i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain);
3070 ret = i915_gem_object_wait_rendering(obj);
3074 /* Ensure that we invalidate the GPU's caches and TLBs. */
3075 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
3080 * Moves a single object to the CPU read, and possibly write domain.
3082 * This function returns when the move is complete, including waiting on
3086 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
3088 uint32_t old_write_domain, old_read_domains;
3091 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
3094 ret = i915_gem_object_flush_gpu_write_domain(obj);
3098 ret = i915_gem_object_wait_rendering(obj);
3102 i915_gem_object_flush_gtt_write_domain(obj);
3104 /* If we have a partially-valid cache of the object in the CPU,
3105 * finish invalidating it and free the per-page flags.
3107 i915_gem_object_set_to_full_cpu_read_domain(obj);
3109 old_write_domain = obj->base.write_domain;
3110 old_read_domains = obj->base.read_domains;
3112 /* Flush the CPU cache if it's still invalid. */
3113 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
3114 i915_gem_clflush_object(obj);
3116 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3119 /* It should now be out of any other write domains, and we can update
3120 * the domain values for our changes.
3122 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3124 /* If we're writing through the CPU, then the GPU read domains will
3125 * need to be invalidated at next use.
3128 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3129 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3132 trace_i915_gem_object_change_domain(obj,
3140 * Moves the object from a partially CPU read to a full one.
3142 * Note that this only resolves i915_gem_object_set_cpu_read_domain_range(),
3143 * and doesn't handle transitioning from !(read_domains & I915_GEM_DOMAIN_CPU).
3146 i915_gem_object_set_to_full_cpu_read_domain(struct drm_i915_gem_object *obj)
3148 if (!obj->page_cpu_valid)
3151 /* If we're partially in the CPU read domain, finish moving it in.
3153 if (obj->base.read_domains & I915_GEM_DOMAIN_CPU) {
3156 for (i = 0; i <= (obj->base.size - 1) / PAGE_SIZE; i++) {
3157 if (obj->page_cpu_valid[i])
3159 drm_clflush_pages(obj->pages + i, 1);
3163 /* Free the page_cpu_valid mappings which are now stale, whether
3164 * or not we've got I915_GEM_DOMAIN_CPU.
3166 kfree(obj->page_cpu_valid);
3167 obj->page_cpu_valid = NULL;
3171 * Set the CPU read domain on a range of the object.
3173 * The object ends up with I915_GEM_DOMAIN_CPU in its read flags although it's
3174 * not entirely valid. The page_cpu_valid member of the object flags which
3175 * pages have been flushed, and will be respected by
3176 * i915_gem_object_set_to_cpu_domain() if it's called on to get a valid mapping
3177 * of the whole object.
3179 * This function returns when the move is complete, including waiting on
3183 i915_gem_object_set_cpu_read_domain_range(struct drm_i915_gem_object *obj,
3184 uint64_t offset, uint64_t size)
3186 uint32_t old_read_domains;
3189 if (offset == 0 && size == obj->base.size)
3190 return i915_gem_object_set_to_cpu_domain(obj, 0);
3192 ret = i915_gem_object_flush_gpu_write_domain(obj);
3196 ret = i915_gem_object_wait_rendering(obj);
3200 i915_gem_object_flush_gtt_write_domain(obj);
3202 /* If we're already fully in the CPU read domain, we're done. */
3203 if (obj->page_cpu_valid == NULL &&
3204 (obj->base.read_domains & I915_GEM_DOMAIN_CPU) != 0)
3207 /* Otherwise, create/clear the per-page CPU read domain flag if we're
3208 * newly adding I915_GEM_DOMAIN_CPU
3210 if (obj->page_cpu_valid == NULL) {
3211 obj->page_cpu_valid = kzalloc(obj->base.size / PAGE_SIZE,
3213 if (obj->page_cpu_valid == NULL)
3215 } else if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0)
3216 memset(obj->page_cpu_valid, 0, obj->base.size / PAGE_SIZE);
3218 /* Flush the cache on any pages that are still invalid from the CPU's
3221 for (i = offset / PAGE_SIZE; i <= (offset + size - 1) / PAGE_SIZE;
3223 if (obj->page_cpu_valid[i])
3226 drm_clflush_pages(obj->pages + i, 1);
3228 obj->page_cpu_valid[i] = 1;
3231 /* It should now be out of any other write domains, and we can update
3232 * the domain values for our changes.
3234 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3236 old_read_domains = obj->base.read_domains;
3237 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3239 trace_i915_gem_object_change_domain(obj,
3241 obj->base.write_domain);
3246 /* Throttle our rendering by waiting until the ring has completed our requests
3247 * emitted over 20 msec ago.
3249 * Note that if we were to use the current jiffies each time around the loop,
3250 * we wouldn't escape the function with any frames outstanding if the time to
3251 * render a frame was over 20ms.
3253 * This should get us reasonable parallelism between CPU and GPU but also
3254 * relatively low latency when blocking on a particular request to finish.
3257 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
3259 struct drm_i915_private *dev_priv = dev->dev_private;
3260 struct drm_i915_file_private *file_priv = file->driver_priv;
3261 unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3262 struct drm_i915_gem_request *request;
3263 struct intel_ring_buffer *ring = NULL;
3267 if (atomic_read(&dev_priv->mm.wedged))
3270 spin_lock(&file_priv->mm.lock);
3271 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
3272 if (time_after_eq(request->emitted_jiffies, recent_enough))
3275 ring = request->ring;
3276 seqno = request->seqno;
3278 spin_unlock(&file_priv->mm.lock);
3284 if (!i915_seqno_passed(ring->get_seqno(ring), seqno)) {
3285 /* And wait for the seqno passing without holding any locks and
3286 * causing extra latency for others. This is safe as the irq
3287 * generation is designed to be run atomically and so is
3290 if (ring->irq_get(ring)) {
3291 ret = wait_event_interruptible(ring->irq_queue,
3292 i915_seqno_passed(ring->get_seqno(ring), seqno)
3293 || atomic_read(&dev_priv->mm.wedged));
3294 ring->irq_put(ring);
3296 if (ret == 0 && atomic_read(&dev_priv->mm.wedged))
3298 } else if (wait_for_atomic(i915_seqno_passed(ring->get_seqno(ring),
3300 atomic_read(&dev_priv->mm.wedged), 3000)) {
3306 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
3312 i915_gem_object_pin(struct drm_i915_gem_object *obj,
3314 bool map_and_fenceable)
3316 struct drm_device *dev = obj->base.dev;
3317 struct drm_i915_private *dev_priv = dev->dev_private;
3320 BUG_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT);
3321 WARN_ON(i915_verify_lists(dev));
3323 if (obj->gtt_space != NULL) {
3324 if ((alignment && obj->gtt_offset & (alignment - 1)) ||
3325 (map_and_fenceable && !obj->map_and_fenceable)) {
3326 WARN(obj->pin_count,
3327 "bo is already pinned with incorrect alignment:"
3328 " offset=%x, req.alignment=%x, req.map_and_fenceable=%d,"
3329 " obj->map_and_fenceable=%d\n",
3330 obj->gtt_offset, alignment,
3332 obj->map_and_fenceable);
3333 ret = i915_gem_object_unbind(obj);
3339 if (obj->gtt_space == NULL) {
3340 ret = i915_gem_object_bind_to_gtt(obj, alignment,
3346 if (obj->pin_count++ == 0) {
3348 list_move_tail(&obj->mm_list,
3349 &dev_priv->mm.pinned_list);
3351 obj->pin_mappable |= map_and_fenceable;
3353 WARN_ON(i915_verify_lists(dev));
3358 i915_gem_object_unpin(struct drm_i915_gem_object *obj)
3360 struct drm_device *dev = obj->base.dev;
3361 drm_i915_private_t *dev_priv = dev->dev_private;
3363 WARN_ON(i915_verify_lists(dev));
3364 BUG_ON(obj->pin_count == 0);
3365 BUG_ON(obj->gtt_space == NULL);
3367 if (--obj->pin_count == 0) {
3369 list_move_tail(&obj->mm_list,
3370 &dev_priv->mm.inactive_list);
3371 obj->pin_mappable = false;
3373 WARN_ON(i915_verify_lists(dev));
3377 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
3378 struct drm_file *file)
3380 struct drm_i915_gem_pin *args = data;
3381 struct drm_i915_gem_object *obj;
3384 ret = i915_mutex_lock_interruptible(dev);
3388 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3389 if (&obj->base == NULL) {
3394 if (obj->madv != I915_MADV_WILLNEED) {
3395 DRM_ERROR("Attempting to pin a purgeable buffer\n");
3400 if (obj->pin_filp != NULL && obj->pin_filp != file) {
3401 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
3407 obj->user_pin_count++;
3408 obj->pin_filp = file;
3409 if (obj->user_pin_count == 1) {
3410 ret = i915_gem_object_pin(obj, args->alignment, true);
3415 /* XXX - flush the CPU caches for pinned objects
3416 * as the X server doesn't manage domains yet
3418 i915_gem_object_flush_cpu_write_domain(obj);
3419 args->offset = obj->gtt_offset;
3421 drm_gem_object_unreference(&obj->base);
3423 mutex_unlock(&dev->struct_mutex);
3428 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
3429 struct drm_file *file)
3431 struct drm_i915_gem_pin *args = data;
3432 struct drm_i915_gem_object *obj;
3435 ret = i915_mutex_lock_interruptible(dev);
3439 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3440 if (&obj->base == NULL) {
3445 if (obj->pin_filp != file) {
3446 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
3451 obj->user_pin_count--;
3452 if (obj->user_pin_count == 0) {
3453 obj->pin_filp = NULL;
3454 i915_gem_object_unpin(obj);
3458 drm_gem_object_unreference(&obj->base);
3460 mutex_unlock(&dev->struct_mutex);
3465 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
3466 struct drm_file *file)
3468 struct drm_i915_gem_busy *args = data;
3469 struct drm_i915_gem_object *obj;
3472 ret = i915_mutex_lock_interruptible(dev);
3476 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3477 if (&obj->base == NULL) {
3482 /* Count all active objects as busy, even if they are currently not used
3483 * by the gpu. Users of this interface expect objects to eventually
3484 * become non-busy without any further actions, therefore emit any
3485 * necessary flushes here.
3487 args->busy = obj->active;
3489 /* Unconditionally flush objects, even when the gpu still uses this
3490 * object. Userspace calling this function indicates that it wants to
3491 * use this buffer rather sooner than later, so issuing the required
3492 * flush earlier is beneficial.
3494 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
3495 ret = i915_gem_flush_ring(obj->ring,
3496 0, obj->base.write_domain);
3498 if (obj->ring->outstanding_lazy_request ==
3499 obj->last_rendering_seqno) {
3500 struct drm_i915_gem_request *request;
3502 /* This ring is not being cleared by active usage,
3503 * so emit a request to do so.
3505 request = kzalloc(sizeof(*request), GFP_KERNEL);
3507 ret = i915_add_request(obj->ring, NULL, request);
3514 /* Update the active list for the hardware's current position.
3515 * Otherwise this only updates on a delayed timer or when irqs
3516 * are actually unmasked, and our working set ends up being
3517 * larger than required.
3519 i915_gem_retire_requests_ring(obj->ring);
3521 args->busy = obj->active;
3524 drm_gem_object_unreference(&obj->base);
3526 mutex_unlock(&dev->struct_mutex);
3531 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
3532 struct drm_file *file_priv)
3534 return i915_gem_ring_throttle(dev, file_priv);
3538 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
3539 struct drm_file *file_priv)
3541 struct drm_i915_gem_madvise *args = data;
3542 struct drm_i915_gem_object *obj;
3545 switch (args->madv) {
3546 case I915_MADV_DONTNEED:
3547 case I915_MADV_WILLNEED:
3553 ret = i915_mutex_lock_interruptible(dev);
3557 obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
3558 if (&obj->base == NULL) {
3563 if (obj->pin_count) {
3568 if (obj->madv != __I915_MADV_PURGED)
3569 obj->madv = args->madv;
3571 /* if the object is no longer bound, discard its backing storage */
3572 if (i915_gem_object_is_purgeable(obj) &&
3573 obj->gtt_space == NULL)
3574 i915_gem_object_truncate(obj);
3576 args->retained = obj->madv != __I915_MADV_PURGED;
3579 drm_gem_object_unreference(&obj->base);
3581 mutex_unlock(&dev->struct_mutex);
3585 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
3588 struct drm_i915_private *dev_priv = dev->dev_private;
3589 struct drm_i915_gem_object *obj;
3590 struct address_space *mapping;
3592 obj = kzalloc(sizeof(*obj), GFP_KERNEL);
3596 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
3601 mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
3602 mapping_set_gfp_mask(mapping, GFP_HIGHUSER | __GFP_RECLAIMABLE);
3604 i915_gem_info_add_obj(dev_priv, size);
3606 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3607 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3610 /* On some devices, we can have the GPU use the LLC (the CPU
3611 * cache) for about a 10% performance improvement
3612 * compared to uncached. Graphics requests other than
3613 * display scanout are coherent with the CPU in
3614 * accessing this cache. This means in this mode we
3615 * don't need to clflush on the CPU side, and on the
3616 * GPU side we only need to flush internal caches to
3617 * get data visible to the CPU.
3619 * However, we maintain the display planes as UC, and so
3620 * need to rebind when first used as such.
3622 obj->cache_level = I915_CACHE_LLC;
3624 obj->cache_level = I915_CACHE_NONE;
3626 obj->base.driver_private = NULL;
3627 obj->fence_reg = I915_FENCE_REG_NONE;
3628 INIT_LIST_HEAD(&obj->mm_list);
3629 INIT_LIST_HEAD(&obj->gtt_list);
3630 INIT_LIST_HEAD(&obj->ring_list);
3631 INIT_LIST_HEAD(&obj->exec_list);
3632 INIT_LIST_HEAD(&obj->gpu_write_list);
3633 obj->madv = I915_MADV_WILLNEED;
3634 /* Avoid an unnecessary call to unbind on the first bind. */
3635 obj->map_and_fenceable = true;
3640 int i915_gem_init_object(struct drm_gem_object *obj)
3647 static void i915_gem_free_object_tail(struct drm_i915_gem_object *obj)
3649 struct drm_device *dev = obj->base.dev;
3650 drm_i915_private_t *dev_priv = dev->dev_private;
3653 ret = i915_gem_object_unbind(obj);
3654 if (ret == -ERESTARTSYS) {
3655 list_move(&obj->mm_list,
3656 &dev_priv->mm.deferred_free_list);
3660 trace_i915_gem_object_destroy(obj);
3662 if (obj->base.map_list.map)
3663 drm_gem_free_mmap_offset(&obj->base);
3665 drm_gem_object_release(&obj->base);
3666 i915_gem_info_remove_obj(dev_priv, obj->base.size);
3668 kfree(obj->page_cpu_valid);
3673 void i915_gem_free_object(struct drm_gem_object *gem_obj)
3675 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
3676 struct drm_device *dev = obj->base.dev;
3678 while (obj->pin_count > 0)
3679 i915_gem_object_unpin(obj);
3682 i915_gem_detach_phys_object(dev, obj);
3684 i915_gem_free_object_tail(obj);
3688 i915_gem_idle(struct drm_device *dev)
3690 drm_i915_private_t *dev_priv = dev->dev_private;
3693 mutex_lock(&dev->struct_mutex);
3695 if (dev_priv->mm.suspended) {
3696 mutex_unlock(&dev->struct_mutex);
3700 ret = i915_gpu_idle(dev, true);
3702 mutex_unlock(&dev->struct_mutex);
3706 /* Under UMS, be paranoid and evict. */
3707 if (!drm_core_check_feature(dev, DRIVER_MODESET)) {
3708 ret = i915_gem_evict_inactive(dev, false);
3710 mutex_unlock(&dev->struct_mutex);
3715 i915_gem_reset_fences(dev);
3717 /* Hack! Don't let anybody do execbuf while we don't control the chip.
3718 * We need to replace this with a semaphore, or something.
3719 * And not confound mm.suspended!
3721 dev_priv->mm.suspended = 1;
3722 del_timer_sync(&dev_priv->hangcheck_timer);
3724 i915_kernel_lost_context(dev);
3725 i915_gem_cleanup_ringbuffer(dev);
3727 mutex_unlock(&dev->struct_mutex);
3729 /* Cancel the retire work handler, which should be idle now. */
3730 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
3735 void i915_gem_init_swizzling(struct drm_device *dev)
3737 drm_i915_private_t *dev_priv = dev->dev_private;
3739 if (INTEL_INFO(dev)->gen < 5 ||
3740 dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
3743 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
3744 DISP_TILE_SURFACE_SWIZZLING);
3749 I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
3751 I915_WRITE(ARB_MODE, ARB_MODE_ENABLE(ARB_MODE_SWIZZLE_SNB));
3753 I915_WRITE(ARB_MODE, ARB_MODE_ENABLE(ARB_MODE_SWIZZLE_IVB));
3756 void i915_gem_init_ppgtt(struct drm_device *dev)
3758 drm_i915_private_t *dev_priv = dev->dev_private;
3760 struct intel_ring_buffer *ring;
3761 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
3762 uint32_t __iomem *pd_addr;
3766 if (!dev_priv->mm.aliasing_ppgtt)
3770 pd_addr = dev_priv->mm.gtt->gtt + ppgtt->pd_offset/sizeof(uint32_t);
3771 for (i = 0; i < ppgtt->num_pd_entries; i++) {
3774 if (dev_priv->mm.gtt->needs_dmar)
3775 pt_addr = ppgtt->pt_dma_addr[i];
3777 pt_addr = page_to_phys(ppgtt->pt_pages[i]);
3779 pd_entry = GEN6_PDE_ADDR_ENCODE(pt_addr);
3780 pd_entry |= GEN6_PDE_VALID;
3782 writel(pd_entry, pd_addr + i);
3786 pd_offset = ppgtt->pd_offset;
3787 pd_offset /= 64; /* in cachelines, */
3790 if (INTEL_INFO(dev)->gen == 6) {
3791 uint32_t ecochk = I915_READ(GAM_ECOCHK);
3792 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT |
3793 ECOCHK_PPGTT_CACHE64B);
3794 I915_WRITE(GFX_MODE, GFX_MODE_ENABLE(GFX_PPGTT_ENABLE));
3795 } else if (INTEL_INFO(dev)->gen >= 7) {
3796 I915_WRITE(GAM_ECOCHK, ECOCHK_PPGTT_CACHE64B);
3797 /* GFX_MODE is per-ring on gen7+ */
3800 for (i = 0; i < I915_NUM_RINGS; i++) {
3801 ring = &dev_priv->ring[i];
3803 if (INTEL_INFO(dev)->gen >= 7)
3804 I915_WRITE(RING_MODE_GEN7(ring),
3805 GFX_MODE_ENABLE(GFX_PPGTT_ENABLE));
3807 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
3808 I915_WRITE(RING_PP_DIR_BASE(ring), pd_offset);
3813 i915_gem_init_hw(struct drm_device *dev)
3815 drm_i915_private_t *dev_priv = dev->dev_private;
3818 i915_gem_init_swizzling(dev);
3820 ret = intel_init_render_ring_buffer(dev);
3825 ret = intel_init_bsd_ring_buffer(dev);
3827 goto cleanup_render_ring;
3831 ret = intel_init_blt_ring_buffer(dev);
3833 goto cleanup_bsd_ring;
3836 dev_priv->next_seqno = 1;
3838 i915_gem_init_ppgtt(dev);
3843 intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
3844 cleanup_render_ring:
3845 intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
3850 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
3852 drm_i915_private_t *dev_priv = dev->dev_private;
3855 for (i = 0; i < I915_NUM_RINGS; i++)
3856 intel_cleanup_ring_buffer(&dev_priv->ring[i]);
3860 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
3861 struct drm_file *file_priv)
3863 drm_i915_private_t *dev_priv = dev->dev_private;
3866 if (drm_core_check_feature(dev, DRIVER_MODESET))
3869 if (atomic_read(&dev_priv->mm.wedged)) {
3870 DRM_ERROR("Reenabling wedged hardware, good luck\n");
3871 atomic_set(&dev_priv->mm.wedged, 0);
3874 mutex_lock(&dev->struct_mutex);
3875 dev_priv->mm.suspended = 0;
3877 ret = i915_gem_init_hw(dev);
3879 mutex_unlock(&dev->struct_mutex);
3883 BUG_ON(!list_empty(&dev_priv->mm.active_list));
3884 BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
3885 BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
3886 for (i = 0; i < I915_NUM_RINGS; i++) {
3887 BUG_ON(!list_empty(&dev_priv->ring[i].active_list));
3888 BUG_ON(!list_empty(&dev_priv->ring[i].request_list));
3890 mutex_unlock(&dev->struct_mutex);
3892 ret = drm_irq_install(dev);
3894 goto cleanup_ringbuffer;
3899 mutex_lock(&dev->struct_mutex);
3900 i915_gem_cleanup_ringbuffer(dev);
3901 dev_priv->mm.suspended = 1;
3902 mutex_unlock(&dev->struct_mutex);
3908 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
3909 struct drm_file *file_priv)
3911 if (drm_core_check_feature(dev, DRIVER_MODESET))
3914 drm_irq_uninstall(dev);
3915 return i915_gem_idle(dev);
3919 i915_gem_lastclose(struct drm_device *dev)
3923 if (drm_core_check_feature(dev, DRIVER_MODESET))
3926 ret = i915_gem_idle(dev);
3928 DRM_ERROR("failed to idle hardware: %d\n", ret);
3932 init_ring_lists(struct intel_ring_buffer *ring)
3934 INIT_LIST_HEAD(&ring->active_list);
3935 INIT_LIST_HEAD(&ring->request_list);
3936 INIT_LIST_HEAD(&ring->gpu_write_list);
3940 i915_gem_load(struct drm_device *dev)
3943 drm_i915_private_t *dev_priv = dev->dev_private;
3945 INIT_LIST_HEAD(&dev_priv->mm.active_list);
3946 INIT_LIST_HEAD(&dev_priv->mm.flushing_list);
3947 INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
3948 INIT_LIST_HEAD(&dev_priv->mm.pinned_list);
3949 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
3950 INIT_LIST_HEAD(&dev_priv->mm.deferred_free_list);
3951 INIT_LIST_HEAD(&dev_priv->mm.gtt_list);
3952 for (i = 0; i < I915_NUM_RINGS; i++)
3953 init_ring_lists(&dev_priv->ring[i]);
3954 for (i = 0; i < I915_MAX_NUM_FENCES; i++)
3955 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
3956 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
3957 i915_gem_retire_work_handler);
3958 init_completion(&dev_priv->error_completion);
3960 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
3962 u32 tmp = I915_READ(MI_ARB_STATE);
3963 if (!(tmp & MI_ARB_C3_LP_WRITE_ENABLE)) {
3964 /* arb state is a masked write, so set bit + bit in mask */
3965 tmp = MI_ARB_C3_LP_WRITE_ENABLE | (MI_ARB_C3_LP_WRITE_ENABLE << MI_ARB_MASK_SHIFT);
3966 I915_WRITE(MI_ARB_STATE, tmp);
3970 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
3972 /* Old X drivers will take 0-2 for front, back, depth buffers */
3973 if (!drm_core_check_feature(dev, DRIVER_MODESET))
3974 dev_priv->fence_reg_start = 3;
3976 if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
3977 dev_priv->num_fence_regs = 16;
3979 dev_priv->num_fence_regs = 8;
3981 /* Initialize fence registers to zero */
3982 for (i = 0; i < dev_priv->num_fence_regs; i++) {
3983 i915_gem_clear_fence_reg(dev, &dev_priv->fence_regs[i]);
3986 i915_gem_detect_bit_6_swizzle(dev);
3987 init_waitqueue_head(&dev_priv->pending_flip_queue);
3989 dev_priv->mm.interruptible = true;
3991 dev_priv->mm.inactive_shrinker.shrink = i915_gem_inactive_shrink;
3992 dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS;
3993 register_shrinker(&dev_priv->mm.inactive_shrinker);
3997 * Create a physically contiguous memory object for this object
3998 * e.g. for cursor + overlay regs
4000 static int i915_gem_init_phys_object(struct drm_device *dev,
4001 int id, int size, int align)
4003 drm_i915_private_t *dev_priv = dev->dev_private;
4004 struct drm_i915_gem_phys_object *phys_obj;
4007 if (dev_priv->mm.phys_objs[id - 1] || !size)
4010 phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL);
4016 phys_obj->handle = drm_pci_alloc(dev, size, align);
4017 if (!phys_obj->handle) {
4022 set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4025 dev_priv->mm.phys_objs[id - 1] = phys_obj;
4033 static void i915_gem_free_phys_object(struct drm_device *dev, int id)
4035 drm_i915_private_t *dev_priv = dev->dev_private;
4036 struct drm_i915_gem_phys_object *phys_obj;
4038 if (!dev_priv->mm.phys_objs[id - 1])
4041 phys_obj = dev_priv->mm.phys_objs[id - 1];
4042 if (phys_obj->cur_obj) {
4043 i915_gem_detach_phys_object(dev, phys_obj->cur_obj);
4047 set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4049 drm_pci_free(dev, phys_obj->handle);
4051 dev_priv->mm.phys_objs[id - 1] = NULL;
4054 void i915_gem_free_all_phys_object(struct drm_device *dev)
4058 for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++)
4059 i915_gem_free_phys_object(dev, i);
4062 void i915_gem_detach_phys_object(struct drm_device *dev,
4063 struct drm_i915_gem_object *obj)
4065 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
4072 vaddr = obj->phys_obj->handle->vaddr;
4074 page_count = obj->base.size / PAGE_SIZE;
4075 for (i = 0; i < page_count; i++) {
4076 struct page *page = shmem_read_mapping_page(mapping, i);
4077 if (!IS_ERR(page)) {
4078 char *dst = kmap_atomic(page);
4079 memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE);
4082 drm_clflush_pages(&page, 1);
4084 set_page_dirty(page);
4085 mark_page_accessed(page);
4086 page_cache_release(page);
4089 intel_gtt_chipset_flush();
4091 obj->phys_obj->cur_obj = NULL;
4092 obj->phys_obj = NULL;
4096 i915_gem_attach_phys_object(struct drm_device *dev,
4097 struct drm_i915_gem_object *obj,
4101 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
4102 drm_i915_private_t *dev_priv = dev->dev_private;
4107 if (id > I915_MAX_PHYS_OBJECT)
4110 if (obj->phys_obj) {
4111 if (obj->phys_obj->id == id)
4113 i915_gem_detach_phys_object(dev, obj);
4116 /* create a new object */
4117 if (!dev_priv->mm.phys_objs[id - 1]) {
4118 ret = i915_gem_init_phys_object(dev, id,
4119 obj->base.size, align);
4121 DRM_ERROR("failed to init phys object %d size: %zu\n",
4122 id, obj->base.size);
4127 /* bind to the object */
4128 obj->phys_obj = dev_priv->mm.phys_objs[id - 1];
4129 obj->phys_obj->cur_obj = obj;
4131 page_count = obj->base.size / PAGE_SIZE;
4133 for (i = 0; i < page_count; i++) {
4137 page = shmem_read_mapping_page(mapping, i);
4139 return PTR_ERR(page);
4141 src = kmap_atomic(page);
4142 dst = obj->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4143 memcpy(dst, src, PAGE_SIZE);
4146 mark_page_accessed(page);
4147 page_cache_release(page);
4154 i915_gem_phys_pwrite(struct drm_device *dev,
4155 struct drm_i915_gem_object *obj,
4156 struct drm_i915_gem_pwrite *args,
4157 struct drm_file *file_priv)
4159 void *vaddr = obj->phys_obj->handle->vaddr + args->offset;
4160 char __user *user_data = (char __user *) (uintptr_t) args->data_ptr;
4162 if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
4163 unsigned long unwritten;
4165 /* The physical object once assigned is fixed for the lifetime
4166 * of the obj, so we can safely drop the lock and continue
4169 mutex_unlock(&dev->struct_mutex);
4170 unwritten = copy_from_user(vaddr, user_data, args->size);
4171 mutex_lock(&dev->struct_mutex);
4176 intel_gtt_chipset_flush();
4180 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
4182 struct drm_i915_file_private *file_priv = file->driver_priv;
4184 /* Clean up our request list when the client is going away, so that
4185 * later retire_requests won't dereference our soon-to-be-gone
4188 spin_lock(&file_priv->mm.lock);
4189 while (!list_empty(&file_priv->mm.request_list)) {
4190 struct drm_i915_gem_request *request;
4192 request = list_first_entry(&file_priv->mm.request_list,
4193 struct drm_i915_gem_request,
4195 list_del(&request->client_list);
4196 request->file_priv = NULL;
4198 spin_unlock(&file_priv->mm.lock);
4202 i915_gpu_is_active(struct drm_device *dev)
4204 drm_i915_private_t *dev_priv = dev->dev_private;
4207 lists_empty = list_empty(&dev_priv->mm.flushing_list) &&
4208 list_empty(&dev_priv->mm.active_list);
4210 return !lists_empty;
4214 i915_gem_inactive_shrink(struct shrinker *shrinker, struct shrink_control *sc)
4216 struct drm_i915_private *dev_priv =
4217 container_of(shrinker,
4218 struct drm_i915_private,
4219 mm.inactive_shrinker);
4220 struct drm_device *dev = dev_priv->dev;
4221 struct drm_i915_gem_object *obj, *next;
4222 int nr_to_scan = sc->nr_to_scan;
4225 if (!mutex_trylock(&dev->struct_mutex))
4228 /* "fast-path" to count number of available objects */
4229 if (nr_to_scan == 0) {
4231 list_for_each_entry(obj,
4232 &dev_priv->mm.inactive_list,
4235 mutex_unlock(&dev->struct_mutex);
4236 return cnt / 100 * sysctl_vfs_cache_pressure;
4240 /* first scan for clean buffers */
4241 i915_gem_retire_requests(dev);
4243 list_for_each_entry_safe(obj, next,
4244 &dev_priv->mm.inactive_list,
4246 if (i915_gem_object_is_purgeable(obj)) {
4247 if (i915_gem_object_unbind(obj) == 0 &&
4253 /* second pass, evict/count anything still on the inactive list */
4255 list_for_each_entry_safe(obj, next,
4256 &dev_priv->mm.inactive_list,
4259 i915_gem_object_unbind(obj) == 0)
4265 if (nr_to_scan && i915_gpu_is_active(dev)) {
4267 * We are desperate for pages, so as a last resort, wait
4268 * for the GPU to finish and discard whatever we can.
4269 * This has a dramatic impact to reduce the number of
4270 * OOM-killer events whilst running the GPU aggressively.
4272 if (i915_gpu_idle(dev, true) == 0)
4275 mutex_unlock(&dev->struct_mutex);
4276 return cnt / 100 * sysctl_vfs_cache_pressure;