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 down_write(¤t->mm->mmap_sem);
1091 addr = do_mmap(obj->filp, 0, args->size,
1092 PROT_READ | PROT_WRITE, MAP_SHARED,
1094 up_write(¤t->mm->mmap_sem);
1095 drm_gem_object_unreference_unlocked(obj);
1096 if (IS_ERR((void *)addr))
1099 args->addr_ptr = (uint64_t) addr;
1105 * i915_gem_fault - fault a page into the GTT
1106 * vma: VMA in question
1109 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1110 * from userspace. The fault handler takes care of binding the object to
1111 * the GTT (if needed), allocating and programming a fence register (again,
1112 * only if needed based on whether the old reg is still valid or the object
1113 * is tiled) and inserting a new PTE into the faulting process.
1115 * Note that the faulting process may involve evicting existing objects
1116 * from the GTT and/or fence registers to make room. So performance may
1117 * suffer if the GTT working set is large or there are few fence registers
1120 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1122 struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1123 struct drm_device *dev = obj->base.dev;
1124 drm_i915_private_t *dev_priv = dev->dev_private;
1125 pgoff_t page_offset;
1128 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1130 /* We don't use vmf->pgoff since that has the fake offset */
1131 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1134 ret = i915_mutex_lock_interruptible(dev);
1138 trace_i915_gem_object_fault(obj, page_offset, true, write);
1140 /* Now bind it into the GTT if needed */
1141 if (!obj->map_and_fenceable) {
1142 ret = i915_gem_object_unbind(obj);
1146 if (!obj->gtt_space) {
1147 ret = i915_gem_object_bind_to_gtt(obj, 0, true);
1151 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1156 if (obj->tiling_mode == I915_TILING_NONE)
1157 ret = i915_gem_object_put_fence(obj);
1159 ret = i915_gem_object_get_fence(obj, NULL);
1163 if (i915_gem_object_is_inactive(obj))
1164 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1166 obj->fault_mappable = true;
1168 pfn = ((dev->agp->base + obj->gtt_offset) >> PAGE_SHIFT) +
1171 /* Finally, remap it using the new GTT offset */
1172 ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1174 mutex_unlock(&dev->struct_mutex);
1179 /* Give the error handler a chance to run and move the
1180 * objects off the GPU active list. Next time we service the
1181 * fault, we should be able to transition the page into the
1182 * GTT without touching the GPU (and so avoid further
1183 * EIO/EGAIN). If the GPU is wedged, then there is no issue
1184 * with coherency, just lost writes.
1190 return VM_FAULT_NOPAGE;
1192 return VM_FAULT_OOM;
1194 return VM_FAULT_SIGBUS;
1199 * i915_gem_release_mmap - remove physical page mappings
1200 * @obj: obj in question
1202 * Preserve the reservation of the mmapping with the DRM core code, but
1203 * relinquish ownership of the pages back to the system.
1205 * It is vital that we remove the page mapping if we have mapped a tiled
1206 * object through the GTT and then lose the fence register due to
1207 * resource pressure. Similarly if the object has been moved out of the
1208 * aperture, than pages mapped into userspace must be revoked. Removing the
1209 * mapping will then trigger a page fault on the next user access, allowing
1210 * fixup by i915_gem_fault().
1213 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1215 if (!obj->fault_mappable)
1218 if (obj->base.dev->dev_mapping)
1219 unmap_mapping_range(obj->base.dev->dev_mapping,
1220 (loff_t)obj->base.map_list.hash.key<<PAGE_SHIFT,
1223 obj->fault_mappable = false;
1227 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1231 if (INTEL_INFO(dev)->gen >= 4 ||
1232 tiling_mode == I915_TILING_NONE)
1235 /* Previous chips need a power-of-two fence region when tiling */
1236 if (INTEL_INFO(dev)->gen == 3)
1237 gtt_size = 1024*1024;
1239 gtt_size = 512*1024;
1241 while (gtt_size < size)
1248 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1249 * @obj: object to check
1251 * Return the required GTT alignment for an object, taking into account
1252 * potential fence register mapping.
1255 i915_gem_get_gtt_alignment(struct drm_device *dev,
1260 * Minimum alignment is 4k (GTT page size), but might be greater
1261 * if a fence register is needed for the object.
1263 if (INTEL_INFO(dev)->gen >= 4 ||
1264 tiling_mode == I915_TILING_NONE)
1268 * Previous chips need to be aligned to the size of the smallest
1269 * fence register that can contain the object.
1271 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1275 * i915_gem_get_unfenced_gtt_alignment - return required GTT alignment for an
1278 * @size: size of the object
1279 * @tiling_mode: tiling mode of the object
1281 * Return the required GTT alignment for an object, only taking into account
1282 * unfenced tiled surface requirements.
1285 i915_gem_get_unfenced_gtt_alignment(struct drm_device *dev,
1290 * Minimum alignment is 4k (GTT page size) for sane hw.
1292 if (INTEL_INFO(dev)->gen >= 4 || IS_G33(dev) ||
1293 tiling_mode == I915_TILING_NONE)
1296 /* Previous hardware however needs to be aligned to a power-of-two
1297 * tile height. The simplest method for determining this is to reuse
1298 * the power-of-tile object size.
1300 return i915_gem_get_gtt_size(dev, size, tiling_mode);
1304 i915_gem_mmap_gtt(struct drm_file *file,
1305 struct drm_device *dev,
1309 struct drm_i915_private *dev_priv = dev->dev_private;
1310 struct drm_i915_gem_object *obj;
1313 if (!(dev->driver->driver_features & DRIVER_GEM))
1316 ret = i915_mutex_lock_interruptible(dev);
1320 obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
1321 if (&obj->base == NULL) {
1326 if (obj->base.size > dev_priv->mm.gtt_mappable_end) {
1331 if (obj->madv != I915_MADV_WILLNEED) {
1332 DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1337 if (!obj->base.map_list.map) {
1338 ret = drm_gem_create_mmap_offset(&obj->base);
1343 *offset = (u64)obj->base.map_list.hash.key << PAGE_SHIFT;
1346 drm_gem_object_unreference(&obj->base);
1348 mutex_unlock(&dev->struct_mutex);
1353 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1355 * @data: GTT mapping ioctl data
1356 * @file: GEM object info
1358 * Simply returns the fake offset to userspace so it can mmap it.
1359 * The mmap call will end up in drm_gem_mmap(), which will set things
1360 * up so we can get faults in the handler above.
1362 * The fault handler will take care of binding the object into the GTT
1363 * (since it may have been evicted to make room for something), allocating
1364 * a fence register, and mapping the appropriate aperture address into
1368 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1369 struct drm_file *file)
1371 struct drm_i915_gem_mmap_gtt *args = data;
1373 if (!(dev->driver->driver_features & DRIVER_GEM))
1376 return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
1381 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj,
1385 struct address_space *mapping;
1386 struct inode *inode;
1389 /* Get the list of pages out of our struct file. They'll be pinned
1390 * at this point until we release them.
1392 page_count = obj->base.size / PAGE_SIZE;
1393 BUG_ON(obj->pages != NULL);
1394 obj->pages = drm_malloc_ab(page_count, sizeof(struct page *));
1395 if (obj->pages == NULL)
1398 inode = obj->base.filp->f_path.dentry->d_inode;
1399 mapping = inode->i_mapping;
1400 gfpmask |= mapping_gfp_mask(mapping);
1402 for (i = 0; i < page_count; i++) {
1403 page = shmem_read_mapping_page_gfp(mapping, i, gfpmask);
1407 obj->pages[i] = page;
1410 if (i915_gem_object_needs_bit17_swizzle(obj))
1411 i915_gem_object_do_bit_17_swizzle(obj);
1417 page_cache_release(obj->pages[i]);
1419 drm_free_large(obj->pages);
1421 return PTR_ERR(page);
1425 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
1427 int page_count = obj->base.size / PAGE_SIZE;
1430 BUG_ON(obj->madv == __I915_MADV_PURGED);
1432 if (i915_gem_object_needs_bit17_swizzle(obj))
1433 i915_gem_object_save_bit_17_swizzle(obj);
1435 if (obj->madv == I915_MADV_DONTNEED)
1438 for (i = 0; i < page_count; i++) {
1440 set_page_dirty(obj->pages[i]);
1442 if (obj->madv == I915_MADV_WILLNEED)
1443 mark_page_accessed(obj->pages[i]);
1445 page_cache_release(obj->pages[i]);
1449 drm_free_large(obj->pages);
1454 i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
1455 struct intel_ring_buffer *ring,
1458 struct drm_device *dev = obj->base.dev;
1459 struct drm_i915_private *dev_priv = dev->dev_private;
1461 BUG_ON(ring == NULL);
1464 /* Add a reference if we're newly entering the active list. */
1466 drm_gem_object_reference(&obj->base);
1470 /* Move from whatever list we were on to the tail of execution. */
1471 list_move_tail(&obj->mm_list, &dev_priv->mm.active_list);
1472 list_move_tail(&obj->ring_list, &ring->active_list);
1474 obj->last_rendering_seqno = seqno;
1475 if (obj->fenced_gpu_access) {
1476 struct drm_i915_fence_reg *reg;
1478 BUG_ON(obj->fence_reg == I915_FENCE_REG_NONE);
1480 obj->last_fenced_seqno = seqno;
1481 obj->last_fenced_ring = ring;
1483 reg = &dev_priv->fence_regs[obj->fence_reg];
1484 list_move_tail(®->lru_list, &dev_priv->mm.fence_list);
1489 i915_gem_object_move_off_active(struct drm_i915_gem_object *obj)
1491 list_del_init(&obj->ring_list);
1492 obj->last_rendering_seqno = 0;
1496 i915_gem_object_move_to_flushing(struct drm_i915_gem_object *obj)
1498 struct drm_device *dev = obj->base.dev;
1499 drm_i915_private_t *dev_priv = dev->dev_private;
1501 BUG_ON(!obj->active);
1502 list_move_tail(&obj->mm_list, &dev_priv->mm.flushing_list);
1504 i915_gem_object_move_off_active(obj);
1508 i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
1510 struct drm_device *dev = obj->base.dev;
1511 struct drm_i915_private *dev_priv = dev->dev_private;
1513 if (obj->pin_count != 0)
1514 list_move_tail(&obj->mm_list, &dev_priv->mm.pinned_list);
1516 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1518 BUG_ON(!list_empty(&obj->gpu_write_list));
1519 BUG_ON(!obj->active);
1522 i915_gem_object_move_off_active(obj);
1523 obj->fenced_gpu_access = false;
1526 obj->pending_gpu_write = false;
1527 drm_gem_object_unreference(&obj->base);
1529 WARN_ON(i915_verify_lists(dev));
1532 /* Immediately discard the backing storage */
1534 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
1536 struct inode *inode;
1538 /* Our goal here is to return as much of the memory as
1539 * is possible back to the system as we are called from OOM.
1540 * To do this we must instruct the shmfs to drop all of its
1541 * backing pages, *now*.
1543 inode = obj->base.filp->f_path.dentry->d_inode;
1544 shmem_truncate_range(inode, 0, (loff_t)-1);
1546 obj->madv = __I915_MADV_PURGED;
1550 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj)
1552 return obj->madv == I915_MADV_DONTNEED;
1556 i915_gem_process_flushing_list(struct intel_ring_buffer *ring,
1557 uint32_t flush_domains)
1559 struct drm_i915_gem_object *obj, *next;
1561 list_for_each_entry_safe(obj, next,
1562 &ring->gpu_write_list,
1564 if (obj->base.write_domain & flush_domains) {
1565 uint32_t old_write_domain = obj->base.write_domain;
1567 obj->base.write_domain = 0;
1568 list_del_init(&obj->gpu_write_list);
1569 i915_gem_object_move_to_active(obj, ring,
1570 i915_gem_next_request_seqno(ring));
1572 trace_i915_gem_object_change_domain(obj,
1573 obj->base.read_domains,
1580 i915_gem_get_seqno(struct drm_device *dev)
1582 drm_i915_private_t *dev_priv = dev->dev_private;
1583 u32 seqno = dev_priv->next_seqno;
1585 /* reserve 0 for non-seqno */
1586 if (++dev_priv->next_seqno == 0)
1587 dev_priv->next_seqno = 1;
1593 i915_gem_next_request_seqno(struct intel_ring_buffer *ring)
1595 if (ring->outstanding_lazy_request == 0)
1596 ring->outstanding_lazy_request = i915_gem_get_seqno(ring->dev);
1598 return ring->outstanding_lazy_request;
1602 i915_add_request(struct intel_ring_buffer *ring,
1603 struct drm_file *file,
1604 struct drm_i915_gem_request *request)
1606 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1608 u32 request_ring_position;
1612 BUG_ON(request == NULL);
1613 seqno = i915_gem_next_request_seqno(ring);
1615 /* Record the position of the start of the request so that
1616 * should we detect the updated seqno part-way through the
1617 * GPU processing the request, we never over-estimate the
1618 * position of the head.
1620 request_ring_position = intel_ring_get_tail(ring);
1622 ret = ring->add_request(ring, &seqno);
1626 trace_i915_gem_request_add(ring, seqno);
1628 request->seqno = seqno;
1629 request->ring = ring;
1630 request->tail = request_ring_position;
1631 request->emitted_jiffies = jiffies;
1632 was_empty = list_empty(&ring->request_list);
1633 list_add_tail(&request->list, &ring->request_list);
1636 struct drm_i915_file_private *file_priv = file->driver_priv;
1638 spin_lock(&file_priv->mm.lock);
1639 request->file_priv = file_priv;
1640 list_add_tail(&request->client_list,
1641 &file_priv->mm.request_list);
1642 spin_unlock(&file_priv->mm.lock);
1645 ring->outstanding_lazy_request = 0;
1647 if (!dev_priv->mm.suspended) {
1648 if (i915_enable_hangcheck) {
1649 mod_timer(&dev_priv->hangcheck_timer,
1651 msecs_to_jiffies(DRM_I915_HANGCHECK_PERIOD));
1654 queue_delayed_work(dev_priv->wq,
1655 &dev_priv->mm.retire_work, HZ);
1661 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
1663 struct drm_i915_file_private *file_priv = request->file_priv;
1668 spin_lock(&file_priv->mm.lock);
1669 if (request->file_priv) {
1670 list_del(&request->client_list);
1671 request->file_priv = NULL;
1673 spin_unlock(&file_priv->mm.lock);
1676 static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv,
1677 struct intel_ring_buffer *ring)
1679 while (!list_empty(&ring->request_list)) {
1680 struct drm_i915_gem_request *request;
1682 request = list_first_entry(&ring->request_list,
1683 struct drm_i915_gem_request,
1686 list_del(&request->list);
1687 i915_gem_request_remove_from_client(request);
1691 while (!list_empty(&ring->active_list)) {
1692 struct drm_i915_gem_object *obj;
1694 obj = list_first_entry(&ring->active_list,
1695 struct drm_i915_gem_object,
1698 obj->base.write_domain = 0;
1699 list_del_init(&obj->gpu_write_list);
1700 i915_gem_object_move_to_inactive(obj);
1704 static void i915_gem_reset_fences(struct drm_device *dev)
1706 struct drm_i915_private *dev_priv = dev->dev_private;
1709 for (i = 0; i < dev_priv->num_fence_regs; i++) {
1710 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
1711 struct drm_i915_gem_object *obj = reg->obj;
1716 if (obj->tiling_mode)
1717 i915_gem_release_mmap(obj);
1719 reg->obj->fence_reg = I915_FENCE_REG_NONE;
1720 reg->obj->fenced_gpu_access = false;
1721 reg->obj->last_fenced_seqno = 0;
1722 reg->obj->last_fenced_ring = NULL;
1723 i915_gem_clear_fence_reg(dev, reg);
1727 void i915_gem_reset(struct drm_device *dev)
1729 struct drm_i915_private *dev_priv = dev->dev_private;
1730 struct drm_i915_gem_object *obj;
1733 for (i = 0; i < I915_NUM_RINGS; i++)
1734 i915_gem_reset_ring_lists(dev_priv, &dev_priv->ring[i]);
1736 /* Remove anything from the flushing lists. The GPU cache is likely
1737 * to be lost on reset along with the data, so simply move the
1738 * lost bo to the inactive list.
1740 while (!list_empty(&dev_priv->mm.flushing_list)) {
1741 obj = list_first_entry(&dev_priv->mm.flushing_list,
1742 struct drm_i915_gem_object,
1745 obj->base.write_domain = 0;
1746 list_del_init(&obj->gpu_write_list);
1747 i915_gem_object_move_to_inactive(obj);
1750 /* Move everything out of the GPU domains to ensure we do any
1751 * necessary invalidation upon reuse.
1753 list_for_each_entry(obj,
1754 &dev_priv->mm.inactive_list,
1757 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
1760 /* The fence registers are invalidated so clear them out */
1761 i915_gem_reset_fences(dev);
1765 * This function clears the request list as sequence numbers are passed.
1768 i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
1773 if (list_empty(&ring->request_list))
1776 WARN_ON(i915_verify_lists(ring->dev));
1778 seqno = ring->get_seqno(ring);
1780 for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++)
1781 if (seqno >= ring->sync_seqno[i])
1782 ring->sync_seqno[i] = 0;
1784 while (!list_empty(&ring->request_list)) {
1785 struct drm_i915_gem_request *request;
1787 request = list_first_entry(&ring->request_list,
1788 struct drm_i915_gem_request,
1791 if (!i915_seqno_passed(seqno, request->seqno))
1794 trace_i915_gem_request_retire(ring, request->seqno);
1795 /* We know the GPU must have read the request to have
1796 * sent us the seqno + interrupt, so use the position
1797 * of tail of the request to update the last known position
1800 ring->last_retired_head = request->tail;
1802 list_del(&request->list);
1803 i915_gem_request_remove_from_client(request);
1807 /* Move any buffers on the active list that are no longer referenced
1808 * by the ringbuffer to the flushing/inactive lists as appropriate.
1810 while (!list_empty(&ring->active_list)) {
1811 struct drm_i915_gem_object *obj;
1813 obj = list_first_entry(&ring->active_list,
1814 struct drm_i915_gem_object,
1817 if (!i915_seqno_passed(seqno, obj->last_rendering_seqno))
1820 if (obj->base.write_domain != 0)
1821 i915_gem_object_move_to_flushing(obj);
1823 i915_gem_object_move_to_inactive(obj);
1826 if (unlikely(ring->trace_irq_seqno &&
1827 i915_seqno_passed(seqno, ring->trace_irq_seqno))) {
1828 ring->irq_put(ring);
1829 ring->trace_irq_seqno = 0;
1832 WARN_ON(i915_verify_lists(ring->dev));
1836 i915_gem_retire_requests(struct drm_device *dev)
1838 drm_i915_private_t *dev_priv = dev->dev_private;
1841 if (!list_empty(&dev_priv->mm.deferred_free_list)) {
1842 struct drm_i915_gem_object *obj, *next;
1844 /* We must be careful that during unbind() we do not
1845 * accidentally infinitely recurse into retire requests.
1847 * retire -> free -> unbind -> wait -> retire_ring
1849 list_for_each_entry_safe(obj, next,
1850 &dev_priv->mm.deferred_free_list,
1852 i915_gem_free_object_tail(obj);
1855 for (i = 0; i < I915_NUM_RINGS; i++)
1856 i915_gem_retire_requests_ring(&dev_priv->ring[i]);
1860 i915_gem_retire_work_handler(struct work_struct *work)
1862 drm_i915_private_t *dev_priv;
1863 struct drm_device *dev;
1867 dev_priv = container_of(work, drm_i915_private_t,
1868 mm.retire_work.work);
1869 dev = dev_priv->dev;
1871 /* Come back later if the device is busy... */
1872 if (!mutex_trylock(&dev->struct_mutex)) {
1873 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1877 i915_gem_retire_requests(dev);
1879 /* Send a periodic flush down the ring so we don't hold onto GEM
1880 * objects indefinitely.
1883 for (i = 0; i < I915_NUM_RINGS; i++) {
1884 struct intel_ring_buffer *ring = &dev_priv->ring[i];
1886 if (!list_empty(&ring->gpu_write_list)) {
1887 struct drm_i915_gem_request *request;
1890 ret = i915_gem_flush_ring(ring,
1891 0, I915_GEM_GPU_DOMAINS);
1892 request = kzalloc(sizeof(*request), GFP_KERNEL);
1893 if (ret || request == NULL ||
1894 i915_add_request(ring, NULL, request))
1898 idle &= list_empty(&ring->request_list);
1901 if (!dev_priv->mm.suspended && !idle)
1902 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1904 mutex_unlock(&dev->struct_mutex);
1908 * Waits for a sequence number to be signaled, and cleans up the
1909 * request and object lists appropriately for that event.
1912 i915_wait_request(struct intel_ring_buffer *ring,
1916 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1922 if (atomic_read(&dev_priv->mm.wedged)) {
1923 struct completion *x = &dev_priv->error_completion;
1924 bool recovery_complete;
1925 unsigned long flags;
1927 /* Give the error handler a chance to run. */
1928 spin_lock_irqsave(&x->wait.lock, flags);
1929 recovery_complete = x->done > 0;
1930 spin_unlock_irqrestore(&x->wait.lock, flags);
1932 return recovery_complete ? -EIO : -EAGAIN;
1935 if (seqno == ring->outstanding_lazy_request) {
1936 struct drm_i915_gem_request *request;
1938 request = kzalloc(sizeof(*request), GFP_KERNEL);
1939 if (request == NULL)
1942 ret = i915_add_request(ring, NULL, request);
1948 seqno = request->seqno;
1951 if (!i915_seqno_passed(ring->get_seqno(ring), seqno)) {
1952 if (HAS_PCH_SPLIT(ring->dev))
1953 ier = I915_READ(DEIER) | I915_READ(GTIER);
1955 ier = I915_READ(IER);
1957 DRM_ERROR("something (likely vbetool) disabled "
1958 "interrupts, re-enabling\n");
1959 ring->dev->driver->irq_preinstall(ring->dev);
1960 ring->dev->driver->irq_postinstall(ring->dev);
1963 trace_i915_gem_request_wait_begin(ring, seqno);
1965 ring->waiting_seqno = seqno;
1966 if (ring->irq_get(ring)) {
1967 if (dev_priv->mm.interruptible)
1968 ret = wait_event_interruptible(ring->irq_queue,
1969 i915_seqno_passed(ring->get_seqno(ring), seqno)
1970 || atomic_read(&dev_priv->mm.wedged));
1972 wait_event(ring->irq_queue,
1973 i915_seqno_passed(ring->get_seqno(ring), seqno)
1974 || atomic_read(&dev_priv->mm.wedged));
1976 ring->irq_put(ring);
1977 } else if (wait_for_atomic(i915_seqno_passed(ring->get_seqno(ring),
1979 atomic_read(&dev_priv->mm.wedged), 3000))
1981 ring->waiting_seqno = 0;
1983 trace_i915_gem_request_wait_end(ring, seqno);
1985 if (atomic_read(&dev_priv->mm.wedged))
1988 /* Directly dispatch request retiring. While we have the work queue
1989 * to handle this, the waiter on a request often wants an associated
1990 * buffer to have made it to the inactive list, and we would need
1991 * a separate wait queue to handle that.
1993 if (ret == 0 && do_retire)
1994 i915_gem_retire_requests_ring(ring);
2000 * Ensures that all rendering to the object has completed and the object is
2001 * safe to unbind from the GTT or access from the CPU.
2004 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj)
2008 /* This function only exists to support waiting for existing rendering,
2009 * not for emitting required flushes.
2011 BUG_ON((obj->base.write_domain & I915_GEM_GPU_DOMAINS) != 0);
2013 /* If there is rendering queued on the buffer being evicted, wait for
2017 ret = i915_wait_request(obj->ring, obj->last_rendering_seqno,
2026 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
2028 u32 old_write_domain, old_read_domains;
2030 /* Act a barrier for all accesses through the GTT */
2033 /* Force a pagefault for domain tracking on next user access */
2034 i915_gem_release_mmap(obj);
2036 if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
2039 old_read_domains = obj->base.read_domains;
2040 old_write_domain = obj->base.write_domain;
2042 obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
2043 obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
2045 trace_i915_gem_object_change_domain(obj,
2051 * Unbinds an object from the GTT aperture.
2054 i915_gem_object_unbind(struct drm_i915_gem_object *obj)
2056 drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
2059 if (obj->gtt_space == NULL)
2062 if (obj->pin_count != 0) {
2063 DRM_ERROR("Attempting to unbind pinned buffer\n");
2067 ret = i915_gem_object_finish_gpu(obj);
2068 if (ret == -ERESTARTSYS)
2070 /* Continue on if we fail due to EIO, the GPU is hung so we
2071 * should be safe and we need to cleanup or else we might
2072 * cause memory corruption through use-after-free.
2075 i915_gem_object_finish_gtt(obj);
2077 /* Move the object to the CPU domain to ensure that
2078 * any possible CPU writes while it's not in the GTT
2079 * are flushed when we go to remap it.
2082 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
2083 if (ret == -ERESTARTSYS)
2086 /* In the event of a disaster, abandon all caches and
2087 * hope for the best.
2089 i915_gem_clflush_object(obj);
2090 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
2093 /* release the fence reg _after_ flushing */
2094 ret = i915_gem_object_put_fence(obj);
2095 if (ret == -ERESTARTSYS)
2098 trace_i915_gem_object_unbind(obj);
2100 i915_gem_gtt_unbind_object(obj);
2101 if (obj->has_aliasing_ppgtt_mapping) {
2102 i915_ppgtt_unbind_object(dev_priv->mm.aliasing_ppgtt, obj);
2103 obj->has_aliasing_ppgtt_mapping = 0;
2106 i915_gem_object_put_pages_gtt(obj);
2108 list_del_init(&obj->gtt_list);
2109 list_del_init(&obj->mm_list);
2110 /* Avoid an unnecessary call to unbind on rebind. */
2111 obj->map_and_fenceable = true;
2113 drm_mm_put_block(obj->gtt_space);
2114 obj->gtt_space = NULL;
2115 obj->gtt_offset = 0;
2117 if (i915_gem_object_is_purgeable(obj))
2118 i915_gem_object_truncate(obj);
2124 i915_gem_flush_ring(struct intel_ring_buffer *ring,
2125 uint32_t invalidate_domains,
2126 uint32_t flush_domains)
2130 if (((invalidate_domains | flush_domains) & I915_GEM_GPU_DOMAINS) == 0)
2133 trace_i915_gem_ring_flush(ring, invalidate_domains, flush_domains);
2135 ret = ring->flush(ring, invalidate_domains, flush_domains);
2139 if (flush_domains & I915_GEM_GPU_DOMAINS)
2140 i915_gem_process_flushing_list(ring, flush_domains);
2145 static int i915_ring_idle(struct intel_ring_buffer *ring, bool do_retire)
2149 if (list_empty(&ring->gpu_write_list) && list_empty(&ring->active_list))
2152 if (!list_empty(&ring->gpu_write_list)) {
2153 ret = i915_gem_flush_ring(ring,
2154 I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
2159 return i915_wait_request(ring, i915_gem_next_request_seqno(ring),
2163 int i915_gpu_idle(struct drm_device *dev, bool do_retire)
2165 drm_i915_private_t *dev_priv = dev->dev_private;
2168 /* Flush everything onto the inactive list. */
2169 for (i = 0; i < I915_NUM_RINGS; i++) {
2170 ret = i915_ring_idle(&dev_priv->ring[i], do_retire);
2178 static int sandybridge_write_fence_reg(struct drm_i915_gem_object *obj,
2179 struct intel_ring_buffer *pipelined)
2181 struct drm_device *dev = obj->base.dev;
2182 drm_i915_private_t *dev_priv = dev->dev_private;
2183 u32 size = obj->gtt_space->size;
2184 int regnum = obj->fence_reg;
2187 val = (uint64_t)((obj->gtt_offset + size - 4096) &
2189 val |= obj->gtt_offset & 0xfffff000;
2190 val |= (uint64_t)((obj->stride / 128) - 1) <<
2191 SANDYBRIDGE_FENCE_PITCH_SHIFT;
2193 if (obj->tiling_mode == I915_TILING_Y)
2194 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2195 val |= I965_FENCE_REG_VALID;
2198 int ret = intel_ring_begin(pipelined, 6);
2202 intel_ring_emit(pipelined, MI_NOOP);
2203 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(2));
2204 intel_ring_emit(pipelined, FENCE_REG_SANDYBRIDGE_0 + regnum*8);
2205 intel_ring_emit(pipelined, (u32)val);
2206 intel_ring_emit(pipelined, FENCE_REG_SANDYBRIDGE_0 + regnum*8 + 4);
2207 intel_ring_emit(pipelined, (u32)(val >> 32));
2208 intel_ring_advance(pipelined);
2210 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + regnum * 8, val);
2215 static int i965_write_fence_reg(struct drm_i915_gem_object *obj,
2216 struct intel_ring_buffer *pipelined)
2218 struct drm_device *dev = obj->base.dev;
2219 drm_i915_private_t *dev_priv = dev->dev_private;
2220 u32 size = obj->gtt_space->size;
2221 int regnum = obj->fence_reg;
2224 val = (uint64_t)((obj->gtt_offset + size - 4096) &
2226 val |= obj->gtt_offset & 0xfffff000;
2227 val |= ((obj->stride / 128) - 1) << I965_FENCE_PITCH_SHIFT;
2228 if (obj->tiling_mode == I915_TILING_Y)
2229 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2230 val |= I965_FENCE_REG_VALID;
2233 int ret = intel_ring_begin(pipelined, 6);
2237 intel_ring_emit(pipelined, MI_NOOP);
2238 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(2));
2239 intel_ring_emit(pipelined, FENCE_REG_965_0 + regnum*8);
2240 intel_ring_emit(pipelined, (u32)val);
2241 intel_ring_emit(pipelined, FENCE_REG_965_0 + regnum*8 + 4);
2242 intel_ring_emit(pipelined, (u32)(val >> 32));
2243 intel_ring_advance(pipelined);
2245 I915_WRITE64(FENCE_REG_965_0 + regnum * 8, val);
2250 static int i915_write_fence_reg(struct drm_i915_gem_object *obj,
2251 struct intel_ring_buffer *pipelined)
2253 struct drm_device *dev = obj->base.dev;
2254 drm_i915_private_t *dev_priv = dev->dev_private;
2255 u32 size = obj->gtt_space->size;
2256 u32 fence_reg, val, pitch_val;
2259 if (WARN((obj->gtt_offset & ~I915_FENCE_START_MASK) ||
2260 (size & -size) != size ||
2261 (obj->gtt_offset & (size - 1)),
2262 "object 0x%08x [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
2263 obj->gtt_offset, obj->map_and_fenceable, size))
2266 if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
2271 /* Note: pitch better be a power of two tile widths */
2272 pitch_val = obj->stride / tile_width;
2273 pitch_val = ffs(pitch_val) - 1;
2275 val = obj->gtt_offset;
2276 if (obj->tiling_mode == I915_TILING_Y)
2277 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2278 val |= I915_FENCE_SIZE_BITS(size);
2279 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2280 val |= I830_FENCE_REG_VALID;
2282 fence_reg = obj->fence_reg;
2284 fence_reg = FENCE_REG_830_0 + fence_reg * 4;
2286 fence_reg = FENCE_REG_945_8 + (fence_reg - 8) * 4;
2289 int ret = intel_ring_begin(pipelined, 4);
2293 intel_ring_emit(pipelined, MI_NOOP);
2294 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(1));
2295 intel_ring_emit(pipelined, fence_reg);
2296 intel_ring_emit(pipelined, val);
2297 intel_ring_advance(pipelined);
2299 I915_WRITE(fence_reg, val);
2304 static int i830_write_fence_reg(struct drm_i915_gem_object *obj,
2305 struct intel_ring_buffer *pipelined)
2307 struct drm_device *dev = obj->base.dev;
2308 drm_i915_private_t *dev_priv = dev->dev_private;
2309 u32 size = obj->gtt_space->size;
2310 int regnum = obj->fence_reg;
2314 if (WARN((obj->gtt_offset & ~I830_FENCE_START_MASK) ||
2315 (size & -size) != size ||
2316 (obj->gtt_offset & (size - 1)),
2317 "object 0x%08x not 512K or pot-size 0x%08x aligned\n",
2318 obj->gtt_offset, size))
2321 pitch_val = obj->stride / 128;
2322 pitch_val = ffs(pitch_val) - 1;
2324 val = obj->gtt_offset;
2325 if (obj->tiling_mode == I915_TILING_Y)
2326 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2327 val |= I830_FENCE_SIZE_BITS(size);
2328 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2329 val |= I830_FENCE_REG_VALID;
2332 int ret = intel_ring_begin(pipelined, 4);
2336 intel_ring_emit(pipelined, MI_NOOP);
2337 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(1));
2338 intel_ring_emit(pipelined, FENCE_REG_830_0 + regnum*4);
2339 intel_ring_emit(pipelined, val);
2340 intel_ring_advance(pipelined);
2342 I915_WRITE(FENCE_REG_830_0 + regnum * 4, val);
2347 static bool ring_passed_seqno(struct intel_ring_buffer *ring, u32 seqno)
2349 return i915_seqno_passed(ring->get_seqno(ring), seqno);
2353 i915_gem_object_flush_fence(struct drm_i915_gem_object *obj,
2354 struct intel_ring_buffer *pipelined)
2358 if (obj->fenced_gpu_access) {
2359 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
2360 ret = i915_gem_flush_ring(obj->last_fenced_ring,
2361 0, obj->base.write_domain);
2366 obj->fenced_gpu_access = false;
2369 if (obj->last_fenced_seqno && pipelined != obj->last_fenced_ring) {
2370 if (!ring_passed_seqno(obj->last_fenced_ring,
2371 obj->last_fenced_seqno)) {
2372 ret = i915_wait_request(obj->last_fenced_ring,
2373 obj->last_fenced_seqno,
2379 obj->last_fenced_seqno = 0;
2380 obj->last_fenced_ring = NULL;
2383 /* Ensure that all CPU reads are completed before installing a fence
2384 * and all writes before removing the fence.
2386 if (obj->base.read_domains & I915_GEM_DOMAIN_GTT)
2393 i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
2397 if (obj->tiling_mode)
2398 i915_gem_release_mmap(obj);
2400 ret = i915_gem_object_flush_fence(obj, NULL);
2404 if (obj->fence_reg != I915_FENCE_REG_NONE) {
2405 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2407 WARN_ON(dev_priv->fence_regs[obj->fence_reg].pin_count);
2408 i915_gem_clear_fence_reg(obj->base.dev,
2409 &dev_priv->fence_regs[obj->fence_reg]);
2411 obj->fence_reg = I915_FENCE_REG_NONE;
2417 static struct drm_i915_fence_reg *
2418 i915_find_fence_reg(struct drm_device *dev,
2419 struct intel_ring_buffer *pipelined)
2421 struct drm_i915_private *dev_priv = dev->dev_private;
2422 struct drm_i915_fence_reg *reg, *first, *avail;
2425 /* First try to find a free reg */
2427 for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2428 reg = &dev_priv->fence_regs[i];
2432 if (!reg->pin_count)
2439 /* None available, try to steal one or wait for a user to finish */
2440 avail = first = NULL;
2441 list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
2449 !reg->obj->last_fenced_ring ||
2450 reg->obj->last_fenced_ring == pipelined) {
2463 * i915_gem_object_get_fence - set up a fence reg for an object
2464 * @obj: object to map through a fence reg
2465 * @pipelined: ring on which to queue the change, or NULL for CPU access
2466 * @interruptible: must we wait uninterruptibly for the register to retire?
2468 * When mapping objects through the GTT, userspace wants to be able to write
2469 * to them without having to worry about swizzling if the object is tiled.
2471 * This function walks the fence regs looking for a free one for @obj,
2472 * stealing one if it can't find any.
2474 * It then sets up the reg based on the object's properties: address, pitch
2475 * and tiling format.
2478 i915_gem_object_get_fence(struct drm_i915_gem_object *obj,
2479 struct intel_ring_buffer *pipelined)
2481 struct drm_device *dev = obj->base.dev;
2482 struct drm_i915_private *dev_priv = dev->dev_private;
2483 struct drm_i915_fence_reg *reg;
2486 /* XXX disable pipelining. There are bugs. Shocking. */
2489 /* Just update our place in the LRU if our fence is getting reused. */
2490 if (obj->fence_reg != I915_FENCE_REG_NONE) {
2491 reg = &dev_priv->fence_regs[obj->fence_reg];
2492 list_move_tail(®->lru_list, &dev_priv->mm.fence_list);
2494 if (obj->tiling_changed) {
2495 ret = i915_gem_object_flush_fence(obj, pipelined);
2499 if (!obj->fenced_gpu_access && !obj->last_fenced_seqno)
2504 i915_gem_next_request_seqno(pipelined);
2505 obj->last_fenced_seqno = reg->setup_seqno;
2506 obj->last_fenced_ring = pipelined;
2513 if (reg->setup_seqno) {
2514 if (!ring_passed_seqno(obj->last_fenced_ring,
2515 reg->setup_seqno)) {
2516 ret = i915_wait_request(obj->last_fenced_ring,
2523 reg->setup_seqno = 0;
2525 } else if (obj->last_fenced_ring &&
2526 obj->last_fenced_ring != pipelined) {
2527 ret = i915_gem_object_flush_fence(obj, pipelined);
2535 reg = i915_find_fence_reg(dev, pipelined);
2539 ret = i915_gem_object_flush_fence(obj, pipelined);
2544 struct drm_i915_gem_object *old = reg->obj;
2546 drm_gem_object_reference(&old->base);
2548 if (old->tiling_mode)
2549 i915_gem_release_mmap(old);
2551 ret = i915_gem_object_flush_fence(old, pipelined);
2553 drm_gem_object_unreference(&old->base);
2557 if (old->last_fenced_seqno == 0 && obj->last_fenced_seqno == 0)
2560 old->fence_reg = I915_FENCE_REG_NONE;
2561 old->last_fenced_ring = pipelined;
2562 old->last_fenced_seqno =
2563 pipelined ? i915_gem_next_request_seqno(pipelined) : 0;
2565 drm_gem_object_unreference(&old->base);
2566 } else if (obj->last_fenced_seqno == 0)
2570 list_move_tail(®->lru_list, &dev_priv->mm.fence_list);
2571 obj->fence_reg = reg - dev_priv->fence_regs;
2572 obj->last_fenced_ring = pipelined;
2575 pipelined ? i915_gem_next_request_seqno(pipelined) : 0;
2576 obj->last_fenced_seqno = reg->setup_seqno;
2579 obj->tiling_changed = false;
2580 switch (INTEL_INFO(dev)->gen) {
2583 ret = sandybridge_write_fence_reg(obj, pipelined);
2587 ret = i965_write_fence_reg(obj, pipelined);
2590 ret = i915_write_fence_reg(obj, pipelined);
2593 ret = i830_write_fence_reg(obj, pipelined);
2601 * i915_gem_clear_fence_reg - clear out fence register info
2602 * @obj: object to clear
2604 * Zeroes out the fence register itself and clears out the associated
2605 * data structures in dev_priv and obj.
2608 i915_gem_clear_fence_reg(struct drm_device *dev,
2609 struct drm_i915_fence_reg *reg)
2611 drm_i915_private_t *dev_priv = dev->dev_private;
2612 uint32_t fence_reg = reg - dev_priv->fence_regs;
2614 switch (INTEL_INFO(dev)->gen) {
2617 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + fence_reg*8, 0);
2621 I915_WRITE64(FENCE_REG_965_0 + fence_reg*8, 0);
2625 fence_reg = FENCE_REG_945_8 + (fence_reg - 8) * 4;
2628 fence_reg = FENCE_REG_830_0 + fence_reg * 4;
2630 I915_WRITE(fence_reg, 0);
2634 list_del_init(®->lru_list);
2636 reg->setup_seqno = 0;
2641 * Finds free space in the GTT aperture and binds the object there.
2644 i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
2646 bool map_and_fenceable)
2648 struct drm_device *dev = obj->base.dev;
2649 drm_i915_private_t *dev_priv = dev->dev_private;
2650 struct drm_mm_node *free_space;
2651 gfp_t gfpmask = __GFP_NORETRY | __GFP_NOWARN;
2652 u32 size, fence_size, fence_alignment, unfenced_alignment;
2653 bool mappable, fenceable;
2656 if (obj->madv != I915_MADV_WILLNEED) {
2657 DRM_ERROR("Attempting to bind a purgeable object\n");
2661 fence_size = i915_gem_get_gtt_size(dev,
2664 fence_alignment = i915_gem_get_gtt_alignment(dev,
2667 unfenced_alignment =
2668 i915_gem_get_unfenced_gtt_alignment(dev,
2673 alignment = map_and_fenceable ? fence_alignment :
2675 if (map_and_fenceable && alignment & (fence_alignment - 1)) {
2676 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
2680 size = map_and_fenceable ? fence_size : obj->base.size;
2682 /* If the object is bigger than the entire aperture, reject it early
2683 * before evicting everything in a vain attempt to find space.
2685 if (obj->base.size >
2686 (map_and_fenceable ? dev_priv->mm.gtt_mappable_end : dev_priv->mm.gtt_total)) {
2687 DRM_ERROR("Attempting to bind an object larger than the aperture\n");
2692 if (map_and_fenceable)
2694 drm_mm_search_free_in_range(&dev_priv->mm.gtt_space,
2696 dev_priv->mm.gtt_mappable_end,
2699 free_space = drm_mm_search_free(&dev_priv->mm.gtt_space,
2700 size, alignment, 0);
2702 if (free_space != NULL) {
2703 if (map_and_fenceable)
2705 drm_mm_get_block_range_generic(free_space,
2707 dev_priv->mm.gtt_mappable_end,
2711 drm_mm_get_block(free_space, size, alignment);
2713 if (obj->gtt_space == NULL) {
2714 /* If the gtt is empty and we're still having trouble
2715 * fitting our object in, we're out of memory.
2717 ret = i915_gem_evict_something(dev, size, alignment,
2725 ret = i915_gem_object_get_pages_gtt(obj, gfpmask);
2727 drm_mm_put_block(obj->gtt_space);
2728 obj->gtt_space = NULL;
2730 if (ret == -ENOMEM) {
2731 /* first try to reclaim some memory by clearing the GTT */
2732 ret = i915_gem_evict_everything(dev, false);
2734 /* now try to shrink everyone else */
2749 ret = i915_gem_gtt_bind_object(obj);
2751 i915_gem_object_put_pages_gtt(obj);
2752 drm_mm_put_block(obj->gtt_space);
2753 obj->gtt_space = NULL;
2755 if (i915_gem_evict_everything(dev, false))
2761 list_add_tail(&obj->gtt_list, &dev_priv->mm.gtt_list);
2762 list_add_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
2764 /* Assert that the object is not currently in any GPU domain. As it
2765 * wasn't in the GTT, there shouldn't be any way it could have been in
2768 BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
2769 BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
2771 obj->gtt_offset = obj->gtt_space->start;
2774 obj->gtt_space->size == fence_size &&
2775 (obj->gtt_space->start & (fence_alignment - 1)) == 0;
2778 obj->gtt_offset + obj->base.size <= dev_priv->mm.gtt_mappable_end;
2780 obj->map_and_fenceable = mappable && fenceable;
2782 trace_i915_gem_object_bind(obj, map_and_fenceable);
2787 i915_gem_clflush_object(struct drm_i915_gem_object *obj)
2789 /* If we don't have a page list set up, then we're not pinned
2790 * to GPU, and we can ignore the cache flush because it'll happen
2791 * again at bind time.
2793 if (obj->pages == NULL)
2796 /* If the GPU is snooping the contents of the CPU cache,
2797 * we do not need to manually clear the CPU cache lines. However,
2798 * the caches are only snooped when the render cache is
2799 * flushed/invalidated. As we always have to emit invalidations
2800 * and flushes when moving into and out of the RENDER domain, correct
2801 * snooping behaviour occurs naturally as the result of our domain
2804 if (obj->cache_level != I915_CACHE_NONE)
2807 trace_i915_gem_object_clflush(obj);
2809 drm_clflush_pages(obj->pages, obj->base.size / PAGE_SIZE);
2812 /** Flushes any GPU write domain for the object if it's dirty. */
2814 i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object *obj)
2816 if ((obj->base.write_domain & I915_GEM_GPU_DOMAINS) == 0)
2819 /* Queue the GPU write cache flushing we need. */
2820 return i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain);
2823 /** Flushes the GTT write domain for the object if it's dirty. */
2825 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
2827 uint32_t old_write_domain;
2829 if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
2832 /* No actual flushing is required for the GTT write domain. Writes
2833 * to it immediately go to main memory as far as we know, so there's
2834 * no chipset flush. It also doesn't land in render cache.
2836 * However, we do have to enforce the order so that all writes through
2837 * the GTT land before any writes to the device, such as updates to
2842 old_write_domain = obj->base.write_domain;
2843 obj->base.write_domain = 0;
2845 trace_i915_gem_object_change_domain(obj,
2846 obj->base.read_domains,
2850 /** Flushes the CPU write domain for the object if it's dirty. */
2852 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj)
2854 uint32_t old_write_domain;
2856 if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
2859 i915_gem_clflush_object(obj);
2860 intel_gtt_chipset_flush();
2861 old_write_domain = obj->base.write_domain;
2862 obj->base.write_domain = 0;
2864 trace_i915_gem_object_change_domain(obj,
2865 obj->base.read_domains,
2870 * Moves a single object to the GTT read, and possibly write domain.
2872 * This function returns when the move is complete, including waiting on
2876 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
2878 uint32_t old_write_domain, old_read_domains;
2881 /* Not valid to be called on unbound objects. */
2882 if (obj->gtt_space == NULL)
2885 if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
2888 ret = i915_gem_object_flush_gpu_write_domain(obj);
2892 if (obj->pending_gpu_write || write) {
2893 ret = i915_gem_object_wait_rendering(obj);
2898 i915_gem_object_flush_cpu_write_domain(obj);
2900 old_write_domain = obj->base.write_domain;
2901 old_read_domains = obj->base.read_domains;
2903 /* It should now be out of any other write domains, and we can update
2904 * the domain values for our changes.
2906 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2907 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
2909 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
2910 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
2914 trace_i915_gem_object_change_domain(obj,
2921 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
2922 enum i915_cache_level cache_level)
2924 struct drm_device *dev = obj->base.dev;
2925 drm_i915_private_t *dev_priv = dev->dev_private;
2928 if (obj->cache_level == cache_level)
2931 if (obj->pin_count) {
2932 DRM_DEBUG("can not change the cache level of pinned objects\n");
2936 if (obj->gtt_space) {
2937 ret = i915_gem_object_finish_gpu(obj);
2941 i915_gem_object_finish_gtt(obj);
2943 /* Before SandyBridge, you could not use tiling or fence
2944 * registers with snooped memory, so relinquish any fences
2945 * currently pointing to our region in the aperture.
2947 if (INTEL_INFO(obj->base.dev)->gen < 6) {
2948 ret = i915_gem_object_put_fence(obj);
2953 i915_gem_gtt_rebind_object(obj, cache_level);
2954 if (obj->has_aliasing_ppgtt_mapping)
2955 i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt,
2959 if (cache_level == I915_CACHE_NONE) {
2960 u32 old_read_domains, old_write_domain;
2962 /* If we're coming from LLC cached, then we haven't
2963 * actually been tracking whether the data is in the
2964 * CPU cache or not, since we only allow one bit set
2965 * in obj->write_domain and have been skipping the clflushes.
2966 * Just set it to the CPU cache for now.
2968 WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU);
2969 WARN_ON(obj->base.read_domains & ~I915_GEM_DOMAIN_CPU);
2971 old_read_domains = obj->base.read_domains;
2972 old_write_domain = obj->base.write_domain;
2974 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
2975 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
2977 trace_i915_gem_object_change_domain(obj,
2982 obj->cache_level = cache_level;
2987 * Prepare buffer for display plane (scanout, cursors, etc).
2988 * Can be called from an uninterruptible phase (modesetting) and allows
2989 * any flushes to be pipelined (for pageflips).
2991 * For the display plane, we want to be in the GTT but out of any write
2992 * domains. So in many ways this looks like set_to_gtt_domain() apart from the
2993 * ability to pipeline the waits, pinning and any additional subtleties
2994 * that may differentiate the display plane from ordinary buffers.
2997 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
2999 struct intel_ring_buffer *pipelined)
3001 u32 old_read_domains, old_write_domain;
3004 ret = i915_gem_object_flush_gpu_write_domain(obj);
3008 if (pipelined != obj->ring) {
3009 ret = i915_gem_object_wait_rendering(obj);
3010 if (ret == -ERESTARTSYS)
3014 /* The display engine is not coherent with the LLC cache on gen6. As
3015 * a result, we make sure that the pinning that is about to occur is
3016 * done with uncached PTEs. This is lowest common denominator for all
3019 * However for gen6+, we could do better by using the GFDT bit instead
3020 * of uncaching, which would allow us to flush all the LLC-cached data
3021 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3023 ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE);
3027 /* As the user may map the buffer once pinned in the display plane
3028 * (e.g. libkms for the bootup splash), we have to ensure that we
3029 * always use map_and_fenceable for all scanout buffers.
3031 ret = i915_gem_object_pin(obj, alignment, true);
3035 i915_gem_object_flush_cpu_write_domain(obj);
3037 old_write_domain = obj->base.write_domain;
3038 old_read_domains = obj->base.read_domains;
3040 /* It should now be out of any other write domains, and we can update
3041 * the domain values for our changes.
3043 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3044 obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3046 trace_i915_gem_object_change_domain(obj,
3054 i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
3058 if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
3061 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
3062 ret = i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain);
3067 ret = i915_gem_object_wait_rendering(obj);
3071 /* Ensure that we invalidate the GPU's caches and TLBs. */
3072 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
3077 * Moves a single object to the CPU read, and possibly write domain.
3079 * This function returns when the move is complete, including waiting on
3083 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
3085 uint32_t old_write_domain, old_read_domains;
3088 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
3091 ret = i915_gem_object_flush_gpu_write_domain(obj);
3095 ret = i915_gem_object_wait_rendering(obj);
3099 i915_gem_object_flush_gtt_write_domain(obj);
3101 /* If we have a partially-valid cache of the object in the CPU,
3102 * finish invalidating it and free the per-page flags.
3104 i915_gem_object_set_to_full_cpu_read_domain(obj);
3106 old_write_domain = obj->base.write_domain;
3107 old_read_domains = obj->base.read_domains;
3109 /* Flush the CPU cache if it's still invalid. */
3110 if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
3111 i915_gem_clflush_object(obj);
3113 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3116 /* It should now be out of any other write domains, and we can update
3117 * the domain values for our changes.
3119 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3121 /* If we're writing through the CPU, then the GPU read domains will
3122 * need to be invalidated at next use.
3125 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3126 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3129 trace_i915_gem_object_change_domain(obj,
3137 * Moves the object from a partially CPU read to a full one.
3139 * Note that this only resolves i915_gem_object_set_cpu_read_domain_range(),
3140 * and doesn't handle transitioning from !(read_domains & I915_GEM_DOMAIN_CPU).
3143 i915_gem_object_set_to_full_cpu_read_domain(struct drm_i915_gem_object *obj)
3145 if (!obj->page_cpu_valid)
3148 /* If we're partially in the CPU read domain, finish moving it in.
3150 if (obj->base.read_domains & I915_GEM_DOMAIN_CPU) {
3153 for (i = 0; i <= (obj->base.size - 1) / PAGE_SIZE; i++) {
3154 if (obj->page_cpu_valid[i])
3156 drm_clflush_pages(obj->pages + i, 1);
3160 /* Free the page_cpu_valid mappings which are now stale, whether
3161 * or not we've got I915_GEM_DOMAIN_CPU.
3163 kfree(obj->page_cpu_valid);
3164 obj->page_cpu_valid = NULL;
3168 * Set the CPU read domain on a range of the object.
3170 * The object ends up with I915_GEM_DOMAIN_CPU in its read flags although it's
3171 * not entirely valid. The page_cpu_valid member of the object flags which
3172 * pages have been flushed, and will be respected by
3173 * i915_gem_object_set_to_cpu_domain() if it's called on to get a valid mapping
3174 * of the whole object.
3176 * This function returns when the move is complete, including waiting on
3180 i915_gem_object_set_cpu_read_domain_range(struct drm_i915_gem_object *obj,
3181 uint64_t offset, uint64_t size)
3183 uint32_t old_read_domains;
3186 if (offset == 0 && size == obj->base.size)
3187 return i915_gem_object_set_to_cpu_domain(obj, 0);
3189 ret = i915_gem_object_flush_gpu_write_domain(obj);
3193 ret = i915_gem_object_wait_rendering(obj);
3197 i915_gem_object_flush_gtt_write_domain(obj);
3199 /* If we're already fully in the CPU read domain, we're done. */
3200 if (obj->page_cpu_valid == NULL &&
3201 (obj->base.read_domains & I915_GEM_DOMAIN_CPU) != 0)
3204 /* Otherwise, create/clear the per-page CPU read domain flag if we're
3205 * newly adding I915_GEM_DOMAIN_CPU
3207 if (obj->page_cpu_valid == NULL) {
3208 obj->page_cpu_valid = kzalloc(obj->base.size / PAGE_SIZE,
3210 if (obj->page_cpu_valid == NULL)
3212 } else if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0)
3213 memset(obj->page_cpu_valid, 0, obj->base.size / PAGE_SIZE);
3215 /* Flush the cache on any pages that are still invalid from the CPU's
3218 for (i = offset / PAGE_SIZE; i <= (offset + size - 1) / PAGE_SIZE;
3220 if (obj->page_cpu_valid[i])
3223 drm_clflush_pages(obj->pages + i, 1);
3225 obj->page_cpu_valid[i] = 1;
3228 /* It should now be out of any other write domains, and we can update
3229 * the domain values for our changes.
3231 BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3233 old_read_domains = obj->base.read_domains;
3234 obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3236 trace_i915_gem_object_change_domain(obj,
3238 obj->base.write_domain);
3243 /* Throttle our rendering by waiting until the ring has completed our requests
3244 * emitted over 20 msec ago.
3246 * Note that if we were to use the current jiffies each time around the loop,
3247 * we wouldn't escape the function with any frames outstanding if the time to
3248 * render a frame was over 20ms.
3250 * This should get us reasonable parallelism between CPU and GPU but also
3251 * relatively low latency when blocking on a particular request to finish.
3254 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
3256 struct drm_i915_private *dev_priv = dev->dev_private;
3257 struct drm_i915_file_private *file_priv = file->driver_priv;
3258 unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3259 struct drm_i915_gem_request *request;
3260 struct intel_ring_buffer *ring = NULL;
3264 if (atomic_read(&dev_priv->mm.wedged))
3267 spin_lock(&file_priv->mm.lock);
3268 list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
3269 if (time_after_eq(request->emitted_jiffies, recent_enough))
3272 ring = request->ring;
3273 seqno = request->seqno;
3275 spin_unlock(&file_priv->mm.lock);
3281 if (!i915_seqno_passed(ring->get_seqno(ring), seqno)) {
3282 /* And wait for the seqno passing without holding any locks and
3283 * causing extra latency for others. This is safe as the irq
3284 * generation is designed to be run atomically and so is
3287 if (ring->irq_get(ring)) {
3288 ret = wait_event_interruptible(ring->irq_queue,
3289 i915_seqno_passed(ring->get_seqno(ring), seqno)
3290 || atomic_read(&dev_priv->mm.wedged));
3291 ring->irq_put(ring);
3293 if (ret == 0 && atomic_read(&dev_priv->mm.wedged))
3295 } else if (wait_for_atomic(i915_seqno_passed(ring->get_seqno(ring),
3297 atomic_read(&dev_priv->mm.wedged), 3000)) {
3303 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
3309 i915_gem_object_pin(struct drm_i915_gem_object *obj,
3311 bool map_and_fenceable)
3313 struct drm_device *dev = obj->base.dev;
3314 struct drm_i915_private *dev_priv = dev->dev_private;
3317 BUG_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT);
3318 WARN_ON(i915_verify_lists(dev));
3320 if (obj->gtt_space != NULL) {
3321 if ((alignment && obj->gtt_offset & (alignment - 1)) ||
3322 (map_and_fenceable && !obj->map_and_fenceable)) {
3323 WARN(obj->pin_count,
3324 "bo is already pinned with incorrect alignment:"
3325 " offset=%x, req.alignment=%x, req.map_and_fenceable=%d,"
3326 " obj->map_and_fenceable=%d\n",
3327 obj->gtt_offset, alignment,
3329 obj->map_and_fenceable);
3330 ret = i915_gem_object_unbind(obj);
3336 if (obj->gtt_space == NULL) {
3337 ret = i915_gem_object_bind_to_gtt(obj, alignment,
3343 if (obj->pin_count++ == 0) {
3345 list_move_tail(&obj->mm_list,
3346 &dev_priv->mm.pinned_list);
3348 obj->pin_mappable |= map_and_fenceable;
3350 WARN_ON(i915_verify_lists(dev));
3355 i915_gem_object_unpin(struct drm_i915_gem_object *obj)
3357 struct drm_device *dev = obj->base.dev;
3358 drm_i915_private_t *dev_priv = dev->dev_private;
3360 WARN_ON(i915_verify_lists(dev));
3361 BUG_ON(obj->pin_count == 0);
3362 BUG_ON(obj->gtt_space == NULL);
3364 if (--obj->pin_count == 0) {
3366 list_move_tail(&obj->mm_list,
3367 &dev_priv->mm.inactive_list);
3368 obj->pin_mappable = false;
3370 WARN_ON(i915_verify_lists(dev));
3374 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
3375 struct drm_file *file)
3377 struct drm_i915_gem_pin *args = data;
3378 struct drm_i915_gem_object *obj;
3381 ret = i915_mutex_lock_interruptible(dev);
3385 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3386 if (&obj->base == NULL) {
3391 if (obj->madv != I915_MADV_WILLNEED) {
3392 DRM_ERROR("Attempting to pin a purgeable buffer\n");
3397 if (obj->pin_filp != NULL && obj->pin_filp != file) {
3398 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
3404 obj->user_pin_count++;
3405 obj->pin_filp = file;
3406 if (obj->user_pin_count == 1) {
3407 ret = i915_gem_object_pin(obj, args->alignment, true);
3412 /* XXX - flush the CPU caches for pinned objects
3413 * as the X server doesn't manage domains yet
3415 i915_gem_object_flush_cpu_write_domain(obj);
3416 args->offset = obj->gtt_offset;
3418 drm_gem_object_unreference(&obj->base);
3420 mutex_unlock(&dev->struct_mutex);
3425 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
3426 struct drm_file *file)
3428 struct drm_i915_gem_pin *args = data;
3429 struct drm_i915_gem_object *obj;
3432 ret = i915_mutex_lock_interruptible(dev);
3436 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3437 if (&obj->base == NULL) {
3442 if (obj->pin_filp != file) {
3443 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
3448 obj->user_pin_count--;
3449 if (obj->user_pin_count == 0) {
3450 obj->pin_filp = NULL;
3451 i915_gem_object_unpin(obj);
3455 drm_gem_object_unreference(&obj->base);
3457 mutex_unlock(&dev->struct_mutex);
3462 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
3463 struct drm_file *file)
3465 struct drm_i915_gem_busy *args = data;
3466 struct drm_i915_gem_object *obj;
3469 ret = i915_mutex_lock_interruptible(dev);
3473 obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3474 if (&obj->base == NULL) {
3479 /* Count all active objects as busy, even if they are currently not used
3480 * by the gpu. Users of this interface expect objects to eventually
3481 * become non-busy without any further actions, therefore emit any
3482 * necessary flushes here.
3484 args->busy = obj->active;
3486 /* Unconditionally flush objects, even when the gpu still uses this
3487 * object. Userspace calling this function indicates that it wants to
3488 * use this buffer rather sooner than later, so issuing the required
3489 * flush earlier is beneficial.
3491 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
3492 ret = i915_gem_flush_ring(obj->ring,
3493 0, obj->base.write_domain);
3494 } else if (obj->ring->outstanding_lazy_request ==
3495 obj->last_rendering_seqno) {
3496 struct drm_i915_gem_request *request;
3498 /* This ring is not being cleared by active usage,
3499 * so emit a request to do so.
3501 request = kzalloc(sizeof(*request), GFP_KERNEL);
3503 ret = i915_add_request(obj->ring, NULL, request);
3510 /* Update the active list for the hardware's current position.
3511 * Otherwise this only updates on a delayed timer or when irqs
3512 * are actually unmasked, and our working set ends up being
3513 * larger than required.
3515 i915_gem_retire_requests_ring(obj->ring);
3517 args->busy = obj->active;
3520 drm_gem_object_unreference(&obj->base);
3522 mutex_unlock(&dev->struct_mutex);
3527 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
3528 struct drm_file *file_priv)
3530 return i915_gem_ring_throttle(dev, file_priv);
3534 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
3535 struct drm_file *file_priv)
3537 struct drm_i915_gem_madvise *args = data;
3538 struct drm_i915_gem_object *obj;
3541 switch (args->madv) {
3542 case I915_MADV_DONTNEED:
3543 case I915_MADV_WILLNEED:
3549 ret = i915_mutex_lock_interruptible(dev);
3553 obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
3554 if (&obj->base == NULL) {
3559 if (obj->pin_count) {
3564 if (obj->madv != __I915_MADV_PURGED)
3565 obj->madv = args->madv;
3567 /* if the object is no longer bound, discard its backing storage */
3568 if (i915_gem_object_is_purgeable(obj) &&
3569 obj->gtt_space == NULL)
3570 i915_gem_object_truncate(obj);
3572 args->retained = obj->madv != __I915_MADV_PURGED;
3575 drm_gem_object_unreference(&obj->base);
3577 mutex_unlock(&dev->struct_mutex);
3581 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
3584 struct drm_i915_private *dev_priv = dev->dev_private;
3585 struct drm_i915_gem_object *obj;
3586 struct address_space *mapping;
3588 obj = kzalloc(sizeof(*obj), GFP_KERNEL);
3592 if (drm_gem_object_init(dev, &obj->base, size) != 0) {
3597 mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
3598 mapping_set_gfp_mask(mapping, GFP_HIGHUSER | __GFP_RECLAIMABLE);
3600 i915_gem_info_add_obj(dev_priv, size);
3602 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3603 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3606 /* On some devices, we can have the GPU use the LLC (the CPU
3607 * cache) for about a 10% performance improvement
3608 * compared to uncached. Graphics requests other than
3609 * display scanout are coherent with the CPU in
3610 * accessing this cache. This means in this mode we
3611 * don't need to clflush on the CPU side, and on the
3612 * GPU side we only need to flush internal caches to
3613 * get data visible to the CPU.
3615 * However, we maintain the display planes as UC, and so
3616 * need to rebind when first used as such.
3618 obj->cache_level = I915_CACHE_LLC;
3620 obj->cache_level = I915_CACHE_NONE;
3622 obj->base.driver_private = NULL;
3623 obj->fence_reg = I915_FENCE_REG_NONE;
3624 INIT_LIST_HEAD(&obj->mm_list);
3625 INIT_LIST_HEAD(&obj->gtt_list);
3626 INIT_LIST_HEAD(&obj->ring_list);
3627 INIT_LIST_HEAD(&obj->exec_list);
3628 INIT_LIST_HEAD(&obj->gpu_write_list);
3629 obj->madv = I915_MADV_WILLNEED;
3630 /* Avoid an unnecessary call to unbind on the first bind. */
3631 obj->map_and_fenceable = true;
3636 int i915_gem_init_object(struct drm_gem_object *obj)
3643 static void i915_gem_free_object_tail(struct drm_i915_gem_object *obj)
3645 struct drm_device *dev = obj->base.dev;
3646 drm_i915_private_t *dev_priv = dev->dev_private;
3649 ret = i915_gem_object_unbind(obj);
3650 if (ret == -ERESTARTSYS) {
3651 list_move(&obj->mm_list,
3652 &dev_priv->mm.deferred_free_list);
3656 trace_i915_gem_object_destroy(obj);
3658 if (obj->base.map_list.map)
3659 drm_gem_free_mmap_offset(&obj->base);
3661 drm_gem_object_release(&obj->base);
3662 i915_gem_info_remove_obj(dev_priv, obj->base.size);
3664 kfree(obj->page_cpu_valid);
3669 void i915_gem_free_object(struct drm_gem_object *gem_obj)
3671 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
3672 struct drm_device *dev = obj->base.dev;
3674 while (obj->pin_count > 0)
3675 i915_gem_object_unpin(obj);
3678 i915_gem_detach_phys_object(dev, obj);
3680 i915_gem_free_object_tail(obj);
3684 i915_gem_idle(struct drm_device *dev)
3686 drm_i915_private_t *dev_priv = dev->dev_private;
3689 mutex_lock(&dev->struct_mutex);
3691 if (dev_priv->mm.suspended) {
3692 mutex_unlock(&dev->struct_mutex);
3696 ret = i915_gpu_idle(dev, true);
3698 mutex_unlock(&dev->struct_mutex);
3702 /* Under UMS, be paranoid and evict. */
3703 if (!drm_core_check_feature(dev, DRIVER_MODESET)) {
3704 ret = i915_gem_evict_inactive(dev, false);
3706 mutex_unlock(&dev->struct_mutex);
3711 i915_gem_reset_fences(dev);
3713 /* Hack! Don't let anybody do execbuf while we don't control the chip.
3714 * We need to replace this with a semaphore, or something.
3715 * And not confound mm.suspended!
3717 dev_priv->mm.suspended = 1;
3718 del_timer_sync(&dev_priv->hangcheck_timer);
3720 i915_kernel_lost_context(dev);
3721 i915_gem_cleanup_ringbuffer(dev);
3723 mutex_unlock(&dev->struct_mutex);
3725 /* Cancel the retire work handler, which should be idle now. */
3726 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
3731 void i915_gem_init_swizzling(struct drm_device *dev)
3733 drm_i915_private_t *dev_priv = dev->dev_private;
3735 if (INTEL_INFO(dev)->gen < 5 ||
3736 dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
3739 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
3740 DISP_TILE_SURFACE_SWIZZLING);
3745 I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
3747 I915_WRITE(ARB_MODE, ARB_MODE_ENABLE(ARB_MODE_SWIZZLE_SNB));
3749 I915_WRITE(ARB_MODE, ARB_MODE_ENABLE(ARB_MODE_SWIZZLE_IVB));
3752 void i915_gem_init_ppgtt(struct drm_device *dev)
3754 drm_i915_private_t *dev_priv = dev->dev_private;
3756 struct intel_ring_buffer *ring;
3759 if (!dev_priv->mm.aliasing_ppgtt)
3762 pd_offset = dev_priv->mm.aliasing_ppgtt->pd_offset;
3763 pd_offset /= 64; /* in cachelines, */
3766 if (INTEL_INFO(dev)->gen == 6) {
3767 uint32_t ecochk = I915_READ(GAM_ECOCHK);
3768 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT |
3769 ECOCHK_PPGTT_CACHE64B);
3770 I915_WRITE(GFX_MODE, GFX_MODE_ENABLE(GFX_PPGTT_ENABLE));
3771 } else if (INTEL_INFO(dev)->gen >= 7) {
3772 I915_WRITE(GAM_ECOCHK, ECOCHK_PPGTT_CACHE64B);
3773 /* GFX_MODE is per-ring on gen7+ */
3776 for (i = 0; i < I915_NUM_RINGS; i++) {
3777 ring = &dev_priv->ring[i];
3779 if (INTEL_INFO(dev)->gen >= 7)
3780 I915_WRITE(RING_MODE_GEN7(ring),
3781 GFX_MODE_ENABLE(GFX_PPGTT_ENABLE));
3783 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
3784 I915_WRITE(RING_PP_DIR_BASE(ring), pd_offset);
3789 i915_gem_init_hw(struct drm_device *dev)
3791 drm_i915_private_t *dev_priv = dev->dev_private;
3794 i915_gem_init_swizzling(dev);
3796 ret = intel_init_render_ring_buffer(dev);
3801 ret = intel_init_bsd_ring_buffer(dev);
3803 goto cleanup_render_ring;
3807 ret = intel_init_blt_ring_buffer(dev);
3809 goto cleanup_bsd_ring;
3812 dev_priv->next_seqno = 1;
3814 i915_gem_init_ppgtt(dev);
3819 intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
3820 cleanup_render_ring:
3821 intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
3826 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
3828 drm_i915_private_t *dev_priv = dev->dev_private;
3831 for (i = 0; i < I915_NUM_RINGS; i++)
3832 intel_cleanup_ring_buffer(&dev_priv->ring[i]);
3836 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
3837 struct drm_file *file_priv)
3839 drm_i915_private_t *dev_priv = dev->dev_private;
3842 if (drm_core_check_feature(dev, DRIVER_MODESET))
3845 if (atomic_read(&dev_priv->mm.wedged)) {
3846 DRM_ERROR("Reenabling wedged hardware, good luck\n");
3847 atomic_set(&dev_priv->mm.wedged, 0);
3850 mutex_lock(&dev->struct_mutex);
3851 dev_priv->mm.suspended = 0;
3853 ret = i915_gem_init_hw(dev);
3855 mutex_unlock(&dev->struct_mutex);
3859 BUG_ON(!list_empty(&dev_priv->mm.active_list));
3860 BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
3861 BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
3862 for (i = 0; i < I915_NUM_RINGS; i++) {
3863 BUG_ON(!list_empty(&dev_priv->ring[i].active_list));
3864 BUG_ON(!list_empty(&dev_priv->ring[i].request_list));
3866 mutex_unlock(&dev->struct_mutex);
3868 ret = drm_irq_install(dev);
3870 goto cleanup_ringbuffer;
3875 mutex_lock(&dev->struct_mutex);
3876 i915_gem_cleanup_ringbuffer(dev);
3877 dev_priv->mm.suspended = 1;
3878 mutex_unlock(&dev->struct_mutex);
3884 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
3885 struct drm_file *file_priv)
3887 if (drm_core_check_feature(dev, DRIVER_MODESET))
3890 drm_irq_uninstall(dev);
3891 return i915_gem_idle(dev);
3895 i915_gem_lastclose(struct drm_device *dev)
3899 if (drm_core_check_feature(dev, DRIVER_MODESET))
3902 ret = i915_gem_idle(dev);
3904 DRM_ERROR("failed to idle hardware: %d\n", ret);
3908 init_ring_lists(struct intel_ring_buffer *ring)
3910 INIT_LIST_HEAD(&ring->active_list);
3911 INIT_LIST_HEAD(&ring->request_list);
3912 INIT_LIST_HEAD(&ring->gpu_write_list);
3916 i915_gem_load(struct drm_device *dev)
3919 drm_i915_private_t *dev_priv = dev->dev_private;
3921 INIT_LIST_HEAD(&dev_priv->mm.active_list);
3922 INIT_LIST_HEAD(&dev_priv->mm.flushing_list);
3923 INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
3924 INIT_LIST_HEAD(&dev_priv->mm.pinned_list);
3925 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
3926 INIT_LIST_HEAD(&dev_priv->mm.deferred_free_list);
3927 INIT_LIST_HEAD(&dev_priv->mm.gtt_list);
3928 for (i = 0; i < I915_NUM_RINGS; i++)
3929 init_ring_lists(&dev_priv->ring[i]);
3930 for (i = 0; i < I915_MAX_NUM_FENCES; i++)
3931 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
3932 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
3933 i915_gem_retire_work_handler);
3934 init_completion(&dev_priv->error_completion);
3936 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
3938 u32 tmp = I915_READ(MI_ARB_STATE);
3939 if (!(tmp & MI_ARB_C3_LP_WRITE_ENABLE)) {
3940 /* arb state is a masked write, so set bit + bit in mask */
3941 tmp = MI_ARB_C3_LP_WRITE_ENABLE | (MI_ARB_C3_LP_WRITE_ENABLE << MI_ARB_MASK_SHIFT);
3942 I915_WRITE(MI_ARB_STATE, tmp);
3946 dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
3948 /* Old X drivers will take 0-2 for front, back, depth buffers */
3949 if (!drm_core_check_feature(dev, DRIVER_MODESET))
3950 dev_priv->fence_reg_start = 3;
3952 if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
3953 dev_priv->num_fence_regs = 16;
3955 dev_priv->num_fence_regs = 8;
3957 /* Initialize fence registers to zero */
3958 for (i = 0; i < dev_priv->num_fence_regs; i++) {
3959 i915_gem_clear_fence_reg(dev, &dev_priv->fence_regs[i]);
3962 i915_gem_detect_bit_6_swizzle(dev);
3963 init_waitqueue_head(&dev_priv->pending_flip_queue);
3965 dev_priv->mm.interruptible = true;
3967 dev_priv->mm.inactive_shrinker.shrink = i915_gem_inactive_shrink;
3968 dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS;
3969 register_shrinker(&dev_priv->mm.inactive_shrinker);
3973 * Create a physically contiguous memory object for this object
3974 * e.g. for cursor + overlay regs
3976 static int i915_gem_init_phys_object(struct drm_device *dev,
3977 int id, int size, int align)
3979 drm_i915_private_t *dev_priv = dev->dev_private;
3980 struct drm_i915_gem_phys_object *phys_obj;
3983 if (dev_priv->mm.phys_objs[id - 1] || !size)
3986 phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL);
3992 phys_obj->handle = drm_pci_alloc(dev, size, align);
3993 if (!phys_obj->handle) {
3998 set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4001 dev_priv->mm.phys_objs[id - 1] = phys_obj;
4009 static void i915_gem_free_phys_object(struct drm_device *dev, int id)
4011 drm_i915_private_t *dev_priv = dev->dev_private;
4012 struct drm_i915_gem_phys_object *phys_obj;
4014 if (!dev_priv->mm.phys_objs[id - 1])
4017 phys_obj = dev_priv->mm.phys_objs[id - 1];
4018 if (phys_obj->cur_obj) {
4019 i915_gem_detach_phys_object(dev, phys_obj->cur_obj);
4023 set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4025 drm_pci_free(dev, phys_obj->handle);
4027 dev_priv->mm.phys_objs[id - 1] = NULL;
4030 void i915_gem_free_all_phys_object(struct drm_device *dev)
4034 for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++)
4035 i915_gem_free_phys_object(dev, i);
4038 void i915_gem_detach_phys_object(struct drm_device *dev,
4039 struct drm_i915_gem_object *obj)
4041 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
4048 vaddr = obj->phys_obj->handle->vaddr;
4050 page_count = obj->base.size / PAGE_SIZE;
4051 for (i = 0; i < page_count; i++) {
4052 struct page *page = shmem_read_mapping_page(mapping, i);
4053 if (!IS_ERR(page)) {
4054 char *dst = kmap_atomic(page);
4055 memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE);
4058 drm_clflush_pages(&page, 1);
4060 set_page_dirty(page);
4061 mark_page_accessed(page);
4062 page_cache_release(page);
4065 intel_gtt_chipset_flush();
4067 obj->phys_obj->cur_obj = NULL;
4068 obj->phys_obj = NULL;
4072 i915_gem_attach_phys_object(struct drm_device *dev,
4073 struct drm_i915_gem_object *obj,
4077 struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
4078 drm_i915_private_t *dev_priv = dev->dev_private;
4083 if (id > I915_MAX_PHYS_OBJECT)
4086 if (obj->phys_obj) {
4087 if (obj->phys_obj->id == id)
4089 i915_gem_detach_phys_object(dev, obj);
4092 /* create a new object */
4093 if (!dev_priv->mm.phys_objs[id - 1]) {
4094 ret = i915_gem_init_phys_object(dev, id,
4095 obj->base.size, align);
4097 DRM_ERROR("failed to init phys object %d size: %zu\n",
4098 id, obj->base.size);
4103 /* bind to the object */
4104 obj->phys_obj = dev_priv->mm.phys_objs[id - 1];
4105 obj->phys_obj->cur_obj = obj;
4107 page_count = obj->base.size / PAGE_SIZE;
4109 for (i = 0; i < page_count; i++) {
4113 page = shmem_read_mapping_page(mapping, i);
4115 return PTR_ERR(page);
4117 src = kmap_atomic(page);
4118 dst = obj->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4119 memcpy(dst, src, PAGE_SIZE);
4122 mark_page_accessed(page);
4123 page_cache_release(page);
4130 i915_gem_phys_pwrite(struct drm_device *dev,
4131 struct drm_i915_gem_object *obj,
4132 struct drm_i915_gem_pwrite *args,
4133 struct drm_file *file_priv)
4135 void *vaddr = obj->phys_obj->handle->vaddr + args->offset;
4136 char __user *user_data = (char __user *) (uintptr_t) args->data_ptr;
4138 if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
4139 unsigned long unwritten;
4141 /* The physical object once assigned is fixed for the lifetime
4142 * of the obj, so we can safely drop the lock and continue
4145 mutex_unlock(&dev->struct_mutex);
4146 unwritten = copy_from_user(vaddr, user_data, args->size);
4147 mutex_lock(&dev->struct_mutex);
4152 intel_gtt_chipset_flush();
4156 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
4158 struct drm_i915_file_private *file_priv = file->driver_priv;
4160 /* Clean up our request list when the client is going away, so that
4161 * later retire_requests won't dereference our soon-to-be-gone
4164 spin_lock(&file_priv->mm.lock);
4165 while (!list_empty(&file_priv->mm.request_list)) {
4166 struct drm_i915_gem_request *request;
4168 request = list_first_entry(&file_priv->mm.request_list,
4169 struct drm_i915_gem_request,
4171 list_del(&request->client_list);
4172 request->file_priv = NULL;
4174 spin_unlock(&file_priv->mm.lock);
4178 i915_gpu_is_active(struct drm_device *dev)
4180 drm_i915_private_t *dev_priv = dev->dev_private;
4183 lists_empty = list_empty(&dev_priv->mm.flushing_list) &&
4184 list_empty(&dev_priv->mm.active_list);
4186 return !lists_empty;
4190 i915_gem_inactive_shrink(struct shrinker *shrinker, struct shrink_control *sc)
4192 struct drm_i915_private *dev_priv =
4193 container_of(shrinker,
4194 struct drm_i915_private,
4195 mm.inactive_shrinker);
4196 struct drm_device *dev = dev_priv->dev;
4197 struct drm_i915_gem_object *obj, *next;
4198 int nr_to_scan = sc->nr_to_scan;
4201 if (!mutex_trylock(&dev->struct_mutex))
4204 /* "fast-path" to count number of available objects */
4205 if (nr_to_scan == 0) {
4207 list_for_each_entry(obj,
4208 &dev_priv->mm.inactive_list,
4211 mutex_unlock(&dev->struct_mutex);
4212 return cnt / 100 * sysctl_vfs_cache_pressure;
4216 /* first scan for clean buffers */
4217 i915_gem_retire_requests(dev);
4219 list_for_each_entry_safe(obj, next,
4220 &dev_priv->mm.inactive_list,
4222 if (i915_gem_object_is_purgeable(obj)) {
4223 if (i915_gem_object_unbind(obj) == 0 &&
4229 /* second pass, evict/count anything still on the inactive list */
4231 list_for_each_entry_safe(obj, next,
4232 &dev_priv->mm.inactive_list,
4235 i915_gem_object_unbind(obj) == 0)
4241 if (nr_to_scan && i915_gpu_is_active(dev)) {
4243 * We are desperate for pages, so as a last resort, wait
4244 * for the GPU to finish and discard whatever we can.
4245 * This has a dramatic impact to reduce the number of
4246 * OOM-killer events whilst running the GPU aggressively.
4248 if (i915_gpu_idle(dev, true) == 0)
4251 mutex_unlock(&dev->struct_mutex);
4252 return cnt / 100 * sysctl_vfs_cache_pressure;