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/swap.h>
35 #include <linux/pci.h>
37 #define I915_GEM_GPU_DOMAINS (~(I915_GEM_DOMAIN_CPU | I915_GEM_DOMAIN_GTT))
39 static void i915_gem_object_flush_gpu_write_domain(struct drm_gem_object *obj);
40 static void i915_gem_object_flush_gtt_write_domain(struct drm_gem_object *obj);
41 static void i915_gem_object_flush_cpu_write_domain(struct drm_gem_object *obj);
42 static int i915_gem_object_set_to_cpu_domain(struct drm_gem_object *obj,
44 static int i915_gem_object_set_cpu_read_domain_range(struct drm_gem_object *obj,
47 static void i915_gem_object_set_to_full_cpu_read_domain(struct drm_gem_object *obj);
48 static int i915_gem_object_wait_rendering(struct drm_gem_object *obj);
49 static int i915_gem_object_bind_to_gtt(struct drm_gem_object *obj,
51 static void i915_gem_clear_fence_reg(struct drm_gem_object *obj);
52 static int i915_gem_evict_something(struct drm_device *dev, int min_size);
53 static int i915_gem_evict_from_inactive_list(struct drm_device *dev);
54 static int i915_gem_phys_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
55 struct drm_i915_gem_pwrite *args,
56 struct drm_file *file_priv);
58 static LIST_HEAD(shrink_list);
59 static DEFINE_SPINLOCK(shrink_list_lock);
61 int i915_gem_do_init(struct drm_device *dev, unsigned long start,
64 drm_i915_private_t *dev_priv = dev->dev_private;
67 (start & (PAGE_SIZE - 1)) != 0 ||
68 (end & (PAGE_SIZE - 1)) != 0) {
72 drm_mm_init(&dev_priv->mm.gtt_space, start,
75 dev->gtt_total = (uint32_t) (end - start);
81 i915_gem_init_ioctl(struct drm_device *dev, void *data,
82 struct drm_file *file_priv)
84 struct drm_i915_gem_init *args = data;
87 mutex_lock(&dev->struct_mutex);
88 ret = i915_gem_do_init(dev, args->gtt_start, args->gtt_end);
89 mutex_unlock(&dev->struct_mutex);
95 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
96 struct drm_file *file_priv)
98 struct drm_i915_gem_get_aperture *args = data;
100 if (!(dev->driver->driver_features & DRIVER_GEM))
103 args->aper_size = dev->gtt_total;
104 args->aper_available_size = (args->aper_size -
105 atomic_read(&dev->pin_memory));
112 * Creates a new mm object and returns a handle to it.
115 i915_gem_create_ioctl(struct drm_device *dev, void *data,
116 struct drm_file *file_priv)
118 struct drm_i915_gem_create *args = data;
119 struct drm_gem_object *obj;
123 args->size = roundup(args->size, PAGE_SIZE);
125 /* Allocate the new object */
126 obj = drm_gem_object_alloc(dev, args->size);
130 ret = drm_gem_handle_create(file_priv, obj, &handle);
131 mutex_lock(&dev->struct_mutex);
132 drm_gem_object_handle_unreference(obj);
133 mutex_unlock(&dev->struct_mutex);
138 args->handle = handle;
144 fast_shmem_read(struct page **pages,
145 loff_t page_base, int page_offset,
152 vaddr = kmap_atomic(pages[page_base >> PAGE_SHIFT], KM_USER0);
155 unwritten = __copy_to_user_inatomic(data, vaddr + page_offset, length);
156 kunmap_atomic(vaddr, KM_USER0);
164 static int i915_gem_object_needs_bit17_swizzle(struct drm_gem_object *obj)
166 drm_i915_private_t *dev_priv = obj->dev->dev_private;
167 struct drm_i915_gem_object *obj_priv = obj->driver_private;
169 return dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_9_10_17 &&
170 obj_priv->tiling_mode != I915_TILING_NONE;
174 slow_shmem_copy(struct page *dst_page,
176 struct page *src_page,
180 char *dst_vaddr, *src_vaddr;
182 dst_vaddr = kmap_atomic(dst_page, KM_USER0);
183 if (dst_vaddr == NULL)
186 src_vaddr = kmap_atomic(src_page, KM_USER1);
187 if (src_vaddr == NULL) {
188 kunmap_atomic(dst_vaddr, KM_USER0);
192 memcpy(dst_vaddr + dst_offset, src_vaddr + src_offset, length);
194 kunmap_atomic(src_vaddr, KM_USER1);
195 kunmap_atomic(dst_vaddr, KM_USER0);
201 slow_shmem_bit17_copy(struct page *gpu_page,
203 struct page *cpu_page,
208 char *gpu_vaddr, *cpu_vaddr;
210 /* Use the unswizzled path if this page isn't affected. */
211 if ((page_to_phys(gpu_page) & (1 << 17)) == 0) {
213 return slow_shmem_copy(cpu_page, cpu_offset,
214 gpu_page, gpu_offset, length);
216 return slow_shmem_copy(gpu_page, gpu_offset,
217 cpu_page, cpu_offset, length);
220 gpu_vaddr = kmap_atomic(gpu_page, KM_USER0);
221 if (gpu_vaddr == NULL)
224 cpu_vaddr = kmap_atomic(cpu_page, KM_USER1);
225 if (cpu_vaddr == NULL) {
226 kunmap_atomic(gpu_vaddr, KM_USER0);
230 /* Copy the data, XORing A6 with A17 (1). The user already knows he's
231 * XORing with the other bits (A9 for Y, A9 and A10 for X)
234 int cacheline_end = ALIGN(gpu_offset + 1, 64);
235 int this_length = min(cacheline_end - gpu_offset, length);
236 int swizzled_gpu_offset = gpu_offset ^ 64;
239 memcpy(cpu_vaddr + cpu_offset,
240 gpu_vaddr + swizzled_gpu_offset,
243 memcpy(gpu_vaddr + swizzled_gpu_offset,
244 cpu_vaddr + cpu_offset,
247 cpu_offset += this_length;
248 gpu_offset += this_length;
249 length -= this_length;
252 kunmap_atomic(cpu_vaddr, KM_USER1);
253 kunmap_atomic(gpu_vaddr, KM_USER0);
259 * This is the fast shmem pread path, which attempts to copy_from_user directly
260 * from the backing pages of the object to the user's address space. On a
261 * fault, it fails so we can fall back to i915_gem_shmem_pwrite_slow().
264 i915_gem_shmem_pread_fast(struct drm_device *dev, struct drm_gem_object *obj,
265 struct drm_i915_gem_pread *args,
266 struct drm_file *file_priv)
268 struct drm_i915_gem_object *obj_priv = obj->driver_private;
270 loff_t offset, page_base;
271 char __user *user_data;
272 int page_offset, page_length;
275 user_data = (char __user *) (uintptr_t) args->data_ptr;
278 mutex_lock(&dev->struct_mutex);
280 ret = i915_gem_object_get_pages(obj, 0);
284 ret = i915_gem_object_set_cpu_read_domain_range(obj, args->offset,
289 obj_priv = obj->driver_private;
290 offset = args->offset;
293 /* Operation in this page
295 * page_base = page offset within aperture
296 * page_offset = offset within page
297 * page_length = bytes to copy for this page
299 page_base = (offset & ~(PAGE_SIZE-1));
300 page_offset = offset & (PAGE_SIZE-1);
301 page_length = remain;
302 if ((page_offset + remain) > PAGE_SIZE)
303 page_length = PAGE_SIZE - page_offset;
305 ret = fast_shmem_read(obj_priv->pages,
306 page_base, page_offset,
307 user_data, page_length);
311 remain -= page_length;
312 user_data += page_length;
313 offset += page_length;
317 i915_gem_object_put_pages(obj);
319 mutex_unlock(&dev->struct_mutex);
325 i915_gem_object_get_pages_or_evict(struct drm_gem_object *obj)
329 ret = i915_gem_object_get_pages(obj, __GFP_NORETRY | __GFP_NOWARN);
331 /* If we've insufficient memory to map in the pages, attempt
332 * to make some space by throwing out some old buffers.
334 if (ret == -ENOMEM) {
335 struct drm_device *dev = obj->dev;
337 ret = i915_gem_evict_something(dev, obj->size);
341 ret = i915_gem_object_get_pages(obj, 0);
348 * This is the fallback shmem pread path, which allocates temporary storage
349 * in kernel space to copy_to_user into outside of the struct_mutex, so we
350 * can copy out of the object's backing pages while holding the struct mutex
351 * and not take page faults.
354 i915_gem_shmem_pread_slow(struct drm_device *dev, struct drm_gem_object *obj,
355 struct drm_i915_gem_pread *args,
356 struct drm_file *file_priv)
358 struct drm_i915_gem_object *obj_priv = obj->driver_private;
359 struct mm_struct *mm = current->mm;
360 struct page **user_pages;
362 loff_t offset, pinned_pages, i;
363 loff_t first_data_page, last_data_page, num_pages;
364 int shmem_page_index, shmem_page_offset;
365 int data_page_index, data_page_offset;
368 uint64_t data_ptr = args->data_ptr;
369 int do_bit17_swizzling;
373 /* Pin the user pages containing the data. We can't fault while
374 * holding the struct mutex, yet we want to hold it while
375 * dereferencing the user data.
377 first_data_page = data_ptr / PAGE_SIZE;
378 last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
379 num_pages = last_data_page - first_data_page + 1;
381 user_pages = drm_calloc_large(num_pages, sizeof(struct page *));
382 if (user_pages == NULL)
385 down_read(&mm->mmap_sem);
386 pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
387 num_pages, 1, 0, user_pages, NULL);
388 up_read(&mm->mmap_sem);
389 if (pinned_pages < num_pages) {
391 goto fail_put_user_pages;
394 do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
396 mutex_lock(&dev->struct_mutex);
398 ret = i915_gem_object_get_pages_or_evict(obj);
402 ret = i915_gem_object_set_cpu_read_domain_range(obj, args->offset,
407 obj_priv = obj->driver_private;
408 offset = args->offset;
411 /* Operation in this page
413 * shmem_page_index = page number within shmem file
414 * shmem_page_offset = offset within page in shmem file
415 * data_page_index = page number in get_user_pages return
416 * data_page_offset = offset with data_page_index page.
417 * page_length = bytes to copy for this page
419 shmem_page_index = offset / PAGE_SIZE;
420 shmem_page_offset = offset & ~PAGE_MASK;
421 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
422 data_page_offset = data_ptr & ~PAGE_MASK;
424 page_length = remain;
425 if ((shmem_page_offset + page_length) > PAGE_SIZE)
426 page_length = PAGE_SIZE - shmem_page_offset;
427 if ((data_page_offset + page_length) > PAGE_SIZE)
428 page_length = PAGE_SIZE - data_page_offset;
430 if (do_bit17_swizzling) {
431 ret = slow_shmem_bit17_copy(obj_priv->pages[shmem_page_index],
433 user_pages[data_page_index],
438 ret = slow_shmem_copy(user_pages[data_page_index],
440 obj_priv->pages[shmem_page_index],
447 remain -= page_length;
448 data_ptr += page_length;
449 offset += page_length;
453 i915_gem_object_put_pages(obj);
455 mutex_unlock(&dev->struct_mutex);
457 for (i = 0; i < pinned_pages; i++) {
458 SetPageDirty(user_pages[i]);
459 page_cache_release(user_pages[i]);
461 drm_free_large(user_pages);
467 * Reads data from the object referenced by handle.
469 * On error, the contents of *data are undefined.
472 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
473 struct drm_file *file_priv)
475 struct drm_i915_gem_pread *args = data;
476 struct drm_gem_object *obj;
477 struct drm_i915_gem_object *obj_priv;
480 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
483 obj_priv = obj->driver_private;
485 /* Bounds check source.
487 * XXX: This could use review for overflow issues...
489 if (args->offset > obj->size || args->size > obj->size ||
490 args->offset + args->size > obj->size) {
491 drm_gem_object_unreference(obj);
495 if (i915_gem_object_needs_bit17_swizzle(obj)) {
496 ret = i915_gem_shmem_pread_slow(dev, obj, args, file_priv);
498 ret = i915_gem_shmem_pread_fast(dev, obj, args, file_priv);
500 ret = i915_gem_shmem_pread_slow(dev, obj, args,
504 drm_gem_object_unreference(obj);
509 /* This is the fast write path which cannot handle
510 * page faults in the source data
514 fast_user_write(struct io_mapping *mapping,
515 loff_t page_base, int page_offset,
516 char __user *user_data,
520 unsigned long unwritten;
522 vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
523 unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + page_offset,
525 io_mapping_unmap_atomic(vaddr_atomic);
531 /* Here's the write path which can sleep for
536 slow_kernel_write(struct io_mapping *mapping,
537 loff_t gtt_base, int gtt_offset,
538 struct page *user_page, int user_offset,
541 char *src_vaddr, *dst_vaddr;
542 unsigned long unwritten;
544 dst_vaddr = io_mapping_map_atomic_wc(mapping, gtt_base);
545 src_vaddr = kmap_atomic(user_page, KM_USER1);
546 unwritten = __copy_from_user_inatomic_nocache(dst_vaddr + gtt_offset,
547 src_vaddr + user_offset,
549 kunmap_atomic(src_vaddr, KM_USER1);
550 io_mapping_unmap_atomic(dst_vaddr);
557 fast_shmem_write(struct page **pages,
558 loff_t page_base, int page_offset,
563 unsigned long unwritten;
565 vaddr = kmap_atomic(pages[page_base >> PAGE_SHIFT], KM_USER0);
568 unwritten = __copy_from_user_inatomic(vaddr + page_offset, data, length);
569 kunmap_atomic(vaddr, KM_USER0);
577 * This is the fast pwrite path, where we copy the data directly from the
578 * user into the GTT, uncached.
581 i915_gem_gtt_pwrite_fast(struct drm_device *dev, struct drm_gem_object *obj,
582 struct drm_i915_gem_pwrite *args,
583 struct drm_file *file_priv)
585 struct drm_i915_gem_object *obj_priv = obj->driver_private;
586 drm_i915_private_t *dev_priv = dev->dev_private;
588 loff_t offset, page_base;
589 char __user *user_data;
590 int page_offset, page_length;
593 user_data = (char __user *) (uintptr_t) args->data_ptr;
595 if (!access_ok(VERIFY_READ, user_data, remain))
599 mutex_lock(&dev->struct_mutex);
600 ret = i915_gem_object_pin(obj, 0);
602 mutex_unlock(&dev->struct_mutex);
605 ret = i915_gem_object_set_to_gtt_domain(obj, 1);
609 obj_priv = obj->driver_private;
610 offset = obj_priv->gtt_offset + args->offset;
613 /* Operation in this page
615 * page_base = page offset within aperture
616 * page_offset = offset within page
617 * page_length = bytes to copy for this page
619 page_base = (offset & ~(PAGE_SIZE-1));
620 page_offset = offset & (PAGE_SIZE-1);
621 page_length = remain;
622 if ((page_offset + remain) > PAGE_SIZE)
623 page_length = PAGE_SIZE - page_offset;
625 ret = fast_user_write (dev_priv->mm.gtt_mapping, page_base,
626 page_offset, user_data, page_length);
628 /* If we get a fault while copying data, then (presumably) our
629 * source page isn't available. Return the error and we'll
630 * retry in the slow path.
635 remain -= page_length;
636 user_data += page_length;
637 offset += page_length;
641 i915_gem_object_unpin(obj);
642 mutex_unlock(&dev->struct_mutex);
648 * This is the fallback GTT pwrite path, which uses get_user_pages to pin
649 * the memory and maps it using kmap_atomic for copying.
651 * This code resulted in x11perf -rgb10text consuming about 10% more CPU
652 * than using i915_gem_gtt_pwrite_fast on a G45 (32-bit).
655 i915_gem_gtt_pwrite_slow(struct drm_device *dev, struct drm_gem_object *obj,
656 struct drm_i915_gem_pwrite *args,
657 struct drm_file *file_priv)
659 struct drm_i915_gem_object *obj_priv = obj->driver_private;
660 drm_i915_private_t *dev_priv = dev->dev_private;
662 loff_t gtt_page_base, offset;
663 loff_t first_data_page, last_data_page, num_pages;
664 loff_t pinned_pages, i;
665 struct page **user_pages;
666 struct mm_struct *mm = current->mm;
667 int gtt_page_offset, data_page_offset, data_page_index, page_length;
669 uint64_t data_ptr = args->data_ptr;
673 /* Pin the user pages containing the data. We can't fault while
674 * holding the struct mutex, and all of the pwrite implementations
675 * want to hold it while dereferencing the user data.
677 first_data_page = data_ptr / PAGE_SIZE;
678 last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
679 num_pages = last_data_page - first_data_page + 1;
681 user_pages = drm_calloc_large(num_pages, sizeof(struct page *));
682 if (user_pages == NULL)
685 down_read(&mm->mmap_sem);
686 pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
687 num_pages, 0, 0, user_pages, NULL);
688 up_read(&mm->mmap_sem);
689 if (pinned_pages < num_pages) {
691 goto out_unpin_pages;
694 mutex_lock(&dev->struct_mutex);
695 ret = i915_gem_object_pin(obj, 0);
699 ret = i915_gem_object_set_to_gtt_domain(obj, 1);
701 goto out_unpin_object;
703 obj_priv = obj->driver_private;
704 offset = obj_priv->gtt_offset + args->offset;
707 /* Operation in this page
709 * gtt_page_base = page offset within aperture
710 * gtt_page_offset = offset within page in aperture
711 * data_page_index = page number in get_user_pages return
712 * data_page_offset = offset with data_page_index page.
713 * page_length = bytes to copy for this page
715 gtt_page_base = offset & PAGE_MASK;
716 gtt_page_offset = offset & ~PAGE_MASK;
717 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
718 data_page_offset = data_ptr & ~PAGE_MASK;
720 page_length = remain;
721 if ((gtt_page_offset + page_length) > PAGE_SIZE)
722 page_length = PAGE_SIZE - gtt_page_offset;
723 if ((data_page_offset + page_length) > PAGE_SIZE)
724 page_length = PAGE_SIZE - data_page_offset;
726 ret = slow_kernel_write(dev_priv->mm.gtt_mapping,
727 gtt_page_base, gtt_page_offset,
728 user_pages[data_page_index],
732 /* If we get a fault while copying data, then (presumably) our
733 * source page isn't available. Return the error and we'll
734 * retry in the slow path.
737 goto out_unpin_object;
739 remain -= page_length;
740 offset += page_length;
741 data_ptr += page_length;
745 i915_gem_object_unpin(obj);
747 mutex_unlock(&dev->struct_mutex);
749 for (i = 0; i < pinned_pages; i++)
750 page_cache_release(user_pages[i]);
751 drm_free_large(user_pages);
757 * This is the fast shmem pwrite path, which attempts to directly
758 * copy_from_user into the kmapped pages backing the object.
761 i915_gem_shmem_pwrite_fast(struct drm_device *dev, struct drm_gem_object *obj,
762 struct drm_i915_gem_pwrite *args,
763 struct drm_file *file_priv)
765 struct drm_i915_gem_object *obj_priv = obj->driver_private;
767 loff_t offset, page_base;
768 char __user *user_data;
769 int page_offset, page_length;
772 user_data = (char __user *) (uintptr_t) args->data_ptr;
775 mutex_lock(&dev->struct_mutex);
777 ret = i915_gem_object_get_pages(obj, 0);
781 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
785 obj_priv = obj->driver_private;
786 offset = args->offset;
790 /* Operation in this page
792 * page_base = page offset within aperture
793 * page_offset = offset within page
794 * page_length = bytes to copy for this page
796 page_base = (offset & ~(PAGE_SIZE-1));
797 page_offset = offset & (PAGE_SIZE-1);
798 page_length = remain;
799 if ((page_offset + remain) > PAGE_SIZE)
800 page_length = PAGE_SIZE - page_offset;
802 ret = fast_shmem_write(obj_priv->pages,
803 page_base, page_offset,
804 user_data, page_length);
808 remain -= page_length;
809 user_data += page_length;
810 offset += page_length;
814 i915_gem_object_put_pages(obj);
816 mutex_unlock(&dev->struct_mutex);
822 * This is the fallback shmem pwrite path, which uses get_user_pages to pin
823 * the memory and maps it using kmap_atomic for copying.
825 * This avoids taking mmap_sem for faulting on the user's address while the
826 * struct_mutex is held.
829 i915_gem_shmem_pwrite_slow(struct drm_device *dev, struct drm_gem_object *obj,
830 struct drm_i915_gem_pwrite *args,
831 struct drm_file *file_priv)
833 struct drm_i915_gem_object *obj_priv = obj->driver_private;
834 struct mm_struct *mm = current->mm;
835 struct page **user_pages;
837 loff_t offset, pinned_pages, i;
838 loff_t first_data_page, last_data_page, num_pages;
839 int shmem_page_index, shmem_page_offset;
840 int data_page_index, data_page_offset;
843 uint64_t data_ptr = args->data_ptr;
844 int do_bit17_swizzling;
848 /* Pin the user pages containing the data. We can't fault while
849 * holding the struct mutex, and all of the pwrite implementations
850 * want to hold it while dereferencing the user data.
852 first_data_page = data_ptr / PAGE_SIZE;
853 last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
854 num_pages = last_data_page - first_data_page + 1;
856 user_pages = drm_calloc_large(num_pages, sizeof(struct page *));
857 if (user_pages == NULL)
860 down_read(&mm->mmap_sem);
861 pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
862 num_pages, 0, 0, user_pages, NULL);
863 up_read(&mm->mmap_sem);
864 if (pinned_pages < num_pages) {
866 goto fail_put_user_pages;
869 do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
871 mutex_lock(&dev->struct_mutex);
873 ret = i915_gem_object_get_pages_or_evict(obj);
877 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
881 obj_priv = obj->driver_private;
882 offset = args->offset;
886 /* Operation in this page
888 * shmem_page_index = page number within shmem file
889 * shmem_page_offset = offset within page in shmem file
890 * data_page_index = page number in get_user_pages return
891 * data_page_offset = offset with data_page_index page.
892 * page_length = bytes to copy for this page
894 shmem_page_index = offset / PAGE_SIZE;
895 shmem_page_offset = offset & ~PAGE_MASK;
896 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
897 data_page_offset = data_ptr & ~PAGE_MASK;
899 page_length = remain;
900 if ((shmem_page_offset + page_length) > PAGE_SIZE)
901 page_length = PAGE_SIZE - shmem_page_offset;
902 if ((data_page_offset + page_length) > PAGE_SIZE)
903 page_length = PAGE_SIZE - data_page_offset;
905 if (do_bit17_swizzling) {
906 ret = slow_shmem_bit17_copy(obj_priv->pages[shmem_page_index],
908 user_pages[data_page_index],
913 ret = slow_shmem_copy(obj_priv->pages[shmem_page_index],
915 user_pages[data_page_index],
922 remain -= page_length;
923 data_ptr += page_length;
924 offset += page_length;
928 i915_gem_object_put_pages(obj);
930 mutex_unlock(&dev->struct_mutex);
932 for (i = 0; i < pinned_pages; i++)
933 page_cache_release(user_pages[i]);
934 drm_free_large(user_pages);
940 * Writes data to the object referenced by handle.
942 * On error, the contents of the buffer that were to be modified are undefined.
945 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
946 struct drm_file *file_priv)
948 struct drm_i915_gem_pwrite *args = data;
949 struct drm_gem_object *obj;
950 struct drm_i915_gem_object *obj_priv;
953 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
956 obj_priv = obj->driver_private;
958 /* Bounds check destination.
960 * XXX: This could use review for overflow issues...
962 if (args->offset > obj->size || args->size > obj->size ||
963 args->offset + args->size > obj->size) {
964 drm_gem_object_unreference(obj);
968 /* We can only do the GTT pwrite on untiled buffers, as otherwise
969 * it would end up going through the fenced access, and we'll get
970 * different detiling behavior between reading and writing.
971 * pread/pwrite currently are reading and writing from the CPU
972 * perspective, requiring manual detiling by the client.
974 if (obj_priv->phys_obj)
975 ret = i915_gem_phys_pwrite(dev, obj, args, file_priv);
976 else if (obj_priv->tiling_mode == I915_TILING_NONE &&
977 dev->gtt_total != 0) {
978 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file_priv);
979 if (ret == -EFAULT) {
980 ret = i915_gem_gtt_pwrite_slow(dev, obj, args,
983 } else if (i915_gem_object_needs_bit17_swizzle(obj)) {
984 ret = i915_gem_shmem_pwrite_slow(dev, obj, args, file_priv);
986 ret = i915_gem_shmem_pwrite_fast(dev, obj, args, file_priv);
987 if (ret == -EFAULT) {
988 ret = i915_gem_shmem_pwrite_slow(dev, obj, args,
995 DRM_INFO("pwrite failed %d\n", ret);
998 drm_gem_object_unreference(obj);
1004 * Called when user space prepares to use an object with the CPU, either
1005 * through the mmap ioctl's mapping or a GTT mapping.
1008 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1009 struct drm_file *file_priv)
1011 struct drm_i915_private *dev_priv = dev->dev_private;
1012 struct drm_i915_gem_set_domain *args = data;
1013 struct drm_gem_object *obj;
1014 struct drm_i915_gem_object *obj_priv;
1015 uint32_t read_domains = args->read_domains;
1016 uint32_t write_domain = args->write_domain;
1019 if (!(dev->driver->driver_features & DRIVER_GEM))
1022 /* Only handle setting domains to types used by the CPU. */
1023 if (write_domain & I915_GEM_GPU_DOMAINS)
1026 if (read_domains & I915_GEM_GPU_DOMAINS)
1029 /* Having something in the write domain implies it's in the read
1030 * domain, and only that read domain. Enforce that in the request.
1032 if (write_domain != 0 && read_domains != write_domain)
1035 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1038 obj_priv = obj->driver_private;
1040 mutex_lock(&dev->struct_mutex);
1042 intel_mark_busy(dev, obj);
1045 DRM_INFO("set_domain_ioctl %p(%zd), %08x %08x\n",
1046 obj, obj->size, read_domains, write_domain);
1048 if (read_domains & I915_GEM_DOMAIN_GTT) {
1049 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1051 /* Update the LRU on the fence for the CPU access that's
1054 if (obj_priv->fence_reg != I915_FENCE_REG_NONE) {
1055 list_move_tail(&obj_priv->fence_list,
1056 &dev_priv->mm.fence_list);
1059 /* Silently promote "you're not bound, there was nothing to do"
1060 * to success, since the client was just asking us to
1061 * make sure everything was done.
1066 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1069 drm_gem_object_unreference(obj);
1070 mutex_unlock(&dev->struct_mutex);
1075 * Called when user space has done writes to this buffer
1078 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1079 struct drm_file *file_priv)
1081 struct drm_i915_gem_sw_finish *args = data;
1082 struct drm_gem_object *obj;
1083 struct drm_i915_gem_object *obj_priv;
1086 if (!(dev->driver->driver_features & DRIVER_GEM))
1089 mutex_lock(&dev->struct_mutex);
1090 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1092 mutex_unlock(&dev->struct_mutex);
1097 DRM_INFO("%s: sw_finish %d (%p %zd)\n",
1098 __func__, args->handle, obj, obj->size);
1100 obj_priv = obj->driver_private;
1102 /* Pinned buffers may be scanout, so flush the cache */
1103 if (obj_priv->pin_count)
1104 i915_gem_object_flush_cpu_write_domain(obj);
1106 drm_gem_object_unreference(obj);
1107 mutex_unlock(&dev->struct_mutex);
1112 * Maps the contents of an object, returning the address it is mapped
1115 * While the mapping holds a reference on the contents of the object, it doesn't
1116 * imply a ref on the object itself.
1119 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1120 struct drm_file *file_priv)
1122 struct drm_i915_gem_mmap *args = data;
1123 struct drm_gem_object *obj;
1127 if (!(dev->driver->driver_features & DRIVER_GEM))
1130 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1134 offset = args->offset;
1136 down_write(¤t->mm->mmap_sem);
1137 addr = do_mmap(obj->filp, 0, args->size,
1138 PROT_READ | PROT_WRITE, MAP_SHARED,
1140 up_write(¤t->mm->mmap_sem);
1141 mutex_lock(&dev->struct_mutex);
1142 drm_gem_object_unreference(obj);
1143 mutex_unlock(&dev->struct_mutex);
1144 if (IS_ERR((void *)addr))
1147 args->addr_ptr = (uint64_t) addr;
1153 * i915_gem_fault - fault a page into the GTT
1154 * vma: VMA in question
1157 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1158 * from userspace. The fault handler takes care of binding the object to
1159 * the GTT (if needed), allocating and programming a fence register (again,
1160 * only if needed based on whether the old reg is still valid or the object
1161 * is tiled) and inserting a new PTE into the faulting process.
1163 * Note that the faulting process may involve evicting existing objects
1164 * from the GTT and/or fence registers to make room. So performance may
1165 * suffer if the GTT working set is large or there are few fence registers
1168 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1170 struct drm_gem_object *obj = vma->vm_private_data;
1171 struct drm_device *dev = obj->dev;
1172 struct drm_i915_private *dev_priv = dev->dev_private;
1173 struct drm_i915_gem_object *obj_priv = obj->driver_private;
1174 pgoff_t page_offset;
1177 bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1179 /* We don't use vmf->pgoff since that has the fake offset */
1180 page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1183 /* Now bind it into the GTT if needed */
1184 mutex_lock(&dev->struct_mutex);
1185 if (!obj_priv->gtt_space) {
1186 ret = i915_gem_object_bind_to_gtt(obj, 0);
1190 list_add_tail(&obj_priv->list, &dev_priv->mm.inactive_list);
1192 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1197 /* Need a new fence register? */
1198 if (obj_priv->tiling_mode != I915_TILING_NONE) {
1199 ret = i915_gem_object_get_fence_reg(obj);
1204 pfn = ((dev->agp->base + obj_priv->gtt_offset) >> PAGE_SHIFT) +
1207 /* Finally, remap it using the new GTT offset */
1208 ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1210 mutex_unlock(&dev->struct_mutex);
1215 return VM_FAULT_NOPAGE;
1218 return VM_FAULT_OOM;
1220 return VM_FAULT_SIGBUS;
1225 * i915_gem_create_mmap_offset - create a fake mmap offset for an object
1226 * @obj: obj in question
1228 * GEM memory mapping works by handing back to userspace a fake mmap offset
1229 * it can use in a subsequent mmap(2) call. The DRM core code then looks
1230 * up the object based on the offset and sets up the various memory mapping
1233 * This routine allocates and attaches a fake offset for @obj.
1236 i915_gem_create_mmap_offset(struct drm_gem_object *obj)
1238 struct drm_device *dev = obj->dev;
1239 struct drm_gem_mm *mm = dev->mm_private;
1240 struct drm_i915_gem_object *obj_priv = obj->driver_private;
1241 struct drm_map_list *list;
1242 struct drm_local_map *map;
1245 /* Set the object up for mmap'ing */
1246 list = &obj->map_list;
1247 list->map = kzalloc(sizeof(struct drm_map_list), GFP_KERNEL);
1252 map->type = _DRM_GEM;
1253 map->size = obj->size;
1256 /* Get a DRM GEM mmap offset allocated... */
1257 list->file_offset_node = drm_mm_search_free(&mm->offset_manager,
1258 obj->size / PAGE_SIZE, 0, 0);
1259 if (!list->file_offset_node) {
1260 DRM_ERROR("failed to allocate offset for bo %d\n", obj->name);
1265 list->file_offset_node = drm_mm_get_block(list->file_offset_node,
1266 obj->size / PAGE_SIZE, 0);
1267 if (!list->file_offset_node) {
1272 list->hash.key = list->file_offset_node->start;
1273 if (drm_ht_insert_item(&mm->offset_hash, &list->hash)) {
1274 DRM_ERROR("failed to add to map hash\n");
1279 /* By now we should be all set, any drm_mmap request on the offset
1280 * below will get to our mmap & fault handler */
1281 obj_priv->mmap_offset = ((uint64_t) list->hash.key) << PAGE_SHIFT;
1286 drm_mm_put_block(list->file_offset_node);
1294 * i915_gem_release_mmap - remove physical page mappings
1295 * @obj: obj in question
1297 * Preserve the reservation of the mmapping with the DRM core code, but
1298 * relinquish ownership of the pages back to the system.
1300 * It is vital that we remove the page mapping if we have mapped a tiled
1301 * object through the GTT and then lose the fence register due to
1302 * resource pressure. Similarly if the object has been moved out of the
1303 * aperture, than pages mapped into userspace must be revoked. Removing the
1304 * mapping will then trigger a page fault on the next user access, allowing
1305 * fixup by i915_gem_fault().
1308 i915_gem_release_mmap(struct drm_gem_object *obj)
1310 struct drm_device *dev = obj->dev;
1311 struct drm_i915_gem_object *obj_priv = obj->driver_private;
1313 if (dev->dev_mapping)
1314 unmap_mapping_range(dev->dev_mapping,
1315 obj_priv->mmap_offset, obj->size, 1);
1319 i915_gem_free_mmap_offset(struct drm_gem_object *obj)
1321 struct drm_device *dev = obj->dev;
1322 struct drm_i915_gem_object *obj_priv = obj->driver_private;
1323 struct drm_gem_mm *mm = dev->mm_private;
1324 struct drm_map_list *list;
1326 list = &obj->map_list;
1327 drm_ht_remove_item(&mm->offset_hash, &list->hash);
1329 if (list->file_offset_node) {
1330 drm_mm_put_block(list->file_offset_node);
1331 list->file_offset_node = NULL;
1339 obj_priv->mmap_offset = 0;
1343 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1344 * @obj: object to check
1346 * Return the required GTT alignment for an object, taking into account
1347 * potential fence register mapping if needed.
1350 i915_gem_get_gtt_alignment(struct drm_gem_object *obj)
1352 struct drm_device *dev = obj->dev;
1353 struct drm_i915_gem_object *obj_priv = obj->driver_private;
1357 * Minimum alignment is 4k (GTT page size), but might be greater
1358 * if a fence register is needed for the object.
1360 if (IS_I965G(dev) || obj_priv->tiling_mode == I915_TILING_NONE)
1364 * Previous chips need to be aligned to the size of the smallest
1365 * fence register that can contain the object.
1372 for (i = start; i < obj->size; i <<= 1)
1379 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1381 * @data: GTT mapping ioctl data
1382 * @file_priv: GEM object info
1384 * Simply returns the fake offset to userspace so it can mmap it.
1385 * The mmap call will end up in drm_gem_mmap(), which will set things
1386 * up so we can get faults in the handler above.
1388 * The fault handler will take care of binding the object into the GTT
1389 * (since it may have been evicted to make room for something), allocating
1390 * a fence register, and mapping the appropriate aperture address into
1394 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1395 struct drm_file *file_priv)
1397 struct drm_i915_gem_mmap_gtt *args = data;
1398 struct drm_i915_private *dev_priv = dev->dev_private;
1399 struct drm_gem_object *obj;
1400 struct drm_i915_gem_object *obj_priv;
1403 if (!(dev->driver->driver_features & DRIVER_GEM))
1406 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1410 mutex_lock(&dev->struct_mutex);
1412 obj_priv = obj->driver_private;
1414 if (obj_priv->madv != I915_MADV_WILLNEED) {
1415 DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1416 drm_gem_object_unreference(obj);
1417 mutex_unlock(&dev->struct_mutex);
1422 if (!obj_priv->mmap_offset) {
1423 ret = i915_gem_create_mmap_offset(obj);
1425 drm_gem_object_unreference(obj);
1426 mutex_unlock(&dev->struct_mutex);
1431 args->offset = obj_priv->mmap_offset;
1434 * Pull it into the GTT so that we have a page list (makes the
1435 * initial fault faster and any subsequent flushing possible).
1437 if (!obj_priv->agp_mem) {
1438 ret = i915_gem_object_bind_to_gtt(obj, 0);
1440 drm_gem_object_unreference(obj);
1441 mutex_unlock(&dev->struct_mutex);
1444 list_add_tail(&obj_priv->list, &dev_priv->mm.inactive_list);
1447 drm_gem_object_unreference(obj);
1448 mutex_unlock(&dev->struct_mutex);
1454 i915_gem_object_put_pages(struct drm_gem_object *obj)
1456 struct drm_i915_gem_object *obj_priv = obj->driver_private;
1457 int page_count = obj->size / PAGE_SIZE;
1460 BUG_ON(obj_priv->pages_refcount == 0);
1461 BUG_ON(obj_priv->madv == __I915_MADV_PURGED);
1463 if (--obj_priv->pages_refcount != 0)
1466 if (obj_priv->tiling_mode != I915_TILING_NONE)
1467 i915_gem_object_save_bit_17_swizzle(obj);
1469 if (obj_priv->madv == I915_MADV_DONTNEED)
1470 obj_priv->dirty = 0;
1472 for (i = 0; i < page_count; i++) {
1473 if (obj_priv->dirty)
1474 set_page_dirty(obj_priv->pages[i]);
1476 if (obj_priv->madv == I915_MADV_WILLNEED)
1477 mark_page_accessed(obj_priv->pages[i]);
1479 page_cache_release(obj_priv->pages[i]);
1481 obj_priv->dirty = 0;
1483 drm_free_large(obj_priv->pages);
1484 obj_priv->pages = NULL;
1488 i915_gem_object_move_to_active(struct drm_gem_object *obj, uint32_t seqno)
1490 struct drm_device *dev = obj->dev;
1491 drm_i915_private_t *dev_priv = dev->dev_private;
1492 struct drm_i915_gem_object *obj_priv = obj->driver_private;
1494 /* Add a reference if we're newly entering the active list. */
1495 if (!obj_priv->active) {
1496 drm_gem_object_reference(obj);
1497 obj_priv->active = 1;
1499 /* Move from whatever list we were on to the tail of execution. */
1500 spin_lock(&dev_priv->mm.active_list_lock);
1501 list_move_tail(&obj_priv->list,
1502 &dev_priv->mm.active_list);
1503 spin_unlock(&dev_priv->mm.active_list_lock);
1504 obj_priv->last_rendering_seqno = seqno;
1508 i915_gem_object_move_to_flushing(struct drm_gem_object *obj)
1510 struct drm_device *dev = obj->dev;
1511 drm_i915_private_t *dev_priv = dev->dev_private;
1512 struct drm_i915_gem_object *obj_priv = obj->driver_private;
1514 BUG_ON(!obj_priv->active);
1515 list_move_tail(&obj_priv->list, &dev_priv->mm.flushing_list);
1516 obj_priv->last_rendering_seqno = 0;
1519 /* Immediately discard the backing storage */
1521 i915_gem_object_truncate(struct drm_gem_object *obj)
1523 struct drm_i915_gem_object *obj_priv = obj->driver_private;
1524 struct inode *inode;
1526 inode = obj->filp->f_path.dentry->d_inode;
1527 if (inode->i_op->truncate)
1528 inode->i_op->truncate (inode);
1530 obj_priv->madv = __I915_MADV_PURGED;
1534 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj_priv)
1536 return obj_priv->madv == I915_MADV_DONTNEED;
1540 i915_gem_object_move_to_inactive(struct drm_gem_object *obj)
1542 struct drm_device *dev = obj->dev;
1543 drm_i915_private_t *dev_priv = dev->dev_private;
1544 struct drm_i915_gem_object *obj_priv = obj->driver_private;
1546 i915_verify_inactive(dev, __FILE__, __LINE__);
1547 if (obj_priv->pin_count != 0)
1548 list_del_init(&obj_priv->list);
1550 list_move_tail(&obj_priv->list, &dev_priv->mm.inactive_list);
1552 BUG_ON(!list_empty(&obj_priv->gpu_write_list));
1554 obj_priv->last_rendering_seqno = 0;
1555 if (obj_priv->active) {
1556 obj_priv->active = 0;
1557 drm_gem_object_unreference(obj);
1559 i915_verify_inactive(dev, __FILE__, __LINE__);
1562 #define PIPE_CONTROL_FLUSH(addr) \
1563 OUT_RING(GFX_OP_PIPE_CONTROL | PIPE_CONTROL_QW_WRITE | \
1564 PIPE_CONTROL_DEPTH_STALL); \
1565 OUT_RING(addr | PIPE_CONTROL_GLOBAL_GTT); \
1570 * Creates a new sequence number, emitting a write of it to the status page
1571 * plus an interrupt, which will trigger i915_user_interrupt_handler.
1573 * Must be called with struct_lock held.
1575 * Returned sequence numbers are nonzero on success.
1578 i915_add_request(struct drm_device *dev, struct drm_file *file_priv,
1579 uint32_t flush_domains)
1581 drm_i915_private_t *dev_priv = dev->dev_private;
1582 struct drm_i915_file_private *i915_file_priv = NULL;
1583 struct drm_i915_gem_request *request;
1588 if (file_priv != NULL)
1589 i915_file_priv = file_priv->driver_priv;
1591 request = kzalloc(sizeof(*request), GFP_KERNEL);
1592 if (request == NULL)
1595 /* Grab the seqno we're going to make this request be, and bump the
1596 * next (skipping 0 so it can be the reserved no-seqno value).
1598 seqno = dev_priv->mm.next_gem_seqno;
1599 dev_priv->mm.next_gem_seqno++;
1600 if (dev_priv->mm.next_gem_seqno == 0)
1601 dev_priv->mm.next_gem_seqno++;
1603 if (HAS_PIPE_CONTROL(dev)) {
1604 u32 scratch_addr = dev_priv->seqno_gfx_addr + 128;
1607 * Workaround qword write incoherence by flushing the
1608 * PIPE_NOTIFY buffers out to memory before requesting
1612 OUT_RING(GFX_OP_PIPE_CONTROL | PIPE_CONTROL_QW_WRITE |
1613 PIPE_CONTROL_WC_FLUSH | PIPE_CONTROL_TC_FLUSH);
1614 OUT_RING(dev_priv->seqno_gfx_addr | PIPE_CONTROL_GLOBAL_GTT);
1617 PIPE_CONTROL_FLUSH(scratch_addr);
1618 scratch_addr += 128; /* write to separate cachelines */
1619 PIPE_CONTROL_FLUSH(scratch_addr);
1620 scratch_addr += 128;
1621 PIPE_CONTROL_FLUSH(scratch_addr);
1622 scratch_addr += 128;
1623 PIPE_CONTROL_FLUSH(scratch_addr);
1624 scratch_addr += 128;
1625 PIPE_CONTROL_FLUSH(scratch_addr);
1626 scratch_addr += 128;
1627 PIPE_CONTROL_FLUSH(scratch_addr);
1628 OUT_RING(GFX_OP_PIPE_CONTROL | PIPE_CONTROL_QW_WRITE |
1629 PIPE_CONTROL_WC_FLUSH | PIPE_CONTROL_TC_FLUSH |
1630 PIPE_CONTROL_NOTIFY);
1631 OUT_RING(dev_priv->seqno_gfx_addr | PIPE_CONTROL_GLOBAL_GTT);
1637 OUT_RING(MI_STORE_DWORD_INDEX);
1638 OUT_RING(I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
1641 OUT_RING(MI_USER_INTERRUPT);
1645 DRM_DEBUG_DRIVER("%d\n", seqno);
1647 request->seqno = seqno;
1648 request->emitted_jiffies = jiffies;
1649 was_empty = list_empty(&dev_priv->mm.request_list);
1650 list_add_tail(&request->list, &dev_priv->mm.request_list);
1651 if (i915_file_priv) {
1652 list_add_tail(&request->client_list,
1653 &i915_file_priv->mm.request_list);
1655 INIT_LIST_HEAD(&request->client_list);
1658 /* Associate any objects on the flushing list matching the write
1659 * domain we're flushing with our flush.
1661 if (flush_domains != 0) {
1662 struct drm_i915_gem_object *obj_priv, *next;
1664 list_for_each_entry_safe(obj_priv, next,
1665 &dev_priv->mm.gpu_write_list,
1667 struct drm_gem_object *obj = obj_priv->obj;
1669 if ((obj->write_domain & flush_domains) ==
1670 obj->write_domain) {
1671 uint32_t old_write_domain = obj->write_domain;
1673 obj->write_domain = 0;
1674 list_del_init(&obj_priv->gpu_write_list);
1675 i915_gem_object_move_to_active(obj, seqno);
1677 trace_i915_gem_object_change_domain(obj,
1685 if (!dev_priv->mm.suspended) {
1686 mod_timer(&dev_priv->hangcheck_timer, jiffies + DRM_I915_HANGCHECK_PERIOD);
1688 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1694 * Command execution barrier
1696 * Ensures that all commands in the ring are finished
1697 * before signalling the CPU
1700 i915_retire_commands(struct drm_device *dev)
1702 drm_i915_private_t *dev_priv = dev->dev_private;
1703 uint32_t cmd = MI_FLUSH | MI_NO_WRITE_FLUSH;
1704 uint32_t flush_domains = 0;
1707 /* The sampler always gets flushed on i965 (sigh) */
1709 flush_domains |= I915_GEM_DOMAIN_SAMPLER;
1712 OUT_RING(0); /* noop */
1714 return flush_domains;
1718 * Moves buffers associated only with the given active seqno from the active
1719 * to inactive list, potentially freeing them.
1722 i915_gem_retire_request(struct drm_device *dev,
1723 struct drm_i915_gem_request *request)
1725 drm_i915_private_t *dev_priv = dev->dev_private;
1727 trace_i915_gem_request_retire(dev, request->seqno);
1729 /* Move any buffers on the active list that are no longer referenced
1730 * by the ringbuffer to the flushing/inactive lists as appropriate.
1732 spin_lock(&dev_priv->mm.active_list_lock);
1733 while (!list_empty(&dev_priv->mm.active_list)) {
1734 struct drm_gem_object *obj;
1735 struct drm_i915_gem_object *obj_priv;
1737 obj_priv = list_first_entry(&dev_priv->mm.active_list,
1738 struct drm_i915_gem_object,
1740 obj = obj_priv->obj;
1742 /* If the seqno being retired doesn't match the oldest in the
1743 * list, then the oldest in the list must still be newer than
1746 if (obj_priv->last_rendering_seqno != request->seqno)
1750 DRM_INFO("%s: retire %d moves to inactive list %p\n",
1751 __func__, request->seqno, obj);
1754 if (obj->write_domain != 0)
1755 i915_gem_object_move_to_flushing(obj);
1757 /* Take a reference on the object so it won't be
1758 * freed while the spinlock is held. The list
1759 * protection for this spinlock is safe when breaking
1760 * the lock like this since the next thing we do
1761 * is just get the head of the list again.
1763 drm_gem_object_reference(obj);
1764 i915_gem_object_move_to_inactive(obj);
1765 spin_unlock(&dev_priv->mm.active_list_lock);
1766 drm_gem_object_unreference(obj);
1767 spin_lock(&dev_priv->mm.active_list_lock);
1771 spin_unlock(&dev_priv->mm.active_list_lock);
1775 * Returns true if seq1 is later than seq2.
1778 i915_seqno_passed(uint32_t seq1, uint32_t seq2)
1780 return (int32_t)(seq1 - seq2) >= 0;
1784 i915_get_gem_seqno(struct drm_device *dev)
1786 drm_i915_private_t *dev_priv = dev->dev_private;
1788 if (HAS_PIPE_CONTROL(dev))
1789 return ((volatile u32 *)(dev_priv->seqno_page))[0];
1791 return READ_HWSP(dev_priv, I915_GEM_HWS_INDEX);
1795 * This function clears the request list as sequence numbers are passed.
1798 i915_gem_retire_requests(struct drm_device *dev)
1800 drm_i915_private_t *dev_priv = dev->dev_private;
1803 if (!dev_priv->hw_status_page || list_empty(&dev_priv->mm.request_list))
1806 seqno = i915_get_gem_seqno(dev);
1808 while (!list_empty(&dev_priv->mm.request_list)) {
1809 struct drm_i915_gem_request *request;
1810 uint32_t retiring_seqno;
1812 request = list_first_entry(&dev_priv->mm.request_list,
1813 struct drm_i915_gem_request,
1815 retiring_seqno = request->seqno;
1817 if (i915_seqno_passed(seqno, retiring_seqno) ||
1818 atomic_read(&dev_priv->mm.wedged)) {
1819 i915_gem_retire_request(dev, request);
1821 list_del(&request->list);
1822 list_del(&request->client_list);
1828 if (unlikely (dev_priv->trace_irq_seqno &&
1829 i915_seqno_passed(dev_priv->trace_irq_seqno, seqno))) {
1830 i915_user_irq_put(dev);
1831 dev_priv->trace_irq_seqno = 0;
1836 i915_gem_retire_work_handler(struct work_struct *work)
1838 drm_i915_private_t *dev_priv;
1839 struct drm_device *dev;
1841 dev_priv = container_of(work, drm_i915_private_t,
1842 mm.retire_work.work);
1843 dev = dev_priv->dev;
1845 mutex_lock(&dev->struct_mutex);
1846 i915_gem_retire_requests(dev);
1847 if (!dev_priv->mm.suspended &&
1848 !list_empty(&dev_priv->mm.request_list))
1849 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1850 mutex_unlock(&dev->struct_mutex);
1854 i915_do_wait_request(struct drm_device *dev, uint32_t seqno, int interruptible)
1856 drm_i915_private_t *dev_priv = dev->dev_private;
1862 if (atomic_read(&dev_priv->mm.wedged))
1865 if (!i915_seqno_passed(i915_get_gem_seqno(dev), seqno)) {
1866 if (HAS_PCH_SPLIT(dev))
1867 ier = I915_READ(DEIER) | I915_READ(GTIER);
1869 ier = I915_READ(IER);
1871 DRM_ERROR("something (likely vbetool) disabled "
1872 "interrupts, re-enabling\n");
1873 i915_driver_irq_preinstall(dev);
1874 i915_driver_irq_postinstall(dev);
1877 trace_i915_gem_request_wait_begin(dev, seqno);
1879 dev_priv->mm.waiting_gem_seqno = seqno;
1880 i915_user_irq_get(dev);
1882 ret = wait_event_interruptible(dev_priv->irq_queue,
1883 i915_seqno_passed(i915_get_gem_seqno(dev), seqno) ||
1884 atomic_read(&dev_priv->mm.wedged));
1886 wait_event(dev_priv->irq_queue,
1887 i915_seqno_passed(i915_get_gem_seqno(dev), seqno) ||
1888 atomic_read(&dev_priv->mm.wedged));
1890 i915_user_irq_put(dev);
1891 dev_priv->mm.waiting_gem_seqno = 0;
1893 trace_i915_gem_request_wait_end(dev, seqno);
1895 if (atomic_read(&dev_priv->mm.wedged))
1898 if (ret && ret != -ERESTARTSYS)
1899 DRM_ERROR("%s returns %d (awaiting %d at %d)\n",
1900 __func__, ret, seqno, i915_get_gem_seqno(dev));
1902 /* Directly dispatch request retiring. While we have the work queue
1903 * to handle this, the waiter on a request often wants an associated
1904 * buffer to have made it to the inactive list, and we would need
1905 * a separate wait queue to handle that.
1908 i915_gem_retire_requests(dev);
1914 * Waits for a sequence number to be signaled, and cleans up the
1915 * request and object lists appropriately for that event.
1918 i915_wait_request(struct drm_device *dev, uint32_t seqno)
1920 return i915_do_wait_request(dev, seqno, 1);
1924 i915_gem_flush(struct drm_device *dev,
1925 uint32_t invalidate_domains,
1926 uint32_t flush_domains)
1928 drm_i915_private_t *dev_priv = dev->dev_private;
1933 DRM_INFO("%s: invalidate %08x flush %08x\n", __func__,
1934 invalidate_domains, flush_domains);
1936 trace_i915_gem_request_flush(dev, dev_priv->mm.next_gem_seqno,
1937 invalidate_domains, flush_domains);
1939 if (flush_domains & I915_GEM_DOMAIN_CPU)
1940 drm_agp_chipset_flush(dev);
1942 if ((invalidate_domains | flush_domains) & I915_GEM_GPU_DOMAINS) {
1944 * read/write caches:
1946 * I915_GEM_DOMAIN_RENDER is always invalidated, but is
1947 * only flushed if MI_NO_WRITE_FLUSH is unset. On 965, it is
1948 * also flushed at 2d versus 3d pipeline switches.
1952 * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
1953 * MI_READ_FLUSH is set, and is always flushed on 965.
1955 * I915_GEM_DOMAIN_COMMAND may not exist?
1957 * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
1958 * invalidated when MI_EXE_FLUSH is set.
1960 * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
1961 * invalidated with every MI_FLUSH.
1965 * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
1966 * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
1967 * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
1968 * are flushed at any MI_FLUSH.
1971 cmd = MI_FLUSH | MI_NO_WRITE_FLUSH;
1972 if ((invalidate_domains|flush_domains) &
1973 I915_GEM_DOMAIN_RENDER)
1974 cmd &= ~MI_NO_WRITE_FLUSH;
1975 if (!IS_I965G(dev)) {
1977 * On the 965, the sampler cache always gets flushed
1978 * and this bit is reserved.
1980 if (invalidate_domains & I915_GEM_DOMAIN_SAMPLER)
1981 cmd |= MI_READ_FLUSH;
1983 if (invalidate_domains & I915_GEM_DOMAIN_INSTRUCTION)
1984 cmd |= MI_EXE_FLUSH;
1987 DRM_INFO("%s: queue flush %08x to ring\n", __func__, cmd);
1997 * Ensures that all rendering to the object has completed and the object is
1998 * safe to unbind from the GTT or access from the CPU.
2001 i915_gem_object_wait_rendering(struct drm_gem_object *obj)
2003 struct drm_device *dev = obj->dev;
2004 struct drm_i915_gem_object *obj_priv = obj->driver_private;
2007 /* This function only exists to support waiting for existing rendering,
2008 * not for emitting required flushes.
2010 BUG_ON((obj->write_domain & I915_GEM_GPU_DOMAINS) != 0);
2012 /* If there is rendering queued on the buffer being evicted, wait for
2015 if (obj_priv->active) {
2017 DRM_INFO("%s: object %p wait for seqno %08x\n",
2018 __func__, obj, obj_priv->last_rendering_seqno);
2020 ret = i915_wait_request(dev, obj_priv->last_rendering_seqno);
2029 * Unbinds an object from the GTT aperture.
2032 i915_gem_object_unbind(struct drm_gem_object *obj)
2034 struct drm_device *dev = obj->dev;
2035 struct drm_i915_gem_object *obj_priv = obj->driver_private;
2039 DRM_INFO("%s:%d %p\n", __func__, __LINE__, obj);
2040 DRM_INFO("gtt_space %p\n", obj_priv->gtt_space);
2042 if (obj_priv->gtt_space == NULL)
2045 if (obj_priv->pin_count != 0) {
2046 DRM_ERROR("Attempting to unbind pinned buffer\n");
2050 /* blow away mappings if mapped through GTT */
2051 i915_gem_release_mmap(obj);
2053 /* Move the object to the CPU domain to ensure that
2054 * any possible CPU writes while it's not in the GTT
2055 * are flushed when we go to remap it. This will
2056 * also ensure that all pending GPU writes are finished
2059 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
2061 if (ret != -ERESTARTSYS)
2062 DRM_ERROR("set_domain failed: %d\n", ret);
2066 BUG_ON(obj_priv->active);
2068 /* release the fence reg _after_ flushing */
2069 if (obj_priv->fence_reg != I915_FENCE_REG_NONE)
2070 i915_gem_clear_fence_reg(obj);
2072 if (obj_priv->agp_mem != NULL) {
2073 drm_unbind_agp(obj_priv->agp_mem);
2074 drm_free_agp(obj_priv->agp_mem, obj->size / PAGE_SIZE);
2075 obj_priv->agp_mem = NULL;
2078 i915_gem_object_put_pages(obj);
2079 BUG_ON(obj_priv->pages_refcount);
2081 if (obj_priv->gtt_space) {
2082 atomic_dec(&dev->gtt_count);
2083 atomic_sub(obj->size, &dev->gtt_memory);
2085 drm_mm_put_block(obj_priv->gtt_space);
2086 obj_priv->gtt_space = NULL;
2089 /* Remove ourselves from the LRU list if present. */
2090 if (!list_empty(&obj_priv->list))
2091 list_del_init(&obj_priv->list);
2093 if (i915_gem_object_is_purgeable(obj_priv))
2094 i915_gem_object_truncate(obj);
2096 trace_i915_gem_object_unbind(obj);
2101 static struct drm_gem_object *
2102 i915_gem_find_inactive_object(struct drm_device *dev, int min_size)
2104 drm_i915_private_t *dev_priv = dev->dev_private;
2105 struct drm_i915_gem_object *obj_priv;
2106 struct drm_gem_object *best = NULL;
2107 struct drm_gem_object *first = NULL;
2109 /* Try to find the smallest clean object */
2110 list_for_each_entry(obj_priv, &dev_priv->mm.inactive_list, list) {
2111 struct drm_gem_object *obj = obj_priv->obj;
2112 if (obj->size >= min_size) {
2113 if ((!obj_priv->dirty ||
2114 i915_gem_object_is_purgeable(obj_priv)) &&
2115 (!best || obj->size < best->size)) {
2117 if (best->size == min_size)
2125 return best ? best : first;
2129 i915_gem_evict_everything(struct drm_device *dev)
2131 drm_i915_private_t *dev_priv = dev->dev_private;
2136 spin_lock(&dev_priv->mm.active_list_lock);
2137 lists_empty = (list_empty(&dev_priv->mm.inactive_list) &&
2138 list_empty(&dev_priv->mm.flushing_list) &&
2139 list_empty(&dev_priv->mm.active_list));
2140 spin_unlock(&dev_priv->mm.active_list_lock);
2145 /* Flush everything (on to the inactive lists) and evict */
2146 i915_gem_flush(dev, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
2147 seqno = i915_add_request(dev, NULL, I915_GEM_GPU_DOMAINS);
2151 ret = i915_wait_request(dev, seqno);
2155 BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
2157 ret = i915_gem_evict_from_inactive_list(dev);
2161 spin_lock(&dev_priv->mm.active_list_lock);
2162 lists_empty = (list_empty(&dev_priv->mm.inactive_list) &&
2163 list_empty(&dev_priv->mm.flushing_list) &&
2164 list_empty(&dev_priv->mm.active_list));
2165 spin_unlock(&dev_priv->mm.active_list_lock);
2166 BUG_ON(!lists_empty);
2172 i915_gem_evict_something(struct drm_device *dev, int min_size)
2174 drm_i915_private_t *dev_priv = dev->dev_private;
2175 struct drm_gem_object *obj;
2179 i915_gem_retire_requests(dev);
2181 /* If there's an inactive buffer available now, grab it
2184 obj = i915_gem_find_inactive_object(dev, min_size);
2186 struct drm_i915_gem_object *obj_priv;
2189 DRM_INFO("%s: evicting %p\n", __func__, obj);
2191 obj_priv = obj->driver_private;
2192 BUG_ON(obj_priv->pin_count != 0);
2193 BUG_ON(obj_priv->active);
2195 /* Wait on the rendering and unbind the buffer. */
2196 return i915_gem_object_unbind(obj);
2199 /* If we didn't get anything, but the ring is still processing
2200 * things, wait for the next to finish and hopefully leave us
2201 * a buffer to evict.
2203 if (!list_empty(&dev_priv->mm.request_list)) {
2204 struct drm_i915_gem_request *request;
2206 request = list_first_entry(&dev_priv->mm.request_list,
2207 struct drm_i915_gem_request,
2210 ret = i915_wait_request(dev, request->seqno);
2217 /* If we didn't have anything on the request list but there
2218 * are buffers awaiting a flush, emit one and try again.
2219 * When we wait on it, those buffers waiting for that flush
2220 * will get moved to inactive.
2222 if (!list_empty(&dev_priv->mm.flushing_list)) {
2223 struct drm_i915_gem_object *obj_priv;
2225 /* Find an object that we can immediately reuse */
2226 list_for_each_entry(obj_priv, &dev_priv->mm.flushing_list, list) {
2227 obj = obj_priv->obj;
2228 if (obj->size >= min_size)
2240 seqno = i915_add_request(dev, NULL, obj->write_domain);
2244 ret = i915_wait_request(dev, seqno);
2252 /* If we didn't do any of the above, there's no single buffer
2253 * large enough to swap out for the new one, so just evict
2254 * everything and start again. (This should be rare.)
2256 if (!list_empty (&dev_priv->mm.inactive_list))
2257 return i915_gem_evict_from_inactive_list(dev);
2259 return i915_gem_evict_everything(dev);
2264 i915_gem_object_get_pages(struct drm_gem_object *obj,
2267 struct drm_i915_gem_object *obj_priv = obj->driver_private;
2269 struct address_space *mapping;
2270 struct inode *inode;
2273 if (obj_priv->pages_refcount++ != 0)
2276 /* Get the list of pages out of our struct file. They'll be pinned
2277 * at this point until we release them.
2279 page_count = obj->size / PAGE_SIZE;
2280 BUG_ON(obj_priv->pages != NULL);
2281 obj_priv->pages = drm_calloc_large(page_count, sizeof(struct page *));
2282 if (obj_priv->pages == NULL) {
2283 obj_priv->pages_refcount--;
2287 inode = obj->filp->f_path.dentry->d_inode;
2288 mapping = inode->i_mapping;
2289 for (i = 0; i < page_count; i++) {
2290 page = read_cache_page_gfp(mapping, i,
2297 obj_priv->pages[i] = page;
2300 if (obj_priv->tiling_mode != I915_TILING_NONE)
2301 i915_gem_object_do_bit_17_swizzle(obj);
2307 page_cache_release(obj_priv->pages[i]);
2309 drm_free_large(obj_priv->pages);
2310 obj_priv->pages = NULL;
2311 obj_priv->pages_refcount--;
2312 return PTR_ERR(page);
2315 static void i965_write_fence_reg(struct drm_i915_fence_reg *reg)
2317 struct drm_gem_object *obj = reg->obj;
2318 struct drm_device *dev = obj->dev;
2319 drm_i915_private_t *dev_priv = dev->dev_private;
2320 struct drm_i915_gem_object *obj_priv = obj->driver_private;
2321 int regnum = obj_priv->fence_reg;
2324 val = (uint64_t)((obj_priv->gtt_offset + obj->size - 4096) &
2326 val |= obj_priv->gtt_offset & 0xfffff000;
2327 val |= ((obj_priv->stride / 128) - 1) << I965_FENCE_PITCH_SHIFT;
2328 if (obj_priv->tiling_mode == I915_TILING_Y)
2329 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2330 val |= I965_FENCE_REG_VALID;
2332 I915_WRITE64(FENCE_REG_965_0 + (regnum * 8), val);
2335 static void i915_write_fence_reg(struct drm_i915_fence_reg *reg)
2337 struct drm_gem_object *obj = reg->obj;
2338 struct drm_device *dev = obj->dev;
2339 drm_i915_private_t *dev_priv = dev->dev_private;
2340 struct drm_i915_gem_object *obj_priv = obj->driver_private;
2341 int regnum = obj_priv->fence_reg;
2343 uint32_t fence_reg, val;
2346 if ((obj_priv->gtt_offset & ~I915_FENCE_START_MASK) ||
2347 (obj_priv->gtt_offset & (obj->size - 1))) {
2348 WARN(1, "%s: object 0x%08x not 1M or size (0x%zx) aligned\n",
2349 __func__, obj_priv->gtt_offset, obj->size);
2353 if (obj_priv->tiling_mode == I915_TILING_Y &&
2354 HAS_128_BYTE_Y_TILING(dev))
2359 /* Note: pitch better be a power of two tile widths */
2360 pitch_val = obj_priv->stride / tile_width;
2361 pitch_val = ffs(pitch_val) - 1;
2363 if (obj_priv->tiling_mode == I915_TILING_Y &&
2364 HAS_128_BYTE_Y_TILING(dev))
2365 WARN_ON(pitch_val > I830_FENCE_MAX_PITCH_VAL);
2367 WARN_ON(pitch_val > I915_FENCE_MAX_PITCH_VAL);
2369 val = obj_priv->gtt_offset;
2370 if (obj_priv->tiling_mode == I915_TILING_Y)
2371 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2372 val |= I915_FENCE_SIZE_BITS(obj->size);
2373 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2374 val |= I830_FENCE_REG_VALID;
2377 fence_reg = FENCE_REG_830_0 + (regnum * 4);
2379 fence_reg = FENCE_REG_945_8 + ((regnum - 8) * 4);
2380 I915_WRITE(fence_reg, val);
2383 static void i830_write_fence_reg(struct drm_i915_fence_reg *reg)
2385 struct drm_gem_object *obj = reg->obj;
2386 struct drm_device *dev = obj->dev;
2387 drm_i915_private_t *dev_priv = dev->dev_private;
2388 struct drm_i915_gem_object *obj_priv = obj->driver_private;
2389 int regnum = obj_priv->fence_reg;
2392 uint32_t fence_size_bits;
2394 if ((obj_priv->gtt_offset & ~I830_FENCE_START_MASK) ||
2395 (obj_priv->gtt_offset & (obj->size - 1))) {
2396 WARN(1, "%s: object 0x%08x not 512K or size aligned\n",
2397 __func__, obj_priv->gtt_offset);
2401 pitch_val = obj_priv->stride / 128;
2402 pitch_val = ffs(pitch_val) - 1;
2403 WARN_ON(pitch_val > I830_FENCE_MAX_PITCH_VAL);
2405 val = obj_priv->gtt_offset;
2406 if (obj_priv->tiling_mode == I915_TILING_Y)
2407 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2408 fence_size_bits = I830_FENCE_SIZE_BITS(obj->size);
2409 WARN_ON(fence_size_bits & ~0x00000f00);
2410 val |= fence_size_bits;
2411 val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2412 val |= I830_FENCE_REG_VALID;
2414 I915_WRITE(FENCE_REG_830_0 + (regnum * 4), val);
2418 * i915_gem_object_get_fence_reg - set up a fence reg for an object
2419 * @obj: object to map through a fence reg
2421 * When mapping objects through the GTT, userspace wants to be able to write
2422 * to them without having to worry about swizzling if the object is tiled.
2424 * This function walks the fence regs looking for a free one for @obj,
2425 * stealing one if it can't find any.
2427 * It then sets up the reg based on the object's properties: address, pitch
2428 * and tiling format.
2431 i915_gem_object_get_fence_reg(struct drm_gem_object *obj)
2433 struct drm_device *dev = obj->dev;
2434 struct drm_i915_private *dev_priv = dev->dev_private;
2435 struct drm_i915_gem_object *obj_priv = obj->driver_private;
2436 struct drm_i915_fence_reg *reg = NULL;
2437 struct drm_i915_gem_object *old_obj_priv = NULL;
2440 /* Just update our place in the LRU if our fence is getting used. */
2441 if (obj_priv->fence_reg != I915_FENCE_REG_NONE) {
2442 list_move_tail(&obj_priv->fence_list, &dev_priv->mm.fence_list);
2446 switch (obj_priv->tiling_mode) {
2447 case I915_TILING_NONE:
2448 WARN(1, "allocating a fence for non-tiled object?\n");
2451 if (!obj_priv->stride)
2453 WARN((obj_priv->stride & (512 - 1)),
2454 "object 0x%08x is X tiled but has non-512B pitch\n",
2455 obj_priv->gtt_offset);
2458 if (!obj_priv->stride)
2460 WARN((obj_priv->stride & (128 - 1)),
2461 "object 0x%08x is Y tiled but has non-128B pitch\n",
2462 obj_priv->gtt_offset);
2466 /* First try to find a free reg */
2468 for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2469 reg = &dev_priv->fence_regs[i];
2473 old_obj_priv = reg->obj->driver_private;
2474 if (!old_obj_priv->pin_count)
2478 /* None available, try to steal one or wait for a user to finish */
2479 if (i == dev_priv->num_fence_regs) {
2480 struct drm_gem_object *old_obj = NULL;
2485 list_for_each_entry(old_obj_priv, &dev_priv->mm.fence_list,
2487 old_obj = old_obj_priv->obj;
2489 if (old_obj_priv->pin_count)
2492 /* Take a reference, as otherwise the wait_rendering
2493 * below may cause the object to get freed out from
2496 drm_gem_object_reference(old_obj);
2498 /* i915 uses fences for GPU access to tiled buffers */
2499 if (IS_I965G(dev) || !old_obj_priv->active)
2502 /* This brings the object to the head of the LRU if it
2503 * had been written to. The only way this should
2504 * result in us waiting longer than the expected
2505 * optimal amount of time is if there was a
2506 * fence-using buffer later that was read-only.
2508 i915_gem_object_flush_gpu_write_domain(old_obj);
2509 ret = i915_gem_object_wait_rendering(old_obj);
2511 drm_gem_object_unreference(old_obj);
2519 * Zap this virtual mapping so we can set up a fence again
2520 * for this object next time we need it.
2522 i915_gem_release_mmap(old_obj);
2524 i = old_obj_priv->fence_reg;
2525 reg = &dev_priv->fence_regs[i];
2527 old_obj_priv->fence_reg = I915_FENCE_REG_NONE;
2528 list_del_init(&old_obj_priv->fence_list);
2530 drm_gem_object_unreference(old_obj);
2533 obj_priv->fence_reg = i;
2534 list_add_tail(&obj_priv->fence_list, &dev_priv->mm.fence_list);
2539 i965_write_fence_reg(reg);
2540 else if (IS_I9XX(dev))
2541 i915_write_fence_reg(reg);
2543 i830_write_fence_reg(reg);
2545 trace_i915_gem_object_get_fence(obj, i, obj_priv->tiling_mode);
2551 * i915_gem_clear_fence_reg - clear out fence register info
2552 * @obj: object to clear
2554 * Zeroes out the fence register itself and clears out the associated
2555 * data structures in dev_priv and obj_priv.
2558 i915_gem_clear_fence_reg(struct drm_gem_object *obj)
2560 struct drm_device *dev = obj->dev;
2561 drm_i915_private_t *dev_priv = dev->dev_private;
2562 struct drm_i915_gem_object *obj_priv = obj->driver_private;
2565 I915_WRITE64(FENCE_REG_965_0 + (obj_priv->fence_reg * 8), 0);
2569 if (obj_priv->fence_reg < 8)
2570 fence_reg = FENCE_REG_830_0 + obj_priv->fence_reg * 4;
2572 fence_reg = FENCE_REG_945_8 + (obj_priv->fence_reg -
2575 I915_WRITE(fence_reg, 0);
2578 dev_priv->fence_regs[obj_priv->fence_reg].obj = NULL;
2579 obj_priv->fence_reg = I915_FENCE_REG_NONE;
2580 list_del_init(&obj_priv->fence_list);
2584 * i915_gem_object_put_fence_reg - waits on outstanding fenced access
2585 * to the buffer to finish, and then resets the fence register.
2586 * @obj: tiled object holding a fence register.
2588 * Zeroes out the fence register itself and clears out the associated
2589 * data structures in dev_priv and obj_priv.
2592 i915_gem_object_put_fence_reg(struct drm_gem_object *obj)
2594 struct drm_device *dev = obj->dev;
2595 struct drm_i915_gem_object *obj_priv = obj->driver_private;
2597 if (obj_priv->fence_reg == I915_FENCE_REG_NONE)
2600 /* On the i915, GPU access to tiled buffers is via a fence,
2601 * therefore we must wait for any outstanding access to complete
2602 * before clearing the fence.
2604 if (!IS_I965G(dev)) {
2607 i915_gem_object_flush_gpu_write_domain(obj);
2608 i915_gem_object_flush_gtt_write_domain(obj);
2609 ret = i915_gem_object_wait_rendering(obj);
2614 i915_gem_clear_fence_reg (obj);
2620 * Finds free space in the GTT aperture and binds the object there.
2623 i915_gem_object_bind_to_gtt(struct drm_gem_object *obj, unsigned alignment)
2625 struct drm_device *dev = obj->dev;
2626 drm_i915_private_t *dev_priv = dev->dev_private;
2627 struct drm_i915_gem_object *obj_priv = obj->driver_private;
2628 struct drm_mm_node *free_space;
2629 gfp_t gfpmask = __GFP_NORETRY | __GFP_NOWARN;
2632 if (obj_priv->madv != I915_MADV_WILLNEED) {
2633 DRM_ERROR("Attempting to bind a purgeable object\n");
2638 alignment = i915_gem_get_gtt_alignment(obj);
2639 if (alignment & (i915_gem_get_gtt_alignment(obj) - 1)) {
2640 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
2644 /* If the object is bigger than the entire aperture, reject it early
2645 * before evicting everything in a vain attempt to find space.
2647 if (obj->size > dev->gtt_total) {
2648 DRM_ERROR("Attempting to bind an object larger than the aperture\n");
2653 free_space = drm_mm_search_free(&dev_priv->mm.gtt_space,
2654 obj->size, alignment, 0);
2655 if (free_space != NULL) {
2656 obj_priv->gtt_space = drm_mm_get_block(free_space, obj->size,
2658 if (obj_priv->gtt_space != NULL) {
2659 obj_priv->gtt_space->private = obj;
2660 obj_priv->gtt_offset = obj_priv->gtt_space->start;
2663 if (obj_priv->gtt_space == NULL) {
2664 /* If the gtt is empty and we're still having trouble
2665 * fitting our object in, we're out of memory.
2668 DRM_INFO("%s: GTT full, evicting something\n", __func__);
2670 ret = i915_gem_evict_something(dev, obj->size);
2678 DRM_INFO("Binding object of size %zd at 0x%08x\n",
2679 obj->size, obj_priv->gtt_offset);
2681 ret = i915_gem_object_get_pages(obj, gfpmask);
2683 drm_mm_put_block(obj_priv->gtt_space);
2684 obj_priv->gtt_space = NULL;
2686 if (ret == -ENOMEM) {
2687 /* first try to clear up some space from the GTT */
2688 ret = i915_gem_evict_something(dev, obj->size);
2690 /* now try to shrink everyone else */
2705 /* Create an AGP memory structure pointing at our pages, and bind it
2708 obj_priv->agp_mem = drm_agp_bind_pages(dev,
2710 obj->size >> PAGE_SHIFT,
2711 obj_priv->gtt_offset,
2712 obj_priv->agp_type);
2713 if (obj_priv->agp_mem == NULL) {
2714 i915_gem_object_put_pages(obj);
2715 drm_mm_put_block(obj_priv->gtt_space);
2716 obj_priv->gtt_space = NULL;
2718 ret = i915_gem_evict_something(dev, obj->size);
2724 atomic_inc(&dev->gtt_count);
2725 atomic_add(obj->size, &dev->gtt_memory);
2727 /* Assert that the object is not currently in any GPU domain. As it
2728 * wasn't in the GTT, there shouldn't be any way it could have been in
2731 BUG_ON(obj->read_domains & I915_GEM_GPU_DOMAINS);
2732 BUG_ON(obj->write_domain & I915_GEM_GPU_DOMAINS);
2734 trace_i915_gem_object_bind(obj, obj_priv->gtt_offset);
2740 i915_gem_clflush_object(struct drm_gem_object *obj)
2742 struct drm_i915_gem_object *obj_priv = obj->driver_private;
2744 /* If we don't have a page list set up, then we're not pinned
2745 * to GPU, and we can ignore the cache flush because it'll happen
2746 * again at bind time.
2748 if (obj_priv->pages == NULL)
2751 trace_i915_gem_object_clflush(obj);
2753 drm_clflush_pages(obj_priv->pages, obj->size / PAGE_SIZE);
2756 /** Flushes any GPU write domain for the object if it's dirty. */
2758 i915_gem_object_flush_gpu_write_domain(struct drm_gem_object *obj)
2760 struct drm_device *dev = obj->dev;
2762 uint32_t old_write_domain;
2764 if ((obj->write_domain & I915_GEM_GPU_DOMAINS) == 0)
2767 /* Queue the GPU write cache flushing we need. */
2768 old_write_domain = obj->write_domain;
2769 i915_gem_flush(dev, 0, obj->write_domain);
2770 seqno = i915_add_request(dev, NULL, obj->write_domain);
2771 BUG_ON(obj->write_domain);
2772 i915_gem_object_move_to_active(obj, seqno);
2774 trace_i915_gem_object_change_domain(obj,
2779 /** Flushes the GTT write domain for the object if it's dirty. */
2781 i915_gem_object_flush_gtt_write_domain(struct drm_gem_object *obj)
2783 uint32_t old_write_domain;
2785 if (obj->write_domain != I915_GEM_DOMAIN_GTT)
2788 /* No actual flushing is required for the GTT write domain. Writes
2789 * to it immediately go to main memory as far as we know, so there's
2790 * no chipset flush. It also doesn't land in render cache.
2792 old_write_domain = obj->write_domain;
2793 obj->write_domain = 0;
2795 trace_i915_gem_object_change_domain(obj,
2800 /** Flushes the CPU write domain for the object if it's dirty. */
2802 i915_gem_object_flush_cpu_write_domain(struct drm_gem_object *obj)
2804 struct drm_device *dev = obj->dev;
2805 uint32_t old_write_domain;
2807 if (obj->write_domain != I915_GEM_DOMAIN_CPU)
2810 i915_gem_clflush_object(obj);
2811 drm_agp_chipset_flush(dev);
2812 old_write_domain = obj->write_domain;
2813 obj->write_domain = 0;
2815 trace_i915_gem_object_change_domain(obj,
2821 i915_gem_object_flush_write_domain(struct drm_gem_object *obj)
2823 switch (obj->write_domain) {
2824 case I915_GEM_DOMAIN_GTT:
2825 i915_gem_object_flush_gtt_write_domain(obj);
2827 case I915_GEM_DOMAIN_CPU:
2828 i915_gem_object_flush_cpu_write_domain(obj);
2831 i915_gem_object_flush_gpu_write_domain(obj);
2837 * Moves a single object to the GTT read, and possibly write domain.
2839 * This function returns when the move is complete, including waiting on
2843 i915_gem_object_set_to_gtt_domain(struct drm_gem_object *obj, int write)
2845 struct drm_i915_gem_object *obj_priv = obj->driver_private;
2846 uint32_t old_write_domain, old_read_domains;
2849 /* Not valid to be called on unbound objects. */
2850 if (obj_priv->gtt_space == NULL)
2853 i915_gem_object_flush_gpu_write_domain(obj);
2854 /* Wait on any GPU rendering and flushing to occur. */
2855 ret = i915_gem_object_wait_rendering(obj);
2859 old_write_domain = obj->write_domain;
2860 old_read_domains = obj->read_domains;
2862 /* If we're writing through the GTT domain, then CPU and GPU caches
2863 * will need to be invalidated at next use.
2866 obj->read_domains &= I915_GEM_DOMAIN_GTT;
2868 i915_gem_object_flush_cpu_write_domain(obj);
2870 /* It should now be out of any other write domains, and we can update
2871 * the domain values for our changes.
2873 BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2874 obj->read_domains |= I915_GEM_DOMAIN_GTT;
2876 obj->write_domain = I915_GEM_DOMAIN_GTT;
2877 obj_priv->dirty = 1;
2880 trace_i915_gem_object_change_domain(obj,
2888 * Prepare buffer for display plane. Use uninterruptible for possible flush
2889 * wait, as in modesetting process we're not supposed to be interrupted.
2892 i915_gem_object_set_to_display_plane(struct drm_gem_object *obj)
2894 struct drm_device *dev = obj->dev;
2895 struct drm_i915_gem_object *obj_priv = obj->driver_private;
2896 uint32_t old_write_domain, old_read_domains;
2899 /* Not valid to be called on unbound objects. */
2900 if (obj_priv->gtt_space == NULL)
2903 i915_gem_object_flush_gpu_write_domain(obj);
2905 /* Wait on any GPU rendering and flushing to occur. */
2906 if (obj_priv->active) {
2908 DRM_INFO("%s: object %p wait for seqno %08x\n",
2909 __func__, obj, obj_priv->last_rendering_seqno);
2911 ret = i915_do_wait_request(dev, obj_priv->last_rendering_seqno, 0);
2916 old_write_domain = obj->write_domain;
2917 old_read_domains = obj->read_domains;
2919 obj->read_domains &= I915_GEM_DOMAIN_GTT;
2921 i915_gem_object_flush_cpu_write_domain(obj);
2923 /* It should now be out of any other write domains, and we can update
2924 * the domain values for our changes.
2926 BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2927 obj->read_domains |= I915_GEM_DOMAIN_GTT;
2928 obj->write_domain = I915_GEM_DOMAIN_GTT;
2929 obj_priv->dirty = 1;
2931 trace_i915_gem_object_change_domain(obj,
2939 * Moves a single object to the CPU read, and possibly write domain.
2941 * This function returns when the move is complete, including waiting on
2945 i915_gem_object_set_to_cpu_domain(struct drm_gem_object *obj, int write)
2947 uint32_t old_write_domain, old_read_domains;
2950 i915_gem_object_flush_gpu_write_domain(obj);
2951 /* Wait on any GPU rendering and flushing to occur. */
2952 ret = i915_gem_object_wait_rendering(obj);
2956 i915_gem_object_flush_gtt_write_domain(obj);
2958 /* If we have a partially-valid cache of the object in the CPU,
2959 * finish invalidating it and free the per-page flags.
2961 i915_gem_object_set_to_full_cpu_read_domain(obj);
2963 old_write_domain = obj->write_domain;
2964 old_read_domains = obj->read_domains;
2966 /* Flush the CPU cache if it's still invalid. */
2967 if ((obj->read_domains & I915_GEM_DOMAIN_CPU) == 0) {
2968 i915_gem_clflush_object(obj);
2970 obj->read_domains |= I915_GEM_DOMAIN_CPU;
2973 /* It should now be out of any other write domains, and we can update
2974 * the domain values for our changes.
2976 BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
2978 /* If we're writing through the CPU, then the GPU read domains will
2979 * need to be invalidated at next use.
2982 obj->read_domains &= I915_GEM_DOMAIN_CPU;
2983 obj->write_domain = I915_GEM_DOMAIN_CPU;
2986 trace_i915_gem_object_change_domain(obj,
2994 * Set the next domain for the specified object. This
2995 * may not actually perform the necessary flushing/invaliding though,
2996 * as that may want to be batched with other set_domain operations
2998 * This is (we hope) the only really tricky part of gem. The goal
2999 * is fairly simple -- track which caches hold bits of the object
3000 * and make sure they remain coherent. A few concrete examples may
3001 * help to explain how it works. For shorthand, we use the notation
3002 * (read_domains, write_domain), e.g. (CPU, CPU) to indicate the
3003 * a pair of read and write domain masks.
3005 * Case 1: the batch buffer
3011 * 5. Unmapped from GTT
3014 * Let's take these a step at a time
3017 * Pages allocated from the kernel may still have
3018 * cache contents, so we set them to (CPU, CPU) always.
3019 * 2. Written by CPU (using pwrite)
3020 * The pwrite function calls set_domain (CPU, CPU) and
3021 * this function does nothing (as nothing changes)
3023 * This function asserts that the object is not
3024 * currently in any GPU-based read or write domains
3026 * i915_gem_execbuffer calls set_domain (COMMAND, 0).
3027 * As write_domain is zero, this function adds in the
3028 * current read domains (CPU+COMMAND, 0).
3029 * flush_domains is set to CPU.
3030 * invalidate_domains is set to COMMAND
3031 * clflush is run to get data out of the CPU caches
3032 * then i915_dev_set_domain calls i915_gem_flush to
3033 * emit an MI_FLUSH and drm_agp_chipset_flush
3034 * 5. Unmapped from GTT
3035 * i915_gem_object_unbind calls set_domain (CPU, CPU)
3036 * flush_domains and invalidate_domains end up both zero
3037 * so no flushing/invalidating happens
3041 * Case 2: The shared render buffer
3045 * 3. Read/written by GPU
3046 * 4. set_domain to (CPU,CPU)
3047 * 5. Read/written by CPU
3048 * 6. Read/written by GPU
3051 * Same as last example, (CPU, CPU)
3053 * Nothing changes (assertions find that it is not in the GPU)
3054 * 3. Read/written by GPU
3055 * execbuffer calls set_domain (RENDER, RENDER)
3056 * flush_domains gets CPU
3057 * invalidate_domains gets GPU
3059 * MI_FLUSH and drm_agp_chipset_flush
3060 * 4. set_domain (CPU, CPU)
3061 * flush_domains gets GPU
3062 * invalidate_domains gets CPU
3063 * wait_rendering (obj) to make sure all drawing is complete.
3064 * This will include an MI_FLUSH to get the data from GPU
3066 * clflush (obj) to invalidate the CPU cache
3067 * Another MI_FLUSH in i915_gem_flush (eliminate this somehow?)
3068 * 5. Read/written by CPU
3069 * cache lines are loaded and dirtied
3070 * 6. Read written by GPU
3071 * Same as last GPU access
3073 * Case 3: The constant buffer
3078 * 4. Updated (written) by CPU again
3087 * flush_domains = CPU
3088 * invalidate_domains = RENDER
3091 * drm_agp_chipset_flush
3092 * 4. Updated (written) by CPU again
3094 * flush_domains = 0 (no previous write domain)
3095 * invalidate_domains = 0 (no new read domains)
3098 * flush_domains = CPU
3099 * invalidate_domains = RENDER
3102 * drm_agp_chipset_flush
3105 i915_gem_object_set_to_gpu_domain(struct drm_gem_object *obj)
3107 struct drm_device *dev = obj->dev;
3108 struct drm_i915_gem_object *obj_priv = obj->driver_private;
3109 uint32_t invalidate_domains = 0;
3110 uint32_t flush_domains = 0;
3111 uint32_t old_read_domains;
3113 BUG_ON(obj->pending_read_domains & I915_GEM_DOMAIN_CPU);
3114 BUG_ON(obj->pending_write_domain == I915_GEM_DOMAIN_CPU);
3116 intel_mark_busy(dev, obj);
3119 DRM_INFO("%s: object %p read %08x -> %08x write %08x -> %08x\n",
3121 obj->read_domains, obj->pending_read_domains,
3122 obj->write_domain, obj->pending_write_domain);
3125 * If the object isn't moving to a new write domain,
3126 * let the object stay in multiple read domains
3128 if (obj->pending_write_domain == 0)
3129 obj->pending_read_domains |= obj->read_domains;
3131 obj_priv->dirty = 1;
3134 * Flush the current write domain if
3135 * the new read domains don't match. Invalidate
3136 * any read domains which differ from the old
3139 if (obj->write_domain &&
3140 obj->write_domain != obj->pending_read_domains) {
3141 flush_domains |= obj->write_domain;
3142 invalidate_domains |=
3143 obj->pending_read_domains & ~obj->write_domain;
3146 * Invalidate any read caches which may have
3147 * stale data. That is, any new read domains.
3149 invalidate_domains |= obj->pending_read_domains & ~obj->read_domains;
3150 if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_CPU) {
3152 DRM_INFO("%s: CPU domain flush %08x invalidate %08x\n",
3153 __func__, flush_domains, invalidate_domains);
3155 i915_gem_clflush_object(obj);
3158 old_read_domains = obj->read_domains;
3160 /* The actual obj->write_domain will be updated with
3161 * pending_write_domain after we emit the accumulated flush for all
3162 * of our domain changes in execbuffers (which clears objects'
3163 * write_domains). So if we have a current write domain that we
3164 * aren't changing, set pending_write_domain to that.
3166 if (flush_domains == 0 && obj->pending_write_domain == 0)
3167 obj->pending_write_domain = obj->write_domain;
3168 obj->read_domains = obj->pending_read_domains;
3170 dev->invalidate_domains |= invalidate_domains;
3171 dev->flush_domains |= flush_domains;
3173 DRM_INFO("%s: read %08x write %08x invalidate %08x flush %08x\n",
3175 obj->read_domains, obj->write_domain,
3176 dev->invalidate_domains, dev->flush_domains);
3179 trace_i915_gem_object_change_domain(obj,
3185 * Moves the object from a partially CPU read to a full one.
3187 * Note that this only resolves i915_gem_object_set_cpu_read_domain_range(),
3188 * and doesn't handle transitioning from !(read_domains & I915_GEM_DOMAIN_CPU).
3191 i915_gem_object_set_to_full_cpu_read_domain(struct drm_gem_object *obj)
3193 struct drm_i915_gem_object *obj_priv = obj->driver_private;
3195 if (!obj_priv->page_cpu_valid)
3198 /* If we're partially in the CPU read domain, finish moving it in.
3200 if (obj->read_domains & I915_GEM_DOMAIN_CPU) {
3203 for (i = 0; i <= (obj->size - 1) / PAGE_SIZE; i++) {
3204 if (obj_priv->page_cpu_valid[i])
3206 drm_clflush_pages(obj_priv->pages + i, 1);
3210 /* Free the page_cpu_valid mappings which are now stale, whether
3211 * or not we've got I915_GEM_DOMAIN_CPU.
3213 kfree(obj_priv->page_cpu_valid);
3214 obj_priv->page_cpu_valid = NULL;
3218 * Set the CPU read domain on a range of the object.
3220 * The object ends up with I915_GEM_DOMAIN_CPU in its read flags although it's
3221 * not entirely valid. The page_cpu_valid member of the object flags which
3222 * pages have been flushed, and will be respected by
3223 * i915_gem_object_set_to_cpu_domain() if it's called on to get a valid mapping
3224 * of the whole object.
3226 * This function returns when the move is complete, including waiting on
3230 i915_gem_object_set_cpu_read_domain_range(struct drm_gem_object *obj,
3231 uint64_t offset, uint64_t size)
3233 struct drm_i915_gem_object *obj_priv = obj->driver_private;
3234 uint32_t old_read_domains;
3237 if (offset == 0 && size == obj->size)
3238 return i915_gem_object_set_to_cpu_domain(obj, 0);
3240 i915_gem_object_flush_gpu_write_domain(obj);
3241 /* Wait on any GPU rendering and flushing to occur. */
3242 ret = i915_gem_object_wait_rendering(obj);
3245 i915_gem_object_flush_gtt_write_domain(obj);
3247 /* If we're already fully in the CPU read domain, we're done. */
3248 if (obj_priv->page_cpu_valid == NULL &&
3249 (obj->read_domains & I915_GEM_DOMAIN_CPU) != 0)
3252 /* Otherwise, create/clear the per-page CPU read domain flag if we're
3253 * newly adding I915_GEM_DOMAIN_CPU
3255 if (obj_priv->page_cpu_valid == NULL) {
3256 obj_priv->page_cpu_valid = kzalloc(obj->size / PAGE_SIZE,
3258 if (obj_priv->page_cpu_valid == NULL)
3260 } else if ((obj->read_domains & I915_GEM_DOMAIN_CPU) == 0)
3261 memset(obj_priv->page_cpu_valid, 0, obj->size / PAGE_SIZE);
3263 /* Flush the cache on any pages that are still invalid from the CPU's
3266 for (i = offset / PAGE_SIZE; i <= (offset + size - 1) / PAGE_SIZE;
3268 if (obj_priv->page_cpu_valid[i])
3271 drm_clflush_pages(obj_priv->pages + i, 1);
3273 obj_priv->page_cpu_valid[i] = 1;
3276 /* It should now be out of any other write domains, and we can update
3277 * the domain values for our changes.
3279 BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3281 old_read_domains = obj->read_domains;
3282 obj->read_domains |= I915_GEM_DOMAIN_CPU;
3284 trace_i915_gem_object_change_domain(obj,
3292 * Pin an object to the GTT and evaluate the relocations landing in it.
3295 i915_gem_object_pin_and_relocate(struct drm_gem_object *obj,
3296 struct drm_file *file_priv,
3297 struct drm_i915_gem_exec_object2 *entry,
3298 struct drm_i915_gem_relocation_entry *relocs)
3300 struct drm_device *dev = obj->dev;
3301 drm_i915_private_t *dev_priv = dev->dev_private;
3302 struct drm_i915_gem_object *obj_priv = obj->driver_private;
3304 void __iomem *reloc_page;
3307 need_fence = entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
3308 obj_priv->tiling_mode != I915_TILING_NONE;
3310 /* Check fence reg constraints and rebind if necessary */
3311 if (need_fence && !i915_obj_fenceable(dev, obj))
3312 i915_gem_object_unbind(obj);
3314 /* Choose the GTT offset for our buffer and put it there. */
3315 ret = i915_gem_object_pin(obj, (uint32_t) entry->alignment);
3320 * Pre-965 chips need a fence register set up in order to
3321 * properly handle blits to/from tiled surfaces.
3324 ret = i915_gem_object_get_fence_reg(obj);
3326 if (ret != -EBUSY && ret != -ERESTARTSYS)
3327 DRM_ERROR("Failure to install fence: %d\n",
3329 i915_gem_object_unpin(obj);
3334 entry->offset = obj_priv->gtt_offset;
3336 /* Apply the relocations, using the GTT aperture to avoid cache
3337 * flushing requirements.
3339 for (i = 0; i < entry->relocation_count; i++) {
3340 struct drm_i915_gem_relocation_entry *reloc= &relocs[i];
3341 struct drm_gem_object *target_obj;
3342 struct drm_i915_gem_object *target_obj_priv;
3343 uint32_t reloc_val, reloc_offset;
3344 uint32_t __iomem *reloc_entry;
3346 target_obj = drm_gem_object_lookup(obj->dev, file_priv,
3347 reloc->target_handle);
3348 if (target_obj == NULL) {
3349 i915_gem_object_unpin(obj);
3352 target_obj_priv = target_obj->driver_private;
3355 DRM_INFO("%s: obj %p offset %08x target %d "
3356 "read %08x write %08x gtt %08x "
3357 "presumed %08x delta %08x\n",
3360 (int) reloc->offset,
3361 (int) reloc->target_handle,
3362 (int) reloc->read_domains,
3363 (int) reloc->write_domain,
3364 (int) target_obj_priv->gtt_offset,
3365 (int) reloc->presumed_offset,
3369 /* The target buffer should have appeared before us in the
3370 * exec_object list, so it should have a GTT space bound by now.
3372 if (target_obj_priv->gtt_space == NULL) {
3373 DRM_ERROR("No GTT space found for object %d\n",
3374 reloc->target_handle);
3375 drm_gem_object_unreference(target_obj);
3376 i915_gem_object_unpin(obj);
3380 /* Validate that the target is in a valid r/w GPU domain */
3381 if (reloc->write_domain & I915_GEM_DOMAIN_CPU ||
3382 reloc->read_domains & I915_GEM_DOMAIN_CPU) {
3383 DRM_ERROR("reloc with read/write CPU domains: "
3384 "obj %p target %d offset %d "
3385 "read %08x write %08x",
3386 obj, reloc->target_handle,
3387 (int) reloc->offset,
3388 reloc->read_domains,
3389 reloc->write_domain);
3390 drm_gem_object_unreference(target_obj);
3391 i915_gem_object_unpin(obj);
3394 if (reloc->write_domain && target_obj->pending_write_domain &&
3395 reloc->write_domain != target_obj->pending_write_domain) {
3396 DRM_ERROR("Write domain conflict: "
3397 "obj %p target %d offset %d "
3398 "new %08x old %08x\n",
3399 obj, reloc->target_handle,
3400 (int) reloc->offset,
3401 reloc->write_domain,
3402 target_obj->pending_write_domain);
3403 drm_gem_object_unreference(target_obj);
3404 i915_gem_object_unpin(obj);
3408 target_obj->pending_read_domains |= reloc->read_domains;
3409 target_obj->pending_write_domain |= reloc->write_domain;
3411 /* If the relocation already has the right value in it, no
3412 * more work needs to be done.
3414 if (target_obj_priv->gtt_offset == reloc->presumed_offset) {
3415 drm_gem_object_unreference(target_obj);
3419 /* Check that the relocation address is valid... */
3420 if (reloc->offset > obj->size - 4) {
3421 DRM_ERROR("Relocation beyond object bounds: "
3422 "obj %p target %d offset %d size %d.\n",
3423 obj, reloc->target_handle,
3424 (int) reloc->offset, (int) obj->size);
3425 drm_gem_object_unreference(target_obj);
3426 i915_gem_object_unpin(obj);
3429 if (reloc->offset & 3) {
3430 DRM_ERROR("Relocation not 4-byte aligned: "
3431 "obj %p target %d offset %d.\n",
3432 obj, reloc->target_handle,
3433 (int) reloc->offset);
3434 drm_gem_object_unreference(target_obj);
3435 i915_gem_object_unpin(obj);
3439 /* and points to somewhere within the target object. */
3440 if (reloc->delta >= target_obj->size) {
3441 DRM_ERROR("Relocation beyond target object bounds: "
3442 "obj %p target %d delta %d size %d.\n",
3443 obj, reloc->target_handle,
3444 (int) reloc->delta, (int) target_obj->size);
3445 drm_gem_object_unreference(target_obj);
3446 i915_gem_object_unpin(obj);
3450 ret = i915_gem_object_set_to_gtt_domain(obj, 1);
3452 drm_gem_object_unreference(target_obj);
3453 i915_gem_object_unpin(obj);
3457 /* Map the page containing the relocation we're going to
3460 reloc_offset = obj_priv->gtt_offset + reloc->offset;
3461 reloc_page = io_mapping_map_atomic_wc(dev_priv->mm.gtt_mapping,
3464 reloc_entry = (uint32_t __iomem *)(reloc_page +
3465 (reloc_offset & (PAGE_SIZE - 1)));
3466 reloc_val = target_obj_priv->gtt_offset + reloc->delta;
3469 DRM_INFO("Applied relocation: %p@0x%08x %08x -> %08x\n",
3470 obj, (unsigned int) reloc->offset,
3471 readl(reloc_entry), reloc_val);
3473 writel(reloc_val, reloc_entry);
3474 io_mapping_unmap_atomic(reloc_page);
3476 /* The updated presumed offset for this entry will be
3477 * copied back out to the user.
3479 reloc->presumed_offset = target_obj_priv->gtt_offset;
3481 drm_gem_object_unreference(target_obj);
3486 i915_gem_dump_object(obj, 128, __func__, ~0);
3491 /** Dispatch a batchbuffer to the ring
3494 i915_dispatch_gem_execbuffer(struct drm_device *dev,
3495 struct drm_i915_gem_execbuffer2 *exec,
3496 struct drm_clip_rect *cliprects,
3497 uint64_t exec_offset)
3499 drm_i915_private_t *dev_priv = dev->dev_private;
3500 int nbox = exec->num_cliprects;
3502 uint32_t exec_start, exec_len;
3505 exec_start = (uint32_t) exec_offset + exec->batch_start_offset;
3506 exec_len = (uint32_t) exec->batch_len;
3508 trace_i915_gem_request_submit(dev, dev_priv->mm.next_gem_seqno + 1);
3510 count = nbox ? nbox : 1;
3512 for (i = 0; i < count; i++) {
3514 int ret = i915_emit_box(dev, cliprects, i,
3515 exec->DR1, exec->DR4);
3520 if (IS_I830(dev) || IS_845G(dev)) {
3522 OUT_RING(MI_BATCH_BUFFER);
3523 OUT_RING(exec_start | MI_BATCH_NON_SECURE);
3524 OUT_RING(exec_start + exec_len - 4);
3529 if (IS_I965G(dev)) {
3530 OUT_RING(MI_BATCH_BUFFER_START |
3532 MI_BATCH_NON_SECURE_I965);
3533 OUT_RING(exec_start);
3535 OUT_RING(MI_BATCH_BUFFER_START |
3537 OUT_RING(exec_start | MI_BATCH_NON_SECURE);
3543 /* XXX breadcrumb */
3547 /* Throttle our rendering by waiting until the ring has completed our requests
3548 * emitted over 20 msec ago.
3550 * Note that if we were to use the current jiffies each time around the loop,
3551 * we wouldn't escape the function with any frames outstanding if the time to
3552 * render a frame was over 20ms.
3554 * This should get us reasonable parallelism between CPU and GPU but also
3555 * relatively low latency when blocking on a particular request to finish.
3558 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file_priv)
3560 struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv;
3562 unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3564 mutex_lock(&dev->struct_mutex);
3565 while (!list_empty(&i915_file_priv->mm.request_list)) {
3566 struct drm_i915_gem_request *request;
3568 request = list_first_entry(&i915_file_priv->mm.request_list,
3569 struct drm_i915_gem_request,
3572 if (time_after_eq(request->emitted_jiffies, recent_enough))
3575 ret = i915_wait_request(dev, request->seqno);
3579 mutex_unlock(&dev->struct_mutex);
3585 i915_gem_get_relocs_from_user(struct drm_i915_gem_exec_object2 *exec_list,
3586 uint32_t buffer_count,
3587 struct drm_i915_gem_relocation_entry **relocs)
3589 uint32_t reloc_count = 0, reloc_index = 0, i;
3593 for (i = 0; i < buffer_count; i++) {
3594 if (reloc_count + exec_list[i].relocation_count < reloc_count)
3596 reloc_count += exec_list[i].relocation_count;
3599 *relocs = drm_calloc_large(reloc_count, sizeof(**relocs));
3600 if (*relocs == NULL) {
3601 DRM_ERROR("failed to alloc relocs, count %d\n", reloc_count);
3605 for (i = 0; i < buffer_count; i++) {
3606 struct drm_i915_gem_relocation_entry __user *user_relocs;
3608 user_relocs = (void __user *)(uintptr_t)exec_list[i].relocs_ptr;
3610 ret = copy_from_user(&(*relocs)[reloc_index],
3612 exec_list[i].relocation_count *
3615 drm_free_large(*relocs);
3620 reloc_index += exec_list[i].relocation_count;
3627 i915_gem_put_relocs_to_user(struct drm_i915_gem_exec_object2 *exec_list,
3628 uint32_t buffer_count,
3629 struct drm_i915_gem_relocation_entry *relocs)
3631 uint32_t reloc_count = 0, i;
3637 for (i = 0; i < buffer_count; i++) {
3638 struct drm_i915_gem_relocation_entry __user *user_relocs;
3641 user_relocs = (void __user *)(uintptr_t)exec_list[i].relocs_ptr;
3643 unwritten = copy_to_user(user_relocs,
3644 &relocs[reloc_count],
3645 exec_list[i].relocation_count *
3653 reloc_count += exec_list[i].relocation_count;
3657 drm_free_large(relocs);
3663 i915_gem_check_execbuffer (struct drm_i915_gem_execbuffer2 *exec,
3664 uint64_t exec_offset)
3666 uint32_t exec_start, exec_len;
3668 exec_start = (uint32_t) exec_offset + exec->batch_start_offset;
3669 exec_len = (uint32_t) exec->batch_len;
3671 if ((exec_start | exec_len) & 0x7)
3681 i915_gem_wait_for_pending_flip(struct drm_device *dev,
3682 struct drm_gem_object **object_list,
3685 drm_i915_private_t *dev_priv = dev->dev_private;
3686 struct drm_i915_gem_object *obj_priv;
3691 prepare_to_wait(&dev_priv->pending_flip_queue,
3692 &wait, TASK_INTERRUPTIBLE);
3693 for (i = 0; i < count; i++) {
3694 obj_priv = object_list[i]->driver_private;
3695 if (atomic_read(&obj_priv->pending_flip) > 0)
3701 if (!signal_pending(current)) {
3702 mutex_unlock(&dev->struct_mutex);
3704 mutex_lock(&dev->struct_mutex);
3710 finish_wait(&dev_priv->pending_flip_queue, &wait);
3716 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
3717 struct drm_file *file_priv,
3718 struct drm_i915_gem_execbuffer2 *args,
3719 struct drm_i915_gem_exec_object2 *exec_list)
3721 drm_i915_private_t *dev_priv = dev->dev_private;
3722 struct drm_gem_object **object_list = NULL;
3723 struct drm_gem_object *batch_obj;
3724 struct drm_i915_gem_object *obj_priv;
3725 struct drm_clip_rect *cliprects = NULL;
3726 struct drm_i915_gem_relocation_entry *relocs = NULL;
3727 int ret = 0, ret2, i, pinned = 0;
3728 uint64_t exec_offset;
3729 uint32_t seqno, flush_domains, reloc_index;
3730 int pin_tries, flips;
3733 DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
3734 (int) args->buffers_ptr, args->buffer_count, args->batch_len);
3737 if (args->buffer_count < 1) {
3738 DRM_ERROR("execbuf with %d buffers\n", args->buffer_count);
3741 object_list = drm_malloc_ab(sizeof(*object_list), args->buffer_count);
3742 if (object_list == NULL) {
3743 DRM_ERROR("Failed to allocate object list for %d buffers\n",
3744 args->buffer_count);
3749 if (args->num_cliprects != 0) {
3750 cliprects = kcalloc(args->num_cliprects, sizeof(*cliprects),
3752 if (cliprects == NULL) {
3757 ret = copy_from_user(cliprects,
3758 (struct drm_clip_rect __user *)
3759 (uintptr_t) args->cliprects_ptr,
3760 sizeof(*cliprects) * args->num_cliprects);
3762 DRM_ERROR("copy %d cliprects failed: %d\n",
3763 args->num_cliprects, ret);
3768 ret = i915_gem_get_relocs_from_user(exec_list, args->buffer_count,
3773 mutex_lock(&dev->struct_mutex);
3775 i915_verify_inactive(dev, __FILE__, __LINE__);
3777 if (atomic_read(&dev_priv->mm.wedged)) {
3778 mutex_unlock(&dev->struct_mutex);
3783 if (dev_priv->mm.suspended) {
3784 mutex_unlock(&dev->struct_mutex);
3789 /* Look up object handles */
3791 for (i = 0; i < args->buffer_count; i++) {
3792 object_list[i] = drm_gem_object_lookup(dev, file_priv,
3793 exec_list[i].handle);
3794 if (object_list[i] == NULL) {
3795 DRM_ERROR("Invalid object handle %d at index %d\n",
3796 exec_list[i].handle, i);
3797 /* prevent error path from reading uninitialized data */
3798 args->buffer_count = i + 1;
3803 obj_priv = object_list[i]->driver_private;
3804 if (obj_priv->in_execbuffer) {
3805 DRM_ERROR("Object %p appears more than once in object list\n",
3807 /* prevent error path from reading uninitialized data */
3808 args->buffer_count = i + 1;
3812 obj_priv->in_execbuffer = true;
3813 flips += atomic_read(&obj_priv->pending_flip);
3817 ret = i915_gem_wait_for_pending_flip(dev, object_list,
3818 args->buffer_count);
3823 /* Pin and relocate */
3824 for (pin_tries = 0; ; pin_tries++) {
3828 for (i = 0; i < args->buffer_count; i++) {
3829 object_list[i]->pending_read_domains = 0;
3830 object_list[i]->pending_write_domain = 0;
3831 ret = i915_gem_object_pin_and_relocate(object_list[i],
3834 &relocs[reloc_index]);
3838 reloc_index += exec_list[i].relocation_count;
3844 /* error other than GTT full, or we've already tried again */
3845 if (ret != -ENOSPC || pin_tries >= 1) {
3846 if (ret != -ERESTARTSYS) {
3847 unsigned long long total_size = 0;
3848 for (i = 0; i < args->buffer_count; i++)
3849 total_size += object_list[i]->size;
3850 DRM_ERROR("Failed to pin buffer %d of %d, total %llu bytes: %d\n",
3851 pinned+1, args->buffer_count,
3853 DRM_ERROR("%d objects [%d pinned], "
3854 "%d object bytes [%d pinned], "
3855 "%d/%d gtt bytes\n",
3856 atomic_read(&dev->object_count),
3857 atomic_read(&dev->pin_count),
3858 atomic_read(&dev->object_memory),
3859 atomic_read(&dev->pin_memory),
3860 atomic_read(&dev->gtt_memory),
3866 /* unpin all of our buffers */
3867 for (i = 0; i < pinned; i++)
3868 i915_gem_object_unpin(object_list[i]);
3871 /* evict everyone we can from the aperture */
3872 ret = i915_gem_evict_everything(dev);
3873 if (ret && ret != -ENOSPC)
3877 /* Set the pending read domains for the batch buffer to COMMAND */
3878 batch_obj = object_list[args->buffer_count-1];
3879 if (batch_obj->pending_write_domain) {
3880 DRM_ERROR("Attempting to use self-modifying batch buffer\n");
3884 batch_obj->pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
3886 /* Sanity check the batch buffer, prior to moving objects */
3887 exec_offset = exec_list[args->buffer_count - 1].offset;
3888 ret = i915_gem_check_execbuffer (args, exec_offset);
3890 DRM_ERROR("execbuf with invalid offset/length\n");
3894 i915_verify_inactive(dev, __FILE__, __LINE__);
3896 /* Zero the global flush/invalidate flags. These
3897 * will be modified as new domains are computed
3900 dev->invalidate_domains = 0;
3901 dev->flush_domains = 0;
3903 for (i = 0; i < args->buffer_count; i++) {
3904 struct drm_gem_object *obj = object_list[i];
3906 /* Compute new gpu domains and update invalidate/flush */
3907 i915_gem_object_set_to_gpu_domain(obj);
3910 i915_verify_inactive(dev, __FILE__, __LINE__);
3912 if (dev->invalidate_domains | dev->flush_domains) {
3914 DRM_INFO("%s: invalidate_domains %08x flush_domains %08x\n",
3916 dev->invalidate_domains,
3917 dev->flush_domains);
3920 dev->invalidate_domains,
3921 dev->flush_domains);
3922 if (dev->flush_domains & I915_GEM_GPU_DOMAINS)
3923 (void)i915_add_request(dev, file_priv,
3924 dev->flush_domains);
3927 for (i = 0; i < args->buffer_count; i++) {
3928 struct drm_gem_object *obj = object_list[i];
3929 struct drm_i915_gem_object *obj_priv = obj->driver_private;
3930 uint32_t old_write_domain = obj->write_domain;
3932 obj->write_domain = obj->pending_write_domain;
3933 if (obj->write_domain)
3934 list_move_tail(&obj_priv->gpu_write_list,
3935 &dev_priv->mm.gpu_write_list);
3937 list_del_init(&obj_priv->gpu_write_list);
3939 trace_i915_gem_object_change_domain(obj,
3944 i915_verify_inactive(dev, __FILE__, __LINE__);
3947 for (i = 0; i < args->buffer_count; i++) {
3948 i915_gem_object_check_coherency(object_list[i],
3949 exec_list[i].handle);
3954 i915_gem_dump_object(batch_obj,
3960 /* Exec the batchbuffer */
3961 ret = i915_dispatch_gem_execbuffer(dev, args, cliprects, exec_offset);
3963 DRM_ERROR("dispatch failed %d\n", ret);
3968 * Ensure that the commands in the batch buffer are
3969 * finished before the interrupt fires
3971 flush_domains = i915_retire_commands(dev);
3973 i915_verify_inactive(dev, __FILE__, __LINE__);
3976 * Get a seqno representing the execution of the current buffer,
3977 * which we can wait on. We would like to mitigate these interrupts,
3978 * likely by only creating seqnos occasionally (so that we have
3979 * *some* interrupts representing completion of buffers that we can
3980 * wait on when trying to clear up gtt space).
3982 seqno = i915_add_request(dev, file_priv, flush_domains);
3984 for (i = 0; i < args->buffer_count; i++) {
3985 struct drm_gem_object *obj = object_list[i];
3987 i915_gem_object_move_to_active(obj, seqno);
3989 DRM_INFO("%s: move to exec list %p\n", __func__, obj);
3993 i915_dump_lru(dev, __func__);
3996 i915_verify_inactive(dev, __FILE__, __LINE__);
3999 for (i = 0; i < pinned; i++)
4000 i915_gem_object_unpin(object_list[i]);
4002 for (i = 0; i < args->buffer_count; i++) {
4003 if (object_list[i]) {
4004 obj_priv = object_list[i]->driver_private;
4005 obj_priv->in_execbuffer = false;
4007 drm_gem_object_unreference(object_list[i]);
4010 mutex_unlock(&dev->struct_mutex);
4013 /* Copy the updated relocations out regardless of current error
4014 * state. Failure to update the relocs would mean that the next
4015 * time userland calls execbuf, it would do so with presumed offset
4016 * state that didn't match the actual object state.
4018 ret2 = i915_gem_put_relocs_to_user(exec_list, args->buffer_count,
4021 DRM_ERROR("Failed to copy relocations back out: %d\n", ret2);
4027 drm_free_large(object_list);
4034 * Legacy execbuffer just creates an exec2 list from the original exec object
4035 * list array and passes it to the real function.
4038 i915_gem_execbuffer(struct drm_device *dev, void *data,
4039 struct drm_file *file_priv)
4041 struct drm_i915_gem_execbuffer *args = data;
4042 struct drm_i915_gem_execbuffer2 exec2;
4043 struct drm_i915_gem_exec_object *exec_list = NULL;
4044 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
4048 DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
4049 (int) args->buffers_ptr, args->buffer_count, args->batch_len);
4052 if (args->buffer_count < 1) {
4053 DRM_ERROR("execbuf with %d buffers\n", args->buffer_count);
4057 /* Copy in the exec list from userland */
4058 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
4059 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
4060 if (exec_list == NULL || exec2_list == NULL) {
4061 DRM_ERROR("Failed to allocate exec list for %d buffers\n",
4062 args->buffer_count);
4063 drm_free_large(exec_list);
4064 drm_free_large(exec2_list);
4067 ret = copy_from_user(exec_list,
4068 (struct drm_i915_relocation_entry __user *)
4069 (uintptr_t) args->buffers_ptr,
4070 sizeof(*exec_list) * args->buffer_count);
4072 DRM_ERROR("copy %d exec entries failed %d\n",
4073 args->buffer_count, ret);
4074 drm_free_large(exec_list);
4075 drm_free_large(exec2_list);
4079 for (i = 0; i < args->buffer_count; i++) {
4080 exec2_list[i].handle = exec_list[i].handle;
4081 exec2_list[i].relocation_count = exec_list[i].relocation_count;
4082 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
4083 exec2_list[i].alignment = exec_list[i].alignment;
4084 exec2_list[i].offset = exec_list[i].offset;
4086 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
4088 exec2_list[i].flags = 0;
4091 exec2.buffers_ptr = args->buffers_ptr;
4092 exec2.buffer_count = args->buffer_count;
4093 exec2.batch_start_offset = args->batch_start_offset;
4094 exec2.batch_len = args->batch_len;
4095 exec2.DR1 = args->DR1;
4096 exec2.DR4 = args->DR4;
4097 exec2.num_cliprects = args->num_cliprects;
4098 exec2.cliprects_ptr = args->cliprects_ptr;
4101 ret = i915_gem_do_execbuffer(dev, data, file_priv, &exec2, exec2_list);
4103 /* Copy the new buffer offsets back to the user's exec list. */
4104 for (i = 0; i < args->buffer_count; i++)
4105 exec_list[i].offset = exec2_list[i].offset;
4106 /* ... and back out to userspace */
4107 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
4108 (uintptr_t) args->buffers_ptr,
4110 sizeof(*exec_list) * args->buffer_count);
4113 DRM_ERROR("failed to copy %d exec entries "
4114 "back to user (%d)\n",
4115 args->buffer_count, ret);
4119 drm_free_large(exec_list);
4120 drm_free_large(exec2_list);
4125 i915_gem_execbuffer2(struct drm_device *dev, void *data,
4126 struct drm_file *file_priv)
4128 struct drm_i915_gem_execbuffer2 *args = data;
4129 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
4133 DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
4134 (int) args->buffers_ptr, args->buffer_count, args->batch_len);
4137 if (args->buffer_count < 1) {
4138 DRM_ERROR("execbuf2 with %d buffers\n", args->buffer_count);
4142 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
4143 if (exec2_list == NULL) {
4144 DRM_ERROR("Failed to allocate exec list for %d buffers\n",
4145 args->buffer_count);
4148 ret = copy_from_user(exec2_list,
4149 (struct drm_i915_relocation_entry __user *)
4150 (uintptr_t) args->buffers_ptr,
4151 sizeof(*exec2_list) * args->buffer_count);
4153 DRM_ERROR("copy %d exec entries failed %d\n",
4154 args->buffer_count, ret);
4155 drm_free_large(exec2_list);
4159 ret = i915_gem_do_execbuffer(dev, data, file_priv, args, exec2_list);
4161 /* Copy the new buffer offsets back to the user's exec list. */
4162 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
4163 (uintptr_t) args->buffers_ptr,
4165 sizeof(*exec2_list) * args->buffer_count);
4168 DRM_ERROR("failed to copy %d exec entries "
4169 "back to user (%d)\n",
4170 args->buffer_count, ret);
4174 drm_free_large(exec2_list);
4179 i915_gem_object_pin(struct drm_gem_object *obj, uint32_t alignment)
4181 struct drm_device *dev = obj->dev;
4182 struct drm_i915_gem_object *obj_priv = obj->driver_private;
4185 i915_verify_inactive(dev, __FILE__, __LINE__);
4187 if (obj_priv->gtt_space != NULL) {
4189 alignment = i915_gem_get_gtt_alignment(obj);
4190 if (obj_priv->gtt_offset & (alignment - 1)) {
4191 ret = i915_gem_object_unbind(obj);
4197 if (obj_priv->gtt_space == NULL) {
4198 ret = i915_gem_object_bind_to_gtt(obj, alignment);
4203 obj_priv->pin_count++;
4205 /* If the object is not active and not pending a flush,
4206 * remove it from the inactive list
4208 if (obj_priv->pin_count == 1) {
4209 atomic_inc(&dev->pin_count);
4210 atomic_add(obj->size, &dev->pin_memory);
4211 if (!obj_priv->active &&
4212 (obj->write_domain & I915_GEM_GPU_DOMAINS) == 0 &&
4213 !list_empty(&obj_priv->list))
4214 list_del_init(&obj_priv->list);
4216 i915_verify_inactive(dev, __FILE__, __LINE__);
4222 i915_gem_object_unpin(struct drm_gem_object *obj)
4224 struct drm_device *dev = obj->dev;
4225 drm_i915_private_t *dev_priv = dev->dev_private;
4226 struct drm_i915_gem_object *obj_priv = obj->driver_private;
4228 i915_verify_inactive(dev, __FILE__, __LINE__);
4229 obj_priv->pin_count--;
4230 BUG_ON(obj_priv->pin_count < 0);
4231 BUG_ON(obj_priv->gtt_space == NULL);
4233 /* If the object is no longer pinned, and is
4234 * neither active nor being flushed, then stick it on
4237 if (obj_priv->pin_count == 0) {
4238 if (!obj_priv->active &&
4239 (obj->write_domain & I915_GEM_GPU_DOMAINS) == 0)
4240 list_move_tail(&obj_priv->list,
4241 &dev_priv->mm.inactive_list);
4242 atomic_dec(&dev->pin_count);
4243 atomic_sub(obj->size, &dev->pin_memory);
4245 i915_verify_inactive(dev, __FILE__, __LINE__);
4249 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
4250 struct drm_file *file_priv)
4252 struct drm_i915_gem_pin *args = data;
4253 struct drm_gem_object *obj;
4254 struct drm_i915_gem_object *obj_priv;
4257 mutex_lock(&dev->struct_mutex);
4259 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4261 DRM_ERROR("Bad handle in i915_gem_pin_ioctl(): %d\n",
4263 mutex_unlock(&dev->struct_mutex);
4266 obj_priv = obj->driver_private;
4268 if (obj_priv->madv != I915_MADV_WILLNEED) {
4269 DRM_ERROR("Attempting to pin a purgeable buffer\n");
4270 drm_gem_object_unreference(obj);
4271 mutex_unlock(&dev->struct_mutex);
4275 if (obj_priv->pin_filp != NULL && obj_priv->pin_filp != file_priv) {
4276 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
4278 drm_gem_object_unreference(obj);
4279 mutex_unlock(&dev->struct_mutex);
4283 obj_priv->user_pin_count++;
4284 obj_priv->pin_filp = file_priv;
4285 if (obj_priv->user_pin_count == 1) {
4286 ret = i915_gem_object_pin(obj, args->alignment);
4288 drm_gem_object_unreference(obj);
4289 mutex_unlock(&dev->struct_mutex);
4294 /* XXX - flush the CPU caches for pinned objects
4295 * as the X server doesn't manage domains yet
4297 i915_gem_object_flush_cpu_write_domain(obj);
4298 args->offset = obj_priv->gtt_offset;
4299 drm_gem_object_unreference(obj);
4300 mutex_unlock(&dev->struct_mutex);
4306 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
4307 struct drm_file *file_priv)
4309 struct drm_i915_gem_pin *args = data;
4310 struct drm_gem_object *obj;
4311 struct drm_i915_gem_object *obj_priv;
4313 mutex_lock(&dev->struct_mutex);
4315 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4317 DRM_ERROR("Bad handle in i915_gem_unpin_ioctl(): %d\n",
4319 mutex_unlock(&dev->struct_mutex);
4323 obj_priv = obj->driver_private;
4324 if (obj_priv->pin_filp != file_priv) {
4325 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
4327 drm_gem_object_unreference(obj);
4328 mutex_unlock(&dev->struct_mutex);
4331 obj_priv->user_pin_count--;
4332 if (obj_priv->user_pin_count == 0) {
4333 obj_priv->pin_filp = NULL;
4334 i915_gem_object_unpin(obj);
4337 drm_gem_object_unreference(obj);
4338 mutex_unlock(&dev->struct_mutex);
4343 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
4344 struct drm_file *file_priv)
4346 struct drm_i915_gem_busy *args = data;
4347 struct drm_gem_object *obj;
4348 struct drm_i915_gem_object *obj_priv;
4350 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4352 DRM_ERROR("Bad handle in i915_gem_busy_ioctl(): %d\n",
4357 mutex_lock(&dev->struct_mutex);
4358 /* Update the active list for the hardware's current position.
4359 * Otherwise this only updates on a delayed timer or when irqs are
4360 * actually unmasked, and our working set ends up being larger than
4363 i915_gem_retire_requests(dev);
4365 obj_priv = obj->driver_private;
4366 /* Don't count being on the flushing list against the object being
4367 * done. Otherwise, a buffer left on the flushing list but not getting
4368 * flushed (because nobody's flushing that domain) won't ever return
4369 * unbusy and get reused by libdrm's bo cache. The other expected
4370 * consumer of this interface, OpenGL's occlusion queries, also specs
4371 * that the objects get unbusy "eventually" without any interference.
4373 args->busy = obj_priv->active && obj_priv->last_rendering_seqno != 0;
4375 drm_gem_object_unreference(obj);
4376 mutex_unlock(&dev->struct_mutex);
4381 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
4382 struct drm_file *file_priv)
4384 return i915_gem_ring_throttle(dev, file_priv);
4388 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
4389 struct drm_file *file_priv)
4391 struct drm_i915_gem_madvise *args = data;
4392 struct drm_gem_object *obj;
4393 struct drm_i915_gem_object *obj_priv;
4395 switch (args->madv) {
4396 case I915_MADV_DONTNEED:
4397 case I915_MADV_WILLNEED:
4403 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4405 DRM_ERROR("Bad handle in i915_gem_madvise_ioctl(): %d\n",
4410 mutex_lock(&dev->struct_mutex);
4411 obj_priv = obj->driver_private;
4413 if (obj_priv->pin_count) {
4414 drm_gem_object_unreference(obj);
4415 mutex_unlock(&dev->struct_mutex);
4417 DRM_ERROR("Attempted i915_gem_madvise_ioctl() on a pinned object\n");
4421 if (obj_priv->madv != __I915_MADV_PURGED)
4422 obj_priv->madv = args->madv;
4424 /* if the object is no longer bound, discard its backing storage */
4425 if (i915_gem_object_is_purgeable(obj_priv) &&
4426 obj_priv->gtt_space == NULL)
4427 i915_gem_object_truncate(obj);
4429 args->retained = obj_priv->madv != __I915_MADV_PURGED;
4431 drm_gem_object_unreference(obj);
4432 mutex_unlock(&dev->struct_mutex);
4437 int i915_gem_init_object(struct drm_gem_object *obj)
4439 struct drm_i915_gem_object *obj_priv;
4441 obj_priv = kzalloc(sizeof(*obj_priv), GFP_KERNEL);
4442 if (obj_priv == NULL)
4446 * We've just allocated pages from the kernel,
4447 * so they've just been written by the CPU with
4448 * zeros. They'll need to be clflushed before we
4449 * use them with the GPU.
4451 obj->write_domain = I915_GEM_DOMAIN_CPU;
4452 obj->read_domains = I915_GEM_DOMAIN_CPU;
4454 obj_priv->agp_type = AGP_USER_MEMORY;
4456 obj->driver_private = obj_priv;
4457 obj_priv->obj = obj;
4458 obj_priv->fence_reg = I915_FENCE_REG_NONE;
4459 INIT_LIST_HEAD(&obj_priv->list);
4460 INIT_LIST_HEAD(&obj_priv->gpu_write_list);
4461 INIT_LIST_HEAD(&obj_priv->fence_list);
4462 obj_priv->madv = I915_MADV_WILLNEED;
4464 trace_i915_gem_object_create(obj);
4469 void i915_gem_free_object(struct drm_gem_object *obj)
4471 struct drm_device *dev = obj->dev;
4472 struct drm_i915_gem_object *obj_priv = obj->driver_private;
4474 trace_i915_gem_object_destroy(obj);
4476 while (obj_priv->pin_count > 0)
4477 i915_gem_object_unpin(obj);
4479 if (obj_priv->phys_obj)
4480 i915_gem_detach_phys_object(dev, obj);
4482 i915_gem_object_unbind(obj);
4484 if (obj_priv->mmap_offset)
4485 i915_gem_free_mmap_offset(obj);
4487 kfree(obj_priv->page_cpu_valid);
4488 kfree(obj_priv->bit_17);
4489 kfree(obj->driver_private);
4492 /** Unbinds all inactive objects. */
4494 i915_gem_evict_from_inactive_list(struct drm_device *dev)
4496 drm_i915_private_t *dev_priv = dev->dev_private;
4498 while (!list_empty(&dev_priv->mm.inactive_list)) {
4499 struct drm_gem_object *obj;
4502 obj = list_first_entry(&dev_priv->mm.inactive_list,
4503 struct drm_i915_gem_object,
4506 ret = i915_gem_object_unbind(obj);
4508 DRM_ERROR("Error unbinding object: %d\n", ret);
4517 i915_gem_idle(struct drm_device *dev)
4519 drm_i915_private_t *dev_priv = dev->dev_private;
4520 uint32_t seqno, cur_seqno, last_seqno;
4523 mutex_lock(&dev->struct_mutex);
4525 if (dev_priv->mm.suspended || dev_priv->ring.ring_obj == NULL) {
4526 mutex_unlock(&dev->struct_mutex);
4530 /* Hack! Don't let anybody do execbuf while we don't control the chip.
4531 * We need to replace this with a semaphore, or something.
4533 dev_priv->mm.suspended = 1;
4534 del_timer(&dev_priv->hangcheck_timer);
4536 /* Cancel the retire work handler, wait for it to finish if running
4538 mutex_unlock(&dev->struct_mutex);
4539 cancel_delayed_work_sync(&dev_priv->mm.retire_work);
4540 mutex_lock(&dev->struct_mutex);
4542 i915_kernel_lost_context(dev);
4544 /* Flush the GPU along with all non-CPU write domains
4546 i915_gem_flush(dev, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
4547 seqno = i915_add_request(dev, NULL, I915_GEM_GPU_DOMAINS);
4550 mutex_unlock(&dev->struct_mutex);
4554 dev_priv->mm.waiting_gem_seqno = seqno;
4558 cur_seqno = i915_get_gem_seqno(dev);
4559 if (i915_seqno_passed(cur_seqno, seqno))
4561 if (last_seqno == cur_seqno) {
4562 if (stuck++ > 100) {
4563 DRM_ERROR("hardware wedged\n");
4564 atomic_set(&dev_priv->mm.wedged, 1);
4565 DRM_WAKEUP(&dev_priv->irq_queue);
4570 last_seqno = cur_seqno;
4572 dev_priv->mm.waiting_gem_seqno = 0;
4574 i915_gem_retire_requests(dev);
4576 spin_lock(&dev_priv->mm.active_list_lock);
4577 if (!atomic_read(&dev_priv->mm.wedged)) {
4578 /* Active and flushing should now be empty as we've
4579 * waited for a sequence higher than any pending execbuffer
4581 WARN_ON(!list_empty(&dev_priv->mm.active_list));
4582 WARN_ON(!list_empty(&dev_priv->mm.flushing_list));
4583 /* Request should now be empty as we've also waited
4584 * for the last request in the list
4586 WARN_ON(!list_empty(&dev_priv->mm.request_list));
4589 /* Empty the active and flushing lists to inactive. If there's
4590 * anything left at this point, it means that we're wedged and
4591 * nothing good's going to happen by leaving them there. So strip
4592 * the GPU domains and just stuff them onto inactive.
4594 while (!list_empty(&dev_priv->mm.active_list)) {
4595 struct drm_gem_object *obj;
4596 uint32_t old_write_domain;
4598 obj = list_first_entry(&dev_priv->mm.active_list,
4599 struct drm_i915_gem_object,
4601 old_write_domain = obj->write_domain;
4602 obj->write_domain &= ~I915_GEM_GPU_DOMAINS;
4603 i915_gem_object_move_to_inactive(obj);
4605 trace_i915_gem_object_change_domain(obj,
4609 spin_unlock(&dev_priv->mm.active_list_lock);
4611 while (!list_empty(&dev_priv->mm.flushing_list)) {
4612 struct drm_gem_object *obj;
4613 uint32_t old_write_domain;
4615 obj = list_first_entry(&dev_priv->mm.flushing_list,
4616 struct drm_i915_gem_object,
4618 old_write_domain = obj->write_domain;
4619 obj->write_domain &= ~I915_GEM_GPU_DOMAINS;
4620 i915_gem_object_move_to_inactive(obj);
4622 trace_i915_gem_object_change_domain(obj,
4628 /* Move all inactive buffers out of the GTT. */
4629 ret = i915_gem_evict_from_inactive_list(dev);
4630 WARN_ON(!list_empty(&dev_priv->mm.inactive_list));
4632 mutex_unlock(&dev->struct_mutex);
4636 i915_gem_cleanup_ringbuffer(dev);
4637 mutex_unlock(&dev->struct_mutex);
4643 * 965+ support PIPE_CONTROL commands, which provide finer grained control
4644 * over cache flushing.
4647 i915_gem_init_pipe_control(struct drm_device *dev)
4649 drm_i915_private_t *dev_priv = dev->dev_private;
4650 struct drm_gem_object *obj;
4651 struct drm_i915_gem_object *obj_priv;
4654 obj = drm_gem_object_alloc(dev, 4096);
4656 DRM_ERROR("Failed to allocate seqno page\n");
4660 obj_priv = obj->driver_private;
4661 obj_priv->agp_type = AGP_USER_CACHED_MEMORY;
4663 ret = i915_gem_object_pin(obj, 4096);
4667 dev_priv->seqno_gfx_addr = obj_priv->gtt_offset;
4668 dev_priv->seqno_page = kmap(obj_priv->pages[0]);
4669 if (dev_priv->seqno_page == NULL)
4672 dev_priv->seqno_obj = obj;
4673 memset(dev_priv->seqno_page, 0, PAGE_SIZE);
4678 i915_gem_object_unpin(obj);
4680 drm_gem_object_unreference(obj);
4686 i915_gem_init_hws(struct drm_device *dev)
4688 drm_i915_private_t *dev_priv = dev->dev_private;
4689 struct drm_gem_object *obj;
4690 struct drm_i915_gem_object *obj_priv;
4693 /* If we need a physical address for the status page, it's already
4694 * initialized at driver load time.
4696 if (!I915_NEED_GFX_HWS(dev))
4699 obj = drm_gem_object_alloc(dev, 4096);
4701 DRM_ERROR("Failed to allocate status page\n");
4705 obj_priv = obj->driver_private;
4706 obj_priv->agp_type = AGP_USER_CACHED_MEMORY;
4708 ret = i915_gem_object_pin(obj, 4096);
4710 drm_gem_object_unreference(obj);
4714 dev_priv->status_gfx_addr = obj_priv->gtt_offset;
4716 dev_priv->hw_status_page = kmap(obj_priv->pages[0]);
4717 if (dev_priv->hw_status_page == NULL) {
4718 DRM_ERROR("Failed to map status page.\n");
4719 memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map));
4724 if (HAS_PIPE_CONTROL(dev)) {
4725 ret = i915_gem_init_pipe_control(dev);
4730 dev_priv->hws_obj = obj;
4731 memset(dev_priv->hw_status_page, 0, PAGE_SIZE);
4732 I915_WRITE(HWS_PGA, dev_priv->status_gfx_addr);
4733 I915_READ(HWS_PGA); /* posting read */
4734 DRM_DEBUG_DRIVER("hws offset: 0x%08x\n", dev_priv->status_gfx_addr);
4739 i915_gem_object_unpin(obj);
4741 drm_gem_object_unreference(obj);
4747 i915_gem_cleanup_pipe_control(struct drm_device *dev)
4749 drm_i915_private_t *dev_priv = dev->dev_private;
4750 struct drm_gem_object *obj;
4751 struct drm_i915_gem_object *obj_priv;
4753 obj = dev_priv->seqno_obj;
4754 obj_priv = obj->driver_private;
4755 kunmap(obj_priv->pages[0]);
4756 i915_gem_object_unpin(obj);
4757 drm_gem_object_unreference(obj);
4758 dev_priv->seqno_obj = NULL;
4760 dev_priv->seqno_page = NULL;
4764 i915_gem_cleanup_hws(struct drm_device *dev)
4766 drm_i915_private_t *dev_priv = dev->dev_private;
4767 struct drm_gem_object *obj;
4768 struct drm_i915_gem_object *obj_priv;
4770 if (dev_priv->hws_obj == NULL)
4773 obj = dev_priv->hws_obj;
4774 obj_priv = obj->driver_private;
4776 kunmap(obj_priv->pages[0]);
4777 i915_gem_object_unpin(obj);
4778 drm_gem_object_unreference(obj);
4779 dev_priv->hws_obj = NULL;
4781 memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map));
4782 dev_priv->hw_status_page = NULL;
4784 if (HAS_PIPE_CONTROL(dev))
4785 i915_gem_cleanup_pipe_control(dev);
4787 /* Write high address into HWS_PGA when disabling. */
4788 I915_WRITE(HWS_PGA, 0x1ffff000);
4792 i915_gem_init_ringbuffer(struct drm_device *dev)
4794 drm_i915_private_t *dev_priv = dev->dev_private;
4795 struct drm_gem_object *obj;
4796 struct drm_i915_gem_object *obj_priv;
4797 drm_i915_ring_buffer_t *ring = &dev_priv->ring;
4801 ret = i915_gem_init_hws(dev);
4805 obj = drm_gem_object_alloc(dev, 128 * 1024);
4807 DRM_ERROR("Failed to allocate ringbuffer\n");
4808 i915_gem_cleanup_hws(dev);
4811 obj_priv = obj->driver_private;
4813 ret = i915_gem_object_pin(obj, 4096);
4815 drm_gem_object_unreference(obj);
4816 i915_gem_cleanup_hws(dev);
4820 /* Set up the kernel mapping for the ring. */
4821 ring->Size = obj->size;
4823 ring->map.offset = dev->agp->base + obj_priv->gtt_offset;
4824 ring->map.size = obj->size;
4826 ring->map.flags = 0;
4829 drm_core_ioremap_wc(&ring->map, dev);
4830 if (ring->map.handle == NULL) {
4831 DRM_ERROR("Failed to map ringbuffer.\n");
4832 memset(&dev_priv->ring, 0, sizeof(dev_priv->ring));
4833 i915_gem_object_unpin(obj);
4834 drm_gem_object_unreference(obj);
4835 i915_gem_cleanup_hws(dev);
4838 ring->ring_obj = obj;
4839 ring->virtual_start = ring->map.handle;
4841 /* Stop the ring if it's running. */
4842 I915_WRITE(PRB0_CTL, 0);
4843 I915_WRITE(PRB0_TAIL, 0);
4844 I915_WRITE(PRB0_HEAD, 0);
4846 /* Initialize the ring. */
4847 I915_WRITE(PRB0_START, obj_priv->gtt_offset);
4848 head = I915_READ(PRB0_HEAD) & HEAD_ADDR;
4850 /* G45 ring initialization fails to reset head to zero */
4852 DRM_ERROR("Ring head not reset to zero "
4853 "ctl %08x head %08x tail %08x start %08x\n",
4854 I915_READ(PRB0_CTL),
4855 I915_READ(PRB0_HEAD),
4856 I915_READ(PRB0_TAIL),
4857 I915_READ(PRB0_START));
4858 I915_WRITE(PRB0_HEAD, 0);
4860 DRM_ERROR("Ring head forced to zero "
4861 "ctl %08x head %08x tail %08x start %08x\n",
4862 I915_READ(PRB0_CTL),
4863 I915_READ(PRB0_HEAD),
4864 I915_READ(PRB0_TAIL),
4865 I915_READ(PRB0_START));
4868 I915_WRITE(PRB0_CTL,
4869 ((obj->size - 4096) & RING_NR_PAGES) |
4873 head = I915_READ(PRB0_HEAD) & HEAD_ADDR;
4875 /* If the head is still not zero, the ring is dead */
4877 DRM_ERROR("Ring initialization failed "
4878 "ctl %08x head %08x tail %08x start %08x\n",
4879 I915_READ(PRB0_CTL),
4880 I915_READ(PRB0_HEAD),
4881 I915_READ(PRB0_TAIL),
4882 I915_READ(PRB0_START));
4886 /* Update our cache of the ring state */
4887 if (!drm_core_check_feature(dev, DRIVER_MODESET))
4888 i915_kernel_lost_context(dev);
4890 ring->head = I915_READ(PRB0_HEAD) & HEAD_ADDR;
4891 ring->tail = I915_READ(PRB0_TAIL) & TAIL_ADDR;
4892 ring->space = ring->head - (ring->tail + 8);
4893 if (ring->space < 0)
4894 ring->space += ring->Size;
4901 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
4903 drm_i915_private_t *dev_priv = dev->dev_private;
4905 if (dev_priv->ring.ring_obj == NULL)
4908 drm_core_ioremapfree(&dev_priv->ring.map, dev);
4910 i915_gem_object_unpin(dev_priv->ring.ring_obj);
4911 drm_gem_object_unreference(dev_priv->ring.ring_obj);
4912 dev_priv->ring.ring_obj = NULL;
4913 memset(&dev_priv->ring, 0, sizeof(dev_priv->ring));
4915 i915_gem_cleanup_hws(dev);
4919 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
4920 struct drm_file *file_priv)
4922 drm_i915_private_t *dev_priv = dev->dev_private;
4925 if (drm_core_check_feature(dev, DRIVER_MODESET))
4928 if (atomic_read(&dev_priv->mm.wedged)) {
4929 DRM_ERROR("Reenabling wedged hardware, good luck\n");
4930 atomic_set(&dev_priv->mm.wedged, 0);
4933 mutex_lock(&dev->struct_mutex);
4934 dev_priv->mm.suspended = 0;
4936 ret = i915_gem_init_ringbuffer(dev);
4938 mutex_unlock(&dev->struct_mutex);
4942 spin_lock(&dev_priv->mm.active_list_lock);
4943 BUG_ON(!list_empty(&dev_priv->mm.active_list));
4944 spin_unlock(&dev_priv->mm.active_list_lock);
4946 BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
4947 BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
4948 BUG_ON(!list_empty(&dev_priv->mm.request_list));
4949 mutex_unlock(&dev->struct_mutex);
4951 drm_irq_install(dev);
4957 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
4958 struct drm_file *file_priv)
4960 if (drm_core_check_feature(dev, DRIVER_MODESET))
4963 drm_irq_uninstall(dev);
4964 return i915_gem_idle(dev);
4968 i915_gem_lastclose(struct drm_device *dev)
4972 if (drm_core_check_feature(dev, DRIVER_MODESET))
4975 ret = i915_gem_idle(dev);
4977 DRM_ERROR("failed to idle hardware: %d\n", ret);
4981 i915_gem_load(struct drm_device *dev)
4984 drm_i915_private_t *dev_priv = dev->dev_private;
4986 spin_lock_init(&dev_priv->mm.active_list_lock);
4987 INIT_LIST_HEAD(&dev_priv->mm.active_list);
4988 INIT_LIST_HEAD(&dev_priv->mm.flushing_list);
4989 INIT_LIST_HEAD(&dev_priv->mm.gpu_write_list);
4990 INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
4991 INIT_LIST_HEAD(&dev_priv->mm.request_list);
4992 INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4993 INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
4994 i915_gem_retire_work_handler);
4995 dev_priv->mm.next_gem_seqno = 1;
4997 spin_lock(&shrink_list_lock);
4998 list_add(&dev_priv->mm.shrink_list, &shrink_list);
4999 spin_unlock(&shrink_list_lock);
5001 /* Old X drivers will take 0-2 for front, back, depth buffers */
5002 dev_priv->fence_reg_start = 3;
5004 if (IS_I965G(dev) || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
5005 dev_priv->num_fence_regs = 16;
5007 dev_priv->num_fence_regs = 8;
5009 /* Initialize fence registers to zero */
5010 if (IS_I965G(dev)) {
5011 for (i = 0; i < 16; i++)
5012 I915_WRITE64(FENCE_REG_965_0 + (i * 8), 0);
5014 for (i = 0; i < 8; i++)
5015 I915_WRITE(FENCE_REG_830_0 + (i * 4), 0);
5016 if (IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
5017 for (i = 0; i < 8; i++)
5018 I915_WRITE(FENCE_REG_945_8 + (i * 4), 0);
5020 i915_gem_detect_bit_6_swizzle(dev);
5021 init_waitqueue_head(&dev_priv->pending_flip_queue);
5025 * Create a physically contiguous memory object for this object
5026 * e.g. for cursor + overlay regs
5028 int i915_gem_init_phys_object(struct drm_device *dev,
5031 drm_i915_private_t *dev_priv = dev->dev_private;
5032 struct drm_i915_gem_phys_object *phys_obj;
5035 if (dev_priv->mm.phys_objs[id - 1] || !size)
5038 phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL);
5044 phys_obj->handle = drm_pci_alloc(dev, size, 0);
5045 if (!phys_obj->handle) {
5050 set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
5053 dev_priv->mm.phys_objs[id - 1] = phys_obj;
5061 void i915_gem_free_phys_object(struct drm_device *dev, int id)
5063 drm_i915_private_t *dev_priv = dev->dev_private;
5064 struct drm_i915_gem_phys_object *phys_obj;
5066 if (!dev_priv->mm.phys_objs[id - 1])
5069 phys_obj = dev_priv->mm.phys_objs[id - 1];
5070 if (phys_obj->cur_obj) {
5071 i915_gem_detach_phys_object(dev, phys_obj->cur_obj);
5075 set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
5077 drm_pci_free(dev, phys_obj->handle);
5079 dev_priv->mm.phys_objs[id - 1] = NULL;
5082 void i915_gem_free_all_phys_object(struct drm_device *dev)
5086 for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++)
5087 i915_gem_free_phys_object(dev, i);
5090 void i915_gem_detach_phys_object(struct drm_device *dev,
5091 struct drm_gem_object *obj)
5093 struct drm_i915_gem_object *obj_priv;
5098 obj_priv = obj->driver_private;
5099 if (!obj_priv->phys_obj)
5102 ret = i915_gem_object_get_pages(obj, 0);
5106 page_count = obj->size / PAGE_SIZE;
5108 for (i = 0; i < page_count; i++) {
5109 char *dst = kmap_atomic(obj_priv->pages[i], KM_USER0);
5110 char *src = obj_priv->phys_obj->handle->vaddr + (i * PAGE_SIZE);
5112 memcpy(dst, src, PAGE_SIZE);
5113 kunmap_atomic(dst, KM_USER0);
5115 drm_clflush_pages(obj_priv->pages, page_count);
5116 drm_agp_chipset_flush(dev);
5118 i915_gem_object_put_pages(obj);
5120 obj_priv->phys_obj->cur_obj = NULL;
5121 obj_priv->phys_obj = NULL;
5125 i915_gem_attach_phys_object(struct drm_device *dev,
5126 struct drm_gem_object *obj, int id)
5128 drm_i915_private_t *dev_priv = dev->dev_private;
5129 struct drm_i915_gem_object *obj_priv;
5134 if (id > I915_MAX_PHYS_OBJECT)
5137 obj_priv = obj->driver_private;
5139 if (obj_priv->phys_obj) {
5140 if (obj_priv->phys_obj->id == id)
5142 i915_gem_detach_phys_object(dev, obj);
5146 /* create a new object */
5147 if (!dev_priv->mm.phys_objs[id - 1]) {
5148 ret = i915_gem_init_phys_object(dev, id,
5151 DRM_ERROR("failed to init phys object %d size: %zu\n", id, obj->size);
5156 /* bind to the object */
5157 obj_priv->phys_obj = dev_priv->mm.phys_objs[id - 1];
5158 obj_priv->phys_obj->cur_obj = obj;
5160 ret = i915_gem_object_get_pages(obj, 0);
5162 DRM_ERROR("failed to get page list\n");
5166 page_count = obj->size / PAGE_SIZE;
5168 for (i = 0; i < page_count; i++) {
5169 char *src = kmap_atomic(obj_priv->pages[i], KM_USER0);
5170 char *dst = obj_priv->phys_obj->handle->vaddr + (i * PAGE_SIZE);
5172 memcpy(dst, src, PAGE_SIZE);
5173 kunmap_atomic(src, KM_USER0);
5176 i915_gem_object_put_pages(obj);
5184 i915_gem_phys_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
5185 struct drm_i915_gem_pwrite *args,
5186 struct drm_file *file_priv)
5188 struct drm_i915_gem_object *obj_priv = obj->driver_private;
5191 char __user *user_data;
5193 user_data = (char __user *) (uintptr_t) args->data_ptr;
5194 obj_addr = obj_priv->phys_obj->handle->vaddr + args->offset;
5196 DRM_DEBUG_DRIVER("obj_addr %p, %lld\n", obj_addr, args->size);
5197 ret = copy_from_user(obj_addr, user_data, args->size);
5201 drm_agp_chipset_flush(dev);
5205 void i915_gem_release(struct drm_device * dev, struct drm_file *file_priv)
5207 struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv;
5209 /* Clean up our request list when the client is going away, so that
5210 * later retire_requests won't dereference our soon-to-be-gone
5213 mutex_lock(&dev->struct_mutex);
5214 while (!list_empty(&i915_file_priv->mm.request_list))
5215 list_del_init(i915_file_priv->mm.request_list.next);
5216 mutex_unlock(&dev->struct_mutex);
5220 i915_gem_shrink(int nr_to_scan, gfp_t gfp_mask)
5222 drm_i915_private_t *dev_priv, *next_dev;
5223 struct drm_i915_gem_object *obj_priv, *next_obj;
5225 int would_deadlock = 1;
5227 /* "fast-path" to count number of available objects */
5228 if (nr_to_scan == 0) {
5229 spin_lock(&shrink_list_lock);
5230 list_for_each_entry(dev_priv, &shrink_list, mm.shrink_list) {
5231 struct drm_device *dev = dev_priv->dev;
5233 if (mutex_trylock(&dev->struct_mutex)) {
5234 list_for_each_entry(obj_priv,
5235 &dev_priv->mm.inactive_list,
5238 mutex_unlock(&dev->struct_mutex);
5241 spin_unlock(&shrink_list_lock);
5243 return (cnt / 100) * sysctl_vfs_cache_pressure;
5246 spin_lock(&shrink_list_lock);
5248 /* first scan for clean buffers */
5249 list_for_each_entry_safe(dev_priv, next_dev,
5250 &shrink_list, mm.shrink_list) {
5251 struct drm_device *dev = dev_priv->dev;
5253 if (! mutex_trylock(&dev->struct_mutex))
5256 spin_unlock(&shrink_list_lock);
5258 i915_gem_retire_requests(dev);
5260 list_for_each_entry_safe(obj_priv, next_obj,
5261 &dev_priv->mm.inactive_list,
5263 if (i915_gem_object_is_purgeable(obj_priv)) {
5264 i915_gem_object_unbind(obj_priv->obj);
5265 if (--nr_to_scan <= 0)
5270 spin_lock(&shrink_list_lock);
5271 mutex_unlock(&dev->struct_mutex);
5275 if (nr_to_scan <= 0)
5279 /* second pass, evict/count anything still on the inactive list */
5280 list_for_each_entry_safe(dev_priv, next_dev,
5281 &shrink_list, mm.shrink_list) {
5282 struct drm_device *dev = dev_priv->dev;
5284 if (! mutex_trylock(&dev->struct_mutex))
5287 spin_unlock(&shrink_list_lock);
5289 list_for_each_entry_safe(obj_priv, next_obj,
5290 &dev_priv->mm.inactive_list,
5292 if (nr_to_scan > 0) {
5293 i915_gem_object_unbind(obj_priv->obj);
5299 spin_lock(&shrink_list_lock);
5300 mutex_unlock(&dev->struct_mutex);
5305 spin_unlock(&shrink_list_lock);
5310 return (cnt / 100) * sysctl_vfs_cache_pressure;
5315 static struct shrinker shrinker = {
5316 .shrink = i915_gem_shrink,
5317 .seeks = DEFAULT_SEEKS,
5321 i915_gem_shrinker_init(void)
5323 register_shrinker(&shrinker);
5327 i915_gem_shrinker_exit(void)
5329 unregister_shrinker(&shrinker);