2 * Copyright © 2008-2015 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>
29 #include <drm/drm_vma_manager.h>
30 #include <drm/i915_drm.h>
32 #include "i915_vgpu.h"
33 #include "i915_trace.h"
34 #include "intel_drv.h"
35 #include "intel_mocs.h"
36 #include <linux/shmem_fs.h>
37 #include <linux/slab.h>
38 #include <linux/swap.h>
39 #include <linux/pci.h>
40 #include <linux/dma-buf.h>
42 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object
*obj
);
43 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object
*obj
);
45 i915_gem_object_retire__write(struct drm_i915_gem_object
*obj
);
47 i915_gem_object_retire__read(struct drm_i915_gem_object
*obj
, int ring
);
49 static bool cpu_cache_is_coherent(struct drm_device
*dev
,
50 enum i915_cache_level level
)
52 return HAS_LLC(dev
) || level
!= I915_CACHE_NONE
;
55 static bool cpu_write_needs_clflush(struct drm_i915_gem_object
*obj
)
57 if (!cpu_cache_is_coherent(obj
->base
.dev
, obj
->cache_level
))
60 return obj
->pin_display
;
63 /* some bookkeeping */
64 static void i915_gem_info_add_obj(struct drm_i915_private
*dev_priv
,
67 spin_lock(&dev_priv
->mm
.object_stat_lock
);
68 dev_priv
->mm
.object_count
++;
69 dev_priv
->mm
.object_memory
+= size
;
70 spin_unlock(&dev_priv
->mm
.object_stat_lock
);
73 static void i915_gem_info_remove_obj(struct drm_i915_private
*dev_priv
,
76 spin_lock(&dev_priv
->mm
.object_stat_lock
);
77 dev_priv
->mm
.object_count
--;
78 dev_priv
->mm
.object_memory
-= size
;
79 spin_unlock(&dev_priv
->mm
.object_stat_lock
);
83 i915_gem_wait_for_error(struct i915_gpu_error
*error
)
87 if (!i915_reset_in_progress(error
))
91 * Only wait 10 seconds for the gpu reset to complete to avoid hanging
92 * userspace. If it takes that long something really bad is going on and
93 * we should simply try to bail out and fail as gracefully as possible.
95 ret
= wait_event_interruptible_timeout(error
->reset_queue
,
96 !i915_reset_in_progress(error
),
99 DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
101 } else if (ret
< 0) {
108 int i915_mutex_lock_interruptible(struct drm_device
*dev
)
110 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
113 ret
= i915_gem_wait_for_error(&dev_priv
->gpu_error
);
117 ret
= mutex_lock_interruptible(&dev
->struct_mutex
);
121 WARN_ON(i915_verify_lists(dev
));
126 i915_gem_get_aperture_ioctl(struct drm_device
*dev
, void *data
,
127 struct drm_file
*file
)
129 struct drm_i915_private
*dev_priv
= to_i915(dev
);
130 struct i915_ggtt
*ggtt
= &dev_priv
->ggtt
;
131 struct drm_i915_gem_get_aperture
*args
= data
;
132 struct i915_vma
*vma
;
136 mutex_lock(&dev
->struct_mutex
);
137 list_for_each_entry(vma
, &ggtt
->base
.active_list
, vm_link
)
139 pinned
+= vma
->node
.size
;
140 list_for_each_entry(vma
, &ggtt
->base
.inactive_list
, vm_link
)
142 pinned
+= vma
->node
.size
;
143 mutex_unlock(&dev
->struct_mutex
);
145 args
->aper_size
= ggtt
->base
.total
;
146 args
->aper_available_size
= args
->aper_size
- pinned
;
152 i915_gem_object_get_pages_phys(struct drm_i915_gem_object
*obj
)
154 struct address_space
*mapping
= file_inode(obj
->base
.filp
)->i_mapping
;
155 char *vaddr
= obj
->phys_handle
->vaddr
;
157 struct scatterlist
*sg
;
160 if (WARN_ON(i915_gem_object_needs_bit17_swizzle(obj
)))
163 for (i
= 0; i
< obj
->base
.size
/ PAGE_SIZE
; i
++) {
167 page
= shmem_read_mapping_page(mapping
, i
);
169 return PTR_ERR(page
);
171 src
= kmap_atomic(page
);
172 memcpy(vaddr
, src
, PAGE_SIZE
);
173 drm_clflush_virt_range(vaddr
, PAGE_SIZE
);
180 i915_gem_chipset_flush(obj
->base
.dev
);
182 st
= kmalloc(sizeof(*st
), GFP_KERNEL
);
186 if (sg_alloc_table(st
, 1, GFP_KERNEL
)) {
193 sg
->length
= obj
->base
.size
;
195 sg_dma_address(sg
) = obj
->phys_handle
->busaddr
;
196 sg_dma_len(sg
) = obj
->base
.size
;
203 i915_gem_object_put_pages_phys(struct drm_i915_gem_object
*obj
)
207 BUG_ON(obj
->madv
== __I915_MADV_PURGED
);
209 ret
= i915_gem_object_set_to_cpu_domain(obj
, true);
211 /* In the event of a disaster, abandon all caches and
214 obj
->base
.read_domains
= obj
->base
.write_domain
= I915_GEM_DOMAIN_CPU
;
217 if (obj
->madv
== I915_MADV_DONTNEED
)
221 struct address_space
*mapping
= file_inode(obj
->base
.filp
)->i_mapping
;
222 char *vaddr
= obj
->phys_handle
->vaddr
;
225 for (i
= 0; i
< obj
->base
.size
/ PAGE_SIZE
; i
++) {
229 page
= shmem_read_mapping_page(mapping
, i
);
233 dst
= kmap_atomic(page
);
234 drm_clflush_virt_range(vaddr
, PAGE_SIZE
);
235 memcpy(dst
, vaddr
, PAGE_SIZE
);
238 set_page_dirty(page
);
239 if (obj
->madv
== I915_MADV_WILLNEED
)
240 mark_page_accessed(page
);
247 sg_free_table(obj
->pages
);
252 i915_gem_object_release_phys(struct drm_i915_gem_object
*obj
)
254 drm_pci_free(obj
->base
.dev
, obj
->phys_handle
);
257 static const struct drm_i915_gem_object_ops i915_gem_phys_ops
= {
258 .get_pages
= i915_gem_object_get_pages_phys
,
259 .put_pages
= i915_gem_object_put_pages_phys
,
260 .release
= i915_gem_object_release_phys
,
264 drop_pages(struct drm_i915_gem_object
*obj
)
266 struct i915_vma
*vma
, *next
;
269 drm_gem_object_reference(&obj
->base
);
270 list_for_each_entry_safe(vma
, next
, &obj
->vma_list
, obj_link
)
271 if (i915_vma_unbind(vma
))
274 ret
= i915_gem_object_put_pages(obj
);
275 drm_gem_object_unreference(&obj
->base
);
281 i915_gem_object_attach_phys(struct drm_i915_gem_object
*obj
,
284 drm_dma_handle_t
*phys
;
287 if (obj
->phys_handle
) {
288 if ((unsigned long)obj
->phys_handle
->vaddr
& (align
-1))
294 if (obj
->madv
!= I915_MADV_WILLNEED
)
297 if (obj
->base
.filp
== NULL
)
300 ret
= drop_pages(obj
);
304 /* create a new object */
305 phys
= drm_pci_alloc(obj
->base
.dev
, obj
->base
.size
, align
);
309 obj
->phys_handle
= phys
;
310 obj
->ops
= &i915_gem_phys_ops
;
312 return i915_gem_object_get_pages(obj
);
316 i915_gem_phys_pwrite(struct drm_i915_gem_object
*obj
,
317 struct drm_i915_gem_pwrite
*args
,
318 struct drm_file
*file_priv
)
320 struct drm_device
*dev
= obj
->base
.dev
;
321 void *vaddr
= obj
->phys_handle
->vaddr
+ args
->offset
;
322 char __user
*user_data
= to_user_ptr(args
->data_ptr
);
325 /* We manually control the domain here and pretend that it
326 * remains coherent i.e. in the GTT domain, like shmem_pwrite.
328 ret
= i915_gem_object_wait_rendering(obj
, false);
332 intel_fb_obj_invalidate(obj
, ORIGIN_CPU
);
333 if (__copy_from_user_inatomic_nocache(vaddr
, user_data
, args
->size
)) {
334 unsigned long unwritten
;
336 /* The physical object once assigned is fixed for the lifetime
337 * of the obj, so we can safely drop the lock and continue
340 mutex_unlock(&dev
->struct_mutex
);
341 unwritten
= copy_from_user(vaddr
, user_data
, args
->size
);
342 mutex_lock(&dev
->struct_mutex
);
349 drm_clflush_virt_range(vaddr
, args
->size
);
350 i915_gem_chipset_flush(dev
);
353 intel_fb_obj_flush(obj
, false, ORIGIN_CPU
);
357 void *i915_gem_object_alloc(struct drm_device
*dev
)
359 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
360 return kmem_cache_zalloc(dev_priv
->objects
, GFP_KERNEL
);
363 void i915_gem_object_free(struct drm_i915_gem_object
*obj
)
365 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
366 kmem_cache_free(dev_priv
->objects
, obj
);
370 i915_gem_create(struct drm_file
*file
,
371 struct drm_device
*dev
,
375 struct drm_i915_gem_object
*obj
;
379 size
= roundup(size
, PAGE_SIZE
);
383 /* Allocate the new object */
384 obj
= i915_gem_alloc_object(dev
, size
);
388 ret
= drm_gem_handle_create(file
, &obj
->base
, &handle
);
389 /* drop reference from allocate - handle holds it now */
390 drm_gem_object_unreference_unlocked(&obj
->base
);
399 i915_gem_dumb_create(struct drm_file
*file
,
400 struct drm_device
*dev
,
401 struct drm_mode_create_dumb
*args
)
403 /* have to work out size/pitch and return them */
404 args
->pitch
= ALIGN(args
->width
* DIV_ROUND_UP(args
->bpp
, 8), 64);
405 args
->size
= args
->pitch
* args
->height
;
406 return i915_gem_create(file
, dev
,
407 args
->size
, &args
->handle
);
411 * Creates a new mm object and returns a handle to it.
414 i915_gem_create_ioctl(struct drm_device
*dev
, void *data
,
415 struct drm_file
*file
)
417 struct drm_i915_gem_create
*args
= data
;
419 return i915_gem_create(file
, dev
,
420 args
->size
, &args
->handle
);
424 __copy_to_user_swizzled(char __user
*cpu_vaddr
,
425 const char *gpu_vaddr
, int gpu_offset
,
428 int ret
, cpu_offset
= 0;
431 int cacheline_end
= ALIGN(gpu_offset
+ 1, 64);
432 int this_length
= min(cacheline_end
- gpu_offset
, length
);
433 int swizzled_gpu_offset
= gpu_offset
^ 64;
435 ret
= __copy_to_user(cpu_vaddr
+ cpu_offset
,
436 gpu_vaddr
+ swizzled_gpu_offset
,
441 cpu_offset
+= this_length
;
442 gpu_offset
+= this_length
;
443 length
-= this_length
;
450 __copy_from_user_swizzled(char *gpu_vaddr
, int gpu_offset
,
451 const char __user
*cpu_vaddr
,
454 int ret
, cpu_offset
= 0;
457 int cacheline_end
= ALIGN(gpu_offset
+ 1, 64);
458 int this_length
= min(cacheline_end
- gpu_offset
, length
);
459 int swizzled_gpu_offset
= gpu_offset
^ 64;
461 ret
= __copy_from_user(gpu_vaddr
+ swizzled_gpu_offset
,
462 cpu_vaddr
+ cpu_offset
,
467 cpu_offset
+= this_length
;
468 gpu_offset
+= this_length
;
469 length
-= this_length
;
476 * Pins the specified object's pages and synchronizes the object with
477 * GPU accesses. Sets needs_clflush to non-zero if the caller should
478 * flush the object from the CPU cache.
480 int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object
*obj
,
487 if (WARN_ON((obj
->ops
->flags
& I915_GEM_OBJECT_HAS_STRUCT_PAGE
) == 0))
490 if (!(obj
->base
.read_domains
& I915_GEM_DOMAIN_CPU
)) {
491 /* If we're not in the cpu read domain, set ourself into the gtt
492 * read domain and manually flush cachelines (if required). This
493 * optimizes for the case when the gpu will dirty the data
494 * anyway again before the next pread happens. */
495 *needs_clflush
= !cpu_cache_is_coherent(obj
->base
.dev
,
497 ret
= i915_gem_object_wait_rendering(obj
, true);
502 ret
= i915_gem_object_get_pages(obj
);
506 i915_gem_object_pin_pages(obj
);
511 /* Per-page copy function for the shmem pread fastpath.
512 * Flushes invalid cachelines before reading the target if
513 * needs_clflush is set. */
515 shmem_pread_fast(struct page
*page
, int shmem_page_offset
, int page_length
,
516 char __user
*user_data
,
517 bool page_do_bit17_swizzling
, bool needs_clflush
)
522 if (unlikely(page_do_bit17_swizzling
))
525 vaddr
= kmap_atomic(page
);
527 drm_clflush_virt_range(vaddr
+ shmem_page_offset
,
529 ret
= __copy_to_user_inatomic(user_data
,
530 vaddr
+ shmem_page_offset
,
532 kunmap_atomic(vaddr
);
534 return ret
? -EFAULT
: 0;
538 shmem_clflush_swizzled_range(char *addr
, unsigned long length
,
541 if (unlikely(swizzled
)) {
542 unsigned long start
= (unsigned long) addr
;
543 unsigned long end
= (unsigned long) addr
+ length
;
545 /* For swizzling simply ensure that we always flush both
546 * channels. Lame, but simple and it works. Swizzled
547 * pwrite/pread is far from a hotpath - current userspace
548 * doesn't use it at all. */
549 start
= round_down(start
, 128);
550 end
= round_up(end
, 128);
552 drm_clflush_virt_range((void *)start
, end
- start
);
554 drm_clflush_virt_range(addr
, length
);
559 /* Only difference to the fast-path function is that this can handle bit17
560 * and uses non-atomic copy and kmap functions. */
562 shmem_pread_slow(struct page
*page
, int shmem_page_offset
, int page_length
,
563 char __user
*user_data
,
564 bool page_do_bit17_swizzling
, bool needs_clflush
)
571 shmem_clflush_swizzled_range(vaddr
+ shmem_page_offset
,
573 page_do_bit17_swizzling
);
575 if (page_do_bit17_swizzling
)
576 ret
= __copy_to_user_swizzled(user_data
,
577 vaddr
, shmem_page_offset
,
580 ret
= __copy_to_user(user_data
,
581 vaddr
+ shmem_page_offset
,
585 return ret
? - EFAULT
: 0;
589 i915_gem_shmem_pread(struct drm_device
*dev
,
590 struct drm_i915_gem_object
*obj
,
591 struct drm_i915_gem_pread
*args
,
592 struct drm_file
*file
)
594 char __user
*user_data
;
597 int shmem_page_offset
, page_length
, ret
= 0;
598 int obj_do_bit17_swizzling
, page_do_bit17_swizzling
;
600 int needs_clflush
= 0;
601 struct sg_page_iter sg_iter
;
603 user_data
= to_user_ptr(args
->data_ptr
);
606 obj_do_bit17_swizzling
= i915_gem_object_needs_bit17_swizzle(obj
);
608 ret
= i915_gem_obj_prepare_shmem_read(obj
, &needs_clflush
);
612 offset
= args
->offset
;
614 for_each_sg_page(obj
->pages
->sgl
, &sg_iter
, obj
->pages
->nents
,
615 offset
>> PAGE_SHIFT
) {
616 struct page
*page
= sg_page_iter_page(&sg_iter
);
621 /* Operation in this page
623 * shmem_page_offset = offset within page in shmem file
624 * page_length = bytes to copy for this page
626 shmem_page_offset
= offset_in_page(offset
);
627 page_length
= remain
;
628 if ((shmem_page_offset
+ page_length
) > PAGE_SIZE
)
629 page_length
= PAGE_SIZE
- shmem_page_offset
;
631 page_do_bit17_swizzling
= obj_do_bit17_swizzling
&&
632 (page_to_phys(page
) & (1 << 17)) != 0;
634 ret
= shmem_pread_fast(page
, shmem_page_offset
, page_length
,
635 user_data
, page_do_bit17_swizzling
,
640 mutex_unlock(&dev
->struct_mutex
);
642 if (likely(!i915
.prefault_disable
) && !prefaulted
) {
643 ret
= fault_in_multipages_writeable(user_data
, remain
);
644 /* Userspace is tricking us, but we've already clobbered
645 * its pages with the prefault and promised to write the
646 * data up to the first fault. Hence ignore any errors
647 * and just continue. */
652 ret
= shmem_pread_slow(page
, shmem_page_offset
, page_length
,
653 user_data
, page_do_bit17_swizzling
,
656 mutex_lock(&dev
->struct_mutex
);
662 remain
-= page_length
;
663 user_data
+= page_length
;
664 offset
+= page_length
;
668 i915_gem_object_unpin_pages(obj
);
674 * Reads data from the object referenced by handle.
676 * On error, the contents of *data are undefined.
679 i915_gem_pread_ioctl(struct drm_device
*dev
, void *data
,
680 struct drm_file
*file
)
682 struct drm_i915_gem_pread
*args
= data
;
683 struct drm_i915_gem_object
*obj
;
689 if (!access_ok(VERIFY_WRITE
,
690 to_user_ptr(args
->data_ptr
),
694 ret
= i915_mutex_lock_interruptible(dev
);
698 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
699 if (&obj
->base
== NULL
) {
704 /* Bounds check source. */
705 if (args
->offset
> obj
->base
.size
||
706 args
->size
> obj
->base
.size
- args
->offset
) {
711 /* prime objects have no backing filp to GEM pread/pwrite
714 if (!obj
->base
.filp
) {
719 trace_i915_gem_object_pread(obj
, args
->offset
, args
->size
);
721 ret
= i915_gem_shmem_pread(dev
, obj
, args
, file
);
724 drm_gem_object_unreference(&obj
->base
);
726 mutex_unlock(&dev
->struct_mutex
);
730 /* This is the fast write path which cannot handle
731 * page faults in the source data
735 fast_user_write(struct io_mapping
*mapping
,
736 loff_t page_base
, int page_offset
,
737 char __user
*user_data
,
740 void __iomem
*vaddr_atomic
;
742 unsigned long unwritten
;
744 vaddr_atomic
= io_mapping_map_atomic_wc(mapping
, page_base
);
745 /* We can use the cpu mem copy function because this is X86. */
746 vaddr
= (void __force
*)vaddr_atomic
+ page_offset
;
747 unwritten
= __copy_from_user_inatomic_nocache(vaddr
,
749 io_mapping_unmap_atomic(vaddr_atomic
);
754 * This is the fast pwrite path, where we copy the data directly from the
755 * user into the GTT, uncached.
758 i915_gem_gtt_pwrite_fast(struct drm_device
*dev
,
759 struct drm_i915_gem_object
*obj
,
760 struct drm_i915_gem_pwrite
*args
,
761 struct drm_file
*file
)
763 struct drm_i915_private
*dev_priv
= to_i915(dev
);
764 struct i915_ggtt
*ggtt
= &dev_priv
->ggtt
;
766 loff_t offset
, page_base
;
767 char __user
*user_data
;
768 int page_offset
, page_length
, ret
;
770 ret
= i915_gem_obj_ggtt_pin(obj
, 0, PIN_MAPPABLE
| PIN_NONBLOCK
);
774 ret
= i915_gem_object_set_to_gtt_domain(obj
, true);
778 ret
= i915_gem_object_put_fence(obj
);
782 user_data
= to_user_ptr(args
->data_ptr
);
785 offset
= i915_gem_obj_ggtt_offset(obj
) + args
->offset
;
787 intel_fb_obj_invalidate(obj
, ORIGIN_GTT
);
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_MASK
;
797 page_offset
= offset_in_page(offset
);
798 page_length
= remain
;
799 if ((page_offset
+ remain
) > PAGE_SIZE
)
800 page_length
= PAGE_SIZE
- page_offset
;
802 /* If we get a fault while copying data, then (presumably) our
803 * source page isn't available. Return the error and we'll
804 * retry in the slow path.
806 if (fast_user_write(ggtt
->mappable
, page_base
,
807 page_offset
, user_data
, page_length
)) {
812 remain
-= page_length
;
813 user_data
+= page_length
;
814 offset
+= page_length
;
818 intel_fb_obj_flush(obj
, false, ORIGIN_GTT
);
820 i915_gem_object_ggtt_unpin(obj
);
825 /* Per-page copy function for the shmem pwrite fastpath.
826 * Flushes invalid cachelines before writing to the target if
827 * needs_clflush_before is set and flushes out any written cachelines after
828 * writing if needs_clflush is set. */
830 shmem_pwrite_fast(struct page
*page
, int shmem_page_offset
, int page_length
,
831 char __user
*user_data
,
832 bool page_do_bit17_swizzling
,
833 bool needs_clflush_before
,
834 bool needs_clflush_after
)
839 if (unlikely(page_do_bit17_swizzling
))
842 vaddr
= kmap_atomic(page
);
843 if (needs_clflush_before
)
844 drm_clflush_virt_range(vaddr
+ shmem_page_offset
,
846 ret
= __copy_from_user_inatomic(vaddr
+ shmem_page_offset
,
847 user_data
, page_length
);
848 if (needs_clflush_after
)
849 drm_clflush_virt_range(vaddr
+ shmem_page_offset
,
851 kunmap_atomic(vaddr
);
853 return ret
? -EFAULT
: 0;
856 /* Only difference to the fast-path function is that this can handle bit17
857 * and uses non-atomic copy and kmap functions. */
859 shmem_pwrite_slow(struct page
*page
, int shmem_page_offset
, int page_length
,
860 char __user
*user_data
,
861 bool page_do_bit17_swizzling
,
862 bool needs_clflush_before
,
863 bool needs_clflush_after
)
869 if (unlikely(needs_clflush_before
|| page_do_bit17_swizzling
))
870 shmem_clflush_swizzled_range(vaddr
+ shmem_page_offset
,
872 page_do_bit17_swizzling
);
873 if (page_do_bit17_swizzling
)
874 ret
= __copy_from_user_swizzled(vaddr
, shmem_page_offset
,
878 ret
= __copy_from_user(vaddr
+ shmem_page_offset
,
881 if (needs_clflush_after
)
882 shmem_clflush_swizzled_range(vaddr
+ shmem_page_offset
,
884 page_do_bit17_swizzling
);
887 return ret
? -EFAULT
: 0;
891 i915_gem_shmem_pwrite(struct drm_device
*dev
,
892 struct drm_i915_gem_object
*obj
,
893 struct drm_i915_gem_pwrite
*args
,
894 struct drm_file
*file
)
898 char __user
*user_data
;
899 int shmem_page_offset
, page_length
, ret
= 0;
900 int obj_do_bit17_swizzling
, page_do_bit17_swizzling
;
901 int hit_slowpath
= 0;
902 int needs_clflush_after
= 0;
903 int needs_clflush_before
= 0;
904 struct sg_page_iter sg_iter
;
906 user_data
= to_user_ptr(args
->data_ptr
);
909 obj_do_bit17_swizzling
= i915_gem_object_needs_bit17_swizzle(obj
);
911 if (obj
->base
.write_domain
!= I915_GEM_DOMAIN_CPU
) {
912 /* If we're not in the cpu write domain, set ourself into the gtt
913 * write domain and manually flush cachelines (if required). This
914 * optimizes for the case when the gpu will use the data
915 * right away and we therefore have to clflush anyway. */
916 needs_clflush_after
= cpu_write_needs_clflush(obj
);
917 ret
= i915_gem_object_wait_rendering(obj
, false);
921 /* Same trick applies to invalidate partially written cachelines read
923 if ((obj
->base
.read_domains
& I915_GEM_DOMAIN_CPU
) == 0)
924 needs_clflush_before
=
925 !cpu_cache_is_coherent(dev
, obj
->cache_level
);
927 ret
= i915_gem_object_get_pages(obj
);
931 intel_fb_obj_invalidate(obj
, ORIGIN_CPU
);
933 i915_gem_object_pin_pages(obj
);
935 offset
= args
->offset
;
938 for_each_sg_page(obj
->pages
->sgl
, &sg_iter
, obj
->pages
->nents
,
939 offset
>> PAGE_SHIFT
) {
940 struct page
*page
= sg_page_iter_page(&sg_iter
);
941 int partial_cacheline_write
;
946 /* Operation in this page
948 * shmem_page_offset = offset within page in shmem file
949 * page_length = bytes to copy for this page
951 shmem_page_offset
= offset_in_page(offset
);
953 page_length
= remain
;
954 if ((shmem_page_offset
+ page_length
) > PAGE_SIZE
)
955 page_length
= PAGE_SIZE
- shmem_page_offset
;
957 /* If we don't overwrite a cacheline completely we need to be
958 * careful to have up-to-date data by first clflushing. Don't
959 * overcomplicate things and flush the entire patch. */
960 partial_cacheline_write
= needs_clflush_before
&&
961 ((shmem_page_offset
| page_length
)
962 & (boot_cpu_data
.x86_clflush_size
- 1));
964 page_do_bit17_swizzling
= obj_do_bit17_swizzling
&&
965 (page_to_phys(page
) & (1 << 17)) != 0;
967 ret
= shmem_pwrite_fast(page
, shmem_page_offset
, page_length
,
968 user_data
, page_do_bit17_swizzling
,
969 partial_cacheline_write
,
970 needs_clflush_after
);
975 mutex_unlock(&dev
->struct_mutex
);
976 ret
= shmem_pwrite_slow(page
, shmem_page_offset
, page_length
,
977 user_data
, page_do_bit17_swizzling
,
978 partial_cacheline_write
,
979 needs_clflush_after
);
981 mutex_lock(&dev
->struct_mutex
);
987 remain
-= page_length
;
988 user_data
+= page_length
;
989 offset
+= page_length
;
993 i915_gem_object_unpin_pages(obj
);
997 * Fixup: Flush cpu caches in case we didn't flush the dirty
998 * cachelines in-line while writing and the object moved
999 * out of the cpu write domain while we've dropped the lock.
1001 if (!needs_clflush_after
&&
1002 obj
->base
.write_domain
!= I915_GEM_DOMAIN_CPU
) {
1003 if (i915_gem_clflush_object(obj
, obj
->pin_display
))
1004 needs_clflush_after
= true;
1008 if (needs_clflush_after
)
1009 i915_gem_chipset_flush(dev
);
1011 obj
->cache_dirty
= true;
1013 intel_fb_obj_flush(obj
, false, ORIGIN_CPU
);
1018 * Writes data to the object referenced by handle.
1020 * On error, the contents of the buffer that were to be modified are undefined.
1023 i915_gem_pwrite_ioctl(struct drm_device
*dev
, void *data
,
1024 struct drm_file
*file
)
1026 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1027 struct drm_i915_gem_pwrite
*args
= data
;
1028 struct drm_i915_gem_object
*obj
;
1031 if (args
->size
== 0)
1034 if (!access_ok(VERIFY_READ
,
1035 to_user_ptr(args
->data_ptr
),
1039 if (likely(!i915
.prefault_disable
)) {
1040 ret
= fault_in_multipages_readable(to_user_ptr(args
->data_ptr
),
1046 intel_runtime_pm_get(dev_priv
);
1048 ret
= i915_mutex_lock_interruptible(dev
);
1052 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
1053 if (&obj
->base
== NULL
) {
1058 /* Bounds check destination. */
1059 if (args
->offset
> obj
->base
.size
||
1060 args
->size
> obj
->base
.size
- args
->offset
) {
1065 /* prime objects have no backing filp to GEM pread/pwrite
1068 if (!obj
->base
.filp
) {
1073 trace_i915_gem_object_pwrite(obj
, args
->offset
, args
->size
);
1076 /* We can only do the GTT pwrite on untiled buffers, as otherwise
1077 * it would end up going through the fenced access, and we'll get
1078 * different detiling behavior between reading and writing.
1079 * pread/pwrite currently are reading and writing from the CPU
1080 * perspective, requiring manual detiling by the client.
1082 if (obj
->tiling_mode
== I915_TILING_NONE
&&
1083 obj
->base
.write_domain
!= I915_GEM_DOMAIN_CPU
&&
1084 cpu_write_needs_clflush(obj
)) {
1085 ret
= i915_gem_gtt_pwrite_fast(dev
, obj
, args
, file
);
1086 /* Note that the gtt paths might fail with non-page-backed user
1087 * pointers (e.g. gtt mappings when moving data between
1088 * textures). Fallback to the shmem path in that case. */
1091 if (ret
== -EFAULT
|| ret
== -ENOSPC
) {
1092 if (obj
->phys_handle
)
1093 ret
= i915_gem_phys_pwrite(obj
, args
, file
);
1095 ret
= i915_gem_shmem_pwrite(dev
, obj
, args
, file
);
1099 drm_gem_object_unreference(&obj
->base
);
1101 mutex_unlock(&dev
->struct_mutex
);
1103 intel_runtime_pm_put(dev_priv
);
1109 i915_gem_check_wedge(unsigned reset_counter
, bool interruptible
)
1111 if (__i915_terminally_wedged(reset_counter
))
1114 if (__i915_reset_in_progress(reset_counter
)) {
1115 /* Non-interruptible callers can't handle -EAGAIN, hence return
1116 * -EIO unconditionally for these. */
1126 static void fake_irq(unsigned long data
)
1128 wake_up_process((struct task_struct
*)data
);
1131 static bool missed_irq(struct drm_i915_private
*dev_priv
,
1132 struct intel_engine_cs
*engine
)
1134 return test_bit(engine
->id
, &dev_priv
->gpu_error
.missed_irq_rings
);
1137 static unsigned long local_clock_us(unsigned *cpu
)
1141 /* Cheaply and approximately convert from nanoseconds to microseconds.
1142 * The result and subsequent calculations are also defined in the same
1143 * approximate microseconds units. The principal source of timing
1144 * error here is from the simple truncation.
1146 * Note that local_clock() is only defined wrt to the current CPU;
1147 * the comparisons are no longer valid if we switch CPUs. Instead of
1148 * blocking preemption for the entire busywait, we can detect the CPU
1149 * switch and use that as indicator of system load and a reason to
1150 * stop busywaiting, see busywait_stop().
1153 t
= local_clock() >> 10;
1159 static bool busywait_stop(unsigned long timeout
, unsigned cpu
)
1163 if (time_after(local_clock_us(&this_cpu
), timeout
))
1166 return this_cpu
!= cpu
;
1169 static int __i915_spin_request(struct drm_i915_gem_request
*req
, int state
)
1171 unsigned long timeout
;
1174 /* When waiting for high frequency requests, e.g. during synchronous
1175 * rendering split between the CPU and GPU, the finite amount of time
1176 * required to set up the irq and wait upon it limits the response
1177 * rate. By busywaiting on the request completion for a short while we
1178 * can service the high frequency waits as quick as possible. However,
1179 * if it is a slow request, we want to sleep as quickly as possible.
1180 * The tradeoff between waiting and sleeping is roughly the time it
1181 * takes to sleep on a request, on the order of a microsecond.
1184 if (req
->engine
->irq_refcount
)
1187 /* Only spin if we know the GPU is processing this request */
1188 if (!i915_gem_request_started(req
, true))
1191 timeout
= local_clock_us(&cpu
) + 5;
1192 while (!need_resched()) {
1193 if (i915_gem_request_completed(req
, true))
1196 if (signal_pending_state(state
, current
))
1199 if (busywait_stop(timeout
, cpu
))
1202 cpu_relax_lowlatency();
1205 if (i915_gem_request_completed(req
, false))
1212 * __i915_wait_request - wait until execution of request has finished
1214 * @interruptible: do an interruptible wait (normally yes)
1215 * @timeout: in - how long to wait (NULL forever); out - how much time remaining
1217 * Note: It is of utmost importance that the passed in seqno and reset_counter
1218 * values have been read by the caller in an smp safe manner. Where read-side
1219 * locks are involved, it is sufficient to read the reset_counter before
1220 * unlocking the lock that protects the seqno. For lockless tricks, the
1221 * reset_counter _must_ be read before, and an appropriate smp_rmb must be
1224 * Returns 0 if the request was found within the alloted time. Else returns the
1225 * errno with remaining time filled in timeout argument.
1227 int __i915_wait_request(struct drm_i915_gem_request
*req
,
1230 struct intel_rps_client
*rps
)
1232 struct intel_engine_cs
*engine
= i915_gem_request_get_engine(req
);
1233 struct drm_device
*dev
= engine
->dev
;
1234 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1235 const bool irq_test_in_progress
=
1236 ACCESS_ONCE(dev_priv
->gpu_error
.test_irq_rings
) & intel_engine_flag(engine
);
1237 int state
= interruptible
? TASK_INTERRUPTIBLE
: TASK_UNINTERRUPTIBLE
;
1239 unsigned long timeout_expire
;
1240 s64 before
= 0; /* Only to silence a compiler warning. */
1243 WARN(!intel_irqs_enabled(dev_priv
), "IRQs disabled");
1245 if (list_empty(&req
->list
))
1248 if (i915_gem_request_completed(req
, true))
1253 if (WARN_ON(*timeout
< 0))
1259 timeout_expire
= jiffies
+ nsecs_to_jiffies_timeout(*timeout
);
1262 * Record current time in case interrupted by signal, or wedged.
1264 before
= ktime_get_raw_ns();
1267 if (INTEL_INFO(dev_priv
)->gen
>= 6)
1268 gen6_rps_boost(dev_priv
, rps
, req
->emitted_jiffies
);
1270 trace_i915_gem_request_wait_begin(req
);
1272 /* Optimistic spin for the next jiffie before touching IRQs */
1273 ret
= __i915_spin_request(req
, state
);
1277 if (!irq_test_in_progress
&& WARN_ON(!engine
->irq_get(engine
))) {
1283 struct timer_list timer
;
1285 prepare_to_wait(&engine
->irq_queue
, &wait
, state
);
1287 /* We need to check whether any gpu reset happened in between
1288 * the request being submitted and now. If a reset has occurred,
1289 * the request is effectively complete (we either are in the
1290 * process of or have discarded the rendering and completely
1291 * reset the GPU. The results of the request are lost and we
1292 * are free to continue on with the original operation.
1294 if (req
->reset_counter
!= i915_reset_counter(&dev_priv
->gpu_error
)) {
1299 if (i915_gem_request_completed(req
, false)) {
1304 if (signal_pending_state(state
, current
)) {
1309 if (timeout
&& time_after_eq(jiffies
, timeout_expire
)) {
1314 timer
.function
= NULL
;
1315 if (timeout
|| missed_irq(dev_priv
, engine
)) {
1316 unsigned long expire
;
1318 setup_timer_on_stack(&timer
, fake_irq
, (unsigned long)current
);
1319 expire
= missed_irq(dev_priv
, engine
) ? jiffies
+ 1 : timeout_expire
;
1320 mod_timer(&timer
, expire
);
1325 if (timer
.function
) {
1326 del_singleshot_timer_sync(&timer
);
1327 destroy_timer_on_stack(&timer
);
1330 if (!irq_test_in_progress
)
1331 engine
->irq_put(engine
);
1333 finish_wait(&engine
->irq_queue
, &wait
);
1336 trace_i915_gem_request_wait_end(req
);
1339 s64 tres
= *timeout
- (ktime_get_raw_ns() - before
);
1341 *timeout
= tres
< 0 ? 0 : tres
;
1344 * Apparently ktime isn't accurate enough and occasionally has a
1345 * bit of mismatch in the jiffies<->nsecs<->ktime loop. So patch
1346 * things up to make the test happy. We allow up to 1 jiffy.
1348 * This is a regrssion from the timespec->ktime conversion.
1350 if (ret
== -ETIME
&& *timeout
< jiffies_to_usecs(1)*1000)
1357 int i915_gem_request_add_to_client(struct drm_i915_gem_request
*req
,
1358 struct drm_file
*file
)
1360 struct drm_i915_file_private
*file_priv
;
1362 WARN_ON(!req
|| !file
|| req
->file_priv
);
1370 file_priv
= file
->driver_priv
;
1372 spin_lock(&file_priv
->mm
.lock
);
1373 req
->file_priv
= file_priv
;
1374 list_add_tail(&req
->client_list
, &file_priv
->mm
.request_list
);
1375 spin_unlock(&file_priv
->mm
.lock
);
1377 req
->pid
= get_pid(task_pid(current
));
1383 i915_gem_request_remove_from_client(struct drm_i915_gem_request
*request
)
1385 struct drm_i915_file_private
*file_priv
= request
->file_priv
;
1390 spin_lock(&file_priv
->mm
.lock
);
1391 list_del(&request
->client_list
);
1392 request
->file_priv
= NULL
;
1393 spin_unlock(&file_priv
->mm
.lock
);
1395 put_pid(request
->pid
);
1396 request
->pid
= NULL
;
1399 static void i915_gem_request_retire(struct drm_i915_gem_request
*request
)
1401 trace_i915_gem_request_retire(request
);
1403 /* We know the GPU must have read the request to have
1404 * sent us the seqno + interrupt, so use the position
1405 * of tail of the request to update the last known position
1408 * Note this requires that we are always called in request
1411 request
->ringbuf
->last_retired_head
= request
->postfix
;
1413 list_del_init(&request
->list
);
1414 i915_gem_request_remove_from_client(request
);
1416 i915_gem_request_unreference(request
);
1420 __i915_gem_request_retire__upto(struct drm_i915_gem_request
*req
)
1422 struct intel_engine_cs
*engine
= req
->engine
;
1423 struct drm_i915_gem_request
*tmp
;
1425 lockdep_assert_held(&engine
->dev
->struct_mutex
);
1427 if (list_empty(&req
->list
))
1431 tmp
= list_first_entry(&engine
->request_list
,
1432 typeof(*tmp
), list
);
1434 i915_gem_request_retire(tmp
);
1435 } while (tmp
!= req
);
1437 WARN_ON(i915_verify_lists(engine
->dev
));
1441 * Waits for a request to be signaled, and cleans up the
1442 * request and object lists appropriately for that event.
1445 i915_wait_request(struct drm_i915_gem_request
*req
)
1447 struct drm_i915_private
*dev_priv
= req
->i915
;
1451 interruptible
= dev_priv
->mm
.interruptible
;
1453 BUG_ON(!mutex_is_locked(&dev_priv
->dev
->struct_mutex
));
1455 ret
= __i915_wait_request(req
, interruptible
, NULL
, NULL
);
1459 /* If the GPU hung, we want to keep the requests to find the guilty. */
1460 if (req
->reset_counter
== i915_reset_counter(&dev_priv
->gpu_error
))
1461 __i915_gem_request_retire__upto(req
);
1467 * Ensures that all rendering to the object has completed and the object is
1468 * safe to unbind from the GTT or access from the CPU.
1471 i915_gem_object_wait_rendering(struct drm_i915_gem_object
*obj
,
1480 if (obj
->last_write_req
!= NULL
) {
1481 ret
= i915_wait_request(obj
->last_write_req
);
1485 i
= obj
->last_write_req
->engine
->id
;
1486 if (obj
->last_read_req
[i
] == obj
->last_write_req
)
1487 i915_gem_object_retire__read(obj
, i
);
1489 i915_gem_object_retire__write(obj
);
1492 for (i
= 0; i
< I915_NUM_ENGINES
; i
++) {
1493 if (obj
->last_read_req
[i
] == NULL
)
1496 ret
= i915_wait_request(obj
->last_read_req
[i
]);
1500 i915_gem_object_retire__read(obj
, i
);
1502 GEM_BUG_ON(obj
->active
);
1509 i915_gem_object_retire_request(struct drm_i915_gem_object
*obj
,
1510 struct drm_i915_gem_request
*req
)
1512 int ring
= req
->engine
->id
;
1514 if (obj
->last_read_req
[ring
] == req
)
1515 i915_gem_object_retire__read(obj
, ring
);
1516 else if (obj
->last_write_req
== req
)
1517 i915_gem_object_retire__write(obj
);
1519 if (req
->reset_counter
== i915_reset_counter(&req
->i915
->gpu_error
))
1520 __i915_gem_request_retire__upto(req
);
1523 /* A nonblocking variant of the above wait. This is a highly dangerous routine
1524 * as the object state may change during this call.
1526 static __must_check
int
1527 i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object
*obj
,
1528 struct intel_rps_client
*rps
,
1531 struct drm_device
*dev
= obj
->base
.dev
;
1532 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1533 struct drm_i915_gem_request
*requests
[I915_NUM_ENGINES
];
1536 BUG_ON(!mutex_is_locked(&dev
->struct_mutex
));
1537 BUG_ON(!dev_priv
->mm
.interruptible
);
1543 struct drm_i915_gem_request
*req
;
1545 req
= obj
->last_write_req
;
1549 requests
[n
++] = i915_gem_request_reference(req
);
1551 for (i
= 0; i
< I915_NUM_ENGINES
; i
++) {
1552 struct drm_i915_gem_request
*req
;
1554 req
= obj
->last_read_req
[i
];
1558 requests
[n
++] = i915_gem_request_reference(req
);
1562 mutex_unlock(&dev
->struct_mutex
);
1564 for (i
= 0; ret
== 0 && i
< n
; i
++)
1565 ret
= __i915_wait_request(requests
[i
], true, NULL
, rps
);
1566 mutex_lock(&dev
->struct_mutex
);
1568 for (i
= 0; i
< n
; i
++) {
1570 i915_gem_object_retire_request(obj
, requests
[i
]);
1571 i915_gem_request_unreference(requests
[i
]);
1577 static struct intel_rps_client
*to_rps_client(struct drm_file
*file
)
1579 struct drm_i915_file_private
*fpriv
= file
->driver_priv
;
1584 * Called when user space prepares to use an object with the CPU, either
1585 * through the mmap ioctl's mapping or a GTT mapping.
1588 i915_gem_set_domain_ioctl(struct drm_device
*dev
, void *data
,
1589 struct drm_file
*file
)
1591 struct drm_i915_gem_set_domain
*args
= data
;
1592 struct drm_i915_gem_object
*obj
;
1593 uint32_t read_domains
= args
->read_domains
;
1594 uint32_t write_domain
= args
->write_domain
;
1597 /* Only handle setting domains to types used by the CPU. */
1598 if (write_domain
& I915_GEM_GPU_DOMAINS
)
1601 if (read_domains
& I915_GEM_GPU_DOMAINS
)
1604 /* Having something in the write domain implies it's in the read
1605 * domain, and only that read domain. Enforce that in the request.
1607 if (write_domain
!= 0 && read_domains
!= write_domain
)
1610 ret
= i915_mutex_lock_interruptible(dev
);
1614 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
1615 if (&obj
->base
== NULL
) {
1620 /* Try to flush the object off the GPU without holding the lock.
1621 * We will repeat the flush holding the lock in the normal manner
1622 * to catch cases where we are gazumped.
1624 ret
= i915_gem_object_wait_rendering__nonblocking(obj
,
1625 to_rps_client(file
),
1630 if (read_domains
& I915_GEM_DOMAIN_GTT
)
1631 ret
= i915_gem_object_set_to_gtt_domain(obj
, write_domain
!= 0);
1633 ret
= i915_gem_object_set_to_cpu_domain(obj
, write_domain
!= 0);
1635 if (write_domain
!= 0)
1636 intel_fb_obj_invalidate(obj
,
1637 write_domain
== I915_GEM_DOMAIN_GTT
?
1638 ORIGIN_GTT
: ORIGIN_CPU
);
1641 drm_gem_object_unreference(&obj
->base
);
1643 mutex_unlock(&dev
->struct_mutex
);
1648 * Called when user space has done writes to this buffer
1651 i915_gem_sw_finish_ioctl(struct drm_device
*dev
, void *data
,
1652 struct drm_file
*file
)
1654 struct drm_i915_gem_sw_finish
*args
= data
;
1655 struct drm_i915_gem_object
*obj
;
1658 ret
= i915_mutex_lock_interruptible(dev
);
1662 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
1663 if (&obj
->base
== NULL
) {
1668 /* Pinned buffers may be scanout, so flush the cache */
1669 if (obj
->pin_display
)
1670 i915_gem_object_flush_cpu_write_domain(obj
);
1672 drm_gem_object_unreference(&obj
->base
);
1674 mutex_unlock(&dev
->struct_mutex
);
1679 * Maps the contents of an object, returning the address it is mapped
1682 * While the mapping holds a reference on the contents of the object, it doesn't
1683 * imply a ref on the object itself.
1687 * DRM driver writers who look a this function as an example for how to do GEM
1688 * mmap support, please don't implement mmap support like here. The modern way
1689 * to implement DRM mmap support is with an mmap offset ioctl (like
1690 * i915_gem_mmap_gtt) and then using the mmap syscall on the DRM fd directly.
1691 * That way debug tooling like valgrind will understand what's going on, hiding
1692 * the mmap call in a driver private ioctl will break that. The i915 driver only
1693 * does cpu mmaps this way because we didn't know better.
1696 i915_gem_mmap_ioctl(struct drm_device
*dev
, void *data
,
1697 struct drm_file
*file
)
1699 struct drm_i915_gem_mmap
*args
= data
;
1700 struct drm_gem_object
*obj
;
1703 if (args
->flags
& ~(I915_MMAP_WC
))
1706 if (args
->flags
& I915_MMAP_WC
&& !cpu_has_pat
)
1709 obj
= drm_gem_object_lookup(dev
, file
, args
->handle
);
1713 /* prime objects have no backing filp to GEM mmap
1717 drm_gem_object_unreference_unlocked(obj
);
1721 addr
= vm_mmap(obj
->filp
, 0, args
->size
,
1722 PROT_READ
| PROT_WRITE
, MAP_SHARED
,
1724 if (args
->flags
& I915_MMAP_WC
) {
1725 struct mm_struct
*mm
= current
->mm
;
1726 struct vm_area_struct
*vma
;
1728 down_write(&mm
->mmap_sem
);
1729 vma
= find_vma(mm
, addr
);
1732 pgprot_writecombine(vm_get_page_prot(vma
->vm_flags
));
1735 up_write(&mm
->mmap_sem
);
1737 drm_gem_object_unreference_unlocked(obj
);
1738 if (IS_ERR((void *)addr
))
1741 args
->addr_ptr
= (uint64_t) addr
;
1747 * i915_gem_fault - fault a page into the GTT
1748 * @vma: VMA in question
1751 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1752 * from userspace. The fault handler takes care of binding the object to
1753 * the GTT (if needed), allocating and programming a fence register (again,
1754 * only if needed based on whether the old reg is still valid or the object
1755 * is tiled) and inserting a new PTE into the faulting process.
1757 * Note that the faulting process may involve evicting existing objects
1758 * from the GTT and/or fence registers to make room. So performance may
1759 * suffer if the GTT working set is large or there are few fence registers
1762 int i915_gem_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1764 struct drm_i915_gem_object
*obj
= to_intel_bo(vma
->vm_private_data
);
1765 struct drm_device
*dev
= obj
->base
.dev
;
1766 struct drm_i915_private
*dev_priv
= to_i915(dev
);
1767 struct i915_ggtt
*ggtt
= &dev_priv
->ggtt
;
1768 struct i915_ggtt_view view
= i915_ggtt_view_normal
;
1769 pgoff_t page_offset
;
1772 bool write
= !!(vmf
->flags
& FAULT_FLAG_WRITE
);
1774 intel_runtime_pm_get(dev_priv
);
1776 /* We don't use vmf->pgoff since that has the fake offset */
1777 page_offset
= ((unsigned long)vmf
->virtual_address
- vma
->vm_start
) >>
1780 ret
= i915_mutex_lock_interruptible(dev
);
1784 trace_i915_gem_object_fault(obj
, page_offset
, true, write
);
1786 /* Try to flush the object off the GPU first without holding the lock.
1787 * Upon reacquiring the lock, we will perform our sanity checks and then
1788 * repeat the flush holding the lock in the normal manner to catch cases
1789 * where we are gazumped.
1791 ret
= i915_gem_object_wait_rendering__nonblocking(obj
, NULL
, !write
);
1795 /* Access to snoopable pages through the GTT is incoherent. */
1796 if (obj
->cache_level
!= I915_CACHE_NONE
&& !HAS_LLC(dev
)) {
1801 /* Use a partial view if the object is bigger than the aperture. */
1802 if (obj
->base
.size
>= ggtt
->mappable_end
&&
1803 obj
->tiling_mode
== I915_TILING_NONE
) {
1804 static const unsigned int chunk_size
= 256; // 1 MiB
1806 memset(&view
, 0, sizeof(view
));
1807 view
.type
= I915_GGTT_VIEW_PARTIAL
;
1808 view
.params
.partial
.offset
= rounddown(page_offset
, chunk_size
);
1809 view
.params
.partial
.size
=
1812 (vma
->vm_end
- vma
->vm_start
)/PAGE_SIZE
-
1813 view
.params
.partial
.offset
);
1816 /* Now pin it into the GTT if needed */
1817 ret
= i915_gem_object_ggtt_pin(obj
, &view
, 0, PIN_MAPPABLE
);
1821 ret
= i915_gem_object_set_to_gtt_domain(obj
, write
);
1825 ret
= i915_gem_object_get_fence(obj
);
1829 /* Finally, remap it using the new GTT offset */
1830 pfn
= ggtt
->mappable_base
+
1831 i915_gem_obj_ggtt_offset_view(obj
, &view
);
1834 if (unlikely(view
.type
== I915_GGTT_VIEW_PARTIAL
)) {
1835 /* Overriding existing pages in partial view does not cause
1836 * us any trouble as TLBs are still valid because the fault
1837 * is due to userspace losing part of the mapping or never
1838 * having accessed it before (at this partials' range).
1840 unsigned long base
= vma
->vm_start
+
1841 (view
.params
.partial
.offset
<< PAGE_SHIFT
);
1844 for (i
= 0; i
< view
.params
.partial
.size
; i
++) {
1845 ret
= vm_insert_pfn(vma
, base
+ i
* PAGE_SIZE
, pfn
+ i
);
1850 obj
->fault_mappable
= true;
1852 if (!obj
->fault_mappable
) {
1853 unsigned long size
= min_t(unsigned long,
1854 vma
->vm_end
- vma
->vm_start
,
1858 for (i
= 0; i
< size
>> PAGE_SHIFT
; i
++) {
1859 ret
= vm_insert_pfn(vma
,
1860 (unsigned long)vma
->vm_start
+ i
* PAGE_SIZE
,
1866 obj
->fault_mappable
= true;
1868 ret
= vm_insert_pfn(vma
,
1869 (unsigned long)vmf
->virtual_address
,
1873 i915_gem_object_ggtt_unpin_view(obj
, &view
);
1875 mutex_unlock(&dev
->struct_mutex
);
1880 * We eat errors when the gpu is terminally wedged to avoid
1881 * userspace unduly crashing (gl has no provisions for mmaps to
1882 * fail). But any other -EIO isn't ours (e.g. swap in failure)
1883 * and so needs to be reported.
1885 if (!i915_terminally_wedged(&dev_priv
->gpu_error
)) {
1886 ret
= VM_FAULT_SIGBUS
;
1891 * EAGAIN means the gpu is hung and we'll wait for the error
1892 * handler to reset everything when re-faulting in
1893 * i915_mutex_lock_interruptible.
1900 * EBUSY is ok: this just means that another thread
1901 * already did the job.
1903 ret
= VM_FAULT_NOPAGE
;
1910 ret
= VM_FAULT_SIGBUS
;
1913 WARN_ONCE(ret
, "unhandled error in i915_gem_fault: %i\n", ret
);
1914 ret
= VM_FAULT_SIGBUS
;
1918 intel_runtime_pm_put(dev_priv
);
1923 * i915_gem_release_mmap - remove physical page mappings
1924 * @obj: obj in question
1926 * Preserve the reservation of the mmapping with the DRM core code, but
1927 * relinquish ownership of the pages back to the system.
1929 * It is vital that we remove the page mapping if we have mapped a tiled
1930 * object through the GTT and then lose the fence register due to
1931 * resource pressure. Similarly if the object has been moved out of the
1932 * aperture, than pages mapped into userspace must be revoked. Removing the
1933 * mapping will then trigger a page fault on the next user access, allowing
1934 * fixup by i915_gem_fault().
1937 i915_gem_release_mmap(struct drm_i915_gem_object
*obj
)
1939 /* Serialisation between user GTT access and our code depends upon
1940 * revoking the CPU's PTE whilst the mutex is held. The next user
1941 * pagefault then has to wait until we release the mutex.
1943 lockdep_assert_held(&obj
->base
.dev
->struct_mutex
);
1945 if (!obj
->fault_mappable
)
1948 drm_vma_node_unmap(&obj
->base
.vma_node
,
1949 obj
->base
.dev
->anon_inode
->i_mapping
);
1951 /* Ensure that the CPU's PTE are revoked and there are not outstanding
1952 * memory transactions from userspace before we return. The TLB
1953 * flushing implied above by changing the PTE above *should* be
1954 * sufficient, an extra barrier here just provides us with a bit
1955 * of paranoid documentation about our requirement to serialise
1956 * memory writes before touching registers / GSM.
1960 obj
->fault_mappable
= false;
1964 i915_gem_release_all_mmaps(struct drm_i915_private
*dev_priv
)
1966 struct drm_i915_gem_object
*obj
;
1968 list_for_each_entry(obj
, &dev_priv
->mm
.bound_list
, global_list
)
1969 i915_gem_release_mmap(obj
);
1973 i915_gem_get_gtt_size(struct drm_device
*dev
, uint32_t size
, int tiling_mode
)
1977 if (INTEL_INFO(dev
)->gen
>= 4 ||
1978 tiling_mode
== I915_TILING_NONE
)
1981 /* Previous chips need a power-of-two fence region when tiling */
1982 if (INTEL_INFO(dev
)->gen
== 3)
1983 gtt_size
= 1024*1024;
1985 gtt_size
= 512*1024;
1987 while (gtt_size
< size
)
1994 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1995 * @obj: object to check
1997 * Return the required GTT alignment for an object, taking into account
1998 * potential fence register mapping.
2001 i915_gem_get_gtt_alignment(struct drm_device
*dev
, uint32_t size
,
2002 int tiling_mode
, bool fenced
)
2005 * Minimum alignment is 4k (GTT page size), but might be greater
2006 * if a fence register is needed for the object.
2008 if (INTEL_INFO(dev
)->gen
>= 4 || (!fenced
&& IS_G33(dev
)) ||
2009 tiling_mode
== I915_TILING_NONE
)
2013 * Previous chips need to be aligned to the size of the smallest
2014 * fence register that can contain the object.
2016 return i915_gem_get_gtt_size(dev
, size
, tiling_mode
);
2019 static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object
*obj
)
2021 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
2024 if (drm_vma_node_has_offset(&obj
->base
.vma_node
))
2027 dev_priv
->mm
.shrinker_no_lock_stealing
= true;
2029 ret
= drm_gem_create_mmap_offset(&obj
->base
);
2033 /* Badly fragmented mmap space? The only way we can recover
2034 * space is by destroying unwanted objects. We can't randomly release
2035 * mmap_offsets as userspace expects them to be persistent for the
2036 * lifetime of the objects. The closest we can is to release the
2037 * offsets on purgeable objects by truncating it and marking it purged,
2038 * which prevents userspace from ever using that object again.
2040 i915_gem_shrink(dev_priv
,
2041 obj
->base
.size
>> PAGE_SHIFT
,
2043 I915_SHRINK_UNBOUND
|
2044 I915_SHRINK_PURGEABLE
);
2045 ret
= drm_gem_create_mmap_offset(&obj
->base
);
2049 i915_gem_shrink_all(dev_priv
);
2050 ret
= drm_gem_create_mmap_offset(&obj
->base
);
2052 dev_priv
->mm
.shrinker_no_lock_stealing
= false;
2057 static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object
*obj
)
2059 drm_gem_free_mmap_offset(&obj
->base
);
2063 i915_gem_mmap_gtt(struct drm_file
*file
,
2064 struct drm_device
*dev
,
2068 struct drm_i915_gem_object
*obj
;
2071 ret
= i915_mutex_lock_interruptible(dev
);
2075 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, handle
));
2076 if (&obj
->base
== NULL
) {
2081 if (obj
->madv
!= I915_MADV_WILLNEED
) {
2082 DRM_DEBUG("Attempting to mmap a purgeable buffer\n");
2087 ret
= i915_gem_object_create_mmap_offset(obj
);
2091 *offset
= drm_vma_node_offset_addr(&obj
->base
.vma_node
);
2094 drm_gem_object_unreference(&obj
->base
);
2096 mutex_unlock(&dev
->struct_mutex
);
2101 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
2103 * @data: GTT mapping ioctl data
2104 * @file: GEM object info
2106 * Simply returns the fake offset to userspace so it can mmap it.
2107 * The mmap call will end up in drm_gem_mmap(), which will set things
2108 * up so we can get faults in the handler above.
2110 * The fault handler will take care of binding the object into the GTT
2111 * (since it may have been evicted to make room for something), allocating
2112 * a fence register, and mapping the appropriate aperture address into
2116 i915_gem_mmap_gtt_ioctl(struct drm_device
*dev
, void *data
,
2117 struct drm_file
*file
)
2119 struct drm_i915_gem_mmap_gtt
*args
= data
;
2121 return i915_gem_mmap_gtt(file
, dev
, args
->handle
, &args
->offset
);
2124 /* Immediately discard the backing storage */
2126 i915_gem_object_truncate(struct drm_i915_gem_object
*obj
)
2128 i915_gem_object_free_mmap_offset(obj
);
2130 if (obj
->base
.filp
== NULL
)
2133 /* Our goal here is to return as much of the memory as
2134 * is possible back to the system as we are called from OOM.
2135 * To do this we must instruct the shmfs to drop all of its
2136 * backing pages, *now*.
2138 shmem_truncate_range(file_inode(obj
->base
.filp
), 0, (loff_t
)-1);
2139 obj
->madv
= __I915_MADV_PURGED
;
2142 /* Try to discard unwanted pages */
2144 i915_gem_object_invalidate(struct drm_i915_gem_object
*obj
)
2146 struct address_space
*mapping
;
2148 switch (obj
->madv
) {
2149 case I915_MADV_DONTNEED
:
2150 i915_gem_object_truncate(obj
);
2151 case __I915_MADV_PURGED
:
2155 if (obj
->base
.filp
== NULL
)
2158 mapping
= file_inode(obj
->base
.filp
)->i_mapping
,
2159 invalidate_mapping_pages(mapping
, 0, (loff_t
)-1);
2163 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object
*obj
)
2165 struct sg_page_iter sg_iter
;
2168 BUG_ON(obj
->madv
== __I915_MADV_PURGED
);
2170 ret
= i915_gem_object_set_to_cpu_domain(obj
, true);
2172 /* In the event of a disaster, abandon all caches and
2173 * hope for the best.
2175 i915_gem_clflush_object(obj
, true);
2176 obj
->base
.read_domains
= obj
->base
.write_domain
= I915_GEM_DOMAIN_CPU
;
2179 i915_gem_gtt_finish_object(obj
);
2181 if (i915_gem_object_needs_bit17_swizzle(obj
))
2182 i915_gem_object_save_bit_17_swizzle(obj
);
2184 if (obj
->madv
== I915_MADV_DONTNEED
)
2187 for_each_sg_page(obj
->pages
->sgl
, &sg_iter
, obj
->pages
->nents
, 0) {
2188 struct page
*page
= sg_page_iter_page(&sg_iter
);
2191 set_page_dirty(page
);
2193 if (obj
->madv
== I915_MADV_WILLNEED
)
2194 mark_page_accessed(page
);
2200 sg_free_table(obj
->pages
);
2205 i915_gem_object_put_pages(struct drm_i915_gem_object
*obj
)
2207 const struct drm_i915_gem_object_ops
*ops
= obj
->ops
;
2209 if (obj
->pages
== NULL
)
2212 if (obj
->pages_pin_count
)
2215 BUG_ON(i915_gem_obj_bound_any(obj
));
2217 /* ->put_pages might need to allocate memory for the bit17 swizzle
2218 * array, hence protect them from being reaped by removing them from gtt
2220 list_del(&obj
->global_list
);
2223 if (is_vmalloc_addr(obj
->mapping
))
2224 vunmap(obj
->mapping
);
2226 kunmap(kmap_to_page(obj
->mapping
));
2227 obj
->mapping
= NULL
;
2230 ops
->put_pages(obj
);
2233 i915_gem_object_invalidate(obj
);
2239 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object
*obj
)
2241 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
2243 struct address_space
*mapping
;
2244 struct sg_table
*st
;
2245 struct scatterlist
*sg
;
2246 struct sg_page_iter sg_iter
;
2248 unsigned long last_pfn
= 0; /* suppress gcc warning */
2252 /* Assert that the object is not currently in any GPU domain. As it
2253 * wasn't in the GTT, there shouldn't be any way it could have been in
2256 BUG_ON(obj
->base
.read_domains
& I915_GEM_GPU_DOMAINS
);
2257 BUG_ON(obj
->base
.write_domain
& I915_GEM_GPU_DOMAINS
);
2259 st
= kmalloc(sizeof(*st
), GFP_KERNEL
);
2263 page_count
= obj
->base
.size
/ PAGE_SIZE
;
2264 if (sg_alloc_table(st
, page_count
, GFP_KERNEL
)) {
2269 /* Get the list of pages out of our struct file. They'll be pinned
2270 * at this point until we release them.
2272 * Fail silently without starting the shrinker
2274 mapping
= file_inode(obj
->base
.filp
)->i_mapping
;
2275 gfp
= mapping_gfp_constraint(mapping
, ~(__GFP_IO
| __GFP_RECLAIM
));
2276 gfp
|= __GFP_NORETRY
| __GFP_NOWARN
;
2279 for (i
= 0; i
< page_count
; i
++) {
2280 page
= shmem_read_mapping_page_gfp(mapping
, i
, gfp
);
2282 i915_gem_shrink(dev_priv
,
2285 I915_SHRINK_UNBOUND
|
2286 I915_SHRINK_PURGEABLE
);
2287 page
= shmem_read_mapping_page_gfp(mapping
, i
, gfp
);
2290 /* We've tried hard to allocate the memory by reaping
2291 * our own buffer, now let the real VM do its job and
2292 * go down in flames if truly OOM.
2294 i915_gem_shrink_all(dev_priv
);
2295 page
= shmem_read_mapping_page(mapping
, i
);
2297 ret
= PTR_ERR(page
);
2301 #ifdef CONFIG_SWIOTLB
2302 if (swiotlb_nr_tbl()) {
2304 sg_set_page(sg
, page
, PAGE_SIZE
, 0);
2309 if (!i
|| page_to_pfn(page
) != last_pfn
+ 1) {
2313 sg_set_page(sg
, page
, PAGE_SIZE
, 0);
2315 sg
->length
+= PAGE_SIZE
;
2317 last_pfn
= page_to_pfn(page
);
2319 /* Check that the i965g/gm workaround works. */
2320 WARN_ON((gfp
& __GFP_DMA32
) && (last_pfn
>= 0x00100000UL
));
2322 #ifdef CONFIG_SWIOTLB
2323 if (!swiotlb_nr_tbl())
2328 ret
= i915_gem_gtt_prepare_object(obj
);
2332 if (i915_gem_object_needs_bit17_swizzle(obj
))
2333 i915_gem_object_do_bit_17_swizzle(obj
);
2335 if (obj
->tiling_mode
!= I915_TILING_NONE
&&
2336 dev_priv
->quirks
& QUIRK_PIN_SWIZZLED_PAGES
)
2337 i915_gem_object_pin_pages(obj
);
2343 for_each_sg_page(st
->sgl
, &sg_iter
, st
->nents
, 0)
2344 put_page(sg_page_iter_page(&sg_iter
));
2348 /* shmemfs first checks if there is enough memory to allocate the page
2349 * and reports ENOSPC should there be insufficient, along with the usual
2350 * ENOMEM for a genuine allocation failure.
2352 * We use ENOSPC in our driver to mean that we have run out of aperture
2353 * space and so want to translate the error from shmemfs back to our
2354 * usual understanding of ENOMEM.
2362 /* Ensure that the associated pages are gathered from the backing storage
2363 * and pinned into our object. i915_gem_object_get_pages() may be called
2364 * multiple times before they are released by a single call to
2365 * i915_gem_object_put_pages() - once the pages are no longer referenced
2366 * either as a result of memory pressure (reaping pages under the shrinker)
2367 * or as the object is itself released.
2370 i915_gem_object_get_pages(struct drm_i915_gem_object
*obj
)
2372 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
2373 const struct drm_i915_gem_object_ops
*ops
= obj
->ops
;
2379 if (obj
->madv
!= I915_MADV_WILLNEED
) {
2380 DRM_DEBUG("Attempting to obtain a purgeable object\n");
2384 BUG_ON(obj
->pages_pin_count
);
2386 ret
= ops
->get_pages(obj
);
2390 list_add_tail(&obj
->global_list
, &dev_priv
->mm
.unbound_list
);
2392 obj
->get_page
.sg
= obj
->pages
->sgl
;
2393 obj
->get_page
.last
= 0;
2398 void *i915_gem_object_pin_map(struct drm_i915_gem_object
*obj
)
2402 lockdep_assert_held(&obj
->base
.dev
->struct_mutex
);
2404 ret
= i915_gem_object_get_pages(obj
);
2406 return ERR_PTR(ret
);
2408 i915_gem_object_pin_pages(obj
);
2410 if (obj
->mapping
== NULL
) {
2411 struct page
**pages
;
2414 if (obj
->base
.size
== PAGE_SIZE
)
2415 obj
->mapping
= kmap(sg_page(obj
->pages
->sgl
));
2417 pages
= drm_malloc_gfp(obj
->base
.size
>> PAGE_SHIFT
,
2420 if (pages
!= NULL
) {
2421 struct sg_page_iter sg_iter
;
2425 for_each_sg_page(obj
->pages
->sgl
, &sg_iter
,
2426 obj
->pages
->nents
, 0)
2427 pages
[n
++] = sg_page_iter_page(&sg_iter
);
2429 obj
->mapping
= vmap(pages
, n
, 0, PAGE_KERNEL
);
2430 drm_free_large(pages
);
2432 if (obj
->mapping
== NULL
) {
2433 i915_gem_object_unpin_pages(obj
);
2434 return ERR_PTR(-ENOMEM
);
2438 return obj
->mapping
;
2441 void i915_vma_move_to_active(struct i915_vma
*vma
,
2442 struct drm_i915_gem_request
*req
)
2444 struct drm_i915_gem_object
*obj
= vma
->obj
;
2445 struct intel_engine_cs
*engine
;
2447 engine
= i915_gem_request_get_engine(req
);
2449 /* Add a reference if we're newly entering the active list. */
2450 if (obj
->active
== 0)
2451 drm_gem_object_reference(&obj
->base
);
2452 obj
->active
|= intel_engine_flag(engine
);
2454 list_move_tail(&obj
->engine_list
[engine
->id
], &engine
->active_list
);
2455 i915_gem_request_assign(&obj
->last_read_req
[engine
->id
], req
);
2457 list_move_tail(&vma
->vm_link
, &vma
->vm
->active_list
);
2461 i915_gem_object_retire__write(struct drm_i915_gem_object
*obj
)
2463 GEM_BUG_ON(obj
->last_write_req
== NULL
);
2464 GEM_BUG_ON(!(obj
->active
& intel_engine_flag(obj
->last_write_req
->engine
)));
2466 i915_gem_request_assign(&obj
->last_write_req
, NULL
);
2467 intel_fb_obj_flush(obj
, true, ORIGIN_CS
);
2471 i915_gem_object_retire__read(struct drm_i915_gem_object
*obj
, int ring
)
2473 struct i915_vma
*vma
;
2475 GEM_BUG_ON(obj
->last_read_req
[ring
] == NULL
);
2476 GEM_BUG_ON(!(obj
->active
& (1 << ring
)));
2478 list_del_init(&obj
->engine_list
[ring
]);
2479 i915_gem_request_assign(&obj
->last_read_req
[ring
], NULL
);
2481 if (obj
->last_write_req
&& obj
->last_write_req
->engine
->id
== ring
)
2482 i915_gem_object_retire__write(obj
);
2484 obj
->active
&= ~(1 << ring
);
2488 /* Bump our place on the bound list to keep it roughly in LRU order
2489 * so that we don't steal from recently used but inactive objects
2490 * (unless we are forced to ofc!)
2492 list_move_tail(&obj
->global_list
,
2493 &to_i915(obj
->base
.dev
)->mm
.bound_list
);
2495 list_for_each_entry(vma
, &obj
->vma_list
, obj_link
) {
2496 if (!list_empty(&vma
->vm_link
))
2497 list_move_tail(&vma
->vm_link
, &vma
->vm
->inactive_list
);
2500 i915_gem_request_assign(&obj
->last_fenced_req
, NULL
);
2501 drm_gem_object_unreference(&obj
->base
);
2505 i915_gem_init_seqno(struct drm_device
*dev
, u32 seqno
)
2507 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2508 struct intel_engine_cs
*engine
;
2511 /* Carefully retire all requests without writing to the rings */
2512 for_each_engine(engine
, dev_priv
) {
2513 ret
= intel_engine_idle(engine
);
2517 i915_gem_retire_requests(dev
);
2519 /* Finally reset hw state */
2520 for_each_engine(engine
, dev_priv
)
2521 intel_ring_init_seqno(engine
, seqno
);
2526 int i915_gem_set_seqno(struct drm_device
*dev
, u32 seqno
)
2528 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2534 /* HWS page needs to be set less than what we
2535 * will inject to ring
2537 ret
= i915_gem_init_seqno(dev
, seqno
- 1);
2541 /* Carefully set the last_seqno value so that wrap
2542 * detection still works
2544 dev_priv
->next_seqno
= seqno
;
2545 dev_priv
->last_seqno
= seqno
- 1;
2546 if (dev_priv
->last_seqno
== 0)
2547 dev_priv
->last_seqno
--;
2553 i915_gem_get_seqno(struct drm_device
*dev
, u32
*seqno
)
2555 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2557 /* reserve 0 for non-seqno */
2558 if (dev_priv
->next_seqno
== 0) {
2559 int ret
= i915_gem_init_seqno(dev
, 0);
2563 dev_priv
->next_seqno
= 1;
2566 *seqno
= dev_priv
->last_seqno
= dev_priv
->next_seqno
++;
2571 * NB: This function is not allowed to fail. Doing so would mean the the
2572 * request is not being tracked for completion but the work itself is
2573 * going to happen on the hardware. This would be a Bad Thing(tm).
2575 void __i915_add_request(struct drm_i915_gem_request
*request
,
2576 struct drm_i915_gem_object
*obj
,
2579 struct intel_engine_cs
*engine
;
2580 struct drm_i915_private
*dev_priv
;
2581 struct intel_ringbuffer
*ringbuf
;
2585 if (WARN_ON(request
== NULL
))
2588 engine
= request
->engine
;
2589 dev_priv
= request
->i915
;
2590 ringbuf
= request
->ringbuf
;
2593 * To ensure that this call will not fail, space for its emissions
2594 * should already have been reserved in the ring buffer. Let the ring
2595 * know that it is time to use that space up.
2597 intel_ring_reserved_space_use(ringbuf
);
2599 request_start
= intel_ring_get_tail(ringbuf
);
2601 * Emit any outstanding flushes - execbuf can fail to emit the flush
2602 * after having emitted the batchbuffer command. Hence we need to fix
2603 * things up similar to emitting the lazy request. The difference here
2604 * is that the flush _must_ happen before the next request, no matter
2608 if (i915
.enable_execlists
)
2609 ret
= logical_ring_flush_all_caches(request
);
2611 ret
= intel_ring_flush_all_caches(request
);
2612 /* Not allowed to fail! */
2613 WARN(ret
, "*_ring_flush_all_caches failed: %d!\n", ret
);
2616 trace_i915_gem_request_add(request
);
2618 request
->head
= request_start
;
2620 /* Whilst this request exists, batch_obj will be on the
2621 * active_list, and so will hold the active reference. Only when this
2622 * request is retired will the the batch_obj be moved onto the
2623 * inactive_list and lose its active reference. Hence we do not need
2624 * to explicitly hold another reference here.
2626 request
->batch_obj
= obj
;
2628 /* Seal the request and mark it as pending execution. Note that
2629 * we may inspect this state, without holding any locks, during
2630 * hangcheck. Hence we apply the barrier to ensure that we do not
2631 * see a more recent value in the hws than we are tracking.
2633 request
->emitted_jiffies
= jiffies
;
2634 request
->previous_seqno
= engine
->last_submitted_seqno
;
2635 smp_store_mb(engine
->last_submitted_seqno
, request
->seqno
);
2636 list_add_tail(&request
->list
, &engine
->request_list
);
2638 /* Record the position of the start of the request so that
2639 * should we detect the updated seqno part-way through the
2640 * GPU processing the request, we never over-estimate the
2641 * position of the head.
2643 request
->postfix
= intel_ring_get_tail(ringbuf
);
2645 if (i915
.enable_execlists
)
2646 ret
= engine
->emit_request(request
);
2648 ret
= engine
->add_request(request
);
2650 request
->tail
= intel_ring_get_tail(ringbuf
);
2652 /* Not allowed to fail! */
2653 WARN(ret
, "emit|add_request failed: %d!\n", ret
);
2655 i915_queue_hangcheck(engine
->dev
);
2657 queue_delayed_work(dev_priv
->wq
,
2658 &dev_priv
->mm
.retire_work
,
2659 round_jiffies_up_relative(HZ
));
2660 intel_mark_busy(dev_priv
->dev
);
2662 /* Sanity check that the reserved size was large enough. */
2663 intel_ring_reserved_space_end(ringbuf
);
2666 static bool i915_context_is_banned(struct drm_i915_private
*dev_priv
,
2667 const struct intel_context
*ctx
)
2669 unsigned long elapsed
;
2671 elapsed
= get_seconds() - ctx
->hang_stats
.guilty_ts
;
2673 if (ctx
->hang_stats
.banned
)
2676 if (ctx
->hang_stats
.ban_period_seconds
&&
2677 elapsed
<= ctx
->hang_stats
.ban_period_seconds
) {
2678 if (!i915_gem_context_is_default(ctx
)) {
2679 DRM_DEBUG("context hanging too fast, banning!\n");
2681 } else if (i915_stop_ring_allow_ban(dev_priv
)) {
2682 if (i915_stop_ring_allow_warn(dev_priv
))
2683 DRM_ERROR("gpu hanging too fast, banning!\n");
2691 static void i915_set_reset_status(struct drm_i915_private
*dev_priv
,
2692 struct intel_context
*ctx
,
2695 struct i915_ctx_hang_stats
*hs
;
2700 hs
= &ctx
->hang_stats
;
2703 hs
->banned
= i915_context_is_banned(dev_priv
, ctx
);
2705 hs
->guilty_ts
= get_seconds();
2707 hs
->batch_pending
++;
2711 void i915_gem_request_free(struct kref
*req_ref
)
2713 struct drm_i915_gem_request
*req
= container_of(req_ref
,
2715 struct intel_context
*ctx
= req
->ctx
;
2718 i915_gem_request_remove_from_client(req
);
2721 if (i915
.enable_execlists
&& ctx
!= req
->i915
->kernel_context
)
2722 intel_lr_context_unpin(ctx
, req
->engine
);
2724 i915_gem_context_unreference(ctx
);
2727 kmem_cache_free(req
->i915
->requests
, req
);
2731 __i915_gem_request_alloc(struct intel_engine_cs
*engine
,
2732 struct intel_context
*ctx
,
2733 struct drm_i915_gem_request
**req_out
)
2735 struct drm_i915_private
*dev_priv
= to_i915(engine
->dev
);
2736 unsigned reset_counter
= i915_reset_counter(&dev_priv
->gpu_error
);
2737 struct drm_i915_gem_request
*req
;
2745 /* ABI: Before userspace accesses the GPU (e.g. execbuffer), report
2746 * EIO if the GPU is already wedged, or EAGAIN to drop the struct_mutex
2749 ret
= i915_gem_check_wedge(reset_counter
, dev_priv
->mm
.interruptible
);
2753 req
= kmem_cache_zalloc(dev_priv
->requests
, GFP_KERNEL
);
2757 ret
= i915_gem_get_seqno(engine
->dev
, &req
->seqno
);
2761 kref_init(&req
->ref
);
2762 req
->i915
= dev_priv
;
2763 req
->engine
= engine
;
2764 req
->reset_counter
= reset_counter
;
2766 i915_gem_context_reference(req
->ctx
);
2768 if (i915
.enable_execlists
)
2769 ret
= intel_logical_ring_alloc_request_extras(req
);
2771 ret
= intel_ring_alloc_request_extras(req
);
2773 i915_gem_context_unreference(req
->ctx
);
2778 * Reserve space in the ring buffer for all the commands required to
2779 * eventually emit this request. This is to guarantee that the
2780 * i915_add_request() call can't fail. Note that the reserve may need
2781 * to be redone if the request is not actually submitted straight
2782 * away, e.g. because a GPU scheduler has deferred it.
2784 if (i915
.enable_execlists
)
2785 ret
= intel_logical_ring_reserve_space(req
);
2787 ret
= intel_ring_reserve_space(req
);
2790 * At this point, the request is fully allocated even if not
2791 * fully prepared. Thus it can be cleaned up using the proper
2794 intel_ring_reserved_space_cancel(req
->ringbuf
);
2795 i915_gem_request_unreference(req
);
2803 kmem_cache_free(dev_priv
->requests
, req
);
2808 * i915_gem_request_alloc - allocate a request structure
2810 * @engine: engine that we wish to issue the request on.
2811 * @ctx: context that the request will be associated with.
2812 * This can be NULL if the request is not directly related to
2813 * any specific user context, in which case this function will
2814 * choose an appropriate context to use.
2816 * Returns a pointer to the allocated request if successful,
2817 * or an error code if not.
2819 struct drm_i915_gem_request
*
2820 i915_gem_request_alloc(struct intel_engine_cs
*engine
,
2821 struct intel_context
*ctx
)
2823 struct drm_i915_gem_request
*req
;
2827 ctx
= to_i915(engine
->dev
)->kernel_context
;
2828 err
= __i915_gem_request_alloc(engine
, ctx
, &req
);
2829 return err
? ERR_PTR(err
) : req
;
2832 struct drm_i915_gem_request
*
2833 i915_gem_find_active_request(struct intel_engine_cs
*engine
)
2835 struct drm_i915_gem_request
*request
;
2837 list_for_each_entry(request
, &engine
->request_list
, list
) {
2838 if (i915_gem_request_completed(request
, false))
2847 static void i915_gem_reset_engine_status(struct drm_i915_private
*dev_priv
,
2848 struct intel_engine_cs
*engine
)
2850 struct drm_i915_gem_request
*request
;
2853 request
= i915_gem_find_active_request(engine
);
2855 if (request
== NULL
)
2858 ring_hung
= engine
->hangcheck
.score
>= HANGCHECK_SCORE_RING_HUNG
;
2860 i915_set_reset_status(dev_priv
, request
->ctx
, ring_hung
);
2862 list_for_each_entry_continue(request
, &engine
->request_list
, list
)
2863 i915_set_reset_status(dev_priv
, request
->ctx
, false);
2866 static void i915_gem_reset_engine_cleanup(struct drm_i915_private
*dev_priv
,
2867 struct intel_engine_cs
*engine
)
2869 struct intel_ringbuffer
*buffer
;
2871 while (!list_empty(&engine
->active_list
)) {
2872 struct drm_i915_gem_object
*obj
;
2874 obj
= list_first_entry(&engine
->active_list
,
2875 struct drm_i915_gem_object
,
2876 engine_list
[engine
->id
]);
2878 i915_gem_object_retire__read(obj
, engine
->id
);
2882 * Clear the execlists queue up before freeing the requests, as those
2883 * are the ones that keep the context and ringbuffer backing objects
2887 if (i915
.enable_execlists
) {
2888 /* Ensure irq handler finishes or is cancelled. */
2889 tasklet_kill(&engine
->irq_tasklet
);
2891 spin_lock_bh(&engine
->execlist_lock
);
2892 /* list_splice_tail_init checks for empty lists */
2893 list_splice_tail_init(&engine
->execlist_queue
,
2894 &engine
->execlist_retired_req_list
);
2895 spin_unlock_bh(&engine
->execlist_lock
);
2897 intel_execlists_retire_requests(engine
);
2901 * We must free the requests after all the corresponding objects have
2902 * been moved off active lists. Which is the same order as the normal
2903 * retire_requests function does. This is important if object hold
2904 * implicit references on things like e.g. ppgtt address spaces through
2907 while (!list_empty(&engine
->request_list
)) {
2908 struct drm_i915_gem_request
*request
;
2910 request
= list_first_entry(&engine
->request_list
,
2911 struct drm_i915_gem_request
,
2914 i915_gem_request_retire(request
);
2917 /* Having flushed all requests from all queues, we know that all
2918 * ringbuffers must now be empty. However, since we do not reclaim
2919 * all space when retiring the request (to prevent HEADs colliding
2920 * with rapid ringbuffer wraparound) the amount of available space
2921 * upon reset is less than when we start. Do one more pass over
2922 * all the ringbuffers to reset last_retired_head.
2924 list_for_each_entry(buffer
, &engine
->buffers
, link
) {
2925 buffer
->last_retired_head
= buffer
->tail
;
2926 intel_ring_update_space(buffer
);
2929 intel_ring_init_seqno(engine
, engine
->last_submitted_seqno
);
2932 void i915_gem_reset(struct drm_device
*dev
)
2934 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2935 struct intel_engine_cs
*engine
;
2938 * Before we free the objects from the requests, we need to inspect
2939 * them for finding the guilty party. As the requests only borrow
2940 * their reference to the objects, the inspection must be done first.
2942 for_each_engine(engine
, dev_priv
)
2943 i915_gem_reset_engine_status(dev_priv
, engine
);
2945 for_each_engine(engine
, dev_priv
)
2946 i915_gem_reset_engine_cleanup(dev_priv
, engine
);
2948 i915_gem_context_reset(dev
);
2950 i915_gem_restore_fences(dev
);
2952 WARN_ON(i915_verify_lists(dev
));
2956 * This function clears the request list as sequence numbers are passed.
2959 i915_gem_retire_requests_ring(struct intel_engine_cs
*engine
)
2961 WARN_ON(i915_verify_lists(engine
->dev
));
2963 /* Retire requests first as we use it above for the early return.
2964 * If we retire requests last, we may use a later seqno and so clear
2965 * the requests lists without clearing the active list, leading to
2968 while (!list_empty(&engine
->request_list
)) {
2969 struct drm_i915_gem_request
*request
;
2971 request
= list_first_entry(&engine
->request_list
,
2972 struct drm_i915_gem_request
,
2975 if (!i915_gem_request_completed(request
, true))
2978 i915_gem_request_retire(request
);
2981 /* Move any buffers on the active list that are no longer referenced
2982 * by the ringbuffer to the flushing/inactive lists as appropriate,
2983 * before we free the context associated with the requests.
2985 while (!list_empty(&engine
->active_list
)) {
2986 struct drm_i915_gem_object
*obj
;
2988 obj
= list_first_entry(&engine
->active_list
,
2989 struct drm_i915_gem_object
,
2990 engine_list
[engine
->id
]);
2992 if (!list_empty(&obj
->last_read_req
[engine
->id
]->list
))
2995 i915_gem_object_retire__read(obj
, engine
->id
);
2998 if (unlikely(engine
->trace_irq_req
&&
2999 i915_gem_request_completed(engine
->trace_irq_req
, true))) {
3000 engine
->irq_put(engine
);
3001 i915_gem_request_assign(&engine
->trace_irq_req
, NULL
);
3004 WARN_ON(i915_verify_lists(engine
->dev
));
3008 i915_gem_retire_requests(struct drm_device
*dev
)
3010 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3011 struct intel_engine_cs
*engine
;
3014 for_each_engine(engine
, dev_priv
) {
3015 i915_gem_retire_requests_ring(engine
);
3016 idle
&= list_empty(&engine
->request_list
);
3017 if (i915
.enable_execlists
) {
3018 spin_lock_bh(&engine
->execlist_lock
);
3019 idle
&= list_empty(&engine
->execlist_queue
);
3020 spin_unlock_bh(&engine
->execlist_lock
);
3022 intel_execlists_retire_requests(engine
);
3027 mod_delayed_work(dev_priv
->wq
,
3028 &dev_priv
->mm
.idle_work
,
3029 msecs_to_jiffies(100));
3035 i915_gem_retire_work_handler(struct work_struct
*work
)
3037 struct drm_i915_private
*dev_priv
=
3038 container_of(work
, typeof(*dev_priv
), mm
.retire_work
.work
);
3039 struct drm_device
*dev
= dev_priv
->dev
;
3042 /* Come back later if the device is busy... */
3044 if (mutex_trylock(&dev
->struct_mutex
)) {
3045 idle
= i915_gem_retire_requests(dev
);
3046 mutex_unlock(&dev
->struct_mutex
);
3049 queue_delayed_work(dev_priv
->wq
, &dev_priv
->mm
.retire_work
,
3050 round_jiffies_up_relative(HZ
));
3054 i915_gem_idle_work_handler(struct work_struct
*work
)
3056 struct drm_i915_private
*dev_priv
=
3057 container_of(work
, typeof(*dev_priv
), mm
.idle_work
.work
);
3058 struct drm_device
*dev
= dev_priv
->dev
;
3059 struct intel_engine_cs
*engine
;
3061 for_each_engine(engine
, dev_priv
)
3062 if (!list_empty(&engine
->request_list
))
3065 /* we probably should sync with hangcheck here, using cancel_work_sync.
3066 * Also locking seems to be fubar here, engine->request_list is protected
3067 * by dev->struct_mutex. */
3069 intel_mark_idle(dev
);
3071 if (mutex_trylock(&dev
->struct_mutex
)) {
3072 for_each_engine(engine
, dev_priv
)
3073 i915_gem_batch_pool_fini(&engine
->batch_pool
);
3075 mutex_unlock(&dev
->struct_mutex
);
3080 * Ensures that an object will eventually get non-busy by flushing any required
3081 * write domains, emitting any outstanding lazy request and retiring and
3082 * completed requests.
3085 i915_gem_object_flush_active(struct drm_i915_gem_object
*obj
)
3092 for (i
= 0; i
< I915_NUM_ENGINES
; i
++) {
3093 struct drm_i915_gem_request
*req
;
3095 req
= obj
->last_read_req
[i
];
3099 if (list_empty(&req
->list
))
3102 if (i915_gem_request_completed(req
, true)) {
3103 __i915_gem_request_retire__upto(req
);
3105 i915_gem_object_retire__read(obj
, i
);
3113 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
3114 * @DRM_IOCTL_ARGS: standard ioctl arguments
3116 * Returns 0 if successful, else an error is returned with the remaining time in
3117 * the timeout parameter.
3118 * -ETIME: object is still busy after timeout
3119 * -ERESTARTSYS: signal interrupted the wait
3120 * -ENONENT: object doesn't exist
3121 * Also possible, but rare:
3122 * -EAGAIN: GPU wedged
3124 * -ENODEV: Internal IRQ fail
3125 * -E?: The add request failed
3127 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
3128 * non-zero timeout parameter the wait ioctl will wait for the given number of
3129 * nanoseconds on an object becoming unbusy. Since the wait itself does so
3130 * without holding struct_mutex the object may become re-busied before this
3131 * function completes. A similar but shorter * race condition exists in the busy
3135 i915_gem_wait_ioctl(struct drm_device
*dev
, void *data
, struct drm_file
*file
)
3137 struct drm_i915_gem_wait
*args
= data
;
3138 struct drm_i915_gem_object
*obj
;
3139 struct drm_i915_gem_request
*req
[I915_NUM_ENGINES
];
3143 if (args
->flags
!= 0)
3146 ret
= i915_mutex_lock_interruptible(dev
);
3150 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->bo_handle
));
3151 if (&obj
->base
== NULL
) {
3152 mutex_unlock(&dev
->struct_mutex
);
3156 /* Need to make sure the object gets inactive eventually. */
3157 ret
= i915_gem_object_flush_active(obj
);
3164 /* Do this after OLR check to make sure we make forward progress polling
3165 * on this IOCTL with a timeout == 0 (like busy ioctl)
3167 if (args
->timeout_ns
== 0) {
3172 drm_gem_object_unreference(&obj
->base
);
3174 for (i
= 0; i
< I915_NUM_ENGINES
; i
++) {
3175 if (obj
->last_read_req
[i
] == NULL
)
3178 req
[n
++] = i915_gem_request_reference(obj
->last_read_req
[i
]);
3181 mutex_unlock(&dev
->struct_mutex
);
3183 for (i
= 0; i
< n
; i
++) {
3185 ret
= __i915_wait_request(req
[i
], true,
3186 args
->timeout_ns
> 0 ? &args
->timeout_ns
: NULL
,
3187 to_rps_client(file
));
3188 i915_gem_request_unreference__unlocked(req
[i
]);
3193 drm_gem_object_unreference(&obj
->base
);
3194 mutex_unlock(&dev
->struct_mutex
);
3199 __i915_gem_object_sync(struct drm_i915_gem_object
*obj
,
3200 struct intel_engine_cs
*to
,
3201 struct drm_i915_gem_request
*from_req
,
3202 struct drm_i915_gem_request
**to_req
)
3204 struct intel_engine_cs
*from
;
3207 from
= i915_gem_request_get_engine(from_req
);
3211 if (i915_gem_request_completed(from_req
, true))
3214 if (!i915_semaphore_is_enabled(obj
->base
.dev
)) {
3215 struct drm_i915_private
*i915
= to_i915(obj
->base
.dev
);
3216 ret
= __i915_wait_request(from_req
,
3217 i915
->mm
.interruptible
,
3219 &i915
->rps
.semaphores
);
3223 i915_gem_object_retire_request(obj
, from_req
);
3225 int idx
= intel_ring_sync_index(from
, to
);
3226 u32 seqno
= i915_gem_request_get_seqno(from_req
);
3230 if (seqno
<= from
->semaphore
.sync_seqno
[idx
])
3233 if (*to_req
== NULL
) {
3234 struct drm_i915_gem_request
*req
;
3236 req
= i915_gem_request_alloc(to
, NULL
);
3238 return PTR_ERR(req
);
3243 trace_i915_gem_ring_sync_to(*to_req
, from
, from_req
);
3244 ret
= to
->semaphore
.sync_to(*to_req
, from
, seqno
);
3248 /* We use last_read_req because sync_to()
3249 * might have just caused seqno wrap under
3252 from
->semaphore
.sync_seqno
[idx
] =
3253 i915_gem_request_get_seqno(obj
->last_read_req
[from
->id
]);
3260 * i915_gem_object_sync - sync an object to a ring.
3262 * @obj: object which may be in use on another ring.
3263 * @to: ring we wish to use the object on. May be NULL.
3264 * @to_req: request we wish to use the object for. See below.
3265 * This will be allocated and returned if a request is
3266 * required but not passed in.
3268 * This code is meant to abstract object synchronization with the GPU.
3269 * Calling with NULL implies synchronizing the object with the CPU
3270 * rather than a particular GPU ring. Conceptually we serialise writes
3271 * between engines inside the GPU. We only allow one engine to write
3272 * into a buffer at any time, but multiple readers. To ensure each has
3273 * a coherent view of memory, we must:
3275 * - If there is an outstanding write request to the object, the new
3276 * request must wait for it to complete (either CPU or in hw, requests
3277 * on the same ring will be naturally ordered).
3279 * - If we are a write request (pending_write_domain is set), the new
3280 * request must wait for outstanding read requests to complete.
3282 * For CPU synchronisation (NULL to) no request is required. For syncing with
3283 * rings to_req must be non-NULL. However, a request does not have to be
3284 * pre-allocated. If *to_req is NULL and sync commands will be emitted then a
3285 * request will be allocated automatically and returned through *to_req. Note
3286 * that it is not guaranteed that commands will be emitted (because the system
3287 * might already be idle). Hence there is no need to create a request that
3288 * might never have any work submitted. Note further that if a request is
3289 * returned in *to_req, it is the responsibility of the caller to submit
3290 * that request (after potentially adding more work to it).
3292 * Returns 0 if successful, else propagates up the lower layer error.
3295 i915_gem_object_sync(struct drm_i915_gem_object
*obj
,
3296 struct intel_engine_cs
*to
,
3297 struct drm_i915_gem_request
**to_req
)
3299 const bool readonly
= obj
->base
.pending_write_domain
== 0;
3300 struct drm_i915_gem_request
*req
[I915_NUM_ENGINES
];
3307 return i915_gem_object_wait_rendering(obj
, readonly
);
3311 if (obj
->last_write_req
)
3312 req
[n
++] = obj
->last_write_req
;
3314 for (i
= 0; i
< I915_NUM_ENGINES
; i
++)
3315 if (obj
->last_read_req
[i
])
3316 req
[n
++] = obj
->last_read_req
[i
];
3318 for (i
= 0; i
< n
; i
++) {
3319 ret
= __i915_gem_object_sync(obj
, to
, req
[i
], to_req
);
3327 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object
*obj
)
3329 u32 old_write_domain
, old_read_domains
;
3331 /* Force a pagefault for domain tracking on next user access */
3332 i915_gem_release_mmap(obj
);
3334 if ((obj
->base
.read_domains
& I915_GEM_DOMAIN_GTT
) == 0)
3337 old_read_domains
= obj
->base
.read_domains
;
3338 old_write_domain
= obj
->base
.write_domain
;
3340 obj
->base
.read_domains
&= ~I915_GEM_DOMAIN_GTT
;
3341 obj
->base
.write_domain
&= ~I915_GEM_DOMAIN_GTT
;
3343 trace_i915_gem_object_change_domain(obj
,
3348 static int __i915_vma_unbind(struct i915_vma
*vma
, bool wait
)
3350 struct drm_i915_gem_object
*obj
= vma
->obj
;
3351 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
3354 if (list_empty(&vma
->obj_link
))
3357 if (!drm_mm_node_allocated(&vma
->node
)) {
3358 i915_gem_vma_destroy(vma
);
3365 BUG_ON(obj
->pages
== NULL
);
3368 ret
= i915_gem_object_wait_rendering(obj
, false);
3373 if (vma
->is_ggtt
&& vma
->ggtt_view
.type
== I915_GGTT_VIEW_NORMAL
) {
3374 i915_gem_object_finish_gtt(obj
);
3376 /* release the fence reg _after_ flushing */
3377 ret
= i915_gem_object_put_fence(obj
);
3382 trace_i915_vma_unbind(vma
);
3384 vma
->vm
->unbind_vma(vma
);
3387 list_del_init(&vma
->vm_link
);
3389 if (vma
->ggtt_view
.type
== I915_GGTT_VIEW_NORMAL
) {
3390 obj
->map_and_fenceable
= false;
3391 } else if (vma
->ggtt_view
.pages
) {
3392 sg_free_table(vma
->ggtt_view
.pages
);
3393 kfree(vma
->ggtt_view
.pages
);
3395 vma
->ggtt_view
.pages
= NULL
;
3398 drm_mm_remove_node(&vma
->node
);
3399 i915_gem_vma_destroy(vma
);
3401 /* Since the unbound list is global, only move to that list if
3402 * no more VMAs exist. */
3403 if (list_empty(&obj
->vma_list
))
3404 list_move_tail(&obj
->global_list
, &dev_priv
->mm
.unbound_list
);
3406 /* And finally now the object is completely decoupled from this vma,
3407 * we can drop its hold on the backing storage and allow it to be
3408 * reaped by the shrinker.
3410 i915_gem_object_unpin_pages(obj
);
3415 int i915_vma_unbind(struct i915_vma
*vma
)
3417 return __i915_vma_unbind(vma
, true);
3420 int __i915_vma_unbind_no_wait(struct i915_vma
*vma
)
3422 return __i915_vma_unbind(vma
, false);
3425 int i915_gpu_idle(struct drm_device
*dev
)
3427 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3428 struct intel_engine_cs
*engine
;
3431 /* Flush everything onto the inactive list. */
3432 for_each_engine(engine
, dev_priv
) {
3433 if (!i915
.enable_execlists
) {
3434 struct drm_i915_gem_request
*req
;
3436 req
= i915_gem_request_alloc(engine
, NULL
);
3438 return PTR_ERR(req
);
3440 ret
= i915_switch_context(req
);
3441 i915_add_request_no_flush(req
);
3446 ret
= intel_engine_idle(engine
);
3451 WARN_ON(i915_verify_lists(dev
));
3455 static bool i915_gem_valid_gtt_space(struct i915_vma
*vma
,
3456 unsigned long cache_level
)
3458 struct drm_mm_node
*gtt_space
= &vma
->node
;
3459 struct drm_mm_node
*other
;
3462 * On some machines we have to be careful when putting differing types
3463 * of snoopable memory together to avoid the prefetcher crossing memory
3464 * domains and dying. During vm initialisation, we decide whether or not
3465 * these constraints apply and set the drm_mm.color_adjust
3468 if (vma
->vm
->mm
.color_adjust
== NULL
)
3471 if (!drm_mm_node_allocated(gtt_space
))
3474 if (list_empty(>t_space
->node_list
))
3477 other
= list_entry(gtt_space
->node_list
.prev
, struct drm_mm_node
, node_list
);
3478 if (other
->allocated
&& !other
->hole_follows
&& other
->color
!= cache_level
)
3481 other
= list_entry(gtt_space
->node_list
.next
, struct drm_mm_node
, node_list
);
3482 if (other
->allocated
&& !gtt_space
->hole_follows
&& other
->color
!= cache_level
)
3489 * Finds free space in the GTT aperture and binds the object or a view of it
3492 static struct i915_vma
*
3493 i915_gem_object_bind_to_vm(struct drm_i915_gem_object
*obj
,
3494 struct i915_address_space
*vm
,
3495 const struct i915_ggtt_view
*ggtt_view
,
3499 struct drm_device
*dev
= obj
->base
.dev
;
3500 struct drm_i915_private
*dev_priv
= to_i915(dev
);
3501 struct i915_ggtt
*ggtt
= &dev_priv
->ggtt
;
3502 u32 fence_alignment
, unfenced_alignment
;
3503 u32 search_flag
, alloc_flag
;
3505 u64 size
, fence_size
;
3506 struct i915_vma
*vma
;
3509 if (i915_is_ggtt(vm
)) {
3512 if (WARN_ON(!ggtt_view
))
3513 return ERR_PTR(-EINVAL
);
3515 view_size
= i915_ggtt_view_size(obj
, ggtt_view
);
3517 fence_size
= i915_gem_get_gtt_size(dev
,
3520 fence_alignment
= i915_gem_get_gtt_alignment(dev
,
3524 unfenced_alignment
= i915_gem_get_gtt_alignment(dev
,
3528 size
= flags
& PIN_MAPPABLE
? fence_size
: view_size
;
3530 fence_size
= i915_gem_get_gtt_size(dev
,
3533 fence_alignment
= i915_gem_get_gtt_alignment(dev
,
3537 unfenced_alignment
=
3538 i915_gem_get_gtt_alignment(dev
,
3542 size
= flags
& PIN_MAPPABLE
? fence_size
: obj
->base
.size
;
3545 start
= flags
& PIN_OFFSET_BIAS
? flags
& PIN_OFFSET_MASK
: 0;
3547 if (flags
& PIN_MAPPABLE
)
3548 end
= min_t(u64
, end
, ggtt
->mappable_end
);
3549 if (flags
& PIN_ZONE_4G
)
3550 end
= min_t(u64
, end
, (1ULL << 32) - PAGE_SIZE
);
3553 alignment
= flags
& PIN_MAPPABLE
? fence_alignment
:
3555 if (flags
& PIN_MAPPABLE
&& alignment
& (fence_alignment
- 1)) {
3556 DRM_DEBUG("Invalid object (view type=%u) alignment requested %u\n",
3557 ggtt_view
? ggtt_view
->type
: 0,
3559 return ERR_PTR(-EINVAL
);
3562 /* If binding the object/GGTT view requires more space than the entire
3563 * aperture has, reject it early before evicting everything in a vain
3564 * attempt to find space.
3567 DRM_DEBUG("Attempting to bind an object (view type=%u) larger than the aperture: size=%llu > %s aperture=%llu\n",
3568 ggtt_view
? ggtt_view
->type
: 0,
3570 flags
& PIN_MAPPABLE
? "mappable" : "total",
3572 return ERR_PTR(-E2BIG
);
3575 ret
= i915_gem_object_get_pages(obj
);
3577 return ERR_PTR(ret
);
3579 i915_gem_object_pin_pages(obj
);
3581 vma
= ggtt_view
? i915_gem_obj_lookup_or_create_ggtt_vma(obj
, ggtt_view
) :
3582 i915_gem_obj_lookup_or_create_vma(obj
, vm
);
3587 if (flags
& PIN_OFFSET_FIXED
) {
3588 uint64_t offset
= flags
& PIN_OFFSET_MASK
;
3590 if (offset
& (alignment
- 1) || offset
+ size
> end
) {
3594 vma
->node
.start
= offset
;
3595 vma
->node
.size
= size
;
3596 vma
->node
.color
= obj
->cache_level
;
3597 ret
= drm_mm_reserve_node(&vm
->mm
, &vma
->node
);
3599 ret
= i915_gem_evict_for_vma(vma
);
3601 ret
= drm_mm_reserve_node(&vm
->mm
, &vma
->node
);
3606 if (flags
& PIN_HIGH
) {
3607 search_flag
= DRM_MM_SEARCH_BELOW
;
3608 alloc_flag
= DRM_MM_CREATE_TOP
;
3610 search_flag
= DRM_MM_SEARCH_DEFAULT
;
3611 alloc_flag
= DRM_MM_CREATE_DEFAULT
;
3615 ret
= drm_mm_insert_node_in_range_generic(&vm
->mm
, &vma
->node
,
3622 ret
= i915_gem_evict_something(dev
, vm
, size
, alignment
,
3632 if (WARN_ON(!i915_gem_valid_gtt_space(vma
, obj
->cache_level
))) {
3634 goto err_remove_node
;
3637 trace_i915_vma_bind(vma
, flags
);
3638 ret
= i915_vma_bind(vma
, obj
->cache_level
, flags
);
3640 goto err_remove_node
;
3642 list_move_tail(&obj
->global_list
, &dev_priv
->mm
.bound_list
);
3643 list_add_tail(&vma
->vm_link
, &vm
->inactive_list
);
3648 drm_mm_remove_node(&vma
->node
);
3650 i915_gem_vma_destroy(vma
);
3653 i915_gem_object_unpin_pages(obj
);
3658 i915_gem_clflush_object(struct drm_i915_gem_object
*obj
,
3661 /* If we don't have a page list set up, then we're not pinned
3662 * to GPU, and we can ignore the cache flush because it'll happen
3663 * again at bind time.
3665 if (obj
->pages
== NULL
)
3669 * Stolen memory is always coherent with the GPU as it is explicitly
3670 * marked as wc by the system, or the system is cache-coherent.
3672 if (obj
->stolen
|| obj
->phys_handle
)
3675 /* If the GPU is snooping the contents of the CPU cache,
3676 * we do not need to manually clear the CPU cache lines. However,
3677 * the caches are only snooped when the render cache is
3678 * flushed/invalidated. As we always have to emit invalidations
3679 * and flushes when moving into and out of the RENDER domain, correct
3680 * snooping behaviour occurs naturally as the result of our domain
3683 if (!force
&& cpu_cache_is_coherent(obj
->base
.dev
, obj
->cache_level
)) {
3684 obj
->cache_dirty
= true;
3688 trace_i915_gem_object_clflush(obj
);
3689 drm_clflush_sg(obj
->pages
);
3690 obj
->cache_dirty
= false;
3695 /** Flushes the GTT write domain for the object if it's dirty. */
3697 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object
*obj
)
3699 uint32_t old_write_domain
;
3701 if (obj
->base
.write_domain
!= I915_GEM_DOMAIN_GTT
)
3704 /* No actual flushing is required for the GTT write domain. Writes
3705 * to it immediately go to main memory as far as we know, so there's
3706 * no chipset flush. It also doesn't land in render cache.
3708 * However, we do have to enforce the order so that all writes through
3709 * the GTT land before any writes to the device, such as updates to
3714 old_write_domain
= obj
->base
.write_domain
;
3715 obj
->base
.write_domain
= 0;
3717 intel_fb_obj_flush(obj
, false, ORIGIN_GTT
);
3719 trace_i915_gem_object_change_domain(obj
,
3720 obj
->base
.read_domains
,
3724 /** Flushes the CPU write domain for the object if it's dirty. */
3726 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object
*obj
)
3728 uint32_t old_write_domain
;
3730 if (obj
->base
.write_domain
!= I915_GEM_DOMAIN_CPU
)
3733 if (i915_gem_clflush_object(obj
, obj
->pin_display
))
3734 i915_gem_chipset_flush(obj
->base
.dev
);
3736 old_write_domain
= obj
->base
.write_domain
;
3737 obj
->base
.write_domain
= 0;
3739 intel_fb_obj_flush(obj
, false, ORIGIN_CPU
);
3741 trace_i915_gem_object_change_domain(obj
,
3742 obj
->base
.read_domains
,
3747 * Moves a single object to the GTT read, and possibly write domain.
3749 * This function returns when the move is complete, including waiting on
3753 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object
*obj
, bool write
)
3755 struct drm_device
*dev
= obj
->base
.dev
;
3756 struct drm_i915_private
*dev_priv
= to_i915(dev
);
3757 struct i915_ggtt
*ggtt
= &dev_priv
->ggtt
;
3758 uint32_t old_write_domain
, old_read_domains
;
3759 struct i915_vma
*vma
;
3762 if (obj
->base
.write_domain
== I915_GEM_DOMAIN_GTT
)
3765 ret
= i915_gem_object_wait_rendering(obj
, !write
);
3769 /* Flush and acquire obj->pages so that we are coherent through
3770 * direct access in memory with previous cached writes through
3771 * shmemfs and that our cache domain tracking remains valid.
3772 * For example, if the obj->filp was moved to swap without us
3773 * being notified and releasing the pages, we would mistakenly
3774 * continue to assume that the obj remained out of the CPU cached
3777 ret
= i915_gem_object_get_pages(obj
);
3781 i915_gem_object_flush_cpu_write_domain(obj
);
3783 /* Serialise direct access to this object with the barriers for
3784 * coherent writes from the GPU, by effectively invalidating the
3785 * GTT domain upon first access.
3787 if ((obj
->base
.read_domains
& I915_GEM_DOMAIN_GTT
) == 0)
3790 old_write_domain
= obj
->base
.write_domain
;
3791 old_read_domains
= obj
->base
.read_domains
;
3793 /* It should now be out of any other write domains, and we can update
3794 * the domain values for our changes.
3796 BUG_ON((obj
->base
.write_domain
& ~I915_GEM_DOMAIN_GTT
) != 0);
3797 obj
->base
.read_domains
|= I915_GEM_DOMAIN_GTT
;
3799 obj
->base
.read_domains
= I915_GEM_DOMAIN_GTT
;
3800 obj
->base
.write_domain
= I915_GEM_DOMAIN_GTT
;
3804 trace_i915_gem_object_change_domain(obj
,
3808 /* And bump the LRU for this access */
3809 vma
= i915_gem_obj_to_ggtt(obj
);
3810 if (vma
&& drm_mm_node_allocated(&vma
->node
) && !obj
->active
)
3811 list_move_tail(&vma
->vm_link
,
3812 &ggtt
->base
.inactive_list
);
3818 * Changes the cache-level of an object across all VMA.
3820 * After this function returns, the object will be in the new cache-level
3821 * across all GTT and the contents of the backing storage will be coherent,
3822 * with respect to the new cache-level. In order to keep the backing storage
3823 * coherent for all users, we only allow a single cache level to be set
3824 * globally on the object and prevent it from being changed whilst the
3825 * hardware is reading from the object. That is if the object is currently
3826 * on the scanout it will be set to uncached (or equivalent display
3827 * cache coherency) and all non-MOCS GPU access will also be uncached so
3828 * that all direct access to the scanout remains coherent.
3830 int i915_gem_object_set_cache_level(struct drm_i915_gem_object
*obj
,
3831 enum i915_cache_level cache_level
)
3833 struct drm_device
*dev
= obj
->base
.dev
;
3834 struct i915_vma
*vma
, *next
;
3838 if (obj
->cache_level
== cache_level
)
3841 /* Inspect the list of currently bound VMA and unbind any that would
3842 * be invalid given the new cache-level. This is principally to
3843 * catch the issue of the CS prefetch crossing page boundaries and
3844 * reading an invalid PTE on older architectures.
3846 list_for_each_entry_safe(vma
, next
, &obj
->vma_list
, obj_link
) {
3847 if (!drm_mm_node_allocated(&vma
->node
))
3850 if (vma
->pin_count
) {
3851 DRM_DEBUG("can not change the cache level of pinned objects\n");
3855 if (!i915_gem_valid_gtt_space(vma
, cache_level
)) {
3856 ret
= i915_vma_unbind(vma
);
3863 /* We can reuse the existing drm_mm nodes but need to change the
3864 * cache-level on the PTE. We could simply unbind them all and
3865 * rebind with the correct cache-level on next use. However since
3866 * we already have a valid slot, dma mapping, pages etc, we may as
3867 * rewrite the PTE in the belief that doing so tramples upon less
3868 * state and so involves less work.
3871 /* Before we change the PTE, the GPU must not be accessing it.
3872 * If we wait upon the object, we know that all the bound
3873 * VMA are no longer active.
3875 ret
= i915_gem_object_wait_rendering(obj
, false);
3879 if (!HAS_LLC(dev
) && cache_level
!= I915_CACHE_NONE
) {
3880 /* Access to snoopable pages through the GTT is
3881 * incoherent and on some machines causes a hard
3882 * lockup. Relinquish the CPU mmaping to force
3883 * userspace to refault in the pages and we can
3884 * then double check if the GTT mapping is still
3885 * valid for that pointer access.
3887 i915_gem_release_mmap(obj
);
3889 /* As we no longer need a fence for GTT access,
3890 * we can relinquish it now (and so prevent having
3891 * to steal a fence from someone else on the next
3892 * fence request). Note GPU activity would have
3893 * dropped the fence as all snoopable access is
3894 * supposed to be linear.
3896 ret
= i915_gem_object_put_fence(obj
);
3900 /* We either have incoherent backing store and
3901 * so no GTT access or the architecture is fully
3902 * coherent. In such cases, existing GTT mmaps
3903 * ignore the cache bit in the PTE and we can
3904 * rewrite it without confusing the GPU or having
3905 * to force userspace to fault back in its mmaps.
3909 list_for_each_entry(vma
, &obj
->vma_list
, obj_link
) {
3910 if (!drm_mm_node_allocated(&vma
->node
))
3913 ret
= i915_vma_bind(vma
, cache_level
, PIN_UPDATE
);
3919 list_for_each_entry(vma
, &obj
->vma_list
, obj_link
)
3920 vma
->node
.color
= cache_level
;
3921 obj
->cache_level
= cache_level
;
3924 /* Flush the dirty CPU caches to the backing storage so that the
3925 * object is now coherent at its new cache level (with respect
3926 * to the access domain).
3928 if (obj
->cache_dirty
&&
3929 obj
->base
.write_domain
!= I915_GEM_DOMAIN_CPU
&&
3930 cpu_write_needs_clflush(obj
)) {
3931 if (i915_gem_clflush_object(obj
, true))
3932 i915_gem_chipset_flush(obj
->base
.dev
);
3938 int i915_gem_get_caching_ioctl(struct drm_device
*dev
, void *data
,
3939 struct drm_file
*file
)
3941 struct drm_i915_gem_caching
*args
= data
;
3942 struct drm_i915_gem_object
*obj
;
3944 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
3945 if (&obj
->base
== NULL
)
3948 switch (obj
->cache_level
) {
3949 case I915_CACHE_LLC
:
3950 case I915_CACHE_L3_LLC
:
3951 args
->caching
= I915_CACHING_CACHED
;
3955 args
->caching
= I915_CACHING_DISPLAY
;
3959 args
->caching
= I915_CACHING_NONE
;
3963 drm_gem_object_unreference_unlocked(&obj
->base
);
3967 int i915_gem_set_caching_ioctl(struct drm_device
*dev
, void *data
,
3968 struct drm_file
*file
)
3970 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3971 struct drm_i915_gem_caching
*args
= data
;
3972 struct drm_i915_gem_object
*obj
;
3973 enum i915_cache_level level
;
3976 switch (args
->caching
) {
3977 case I915_CACHING_NONE
:
3978 level
= I915_CACHE_NONE
;
3980 case I915_CACHING_CACHED
:
3982 * Due to a HW issue on BXT A stepping, GPU stores via a
3983 * snooped mapping may leave stale data in a corresponding CPU
3984 * cacheline, whereas normally such cachelines would get
3987 if (!HAS_LLC(dev
) && !HAS_SNOOP(dev
))
3990 level
= I915_CACHE_LLC
;
3992 case I915_CACHING_DISPLAY
:
3993 level
= HAS_WT(dev
) ? I915_CACHE_WT
: I915_CACHE_NONE
;
3999 intel_runtime_pm_get(dev_priv
);
4001 ret
= i915_mutex_lock_interruptible(dev
);
4005 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
4006 if (&obj
->base
== NULL
) {
4011 ret
= i915_gem_object_set_cache_level(obj
, level
);
4013 drm_gem_object_unreference(&obj
->base
);
4015 mutex_unlock(&dev
->struct_mutex
);
4017 intel_runtime_pm_put(dev_priv
);
4023 * Prepare buffer for display plane (scanout, cursors, etc).
4024 * Can be called from an uninterruptible phase (modesetting) and allows
4025 * any flushes to be pipelined (for pageflips).
4028 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object
*obj
,
4030 const struct i915_ggtt_view
*view
)
4032 u32 old_read_domains
, old_write_domain
;
4035 /* Mark the pin_display early so that we account for the
4036 * display coherency whilst setting up the cache domains.
4040 /* The display engine is not coherent with the LLC cache on gen6. As
4041 * a result, we make sure that the pinning that is about to occur is
4042 * done with uncached PTEs. This is lowest common denominator for all
4045 * However for gen6+, we could do better by using the GFDT bit instead
4046 * of uncaching, which would allow us to flush all the LLC-cached data
4047 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
4049 ret
= i915_gem_object_set_cache_level(obj
,
4050 HAS_WT(obj
->base
.dev
) ? I915_CACHE_WT
: I915_CACHE_NONE
);
4052 goto err_unpin_display
;
4054 /* As the user may map the buffer once pinned in the display plane
4055 * (e.g. libkms for the bootup splash), we have to ensure that we
4056 * always use map_and_fenceable for all scanout buffers.
4058 ret
= i915_gem_object_ggtt_pin(obj
, view
, alignment
,
4059 view
->type
== I915_GGTT_VIEW_NORMAL
?
4062 goto err_unpin_display
;
4064 i915_gem_object_flush_cpu_write_domain(obj
);
4066 old_write_domain
= obj
->base
.write_domain
;
4067 old_read_domains
= obj
->base
.read_domains
;
4069 /* It should now be out of any other write domains, and we can update
4070 * the domain values for our changes.
4072 obj
->base
.write_domain
= 0;
4073 obj
->base
.read_domains
|= I915_GEM_DOMAIN_GTT
;
4075 trace_i915_gem_object_change_domain(obj
,
4087 i915_gem_object_unpin_from_display_plane(struct drm_i915_gem_object
*obj
,
4088 const struct i915_ggtt_view
*view
)
4090 if (WARN_ON(obj
->pin_display
== 0))
4093 i915_gem_object_ggtt_unpin_view(obj
, view
);
4099 * Moves a single object to the CPU read, and possibly write domain.
4101 * This function returns when the move is complete, including waiting on
4105 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object
*obj
, bool write
)
4107 uint32_t old_write_domain
, old_read_domains
;
4110 if (obj
->base
.write_domain
== I915_GEM_DOMAIN_CPU
)
4113 ret
= i915_gem_object_wait_rendering(obj
, !write
);
4117 i915_gem_object_flush_gtt_write_domain(obj
);
4119 old_write_domain
= obj
->base
.write_domain
;
4120 old_read_domains
= obj
->base
.read_domains
;
4122 /* Flush the CPU cache if it's still invalid. */
4123 if ((obj
->base
.read_domains
& I915_GEM_DOMAIN_CPU
) == 0) {
4124 i915_gem_clflush_object(obj
, false);
4126 obj
->base
.read_domains
|= I915_GEM_DOMAIN_CPU
;
4129 /* It should now be out of any other write domains, and we can update
4130 * the domain values for our changes.
4132 BUG_ON((obj
->base
.write_domain
& ~I915_GEM_DOMAIN_CPU
) != 0);
4134 /* If we're writing through the CPU, then the GPU read domains will
4135 * need to be invalidated at next use.
4138 obj
->base
.read_domains
= I915_GEM_DOMAIN_CPU
;
4139 obj
->base
.write_domain
= I915_GEM_DOMAIN_CPU
;
4142 trace_i915_gem_object_change_domain(obj
,
4149 /* Throttle our rendering by waiting until the ring has completed our requests
4150 * emitted over 20 msec ago.
4152 * Note that if we were to use the current jiffies each time around the loop,
4153 * we wouldn't escape the function with any frames outstanding if the time to
4154 * render a frame was over 20ms.
4156 * This should get us reasonable parallelism between CPU and GPU but also
4157 * relatively low latency when blocking on a particular request to finish.
4160 i915_gem_ring_throttle(struct drm_device
*dev
, struct drm_file
*file
)
4162 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4163 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
4164 unsigned long recent_enough
= jiffies
- DRM_I915_THROTTLE_JIFFIES
;
4165 struct drm_i915_gem_request
*request
, *target
= NULL
;
4168 ret
= i915_gem_wait_for_error(&dev_priv
->gpu_error
);
4172 /* ABI: return -EIO if already wedged */
4173 if (i915_terminally_wedged(&dev_priv
->gpu_error
))
4176 spin_lock(&file_priv
->mm
.lock
);
4177 list_for_each_entry(request
, &file_priv
->mm
.request_list
, client_list
) {
4178 if (time_after_eq(request
->emitted_jiffies
, recent_enough
))
4182 * Note that the request might not have been submitted yet.
4183 * In which case emitted_jiffies will be zero.
4185 if (!request
->emitted_jiffies
)
4191 i915_gem_request_reference(target
);
4192 spin_unlock(&file_priv
->mm
.lock
);
4197 ret
= __i915_wait_request(target
, true, NULL
, NULL
);
4199 queue_delayed_work(dev_priv
->wq
, &dev_priv
->mm
.retire_work
, 0);
4201 i915_gem_request_unreference__unlocked(target
);
4207 i915_vma_misplaced(struct i915_vma
*vma
, uint32_t alignment
, uint64_t flags
)
4209 struct drm_i915_gem_object
*obj
= vma
->obj
;
4212 vma
->node
.start
& (alignment
- 1))
4215 if (flags
& PIN_MAPPABLE
&& !obj
->map_and_fenceable
)
4218 if (flags
& PIN_OFFSET_BIAS
&&
4219 vma
->node
.start
< (flags
& PIN_OFFSET_MASK
))
4222 if (flags
& PIN_OFFSET_FIXED
&&
4223 vma
->node
.start
!= (flags
& PIN_OFFSET_MASK
))
4229 void __i915_vma_set_map_and_fenceable(struct i915_vma
*vma
)
4231 struct drm_i915_gem_object
*obj
= vma
->obj
;
4232 bool mappable
, fenceable
;
4233 u32 fence_size
, fence_alignment
;
4235 fence_size
= i915_gem_get_gtt_size(obj
->base
.dev
,
4238 fence_alignment
= i915_gem_get_gtt_alignment(obj
->base
.dev
,
4243 fenceable
= (vma
->node
.size
== fence_size
&&
4244 (vma
->node
.start
& (fence_alignment
- 1)) == 0);
4246 mappable
= (vma
->node
.start
+ fence_size
<=
4247 to_i915(obj
->base
.dev
)->ggtt
.mappable_end
);
4249 obj
->map_and_fenceable
= mappable
&& fenceable
;
4253 i915_gem_object_do_pin(struct drm_i915_gem_object
*obj
,
4254 struct i915_address_space
*vm
,
4255 const struct i915_ggtt_view
*ggtt_view
,
4259 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
4260 struct i915_vma
*vma
;
4264 if (WARN_ON(vm
== &dev_priv
->mm
.aliasing_ppgtt
->base
))
4267 if (WARN_ON(flags
& (PIN_GLOBAL
| PIN_MAPPABLE
) && !i915_is_ggtt(vm
)))
4270 if (WARN_ON((flags
& (PIN_MAPPABLE
| PIN_GLOBAL
)) == PIN_MAPPABLE
))
4273 if (WARN_ON(i915_is_ggtt(vm
) != !!ggtt_view
))
4276 vma
= ggtt_view
? i915_gem_obj_to_ggtt_view(obj
, ggtt_view
) :
4277 i915_gem_obj_to_vma(obj
, vm
);
4280 if (WARN_ON(vma
->pin_count
== DRM_I915_GEM_OBJECT_MAX_PIN_COUNT
))
4283 if (i915_vma_misplaced(vma
, alignment
, flags
)) {
4284 WARN(vma
->pin_count
,
4285 "bo is already pinned in %s with incorrect alignment:"
4286 " offset=%08x %08x, req.alignment=%x, req.map_and_fenceable=%d,"
4287 " obj->map_and_fenceable=%d\n",
4288 ggtt_view
? "ggtt" : "ppgtt",
4289 upper_32_bits(vma
->node
.start
),
4290 lower_32_bits(vma
->node
.start
),
4292 !!(flags
& PIN_MAPPABLE
),
4293 obj
->map_and_fenceable
);
4294 ret
= i915_vma_unbind(vma
);
4302 bound
= vma
? vma
->bound
: 0;
4303 if (vma
== NULL
|| !drm_mm_node_allocated(&vma
->node
)) {
4304 vma
= i915_gem_object_bind_to_vm(obj
, vm
, ggtt_view
, alignment
,
4307 return PTR_ERR(vma
);
4309 ret
= i915_vma_bind(vma
, obj
->cache_level
, flags
);
4314 if (ggtt_view
&& ggtt_view
->type
== I915_GGTT_VIEW_NORMAL
&&
4315 (bound
^ vma
->bound
) & GLOBAL_BIND
) {
4316 __i915_vma_set_map_and_fenceable(vma
);
4317 WARN_ON(flags
& PIN_MAPPABLE
&& !obj
->map_and_fenceable
);
4325 i915_gem_object_pin(struct drm_i915_gem_object
*obj
,
4326 struct i915_address_space
*vm
,
4330 return i915_gem_object_do_pin(obj
, vm
,
4331 i915_is_ggtt(vm
) ? &i915_ggtt_view_normal
: NULL
,
4336 i915_gem_object_ggtt_pin(struct drm_i915_gem_object
*obj
,
4337 const struct i915_ggtt_view
*view
,
4341 struct drm_device
*dev
= obj
->base
.dev
;
4342 struct drm_i915_private
*dev_priv
= to_i915(dev
);
4343 struct i915_ggtt
*ggtt
= &dev_priv
->ggtt
;
4347 return i915_gem_object_do_pin(obj
, &ggtt
->base
, view
,
4348 alignment
, flags
| PIN_GLOBAL
);
4352 i915_gem_object_ggtt_unpin_view(struct drm_i915_gem_object
*obj
,
4353 const struct i915_ggtt_view
*view
)
4355 struct i915_vma
*vma
= i915_gem_obj_to_ggtt_view(obj
, view
);
4357 WARN_ON(vma
->pin_count
== 0);
4358 WARN_ON(!i915_gem_obj_ggtt_bound_view(obj
, view
));
4364 i915_gem_busy_ioctl(struct drm_device
*dev
, void *data
,
4365 struct drm_file
*file
)
4367 struct drm_i915_gem_busy
*args
= data
;
4368 struct drm_i915_gem_object
*obj
;
4371 ret
= i915_mutex_lock_interruptible(dev
);
4375 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
4376 if (&obj
->base
== NULL
) {
4381 /* Count all active objects as busy, even if they are currently not used
4382 * by the gpu. Users of this interface expect objects to eventually
4383 * become non-busy without any further actions, therefore emit any
4384 * necessary flushes here.
4386 ret
= i915_gem_object_flush_active(obj
);
4394 for (i
= 0; i
< I915_NUM_ENGINES
; i
++) {
4395 struct drm_i915_gem_request
*req
;
4397 req
= obj
->last_read_req
[i
];
4399 args
->busy
|= 1 << (16 + req
->engine
->exec_id
);
4401 if (obj
->last_write_req
)
4402 args
->busy
|= obj
->last_write_req
->engine
->exec_id
;
4406 drm_gem_object_unreference(&obj
->base
);
4408 mutex_unlock(&dev
->struct_mutex
);
4413 i915_gem_throttle_ioctl(struct drm_device
*dev
, void *data
,
4414 struct drm_file
*file_priv
)
4416 return i915_gem_ring_throttle(dev
, file_priv
);
4420 i915_gem_madvise_ioctl(struct drm_device
*dev
, void *data
,
4421 struct drm_file
*file_priv
)
4423 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4424 struct drm_i915_gem_madvise
*args
= data
;
4425 struct drm_i915_gem_object
*obj
;
4428 switch (args
->madv
) {
4429 case I915_MADV_DONTNEED
:
4430 case I915_MADV_WILLNEED
:
4436 ret
= i915_mutex_lock_interruptible(dev
);
4440 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file_priv
, args
->handle
));
4441 if (&obj
->base
== NULL
) {
4446 if (i915_gem_obj_is_pinned(obj
)) {
4452 obj
->tiling_mode
!= I915_TILING_NONE
&&
4453 dev_priv
->quirks
& QUIRK_PIN_SWIZZLED_PAGES
) {
4454 if (obj
->madv
== I915_MADV_WILLNEED
)
4455 i915_gem_object_unpin_pages(obj
);
4456 if (args
->madv
== I915_MADV_WILLNEED
)
4457 i915_gem_object_pin_pages(obj
);
4460 if (obj
->madv
!= __I915_MADV_PURGED
)
4461 obj
->madv
= args
->madv
;
4463 /* if the object is no longer attached, discard its backing storage */
4464 if (obj
->madv
== I915_MADV_DONTNEED
&& obj
->pages
== NULL
)
4465 i915_gem_object_truncate(obj
);
4467 args
->retained
= obj
->madv
!= __I915_MADV_PURGED
;
4470 drm_gem_object_unreference(&obj
->base
);
4472 mutex_unlock(&dev
->struct_mutex
);
4476 void i915_gem_object_init(struct drm_i915_gem_object
*obj
,
4477 const struct drm_i915_gem_object_ops
*ops
)
4481 INIT_LIST_HEAD(&obj
->global_list
);
4482 for (i
= 0; i
< I915_NUM_ENGINES
; i
++)
4483 INIT_LIST_HEAD(&obj
->engine_list
[i
]);
4484 INIT_LIST_HEAD(&obj
->obj_exec_link
);
4485 INIT_LIST_HEAD(&obj
->vma_list
);
4486 INIT_LIST_HEAD(&obj
->batch_pool_link
);
4490 obj
->fence_reg
= I915_FENCE_REG_NONE
;
4491 obj
->madv
= I915_MADV_WILLNEED
;
4493 i915_gem_info_add_obj(obj
->base
.dev
->dev_private
, obj
->base
.size
);
4496 static const struct drm_i915_gem_object_ops i915_gem_object_ops
= {
4497 .flags
= I915_GEM_OBJECT_HAS_STRUCT_PAGE
,
4498 .get_pages
= i915_gem_object_get_pages_gtt
,
4499 .put_pages
= i915_gem_object_put_pages_gtt
,
4502 struct drm_i915_gem_object
*i915_gem_alloc_object(struct drm_device
*dev
,
4505 struct drm_i915_gem_object
*obj
;
4506 struct address_space
*mapping
;
4509 obj
= i915_gem_object_alloc(dev
);
4513 if (drm_gem_object_init(dev
, &obj
->base
, size
) != 0) {
4514 i915_gem_object_free(obj
);
4518 mask
= GFP_HIGHUSER
| __GFP_RECLAIMABLE
;
4519 if (IS_CRESTLINE(dev
) || IS_BROADWATER(dev
)) {
4520 /* 965gm cannot relocate objects above 4GiB. */
4521 mask
&= ~__GFP_HIGHMEM
;
4522 mask
|= __GFP_DMA32
;
4525 mapping
= file_inode(obj
->base
.filp
)->i_mapping
;
4526 mapping_set_gfp_mask(mapping
, mask
);
4528 i915_gem_object_init(obj
, &i915_gem_object_ops
);
4530 obj
->base
.write_domain
= I915_GEM_DOMAIN_CPU
;
4531 obj
->base
.read_domains
= I915_GEM_DOMAIN_CPU
;
4534 /* On some devices, we can have the GPU use the LLC (the CPU
4535 * cache) for about a 10% performance improvement
4536 * compared to uncached. Graphics requests other than
4537 * display scanout are coherent with the CPU in
4538 * accessing this cache. This means in this mode we
4539 * don't need to clflush on the CPU side, and on the
4540 * GPU side we only need to flush internal caches to
4541 * get data visible to the CPU.
4543 * However, we maintain the display planes as UC, and so
4544 * need to rebind when first used as such.
4546 obj
->cache_level
= I915_CACHE_LLC
;
4548 obj
->cache_level
= I915_CACHE_NONE
;
4550 trace_i915_gem_object_create(obj
);
4555 static bool discard_backing_storage(struct drm_i915_gem_object
*obj
)
4557 /* If we are the last user of the backing storage (be it shmemfs
4558 * pages or stolen etc), we know that the pages are going to be
4559 * immediately released. In this case, we can then skip copying
4560 * back the contents from the GPU.
4563 if (obj
->madv
!= I915_MADV_WILLNEED
)
4566 if (obj
->base
.filp
== NULL
)
4569 /* At first glance, this looks racy, but then again so would be
4570 * userspace racing mmap against close. However, the first external
4571 * reference to the filp can only be obtained through the
4572 * i915_gem_mmap_ioctl() which safeguards us against the user
4573 * acquiring such a reference whilst we are in the middle of
4574 * freeing the object.
4576 return atomic_long_read(&obj
->base
.filp
->f_count
) == 1;
4579 void i915_gem_free_object(struct drm_gem_object
*gem_obj
)
4581 struct drm_i915_gem_object
*obj
= to_intel_bo(gem_obj
);
4582 struct drm_device
*dev
= obj
->base
.dev
;
4583 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4584 struct i915_vma
*vma
, *next
;
4586 intel_runtime_pm_get(dev_priv
);
4588 trace_i915_gem_object_destroy(obj
);
4590 list_for_each_entry_safe(vma
, next
, &obj
->vma_list
, obj_link
) {
4594 ret
= i915_vma_unbind(vma
);
4595 if (WARN_ON(ret
== -ERESTARTSYS
)) {
4596 bool was_interruptible
;
4598 was_interruptible
= dev_priv
->mm
.interruptible
;
4599 dev_priv
->mm
.interruptible
= false;
4601 WARN_ON(i915_vma_unbind(vma
));
4603 dev_priv
->mm
.interruptible
= was_interruptible
;
4607 /* Stolen objects don't hold a ref, but do hold pin count. Fix that up
4608 * before progressing. */
4610 i915_gem_object_unpin_pages(obj
);
4612 WARN_ON(obj
->frontbuffer_bits
);
4614 if (obj
->pages
&& obj
->madv
== I915_MADV_WILLNEED
&&
4615 dev_priv
->quirks
& QUIRK_PIN_SWIZZLED_PAGES
&&
4616 obj
->tiling_mode
!= I915_TILING_NONE
)
4617 i915_gem_object_unpin_pages(obj
);
4619 if (WARN_ON(obj
->pages_pin_count
))
4620 obj
->pages_pin_count
= 0;
4621 if (discard_backing_storage(obj
))
4622 obj
->madv
= I915_MADV_DONTNEED
;
4623 i915_gem_object_put_pages(obj
);
4624 i915_gem_object_free_mmap_offset(obj
);
4628 if (obj
->base
.import_attach
)
4629 drm_prime_gem_destroy(&obj
->base
, NULL
);
4631 if (obj
->ops
->release
)
4632 obj
->ops
->release(obj
);
4634 drm_gem_object_release(&obj
->base
);
4635 i915_gem_info_remove_obj(dev_priv
, obj
->base
.size
);
4638 i915_gem_object_free(obj
);
4640 intel_runtime_pm_put(dev_priv
);
4643 struct i915_vma
*i915_gem_obj_to_vma(struct drm_i915_gem_object
*obj
,
4644 struct i915_address_space
*vm
)
4646 struct i915_vma
*vma
;
4647 list_for_each_entry(vma
, &obj
->vma_list
, obj_link
) {
4648 if (vma
->ggtt_view
.type
== I915_GGTT_VIEW_NORMAL
&&
4655 struct i915_vma
*i915_gem_obj_to_ggtt_view(struct drm_i915_gem_object
*obj
,
4656 const struct i915_ggtt_view
*view
)
4658 struct drm_device
*dev
= obj
->base
.dev
;
4659 struct drm_i915_private
*dev_priv
= to_i915(dev
);
4660 struct i915_ggtt
*ggtt
= &dev_priv
->ggtt
;
4661 struct i915_vma
*vma
;
4665 list_for_each_entry(vma
, &obj
->vma_list
, obj_link
)
4666 if (vma
->vm
== &ggtt
->base
&&
4667 i915_ggtt_view_equal(&vma
->ggtt_view
, view
))
4672 void i915_gem_vma_destroy(struct i915_vma
*vma
)
4674 WARN_ON(vma
->node
.allocated
);
4676 /* Keep the vma as a placeholder in the execbuffer reservation lists */
4677 if (!list_empty(&vma
->exec_list
))
4681 i915_ppgtt_put(i915_vm_to_ppgtt(vma
->vm
));
4683 list_del(&vma
->obj_link
);
4685 kmem_cache_free(to_i915(vma
->obj
->base
.dev
)->vmas
, vma
);
4689 i915_gem_stop_engines(struct drm_device
*dev
)
4691 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4692 struct intel_engine_cs
*engine
;
4694 for_each_engine(engine
, dev_priv
)
4695 dev_priv
->gt
.stop_engine(engine
);
4699 i915_gem_suspend(struct drm_device
*dev
)
4701 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4704 mutex_lock(&dev
->struct_mutex
);
4705 ret
= i915_gpu_idle(dev
);
4709 i915_gem_retire_requests(dev
);
4711 i915_gem_stop_engines(dev
);
4712 mutex_unlock(&dev
->struct_mutex
);
4714 cancel_delayed_work_sync(&dev_priv
->gpu_error
.hangcheck_work
);
4715 cancel_delayed_work_sync(&dev_priv
->mm
.retire_work
);
4716 flush_delayed_work(&dev_priv
->mm
.idle_work
);
4718 /* Assert that we sucessfully flushed all the work and
4719 * reset the GPU back to its idle, low power state.
4721 WARN_ON(dev_priv
->mm
.busy
);
4726 mutex_unlock(&dev
->struct_mutex
);
4730 int i915_gem_l3_remap(struct drm_i915_gem_request
*req
, int slice
)
4732 struct intel_engine_cs
*engine
= req
->engine
;
4733 struct drm_device
*dev
= engine
->dev
;
4734 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4735 u32
*remap_info
= dev_priv
->l3_parity
.remap_info
[slice
];
4738 if (!HAS_L3_DPF(dev
) || !remap_info
)
4741 ret
= intel_ring_begin(req
, GEN7_L3LOG_SIZE
/ 4 * 3);
4746 * Note: We do not worry about the concurrent register cacheline hang
4747 * here because no other code should access these registers other than
4748 * at initialization time.
4750 for (i
= 0; i
< GEN7_L3LOG_SIZE
/ 4; i
++) {
4751 intel_ring_emit(engine
, MI_LOAD_REGISTER_IMM(1));
4752 intel_ring_emit_reg(engine
, GEN7_L3LOG(slice
, i
));
4753 intel_ring_emit(engine
, remap_info
[i
]);
4756 intel_ring_advance(engine
);
4761 void i915_gem_init_swizzling(struct drm_device
*dev
)
4763 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4765 if (INTEL_INFO(dev
)->gen
< 5 ||
4766 dev_priv
->mm
.bit_6_swizzle_x
== I915_BIT_6_SWIZZLE_NONE
)
4769 I915_WRITE(DISP_ARB_CTL
, I915_READ(DISP_ARB_CTL
) |
4770 DISP_TILE_SURFACE_SWIZZLING
);
4775 I915_WRITE(TILECTL
, I915_READ(TILECTL
) | TILECTL_SWZCTL
);
4777 I915_WRITE(ARB_MODE
, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB
));
4778 else if (IS_GEN7(dev
))
4779 I915_WRITE(ARB_MODE
, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB
));
4780 else if (IS_GEN8(dev
))
4781 I915_WRITE(GAMTARBMODE
, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW
));
4786 static void init_unused_ring(struct drm_device
*dev
, u32 base
)
4788 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4790 I915_WRITE(RING_CTL(base
), 0);
4791 I915_WRITE(RING_HEAD(base
), 0);
4792 I915_WRITE(RING_TAIL(base
), 0);
4793 I915_WRITE(RING_START(base
), 0);
4796 static void init_unused_rings(struct drm_device
*dev
)
4799 init_unused_ring(dev
, PRB1_BASE
);
4800 init_unused_ring(dev
, SRB0_BASE
);
4801 init_unused_ring(dev
, SRB1_BASE
);
4802 init_unused_ring(dev
, SRB2_BASE
);
4803 init_unused_ring(dev
, SRB3_BASE
);
4804 } else if (IS_GEN2(dev
)) {
4805 init_unused_ring(dev
, SRB0_BASE
);
4806 init_unused_ring(dev
, SRB1_BASE
);
4807 } else if (IS_GEN3(dev
)) {
4808 init_unused_ring(dev
, PRB1_BASE
);
4809 init_unused_ring(dev
, PRB2_BASE
);
4813 int i915_gem_init_engines(struct drm_device
*dev
)
4815 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4818 ret
= intel_init_render_ring_buffer(dev
);
4823 ret
= intel_init_bsd_ring_buffer(dev
);
4825 goto cleanup_render_ring
;
4829 ret
= intel_init_blt_ring_buffer(dev
);
4831 goto cleanup_bsd_ring
;
4834 if (HAS_VEBOX(dev
)) {
4835 ret
= intel_init_vebox_ring_buffer(dev
);
4837 goto cleanup_blt_ring
;
4840 if (HAS_BSD2(dev
)) {
4841 ret
= intel_init_bsd2_ring_buffer(dev
);
4843 goto cleanup_vebox_ring
;
4849 intel_cleanup_engine(&dev_priv
->engine
[VECS
]);
4851 intel_cleanup_engine(&dev_priv
->engine
[BCS
]);
4853 intel_cleanup_engine(&dev_priv
->engine
[VCS
]);
4854 cleanup_render_ring
:
4855 intel_cleanup_engine(&dev_priv
->engine
[RCS
]);
4861 i915_gem_init_hw(struct drm_device
*dev
)
4863 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4864 struct intel_engine_cs
*engine
;
4867 /* Double layer security blanket, see i915_gem_init() */
4868 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4870 if (HAS_EDRAM(dev
) && INTEL_GEN(dev_priv
) < 9)
4871 I915_WRITE(HSW_IDICR
, I915_READ(HSW_IDICR
) | IDIHASHMSK(0xf));
4873 if (IS_HASWELL(dev
))
4874 I915_WRITE(MI_PREDICATE_RESULT_2
, IS_HSW_GT3(dev
) ?
4875 LOWER_SLICE_ENABLED
: LOWER_SLICE_DISABLED
);
4877 if (HAS_PCH_NOP(dev
)) {
4878 if (IS_IVYBRIDGE(dev
)) {
4879 u32 temp
= I915_READ(GEN7_MSG_CTL
);
4880 temp
&= ~(WAIT_FOR_PCH_FLR_ACK
| WAIT_FOR_PCH_RESET_ACK
);
4881 I915_WRITE(GEN7_MSG_CTL
, temp
);
4882 } else if (INTEL_INFO(dev
)->gen
>= 7) {
4883 u32 temp
= I915_READ(HSW_NDE_RSTWRN_OPT
);
4884 temp
&= ~RESET_PCH_HANDSHAKE_ENABLE
;
4885 I915_WRITE(HSW_NDE_RSTWRN_OPT
, temp
);
4889 i915_gem_init_swizzling(dev
);
4892 * At least 830 can leave some of the unused rings
4893 * "active" (ie. head != tail) after resume which
4894 * will prevent c3 entry. Makes sure all unused rings
4897 init_unused_rings(dev
);
4899 BUG_ON(!dev_priv
->kernel_context
);
4901 ret
= i915_ppgtt_init_hw(dev
);
4903 DRM_ERROR("PPGTT enable HW failed %d\n", ret
);
4907 /* Need to do basic initialisation of all rings first: */
4908 for_each_engine(engine
, dev_priv
) {
4909 ret
= engine
->init_hw(engine
);
4914 intel_mocs_init_l3cc_table(dev
);
4916 /* We can't enable contexts until all firmware is loaded */
4917 if (HAS_GUC_UCODE(dev
)) {
4918 ret
= intel_guc_ucode_load(dev
);
4920 DRM_ERROR("Failed to initialize GuC, error %d\n", ret
);
4927 * Increment the next seqno by 0x100 so we have a visible break
4928 * on re-initialisation
4930 ret
= i915_gem_set_seqno(dev
, dev_priv
->next_seqno
+0x100);
4934 /* Now it is safe to go back round and do everything else: */
4935 for_each_engine(engine
, dev_priv
) {
4936 struct drm_i915_gem_request
*req
;
4938 req
= i915_gem_request_alloc(engine
, NULL
);
4944 if (engine
->id
== RCS
) {
4945 for (j
= 0; j
< NUM_L3_SLICES(dev
); j
++) {
4946 ret
= i915_gem_l3_remap(req
, j
);
4952 ret
= i915_ppgtt_init_ring(req
);
4956 ret
= i915_gem_context_enable(req
);
4961 i915_add_request_no_flush(req
);
4963 DRM_ERROR("Failed to enable %s, error=%d\n",
4965 i915_gem_cleanup_engines(dev
);
4971 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4975 int i915_gem_init(struct drm_device
*dev
)
4977 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4980 i915
.enable_execlists
= intel_sanitize_enable_execlists(dev
,
4981 i915
.enable_execlists
);
4983 mutex_lock(&dev
->struct_mutex
);
4985 if (!i915
.enable_execlists
) {
4986 dev_priv
->gt
.execbuf_submit
= i915_gem_ringbuffer_submission
;
4987 dev_priv
->gt
.init_engines
= i915_gem_init_engines
;
4988 dev_priv
->gt
.cleanup_engine
= intel_cleanup_engine
;
4989 dev_priv
->gt
.stop_engine
= intel_stop_engine
;
4991 dev_priv
->gt
.execbuf_submit
= intel_execlists_submission
;
4992 dev_priv
->gt
.init_engines
= intel_logical_rings_init
;
4993 dev_priv
->gt
.cleanup_engine
= intel_logical_ring_cleanup
;
4994 dev_priv
->gt
.stop_engine
= intel_logical_ring_stop
;
4997 /* This is just a security blanket to placate dragons.
4998 * On some systems, we very sporadically observe that the first TLBs
4999 * used by the CS may be stale, despite us poking the TLB reset. If
5000 * we hold the forcewake during initialisation these problems
5001 * just magically go away.
5003 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
5005 ret
= i915_gem_init_userptr(dev
);
5009 i915_gem_init_ggtt(dev
);
5011 ret
= i915_gem_context_init(dev
);
5015 ret
= dev_priv
->gt
.init_engines(dev
);
5019 ret
= i915_gem_init_hw(dev
);
5021 /* Allow ring initialisation to fail by marking the GPU as
5022 * wedged. But we only want to do this where the GPU is angry,
5023 * for all other failure, such as an allocation failure, bail.
5025 DRM_ERROR("Failed to initialize GPU, declaring it wedged\n");
5026 atomic_or(I915_WEDGED
, &dev_priv
->gpu_error
.reset_counter
);
5031 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
5032 mutex_unlock(&dev
->struct_mutex
);
5038 i915_gem_cleanup_engines(struct drm_device
*dev
)
5040 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5041 struct intel_engine_cs
*engine
;
5043 for_each_engine(engine
, dev_priv
)
5044 dev_priv
->gt
.cleanup_engine(engine
);
5046 if (i915
.enable_execlists
)
5048 * Neither the BIOS, ourselves or any other kernel
5049 * expects the system to be in execlists mode on startup,
5050 * so we need to reset the GPU back to legacy mode.
5052 intel_gpu_reset(dev
, ALL_ENGINES
);
5056 init_engine_lists(struct intel_engine_cs
*engine
)
5058 INIT_LIST_HEAD(&engine
->active_list
);
5059 INIT_LIST_HEAD(&engine
->request_list
);
5063 i915_gem_load_init_fences(struct drm_i915_private
*dev_priv
)
5065 struct drm_device
*dev
= dev_priv
->dev
;
5067 if (INTEL_INFO(dev_priv
)->gen
>= 7 && !IS_VALLEYVIEW(dev_priv
) &&
5068 !IS_CHERRYVIEW(dev_priv
))
5069 dev_priv
->num_fence_regs
= 32;
5070 else if (INTEL_INFO(dev_priv
)->gen
>= 4 || IS_I945G(dev_priv
) ||
5071 IS_I945GM(dev_priv
) || IS_G33(dev_priv
))
5072 dev_priv
->num_fence_regs
= 16;
5074 dev_priv
->num_fence_regs
= 8;
5076 if (intel_vgpu_active(dev
))
5077 dev_priv
->num_fence_regs
=
5078 I915_READ(vgtif_reg(avail_rs
.fence_num
));
5080 /* Initialize fence registers to zero */
5081 i915_gem_restore_fences(dev
);
5083 i915_gem_detect_bit_6_swizzle(dev
);
5087 i915_gem_load_init(struct drm_device
*dev
)
5089 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5093 kmem_cache_create("i915_gem_object",
5094 sizeof(struct drm_i915_gem_object
), 0,
5098 kmem_cache_create("i915_gem_vma",
5099 sizeof(struct i915_vma
), 0,
5102 dev_priv
->requests
=
5103 kmem_cache_create("i915_gem_request",
5104 sizeof(struct drm_i915_gem_request
), 0,
5108 INIT_LIST_HEAD(&dev_priv
->vm_list
);
5109 INIT_LIST_HEAD(&dev_priv
->context_list
);
5110 INIT_LIST_HEAD(&dev_priv
->mm
.unbound_list
);
5111 INIT_LIST_HEAD(&dev_priv
->mm
.bound_list
);
5112 INIT_LIST_HEAD(&dev_priv
->mm
.fence_list
);
5113 for (i
= 0; i
< I915_NUM_ENGINES
; i
++)
5114 init_engine_lists(&dev_priv
->engine
[i
]);
5115 for (i
= 0; i
< I915_MAX_NUM_FENCES
; i
++)
5116 INIT_LIST_HEAD(&dev_priv
->fence_regs
[i
].lru_list
);
5117 INIT_DELAYED_WORK(&dev_priv
->mm
.retire_work
,
5118 i915_gem_retire_work_handler
);
5119 INIT_DELAYED_WORK(&dev_priv
->mm
.idle_work
,
5120 i915_gem_idle_work_handler
);
5121 init_waitqueue_head(&dev_priv
->gpu_error
.reset_queue
);
5123 dev_priv
->relative_constants_mode
= I915_EXEC_CONSTANTS_REL_GENERAL
;
5126 * Set initial sequence number for requests.
5127 * Using this number allows the wraparound to happen early,
5128 * catching any obvious problems.
5130 dev_priv
->next_seqno
= ((u32
)~0 - 0x1100);
5131 dev_priv
->last_seqno
= ((u32
)~0 - 0x1101);
5133 INIT_LIST_HEAD(&dev_priv
->mm
.fence_list
);
5135 init_waitqueue_head(&dev_priv
->pending_flip_queue
);
5137 dev_priv
->mm
.interruptible
= true;
5139 mutex_init(&dev_priv
->fb_tracking
.lock
);
5142 void i915_gem_load_cleanup(struct drm_device
*dev
)
5144 struct drm_i915_private
*dev_priv
= to_i915(dev
);
5146 kmem_cache_destroy(dev_priv
->requests
);
5147 kmem_cache_destroy(dev_priv
->vmas
);
5148 kmem_cache_destroy(dev_priv
->objects
);
5151 void i915_gem_release(struct drm_device
*dev
, struct drm_file
*file
)
5153 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
5155 /* Clean up our request list when the client is going away, so that
5156 * later retire_requests won't dereference our soon-to-be-gone
5159 spin_lock(&file_priv
->mm
.lock
);
5160 while (!list_empty(&file_priv
->mm
.request_list
)) {
5161 struct drm_i915_gem_request
*request
;
5163 request
= list_first_entry(&file_priv
->mm
.request_list
,
5164 struct drm_i915_gem_request
,
5166 list_del(&request
->client_list
);
5167 request
->file_priv
= NULL
;
5169 spin_unlock(&file_priv
->mm
.lock
);
5171 if (!list_empty(&file_priv
->rps
.link
)) {
5172 spin_lock(&to_i915(dev
)->rps
.client_lock
);
5173 list_del(&file_priv
->rps
.link
);
5174 spin_unlock(&to_i915(dev
)->rps
.client_lock
);
5178 int i915_gem_open(struct drm_device
*dev
, struct drm_file
*file
)
5180 struct drm_i915_file_private
*file_priv
;
5183 DRM_DEBUG_DRIVER("\n");
5185 file_priv
= kzalloc(sizeof(*file_priv
), GFP_KERNEL
);
5189 file
->driver_priv
= file_priv
;
5190 file_priv
->dev_priv
= dev
->dev_private
;
5191 file_priv
->file
= file
;
5192 INIT_LIST_HEAD(&file_priv
->rps
.link
);
5194 spin_lock_init(&file_priv
->mm
.lock
);
5195 INIT_LIST_HEAD(&file_priv
->mm
.request_list
);
5197 file_priv
->bsd_ring
= -1;
5199 ret
= i915_gem_context_open(dev
, file
);
5207 * i915_gem_track_fb - update frontbuffer tracking
5208 * @old: current GEM buffer for the frontbuffer slots
5209 * @new: new GEM buffer for the frontbuffer slots
5210 * @frontbuffer_bits: bitmask of frontbuffer slots
5212 * This updates the frontbuffer tracking bits @frontbuffer_bits by clearing them
5213 * from @old and setting them in @new. Both @old and @new can be NULL.
5215 void i915_gem_track_fb(struct drm_i915_gem_object
*old
,
5216 struct drm_i915_gem_object
*new,
5217 unsigned frontbuffer_bits
)
5220 WARN_ON(!mutex_is_locked(&old
->base
.dev
->struct_mutex
));
5221 WARN_ON(!(old
->frontbuffer_bits
& frontbuffer_bits
));
5222 old
->frontbuffer_bits
&= ~frontbuffer_bits
;
5226 WARN_ON(!mutex_is_locked(&new->base
.dev
->struct_mutex
));
5227 WARN_ON(new->frontbuffer_bits
& frontbuffer_bits
);
5228 new->frontbuffer_bits
|= frontbuffer_bits
;
5232 /* All the new VM stuff */
5233 u64
i915_gem_obj_offset(struct drm_i915_gem_object
*o
,
5234 struct i915_address_space
*vm
)
5236 struct drm_i915_private
*dev_priv
= o
->base
.dev
->dev_private
;
5237 struct i915_vma
*vma
;
5239 WARN_ON(vm
== &dev_priv
->mm
.aliasing_ppgtt
->base
);
5241 list_for_each_entry(vma
, &o
->vma_list
, obj_link
) {
5243 vma
->ggtt_view
.type
!= I915_GGTT_VIEW_NORMAL
)
5246 return vma
->node
.start
;
5249 WARN(1, "%s vma for this object not found.\n",
5250 i915_is_ggtt(vm
) ? "global" : "ppgtt");
5254 u64
i915_gem_obj_ggtt_offset_view(struct drm_i915_gem_object
*o
,
5255 const struct i915_ggtt_view
*view
)
5257 struct drm_i915_private
*dev_priv
= to_i915(o
->base
.dev
);
5258 struct i915_ggtt
*ggtt
= &dev_priv
->ggtt
;
5259 struct i915_vma
*vma
;
5261 list_for_each_entry(vma
, &o
->vma_list
, obj_link
)
5262 if (vma
->vm
== &ggtt
->base
&&
5263 i915_ggtt_view_equal(&vma
->ggtt_view
, view
))
5264 return vma
->node
.start
;
5266 WARN(1, "global vma for this object not found. (view=%u)\n", view
->type
);
5270 bool i915_gem_obj_bound(struct drm_i915_gem_object
*o
,
5271 struct i915_address_space
*vm
)
5273 struct i915_vma
*vma
;
5275 list_for_each_entry(vma
, &o
->vma_list
, obj_link
) {
5277 vma
->ggtt_view
.type
!= I915_GGTT_VIEW_NORMAL
)
5279 if (vma
->vm
== vm
&& drm_mm_node_allocated(&vma
->node
))
5286 bool i915_gem_obj_ggtt_bound_view(struct drm_i915_gem_object
*o
,
5287 const struct i915_ggtt_view
*view
)
5289 struct drm_i915_private
*dev_priv
= to_i915(o
->base
.dev
);
5290 struct i915_ggtt
*ggtt
= &dev_priv
->ggtt
;
5291 struct i915_vma
*vma
;
5293 list_for_each_entry(vma
, &o
->vma_list
, obj_link
)
5294 if (vma
->vm
== &ggtt
->base
&&
5295 i915_ggtt_view_equal(&vma
->ggtt_view
, view
) &&
5296 drm_mm_node_allocated(&vma
->node
))
5302 bool i915_gem_obj_bound_any(struct drm_i915_gem_object
*o
)
5304 struct i915_vma
*vma
;
5306 list_for_each_entry(vma
, &o
->vma_list
, obj_link
)
5307 if (drm_mm_node_allocated(&vma
->node
))
5313 unsigned long i915_gem_obj_size(struct drm_i915_gem_object
*o
,
5314 struct i915_address_space
*vm
)
5316 struct drm_i915_private
*dev_priv
= o
->base
.dev
->dev_private
;
5317 struct i915_vma
*vma
;
5319 WARN_ON(vm
== &dev_priv
->mm
.aliasing_ppgtt
->base
);
5321 BUG_ON(list_empty(&o
->vma_list
));
5323 list_for_each_entry(vma
, &o
->vma_list
, obj_link
) {
5325 vma
->ggtt_view
.type
!= I915_GGTT_VIEW_NORMAL
)
5328 return vma
->node
.size
;
5333 bool i915_gem_obj_is_pinned(struct drm_i915_gem_object
*obj
)
5335 struct i915_vma
*vma
;
5336 list_for_each_entry(vma
, &obj
->vma_list
, obj_link
)
5337 if (vma
->pin_count
> 0)
5343 /* Like i915_gem_object_get_page(), but mark the returned page dirty */
5345 i915_gem_object_get_dirty_page(struct drm_i915_gem_object
*obj
, int n
)
5349 /* Only default objects have per-page dirty tracking */
5350 if (WARN_ON((obj
->ops
->flags
& I915_GEM_OBJECT_HAS_STRUCT_PAGE
) == 0))
5353 page
= i915_gem_object_get_page(obj
, n
);
5354 set_page_dirty(page
);
5358 /* Allocate a new GEM object and fill it with the supplied data */
5359 struct drm_i915_gem_object
*
5360 i915_gem_object_create_from_data(struct drm_device
*dev
,
5361 const void *data
, size_t size
)
5363 struct drm_i915_gem_object
*obj
;
5364 struct sg_table
*sg
;
5368 obj
= i915_gem_alloc_object(dev
, round_up(size
, PAGE_SIZE
));
5369 if (IS_ERR_OR_NULL(obj
))
5372 ret
= i915_gem_object_set_to_cpu_domain(obj
, true);
5376 ret
= i915_gem_object_get_pages(obj
);
5380 i915_gem_object_pin_pages(obj
);
5382 bytes
= sg_copy_from_buffer(sg
->sgl
, sg
->nents
, (void *)data
, size
);
5383 obj
->dirty
= 1; /* Backing store is now out of date */
5384 i915_gem_object_unpin_pages(obj
);
5386 if (WARN_ON(bytes
!= size
)) {
5387 DRM_ERROR("Incomplete copy, wrote %zu of %zu", bytes
, size
);
5395 drm_gem_object_unreference(&obj
->base
);
5396 return ERR_PTR(ret
);