2 * Copyright © 2008 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eric Anholt <eric@anholt.net>
29 #include <drm/i915_drm.h>
31 #include "i915_trace.h"
32 #include "intel_drv.h"
33 #include <linux/shmem_fs.h>
34 #include <linux/slab.h>
35 #include <linux/swap.h>
36 #include <linux/pci.h>
37 #include <linux/dma-buf.h>
39 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object
*obj
);
40 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object
*obj
);
41 static __must_check
int i915_gem_object_bind_to_gtt(struct drm_i915_gem_object
*obj
,
43 bool map_and_fenceable
,
45 static int i915_gem_phys_pwrite(struct drm_device
*dev
,
46 struct drm_i915_gem_object
*obj
,
47 struct drm_i915_gem_pwrite
*args
,
48 struct drm_file
*file
);
50 static void i915_gem_write_fence(struct drm_device
*dev
, int reg
,
51 struct drm_i915_gem_object
*obj
);
52 static void i915_gem_object_update_fence(struct drm_i915_gem_object
*obj
,
53 struct drm_i915_fence_reg
*fence
,
56 static int i915_gem_inactive_shrink(struct shrinker
*shrinker
,
57 struct shrink_control
*sc
);
58 static long i915_gem_purge(struct drm_i915_private
*dev_priv
, long target
);
59 static void i915_gem_shrink_all(struct drm_i915_private
*dev_priv
);
60 static void i915_gem_object_truncate(struct drm_i915_gem_object
*obj
);
62 static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object
*obj
)
65 i915_gem_release_mmap(obj
);
67 /* As we do not have an associated fence register, we will force
68 * a tiling change if we ever need to acquire one.
70 obj
->fence_dirty
= false;
71 obj
->fence_reg
= I915_FENCE_REG_NONE
;
74 /* some bookkeeping */
75 static void i915_gem_info_add_obj(struct drm_i915_private
*dev_priv
,
78 dev_priv
->mm
.object_count
++;
79 dev_priv
->mm
.object_memory
+= size
;
82 static void i915_gem_info_remove_obj(struct drm_i915_private
*dev_priv
,
85 dev_priv
->mm
.object_count
--;
86 dev_priv
->mm
.object_memory
-= size
;
90 i915_gem_wait_for_error(struct i915_gpu_error
*error
)
94 #define EXIT_COND (!i915_reset_in_progress(error))
98 /* GPU is already declared terminally dead, give up. */
99 if (i915_terminally_wedged(error
))
103 * Only wait 10 seconds for the gpu reset to complete to avoid hanging
104 * userspace. If it takes that long something really bad is going on and
105 * we should simply try to bail out and fail as gracefully as possible.
107 ret
= wait_event_interruptible_timeout(error
->reset_queue
,
111 DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
113 } else if (ret
< 0) {
121 int i915_mutex_lock_interruptible(struct drm_device
*dev
)
123 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
126 ret
= i915_gem_wait_for_error(&dev_priv
->gpu_error
);
130 ret
= mutex_lock_interruptible(&dev
->struct_mutex
);
134 WARN_ON(i915_verify_lists(dev
));
139 i915_gem_object_is_inactive(struct drm_i915_gem_object
*obj
)
141 return obj
->gtt_space
&& !obj
->active
;
145 i915_gem_init_ioctl(struct drm_device
*dev
, void *data
,
146 struct drm_file
*file
)
148 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
149 struct drm_i915_gem_init
*args
= data
;
151 if (drm_core_check_feature(dev
, DRIVER_MODESET
))
154 if (args
->gtt_start
>= args
->gtt_end
||
155 (args
->gtt_end
| args
->gtt_start
) & (PAGE_SIZE
- 1))
158 /* GEM with user mode setting was never supported on ilk and later. */
159 if (INTEL_INFO(dev
)->gen
>= 5)
162 mutex_lock(&dev
->struct_mutex
);
163 i915_gem_setup_global_gtt(dev
, args
->gtt_start
, args
->gtt_end
,
165 dev_priv
->gtt
.mappable_end
= args
->gtt_end
;
166 mutex_unlock(&dev
->struct_mutex
);
172 i915_gem_get_aperture_ioctl(struct drm_device
*dev
, void *data
,
173 struct drm_file
*file
)
175 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
176 struct drm_i915_gem_get_aperture
*args
= data
;
177 struct drm_i915_gem_object
*obj
;
181 mutex_lock(&dev
->struct_mutex
);
182 list_for_each_entry(obj
, &dev_priv
->mm
.bound_list
, global_list
)
184 pinned
+= obj
->gtt_space
->size
;
185 mutex_unlock(&dev
->struct_mutex
);
187 args
->aper_size
= dev_priv
->gtt
.total
;
188 args
->aper_available_size
= args
->aper_size
- pinned
;
193 void *i915_gem_object_alloc(struct drm_device
*dev
)
195 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
196 return kmem_cache_alloc(dev_priv
->slab
, GFP_KERNEL
| __GFP_ZERO
);
199 void i915_gem_object_free(struct drm_i915_gem_object
*obj
)
201 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
202 kmem_cache_free(dev_priv
->slab
, obj
);
206 i915_gem_create(struct drm_file
*file
,
207 struct drm_device
*dev
,
211 struct drm_i915_gem_object
*obj
;
215 size
= roundup(size
, PAGE_SIZE
);
219 /* Allocate the new object */
220 obj
= i915_gem_alloc_object(dev
, size
);
224 ret
= drm_gem_handle_create(file
, &obj
->base
, &handle
);
226 drm_gem_object_release(&obj
->base
);
227 i915_gem_info_remove_obj(dev
->dev_private
, obj
->base
.size
);
228 i915_gem_object_free(obj
);
232 /* drop reference from allocate - handle holds it now */
233 drm_gem_object_unreference(&obj
->base
);
234 trace_i915_gem_object_create(obj
);
241 i915_gem_dumb_create(struct drm_file
*file
,
242 struct drm_device
*dev
,
243 struct drm_mode_create_dumb
*args
)
245 /* have to work out size/pitch and return them */
246 args
->pitch
= ALIGN(args
->width
* ((args
->bpp
+ 7) / 8), 64);
247 args
->size
= args
->pitch
* args
->height
;
248 return i915_gem_create(file
, dev
,
249 args
->size
, &args
->handle
);
252 int i915_gem_dumb_destroy(struct drm_file
*file
,
253 struct drm_device
*dev
,
256 return drm_gem_handle_delete(file
, handle
);
260 * Creates a new mm object and returns a handle to it.
263 i915_gem_create_ioctl(struct drm_device
*dev
, void *data
,
264 struct drm_file
*file
)
266 struct drm_i915_gem_create
*args
= data
;
268 return i915_gem_create(file
, dev
,
269 args
->size
, &args
->handle
);
273 __copy_to_user_swizzled(char __user
*cpu_vaddr
,
274 const char *gpu_vaddr
, int gpu_offset
,
277 int ret
, cpu_offset
= 0;
280 int cacheline_end
= ALIGN(gpu_offset
+ 1, 64);
281 int this_length
= min(cacheline_end
- gpu_offset
, length
);
282 int swizzled_gpu_offset
= gpu_offset
^ 64;
284 ret
= __copy_to_user(cpu_vaddr
+ cpu_offset
,
285 gpu_vaddr
+ swizzled_gpu_offset
,
290 cpu_offset
+= this_length
;
291 gpu_offset
+= this_length
;
292 length
-= this_length
;
299 __copy_from_user_swizzled(char *gpu_vaddr
, int gpu_offset
,
300 const char __user
*cpu_vaddr
,
303 int ret
, cpu_offset
= 0;
306 int cacheline_end
= ALIGN(gpu_offset
+ 1, 64);
307 int this_length
= min(cacheline_end
- gpu_offset
, length
);
308 int swizzled_gpu_offset
= gpu_offset
^ 64;
310 ret
= __copy_from_user(gpu_vaddr
+ swizzled_gpu_offset
,
311 cpu_vaddr
+ cpu_offset
,
316 cpu_offset
+= this_length
;
317 gpu_offset
+= this_length
;
318 length
-= this_length
;
324 /* Per-page copy function for the shmem pread fastpath.
325 * Flushes invalid cachelines before reading the target if
326 * needs_clflush is set. */
328 shmem_pread_fast(struct page
*page
, int shmem_page_offset
, int page_length
,
329 char __user
*user_data
,
330 bool page_do_bit17_swizzling
, bool needs_clflush
)
335 if (unlikely(page_do_bit17_swizzling
))
338 vaddr
= kmap_atomic(page
);
340 drm_clflush_virt_range(vaddr
+ shmem_page_offset
,
342 ret
= __copy_to_user_inatomic(user_data
,
343 vaddr
+ shmem_page_offset
,
345 kunmap_atomic(vaddr
);
347 return ret
? -EFAULT
: 0;
351 shmem_clflush_swizzled_range(char *addr
, unsigned long length
,
354 if (unlikely(swizzled
)) {
355 unsigned long start
= (unsigned long) addr
;
356 unsigned long end
= (unsigned long) addr
+ length
;
358 /* For swizzling simply ensure that we always flush both
359 * channels. Lame, but simple and it works. Swizzled
360 * pwrite/pread is far from a hotpath - current userspace
361 * doesn't use it at all. */
362 start
= round_down(start
, 128);
363 end
= round_up(end
, 128);
365 drm_clflush_virt_range((void *)start
, end
- start
);
367 drm_clflush_virt_range(addr
, length
);
372 /* Only difference to the fast-path function is that this can handle bit17
373 * and uses non-atomic copy and kmap functions. */
375 shmem_pread_slow(struct page
*page
, int shmem_page_offset
, int page_length
,
376 char __user
*user_data
,
377 bool page_do_bit17_swizzling
, bool needs_clflush
)
384 shmem_clflush_swizzled_range(vaddr
+ shmem_page_offset
,
386 page_do_bit17_swizzling
);
388 if (page_do_bit17_swizzling
)
389 ret
= __copy_to_user_swizzled(user_data
,
390 vaddr
, shmem_page_offset
,
393 ret
= __copy_to_user(user_data
,
394 vaddr
+ shmem_page_offset
,
398 return ret
? - EFAULT
: 0;
402 i915_gem_shmem_pread(struct drm_device
*dev
,
403 struct drm_i915_gem_object
*obj
,
404 struct drm_i915_gem_pread
*args
,
405 struct drm_file
*file
)
407 char __user
*user_data
;
410 int shmem_page_offset
, page_length
, ret
= 0;
411 int obj_do_bit17_swizzling
, page_do_bit17_swizzling
;
413 int needs_clflush
= 0;
414 struct sg_page_iter sg_iter
;
416 user_data
= to_user_ptr(args
->data_ptr
);
419 obj_do_bit17_swizzling
= i915_gem_object_needs_bit17_swizzle(obj
);
421 if (!(obj
->base
.read_domains
& I915_GEM_DOMAIN_CPU
)) {
422 /* If we're not in the cpu read domain, set ourself into the gtt
423 * read domain and manually flush cachelines (if required). This
424 * optimizes for the case when the gpu will dirty the data
425 * anyway again before the next pread happens. */
426 if (obj
->cache_level
== I915_CACHE_NONE
)
428 if (obj
->gtt_space
) {
429 ret
= i915_gem_object_set_to_gtt_domain(obj
, false);
435 ret
= i915_gem_object_get_pages(obj
);
439 i915_gem_object_pin_pages(obj
);
441 offset
= args
->offset
;
443 for_each_sg_page(obj
->pages
->sgl
, &sg_iter
, obj
->pages
->nents
,
444 offset
>> PAGE_SHIFT
) {
445 struct page
*page
= sg_page_iter_page(&sg_iter
);
450 /* Operation in this page
452 * shmem_page_offset = offset within page in shmem file
453 * page_length = bytes to copy for this page
455 shmem_page_offset
= offset_in_page(offset
);
456 page_length
= remain
;
457 if ((shmem_page_offset
+ page_length
) > PAGE_SIZE
)
458 page_length
= PAGE_SIZE
- shmem_page_offset
;
460 page_do_bit17_swizzling
= obj_do_bit17_swizzling
&&
461 (page_to_phys(page
) & (1 << 17)) != 0;
463 ret
= shmem_pread_fast(page
, shmem_page_offset
, page_length
,
464 user_data
, page_do_bit17_swizzling
,
469 mutex_unlock(&dev
->struct_mutex
);
472 ret
= fault_in_multipages_writeable(user_data
, remain
);
473 /* Userspace is tricking us, but we've already clobbered
474 * its pages with the prefault and promised to write the
475 * data up to the first fault. Hence ignore any errors
476 * and just continue. */
481 ret
= shmem_pread_slow(page
, shmem_page_offset
, page_length
,
482 user_data
, page_do_bit17_swizzling
,
485 mutex_lock(&dev
->struct_mutex
);
488 mark_page_accessed(page
);
493 remain
-= page_length
;
494 user_data
+= page_length
;
495 offset
+= page_length
;
499 i915_gem_object_unpin_pages(obj
);
505 * Reads data from the object referenced by handle.
507 * On error, the contents of *data are undefined.
510 i915_gem_pread_ioctl(struct drm_device
*dev
, void *data
,
511 struct drm_file
*file
)
513 struct drm_i915_gem_pread
*args
= data
;
514 struct drm_i915_gem_object
*obj
;
520 if (!access_ok(VERIFY_WRITE
,
521 to_user_ptr(args
->data_ptr
),
525 ret
= i915_mutex_lock_interruptible(dev
);
529 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
530 if (&obj
->base
== NULL
) {
535 /* Bounds check source. */
536 if (args
->offset
> obj
->base
.size
||
537 args
->size
> obj
->base
.size
- args
->offset
) {
542 /* prime objects have no backing filp to GEM pread/pwrite
545 if (!obj
->base
.filp
) {
550 trace_i915_gem_object_pread(obj
, args
->offset
, args
->size
);
552 ret
= i915_gem_shmem_pread(dev
, obj
, args
, file
);
555 drm_gem_object_unreference(&obj
->base
);
557 mutex_unlock(&dev
->struct_mutex
);
561 /* This is the fast write path which cannot handle
562 * page faults in the source data
566 fast_user_write(struct io_mapping
*mapping
,
567 loff_t page_base
, int page_offset
,
568 char __user
*user_data
,
571 void __iomem
*vaddr_atomic
;
573 unsigned long unwritten
;
575 vaddr_atomic
= io_mapping_map_atomic_wc(mapping
, page_base
);
576 /* We can use the cpu mem copy function because this is X86. */
577 vaddr
= (void __force
*)vaddr_atomic
+ page_offset
;
578 unwritten
= __copy_from_user_inatomic_nocache(vaddr
,
580 io_mapping_unmap_atomic(vaddr_atomic
);
585 * This is the fast pwrite path, where we copy the data directly from the
586 * user into the GTT, uncached.
589 i915_gem_gtt_pwrite_fast(struct drm_device
*dev
,
590 struct drm_i915_gem_object
*obj
,
591 struct drm_i915_gem_pwrite
*args
,
592 struct drm_file
*file
)
594 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
596 loff_t offset
, page_base
;
597 char __user
*user_data
;
598 int page_offset
, page_length
, ret
;
600 ret
= i915_gem_object_pin(obj
, 0, true, true);
604 ret
= i915_gem_object_set_to_gtt_domain(obj
, true);
608 ret
= i915_gem_object_put_fence(obj
);
612 user_data
= to_user_ptr(args
->data_ptr
);
615 offset
= obj
->gtt_offset
+ args
->offset
;
618 /* Operation in this page
620 * page_base = page offset within aperture
621 * page_offset = offset within page
622 * page_length = bytes to copy for this page
624 page_base
= offset
& PAGE_MASK
;
625 page_offset
= offset_in_page(offset
);
626 page_length
= remain
;
627 if ((page_offset
+ remain
) > PAGE_SIZE
)
628 page_length
= PAGE_SIZE
- page_offset
;
630 /* If we get a fault while copying data, then (presumably) our
631 * source page isn't available. Return the error and we'll
632 * retry in the slow path.
634 if (fast_user_write(dev_priv
->gtt
.mappable
, page_base
,
635 page_offset
, user_data
, page_length
)) {
640 remain
-= page_length
;
641 user_data
+= page_length
;
642 offset
+= page_length
;
646 i915_gem_object_unpin(obj
);
651 /* Per-page copy function for the shmem pwrite fastpath.
652 * Flushes invalid cachelines before writing to the target if
653 * needs_clflush_before is set and flushes out any written cachelines after
654 * writing if needs_clflush is set. */
656 shmem_pwrite_fast(struct page
*page
, int shmem_page_offset
, int page_length
,
657 char __user
*user_data
,
658 bool page_do_bit17_swizzling
,
659 bool needs_clflush_before
,
660 bool needs_clflush_after
)
665 if (unlikely(page_do_bit17_swizzling
))
668 vaddr
= kmap_atomic(page
);
669 if (needs_clflush_before
)
670 drm_clflush_virt_range(vaddr
+ shmem_page_offset
,
672 ret
= __copy_from_user_inatomic_nocache(vaddr
+ shmem_page_offset
,
675 if (needs_clflush_after
)
676 drm_clflush_virt_range(vaddr
+ shmem_page_offset
,
678 kunmap_atomic(vaddr
);
680 return ret
? -EFAULT
: 0;
683 /* Only difference to the fast-path function is that this can handle bit17
684 * and uses non-atomic copy and kmap functions. */
686 shmem_pwrite_slow(struct page
*page
, int shmem_page_offset
, int page_length
,
687 char __user
*user_data
,
688 bool page_do_bit17_swizzling
,
689 bool needs_clflush_before
,
690 bool needs_clflush_after
)
696 if (unlikely(needs_clflush_before
|| page_do_bit17_swizzling
))
697 shmem_clflush_swizzled_range(vaddr
+ shmem_page_offset
,
699 page_do_bit17_swizzling
);
700 if (page_do_bit17_swizzling
)
701 ret
= __copy_from_user_swizzled(vaddr
, shmem_page_offset
,
705 ret
= __copy_from_user(vaddr
+ shmem_page_offset
,
708 if (needs_clflush_after
)
709 shmem_clflush_swizzled_range(vaddr
+ shmem_page_offset
,
711 page_do_bit17_swizzling
);
714 return ret
? -EFAULT
: 0;
718 i915_gem_shmem_pwrite(struct drm_device
*dev
,
719 struct drm_i915_gem_object
*obj
,
720 struct drm_i915_gem_pwrite
*args
,
721 struct drm_file
*file
)
725 char __user
*user_data
;
726 int shmem_page_offset
, page_length
, ret
= 0;
727 int obj_do_bit17_swizzling
, page_do_bit17_swizzling
;
728 int hit_slowpath
= 0;
729 int needs_clflush_after
= 0;
730 int needs_clflush_before
= 0;
731 struct sg_page_iter sg_iter
;
733 user_data
= to_user_ptr(args
->data_ptr
);
736 obj_do_bit17_swizzling
= i915_gem_object_needs_bit17_swizzle(obj
);
738 if (obj
->base
.write_domain
!= I915_GEM_DOMAIN_CPU
) {
739 /* If we're not in the cpu write domain, set ourself into the gtt
740 * write domain and manually flush cachelines (if required). This
741 * optimizes for the case when the gpu will use the data
742 * right away and we therefore have to clflush anyway. */
743 if (obj
->cache_level
== I915_CACHE_NONE
)
744 needs_clflush_after
= 1;
745 if (obj
->gtt_space
) {
746 ret
= i915_gem_object_set_to_gtt_domain(obj
, true);
751 /* Same trick applies for invalidate partially written cachelines before
753 if (!(obj
->base
.read_domains
& I915_GEM_DOMAIN_CPU
)
754 && obj
->cache_level
== I915_CACHE_NONE
)
755 needs_clflush_before
= 1;
757 ret
= i915_gem_object_get_pages(obj
);
761 i915_gem_object_pin_pages(obj
);
763 offset
= args
->offset
;
766 for_each_sg_page(obj
->pages
->sgl
, &sg_iter
, obj
->pages
->nents
,
767 offset
>> PAGE_SHIFT
) {
768 struct page
*page
= sg_page_iter_page(&sg_iter
);
769 int partial_cacheline_write
;
774 /* Operation in this page
776 * shmem_page_offset = offset within page in shmem file
777 * page_length = bytes to copy for this page
779 shmem_page_offset
= offset_in_page(offset
);
781 page_length
= remain
;
782 if ((shmem_page_offset
+ page_length
) > PAGE_SIZE
)
783 page_length
= PAGE_SIZE
- shmem_page_offset
;
785 /* If we don't overwrite a cacheline completely we need to be
786 * careful to have up-to-date data by first clflushing. Don't
787 * overcomplicate things and flush the entire patch. */
788 partial_cacheline_write
= needs_clflush_before
&&
789 ((shmem_page_offset
| page_length
)
790 & (boot_cpu_data
.x86_clflush_size
- 1));
792 page_do_bit17_swizzling
= obj_do_bit17_swizzling
&&
793 (page_to_phys(page
) & (1 << 17)) != 0;
795 ret
= shmem_pwrite_fast(page
, shmem_page_offset
, page_length
,
796 user_data
, page_do_bit17_swizzling
,
797 partial_cacheline_write
,
798 needs_clflush_after
);
803 mutex_unlock(&dev
->struct_mutex
);
804 ret
= shmem_pwrite_slow(page
, shmem_page_offset
, page_length
,
805 user_data
, page_do_bit17_swizzling
,
806 partial_cacheline_write
,
807 needs_clflush_after
);
809 mutex_lock(&dev
->struct_mutex
);
812 set_page_dirty(page
);
813 mark_page_accessed(page
);
818 remain
-= page_length
;
819 user_data
+= page_length
;
820 offset
+= page_length
;
824 i915_gem_object_unpin_pages(obj
);
828 * Fixup: Flush cpu caches in case we didn't flush the dirty
829 * cachelines in-line while writing and the object moved
830 * out of the cpu write domain while we've dropped the lock.
832 if (!needs_clflush_after
&&
833 obj
->base
.write_domain
!= I915_GEM_DOMAIN_CPU
) {
834 i915_gem_clflush_object(obj
);
835 i915_gem_chipset_flush(dev
);
839 if (needs_clflush_after
)
840 i915_gem_chipset_flush(dev
);
846 * Writes data to the object referenced by handle.
848 * On error, the contents of the buffer that were to be modified are undefined.
851 i915_gem_pwrite_ioctl(struct drm_device
*dev
, void *data
,
852 struct drm_file
*file
)
854 struct drm_i915_gem_pwrite
*args
= data
;
855 struct drm_i915_gem_object
*obj
;
861 if (!access_ok(VERIFY_READ
,
862 to_user_ptr(args
->data_ptr
),
866 ret
= fault_in_multipages_readable(to_user_ptr(args
->data_ptr
),
871 ret
= i915_mutex_lock_interruptible(dev
);
875 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
876 if (&obj
->base
== NULL
) {
881 /* Bounds check destination. */
882 if (args
->offset
> obj
->base
.size
||
883 args
->size
> obj
->base
.size
- args
->offset
) {
888 /* prime objects have no backing filp to GEM pread/pwrite
891 if (!obj
->base
.filp
) {
896 trace_i915_gem_object_pwrite(obj
, args
->offset
, args
->size
);
899 /* We can only do the GTT pwrite on untiled buffers, as otherwise
900 * it would end up going through the fenced access, and we'll get
901 * different detiling behavior between reading and writing.
902 * pread/pwrite currently are reading and writing from the CPU
903 * perspective, requiring manual detiling by the client.
906 ret
= i915_gem_phys_pwrite(dev
, obj
, args
, file
);
910 if (obj
->cache_level
== I915_CACHE_NONE
&&
911 obj
->tiling_mode
== I915_TILING_NONE
&&
912 obj
->base
.write_domain
!= I915_GEM_DOMAIN_CPU
) {
913 ret
= i915_gem_gtt_pwrite_fast(dev
, obj
, args
, file
);
914 /* Note that the gtt paths might fail with non-page-backed user
915 * pointers (e.g. gtt mappings when moving data between
916 * textures). Fallback to the shmem path in that case. */
919 if (ret
== -EFAULT
|| ret
== -ENOSPC
)
920 ret
= i915_gem_shmem_pwrite(dev
, obj
, args
, file
);
923 drm_gem_object_unreference(&obj
->base
);
925 mutex_unlock(&dev
->struct_mutex
);
930 i915_gem_check_wedge(struct i915_gpu_error
*error
,
933 if (i915_reset_in_progress(error
)) {
934 /* Non-interruptible callers can't handle -EAGAIN, hence return
935 * -EIO unconditionally for these. */
939 /* Recovery complete, but the reset failed ... */
940 if (i915_terminally_wedged(error
))
950 * Compare seqno against outstanding lazy request. Emit a request if they are
954 i915_gem_check_olr(struct intel_ring_buffer
*ring
, u32 seqno
)
958 BUG_ON(!mutex_is_locked(&ring
->dev
->struct_mutex
));
961 if (seqno
== ring
->outstanding_lazy_request
)
962 ret
= i915_add_request(ring
, NULL
);
968 * __wait_seqno - wait until execution of seqno has finished
969 * @ring: the ring expected to report seqno
971 * @reset_counter: reset sequence associated with the given seqno
972 * @interruptible: do an interruptible wait (normally yes)
973 * @timeout: in - how long to wait (NULL forever); out - how much time remaining
975 * Note: It is of utmost importance that the passed in seqno and reset_counter
976 * values have been read by the caller in an smp safe manner. Where read-side
977 * locks are involved, it is sufficient to read the reset_counter before
978 * unlocking the lock that protects the seqno. For lockless tricks, the
979 * reset_counter _must_ be read before, and an appropriate smp_rmb must be
982 * Returns 0 if the seqno was found within the alloted time. Else returns the
983 * errno with remaining time filled in timeout argument.
985 static int __wait_seqno(struct intel_ring_buffer
*ring
, u32 seqno
,
986 unsigned reset_counter
,
987 bool interruptible
, struct timespec
*timeout
)
989 drm_i915_private_t
*dev_priv
= ring
->dev
->dev_private
;
990 struct timespec before
, now
, wait_time
={1,0};
991 unsigned long timeout_jiffies
;
993 bool wait_forever
= true;
996 if (i915_seqno_passed(ring
->get_seqno(ring
, true), seqno
))
999 trace_i915_gem_request_wait_begin(ring
, seqno
);
1001 if (timeout
!= NULL
) {
1002 wait_time
= *timeout
;
1003 wait_forever
= false;
1006 timeout_jiffies
= timespec_to_jiffies(&wait_time
);
1008 if (WARN_ON(!ring
->irq_get(ring
)))
1011 /* Record current time in case interrupted by signal, or wedged * */
1012 getrawmonotonic(&before
);
1015 (i915_seqno_passed(ring->get_seqno(ring, false), seqno) || \
1016 i915_reset_in_progress(&dev_priv->gpu_error) || \
1017 reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter))
1020 end
= wait_event_interruptible_timeout(ring
->irq_queue
,
1024 end
= wait_event_timeout(ring
->irq_queue
, EXIT_COND
,
1027 /* We need to check whether any gpu reset happened in between
1028 * the caller grabbing the seqno and now ... */
1029 if (reset_counter
!= atomic_read(&dev_priv
->gpu_error
.reset_counter
))
1032 /* ... but upgrade the -EGAIN to an -EIO if the gpu is truely
1034 ret
= i915_gem_check_wedge(&dev_priv
->gpu_error
, interruptible
);
1037 } while (end
== 0 && wait_forever
);
1039 getrawmonotonic(&now
);
1041 ring
->irq_put(ring
);
1042 trace_i915_gem_request_wait_end(ring
, seqno
);
1046 struct timespec sleep_time
= timespec_sub(now
, before
);
1047 *timeout
= timespec_sub(*timeout
, sleep_time
);
1048 if (!timespec_valid(timeout
)) /* i.e. negative time remains */
1049 set_normalized_timespec(timeout
, 0, 0);
1054 case -EAGAIN
: /* Wedged */
1055 case -ERESTARTSYS
: /* Signal */
1057 case 0: /* Timeout */
1059 default: /* Completed */
1060 WARN_ON(end
< 0); /* We're not aware of other errors */
1066 * Waits for a sequence number to be signaled, and cleans up the
1067 * request and object lists appropriately for that event.
1070 i915_wait_seqno(struct intel_ring_buffer
*ring
, uint32_t seqno
)
1072 struct drm_device
*dev
= ring
->dev
;
1073 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1074 bool interruptible
= dev_priv
->mm
.interruptible
;
1077 BUG_ON(!mutex_is_locked(&dev
->struct_mutex
));
1080 ret
= i915_gem_check_wedge(&dev_priv
->gpu_error
, interruptible
);
1084 ret
= i915_gem_check_olr(ring
, seqno
);
1088 return __wait_seqno(ring
, seqno
,
1089 atomic_read(&dev_priv
->gpu_error
.reset_counter
),
1090 interruptible
, NULL
);
1094 * Ensures that all rendering to the object has completed and the object is
1095 * safe to unbind from the GTT or access from the CPU.
1097 static __must_check
int
1098 i915_gem_object_wait_rendering(struct drm_i915_gem_object
*obj
,
1101 struct intel_ring_buffer
*ring
= obj
->ring
;
1105 seqno
= readonly
? obj
->last_write_seqno
: obj
->last_read_seqno
;
1109 ret
= i915_wait_seqno(ring
, seqno
);
1113 i915_gem_retire_requests_ring(ring
);
1115 /* Manually manage the write flush as we may have not yet
1116 * retired the buffer.
1118 if (obj
->last_write_seqno
&&
1119 i915_seqno_passed(seqno
, obj
->last_write_seqno
)) {
1120 obj
->last_write_seqno
= 0;
1121 obj
->base
.write_domain
&= ~I915_GEM_GPU_DOMAINS
;
1127 /* A nonblocking variant of the above wait. This is a highly dangerous routine
1128 * as the object state may change during this call.
1130 static __must_check
int
1131 i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object
*obj
,
1134 struct drm_device
*dev
= obj
->base
.dev
;
1135 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1136 struct intel_ring_buffer
*ring
= obj
->ring
;
1137 unsigned reset_counter
;
1141 BUG_ON(!mutex_is_locked(&dev
->struct_mutex
));
1142 BUG_ON(!dev_priv
->mm
.interruptible
);
1144 seqno
= readonly
? obj
->last_write_seqno
: obj
->last_read_seqno
;
1148 ret
= i915_gem_check_wedge(&dev_priv
->gpu_error
, true);
1152 ret
= i915_gem_check_olr(ring
, seqno
);
1156 reset_counter
= atomic_read(&dev_priv
->gpu_error
.reset_counter
);
1157 mutex_unlock(&dev
->struct_mutex
);
1158 ret
= __wait_seqno(ring
, seqno
, reset_counter
, true, NULL
);
1159 mutex_lock(&dev
->struct_mutex
);
1161 i915_gem_retire_requests_ring(ring
);
1163 /* Manually manage the write flush as we may have not yet
1164 * retired the buffer.
1166 if (obj
->last_write_seqno
&&
1167 i915_seqno_passed(seqno
, obj
->last_write_seqno
)) {
1168 obj
->last_write_seqno
= 0;
1169 obj
->base
.write_domain
&= ~I915_GEM_GPU_DOMAINS
;
1176 * Called when user space prepares to use an object with the CPU, either
1177 * through the mmap ioctl's mapping or a GTT mapping.
1180 i915_gem_set_domain_ioctl(struct drm_device
*dev
, void *data
,
1181 struct drm_file
*file
)
1183 struct drm_i915_gem_set_domain
*args
= data
;
1184 struct drm_i915_gem_object
*obj
;
1185 uint32_t read_domains
= args
->read_domains
;
1186 uint32_t write_domain
= args
->write_domain
;
1189 /* Only handle setting domains to types used by the CPU. */
1190 if (write_domain
& I915_GEM_GPU_DOMAINS
)
1193 if (read_domains
& I915_GEM_GPU_DOMAINS
)
1196 /* Having something in the write domain implies it's in the read
1197 * domain, and only that read domain. Enforce that in the request.
1199 if (write_domain
!= 0 && read_domains
!= write_domain
)
1202 ret
= i915_mutex_lock_interruptible(dev
);
1206 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
1207 if (&obj
->base
== NULL
) {
1212 /* Try to flush the object off the GPU without holding the lock.
1213 * We will repeat the flush holding the lock in the normal manner
1214 * to catch cases where we are gazumped.
1216 ret
= i915_gem_object_wait_rendering__nonblocking(obj
, !write_domain
);
1220 if (read_domains
& I915_GEM_DOMAIN_GTT
) {
1221 ret
= i915_gem_object_set_to_gtt_domain(obj
, write_domain
!= 0);
1223 /* Silently promote "you're not bound, there was nothing to do"
1224 * to success, since the client was just asking us to
1225 * make sure everything was done.
1230 ret
= i915_gem_object_set_to_cpu_domain(obj
, write_domain
!= 0);
1234 drm_gem_object_unreference(&obj
->base
);
1236 mutex_unlock(&dev
->struct_mutex
);
1241 * Called when user space has done writes to this buffer
1244 i915_gem_sw_finish_ioctl(struct drm_device
*dev
, void *data
,
1245 struct drm_file
*file
)
1247 struct drm_i915_gem_sw_finish
*args
= data
;
1248 struct drm_i915_gem_object
*obj
;
1251 ret
= i915_mutex_lock_interruptible(dev
);
1255 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
1256 if (&obj
->base
== NULL
) {
1261 /* Pinned buffers may be scanout, so flush the cache */
1263 i915_gem_object_flush_cpu_write_domain(obj
);
1265 drm_gem_object_unreference(&obj
->base
);
1267 mutex_unlock(&dev
->struct_mutex
);
1272 * Maps the contents of an object, returning the address it is mapped
1275 * While the mapping holds a reference on the contents of the object, it doesn't
1276 * imply a ref on the object itself.
1279 i915_gem_mmap_ioctl(struct drm_device
*dev
, void *data
,
1280 struct drm_file
*file
)
1282 struct drm_i915_gem_mmap
*args
= data
;
1283 struct drm_gem_object
*obj
;
1286 obj
= drm_gem_object_lookup(dev
, file
, args
->handle
);
1290 /* prime objects have no backing filp to GEM mmap
1294 drm_gem_object_unreference_unlocked(obj
);
1298 addr
= vm_mmap(obj
->filp
, 0, args
->size
,
1299 PROT_READ
| PROT_WRITE
, MAP_SHARED
,
1301 drm_gem_object_unreference_unlocked(obj
);
1302 if (IS_ERR((void *)addr
))
1305 args
->addr_ptr
= (uint64_t) addr
;
1311 * i915_gem_fault - fault a page into the GTT
1312 * vma: VMA in question
1315 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1316 * from userspace. The fault handler takes care of binding the object to
1317 * the GTT (if needed), allocating and programming a fence register (again,
1318 * only if needed based on whether the old reg is still valid or the object
1319 * is tiled) and inserting a new PTE into the faulting process.
1321 * Note that the faulting process may involve evicting existing objects
1322 * from the GTT and/or fence registers to make room. So performance may
1323 * suffer if the GTT working set is large or there are few fence registers
1326 int i915_gem_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1328 struct drm_i915_gem_object
*obj
= to_intel_bo(vma
->vm_private_data
);
1329 struct drm_device
*dev
= obj
->base
.dev
;
1330 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
1331 pgoff_t page_offset
;
1334 bool write
= !!(vmf
->flags
& FAULT_FLAG_WRITE
);
1336 /* We don't use vmf->pgoff since that has the fake offset */
1337 page_offset
= ((unsigned long)vmf
->virtual_address
- vma
->vm_start
) >>
1340 ret
= i915_mutex_lock_interruptible(dev
);
1344 trace_i915_gem_object_fault(obj
, page_offset
, true, write
);
1346 /* Access to snoopable pages through the GTT is incoherent. */
1347 if (obj
->cache_level
!= I915_CACHE_NONE
&& !HAS_LLC(dev
)) {
1352 /* Now bind it into the GTT if needed */
1353 ret
= i915_gem_object_pin(obj
, 0, true, false);
1357 ret
= i915_gem_object_set_to_gtt_domain(obj
, write
);
1361 ret
= i915_gem_object_get_fence(obj
);
1365 obj
->fault_mappable
= true;
1367 pfn
= ((dev_priv
->gtt
.mappable_base
+ obj
->gtt_offset
) >> PAGE_SHIFT
) +
1370 /* Finally, remap it using the new GTT offset */
1371 ret
= vm_insert_pfn(vma
, (unsigned long)vmf
->virtual_address
, pfn
);
1373 i915_gem_object_unpin(obj
);
1375 mutex_unlock(&dev
->struct_mutex
);
1379 /* If this -EIO is due to a gpu hang, give the reset code a
1380 * chance to clean up the mess. Otherwise return the proper
1382 if (i915_terminally_wedged(&dev_priv
->gpu_error
))
1383 return VM_FAULT_SIGBUS
;
1385 /* Give the error handler a chance to run and move the
1386 * objects off the GPU active list. Next time we service the
1387 * fault, we should be able to transition the page into the
1388 * GTT without touching the GPU (and so avoid further
1389 * EIO/EGAIN). If the GPU is wedged, then there is no issue
1390 * with coherency, just lost writes.
1398 * EBUSY is ok: this just means that another thread
1399 * already did the job.
1401 return VM_FAULT_NOPAGE
;
1403 return VM_FAULT_OOM
;
1405 return VM_FAULT_SIGBUS
;
1407 WARN_ONCE(ret
, "unhandled error in i915_gem_fault: %i\n", ret
);
1408 return VM_FAULT_SIGBUS
;
1413 * i915_gem_release_mmap - remove physical page mappings
1414 * @obj: obj in question
1416 * Preserve the reservation of the mmapping with the DRM core code, but
1417 * relinquish ownership of the pages back to the system.
1419 * It is vital that we remove the page mapping if we have mapped a tiled
1420 * object through the GTT and then lose the fence register due to
1421 * resource pressure. Similarly if the object has been moved out of the
1422 * aperture, than pages mapped into userspace must be revoked. Removing the
1423 * mapping will then trigger a page fault on the next user access, allowing
1424 * fixup by i915_gem_fault().
1427 i915_gem_release_mmap(struct drm_i915_gem_object
*obj
)
1429 if (!obj
->fault_mappable
)
1432 if (obj
->base
.dev
->dev_mapping
)
1433 unmap_mapping_range(obj
->base
.dev
->dev_mapping
,
1434 (loff_t
)obj
->base
.map_list
.hash
.key
<<PAGE_SHIFT
,
1437 obj
->fault_mappable
= false;
1441 i915_gem_get_gtt_size(struct drm_device
*dev
, uint32_t size
, int tiling_mode
)
1445 if (INTEL_INFO(dev
)->gen
>= 4 ||
1446 tiling_mode
== I915_TILING_NONE
)
1449 /* Previous chips need a power-of-two fence region when tiling */
1450 if (INTEL_INFO(dev
)->gen
== 3)
1451 gtt_size
= 1024*1024;
1453 gtt_size
= 512*1024;
1455 while (gtt_size
< size
)
1462 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1463 * @obj: object to check
1465 * Return the required GTT alignment for an object, taking into account
1466 * potential fence register mapping.
1469 i915_gem_get_gtt_alignment(struct drm_device
*dev
, uint32_t size
,
1470 int tiling_mode
, bool fenced
)
1473 * Minimum alignment is 4k (GTT page size), but might be greater
1474 * if a fence register is needed for the object.
1476 if (INTEL_INFO(dev
)->gen
>= 4 || (!fenced
&& IS_G33(dev
)) ||
1477 tiling_mode
== I915_TILING_NONE
)
1481 * Previous chips need to be aligned to the size of the smallest
1482 * fence register that can contain the object.
1484 return i915_gem_get_gtt_size(dev
, size
, tiling_mode
);
1487 static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object
*obj
)
1489 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
1492 if (obj
->base
.map_list
.map
)
1495 dev_priv
->mm
.shrinker_no_lock_stealing
= true;
1497 ret
= drm_gem_create_mmap_offset(&obj
->base
);
1501 /* Badly fragmented mmap space? The only way we can recover
1502 * space is by destroying unwanted objects. We can't randomly release
1503 * mmap_offsets as userspace expects them to be persistent for the
1504 * lifetime of the objects. The closest we can is to release the
1505 * offsets on purgeable objects by truncating it and marking it purged,
1506 * which prevents userspace from ever using that object again.
1508 i915_gem_purge(dev_priv
, obj
->base
.size
>> PAGE_SHIFT
);
1509 ret
= drm_gem_create_mmap_offset(&obj
->base
);
1513 i915_gem_shrink_all(dev_priv
);
1514 ret
= drm_gem_create_mmap_offset(&obj
->base
);
1516 dev_priv
->mm
.shrinker_no_lock_stealing
= false;
1521 static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object
*obj
)
1523 if (!obj
->base
.map_list
.map
)
1526 drm_gem_free_mmap_offset(&obj
->base
);
1530 i915_gem_mmap_gtt(struct drm_file
*file
,
1531 struct drm_device
*dev
,
1535 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1536 struct drm_i915_gem_object
*obj
;
1539 ret
= i915_mutex_lock_interruptible(dev
);
1543 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, handle
));
1544 if (&obj
->base
== NULL
) {
1549 if (obj
->base
.size
> dev_priv
->gtt
.mappable_end
) {
1554 if (obj
->madv
!= I915_MADV_WILLNEED
) {
1555 DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1560 ret
= i915_gem_object_create_mmap_offset(obj
);
1564 *offset
= (u64
)obj
->base
.map_list
.hash
.key
<< PAGE_SHIFT
;
1567 drm_gem_object_unreference(&obj
->base
);
1569 mutex_unlock(&dev
->struct_mutex
);
1574 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1576 * @data: GTT mapping ioctl data
1577 * @file: GEM object info
1579 * Simply returns the fake offset to userspace so it can mmap it.
1580 * The mmap call will end up in drm_gem_mmap(), which will set things
1581 * up so we can get faults in the handler above.
1583 * The fault handler will take care of binding the object into the GTT
1584 * (since it may have been evicted to make room for something), allocating
1585 * a fence register, and mapping the appropriate aperture address into
1589 i915_gem_mmap_gtt_ioctl(struct drm_device
*dev
, void *data
,
1590 struct drm_file
*file
)
1592 struct drm_i915_gem_mmap_gtt
*args
= data
;
1594 return i915_gem_mmap_gtt(file
, dev
, args
->handle
, &args
->offset
);
1597 /* Immediately discard the backing storage */
1599 i915_gem_object_truncate(struct drm_i915_gem_object
*obj
)
1601 struct inode
*inode
;
1603 i915_gem_object_free_mmap_offset(obj
);
1605 if (obj
->base
.filp
== NULL
)
1608 /* Our goal here is to return as much of the memory as
1609 * is possible back to the system as we are called from OOM.
1610 * To do this we must instruct the shmfs to drop all of its
1611 * backing pages, *now*.
1613 inode
= file_inode(obj
->base
.filp
);
1614 shmem_truncate_range(inode
, 0, (loff_t
)-1);
1616 obj
->madv
= __I915_MADV_PURGED
;
1620 i915_gem_object_is_purgeable(struct drm_i915_gem_object
*obj
)
1622 return obj
->madv
== I915_MADV_DONTNEED
;
1626 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object
*obj
)
1628 struct sg_page_iter sg_iter
;
1631 BUG_ON(obj
->madv
== __I915_MADV_PURGED
);
1633 ret
= i915_gem_object_set_to_cpu_domain(obj
, true);
1635 /* In the event of a disaster, abandon all caches and
1636 * hope for the best.
1638 WARN_ON(ret
!= -EIO
);
1639 i915_gem_clflush_object(obj
);
1640 obj
->base
.read_domains
= obj
->base
.write_domain
= I915_GEM_DOMAIN_CPU
;
1643 if (i915_gem_object_needs_bit17_swizzle(obj
))
1644 i915_gem_object_save_bit_17_swizzle(obj
);
1646 if (obj
->madv
== I915_MADV_DONTNEED
)
1649 for_each_sg_page(obj
->pages
->sgl
, &sg_iter
, obj
->pages
->nents
, 0) {
1650 struct page
*page
= sg_page_iter_page(&sg_iter
);
1653 set_page_dirty(page
);
1655 if (obj
->madv
== I915_MADV_WILLNEED
)
1656 mark_page_accessed(page
);
1658 page_cache_release(page
);
1662 sg_free_table(obj
->pages
);
1667 i915_gem_object_put_pages(struct drm_i915_gem_object
*obj
)
1669 const struct drm_i915_gem_object_ops
*ops
= obj
->ops
;
1671 if (obj
->pages
== NULL
)
1674 BUG_ON(obj
->gtt_space
);
1676 if (obj
->pages_pin_count
)
1679 /* ->put_pages might need to allocate memory for the bit17 swizzle
1680 * array, hence protect them from being reaped by removing them from gtt
1682 list_del(&obj
->global_list
);
1684 ops
->put_pages(obj
);
1687 if (i915_gem_object_is_purgeable(obj
))
1688 i915_gem_object_truncate(obj
);
1694 __i915_gem_shrink(struct drm_i915_private
*dev_priv
, long target
,
1695 bool purgeable_only
)
1697 struct drm_i915_gem_object
*obj
, *next
;
1700 list_for_each_entry_safe(obj
, next
,
1701 &dev_priv
->mm
.unbound_list
,
1703 if ((i915_gem_object_is_purgeable(obj
) || !purgeable_only
) &&
1704 i915_gem_object_put_pages(obj
) == 0) {
1705 count
+= obj
->base
.size
>> PAGE_SHIFT
;
1706 if (count
>= target
)
1711 list_for_each_entry_safe(obj
, next
,
1712 &dev_priv
->mm
.inactive_list
,
1714 if ((i915_gem_object_is_purgeable(obj
) || !purgeable_only
) &&
1715 i915_gem_object_unbind(obj
) == 0 &&
1716 i915_gem_object_put_pages(obj
) == 0) {
1717 count
+= obj
->base
.size
>> PAGE_SHIFT
;
1718 if (count
>= target
)
1727 i915_gem_purge(struct drm_i915_private
*dev_priv
, long target
)
1729 return __i915_gem_shrink(dev_priv
, target
, true);
1733 i915_gem_shrink_all(struct drm_i915_private
*dev_priv
)
1735 struct drm_i915_gem_object
*obj
, *next
;
1737 i915_gem_evict_everything(dev_priv
->dev
);
1739 list_for_each_entry_safe(obj
, next
, &dev_priv
->mm
.unbound_list
,
1741 i915_gem_object_put_pages(obj
);
1745 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object
*obj
)
1747 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
1749 struct address_space
*mapping
;
1750 struct sg_table
*st
;
1751 struct scatterlist
*sg
;
1752 struct sg_page_iter sg_iter
;
1754 unsigned long last_pfn
= 0; /* suppress gcc warning */
1757 /* Assert that the object is not currently in any GPU domain. As it
1758 * wasn't in the GTT, there shouldn't be any way it could have been in
1761 BUG_ON(obj
->base
.read_domains
& I915_GEM_GPU_DOMAINS
);
1762 BUG_ON(obj
->base
.write_domain
& I915_GEM_GPU_DOMAINS
);
1764 st
= kmalloc(sizeof(*st
), GFP_KERNEL
);
1768 page_count
= obj
->base
.size
/ PAGE_SIZE
;
1769 if (sg_alloc_table(st
, page_count
, GFP_KERNEL
)) {
1775 /* Get the list of pages out of our struct file. They'll be pinned
1776 * at this point until we release them.
1778 * Fail silently without starting the shrinker
1780 mapping
= file_inode(obj
->base
.filp
)->i_mapping
;
1781 gfp
= mapping_gfp_mask(mapping
);
1782 gfp
|= __GFP_NORETRY
| __GFP_NOWARN
| __GFP_NO_KSWAPD
;
1783 gfp
&= ~(__GFP_IO
| __GFP_WAIT
);
1786 for (i
= 0; i
< page_count
; i
++) {
1787 page
= shmem_read_mapping_page_gfp(mapping
, i
, gfp
);
1789 i915_gem_purge(dev_priv
, page_count
);
1790 page
= shmem_read_mapping_page_gfp(mapping
, i
, gfp
);
1793 /* We've tried hard to allocate the memory by reaping
1794 * our own buffer, now let the real VM do its job and
1795 * go down in flames if truly OOM.
1797 gfp
&= ~(__GFP_NORETRY
| __GFP_NOWARN
| __GFP_NO_KSWAPD
);
1798 gfp
|= __GFP_IO
| __GFP_WAIT
;
1800 i915_gem_shrink_all(dev_priv
);
1801 page
= shmem_read_mapping_page_gfp(mapping
, i
, gfp
);
1805 gfp
|= __GFP_NORETRY
| __GFP_NOWARN
| __GFP_NO_KSWAPD
;
1806 gfp
&= ~(__GFP_IO
| __GFP_WAIT
);
1809 if (!i
|| page_to_pfn(page
) != last_pfn
+ 1) {
1813 sg_set_page(sg
, page
, PAGE_SIZE
, 0);
1815 sg
->length
+= PAGE_SIZE
;
1817 last_pfn
= page_to_pfn(page
);
1823 if (i915_gem_object_needs_bit17_swizzle(obj
))
1824 i915_gem_object_do_bit_17_swizzle(obj
);
1830 for_each_sg_page(st
->sgl
, &sg_iter
, st
->nents
, 0)
1831 page_cache_release(sg_page_iter_page(&sg_iter
));
1834 return PTR_ERR(page
);
1837 /* Ensure that the associated pages are gathered from the backing storage
1838 * and pinned into our object. i915_gem_object_get_pages() may be called
1839 * multiple times before they are released by a single call to
1840 * i915_gem_object_put_pages() - once the pages are no longer referenced
1841 * either as a result of memory pressure (reaping pages under the shrinker)
1842 * or as the object is itself released.
1845 i915_gem_object_get_pages(struct drm_i915_gem_object
*obj
)
1847 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
1848 const struct drm_i915_gem_object_ops
*ops
= obj
->ops
;
1854 if (obj
->madv
!= I915_MADV_WILLNEED
) {
1855 DRM_ERROR("Attempting to obtain a purgeable object\n");
1859 BUG_ON(obj
->pages_pin_count
);
1861 ret
= ops
->get_pages(obj
);
1865 list_add_tail(&obj
->global_list
, &dev_priv
->mm
.unbound_list
);
1870 i915_gem_object_move_to_active(struct drm_i915_gem_object
*obj
,
1871 struct intel_ring_buffer
*ring
)
1873 struct drm_device
*dev
= obj
->base
.dev
;
1874 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1875 u32 seqno
= intel_ring_get_seqno(ring
);
1877 BUG_ON(ring
== NULL
);
1880 /* Add a reference if we're newly entering the active list. */
1882 drm_gem_object_reference(&obj
->base
);
1886 /* Move from whatever list we were on to the tail of execution. */
1887 list_move_tail(&obj
->mm_list
, &dev_priv
->mm
.active_list
);
1888 list_move_tail(&obj
->ring_list
, &ring
->active_list
);
1890 obj
->last_read_seqno
= seqno
;
1892 if (obj
->fenced_gpu_access
) {
1893 obj
->last_fenced_seqno
= seqno
;
1895 /* Bump MRU to take account of the delayed flush */
1896 if (obj
->fence_reg
!= I915_FENCE_REG_NONE
) {
1897 struct drm_i915_fence_reg
*reg
;
1899 reg
= &dev_priv
->fence_regs
[obj
->fence_reg
];
1900 list_move_tail(®
->lru_list
,
1901 &dev_priv
->mm
.fence_list
);
1907 i915_gem_object_move_to_inactive(struct drm_i915_gem_object
*obj
)
1909 struct drm_device
*dev
= obj
->base
.dev
;
1910 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1912 BUG_ON(obj
->base
.write_domain
& ~I915_GEM_GPU_DOMAINS
);
1913 BUG_ON(!obj
->active
);
1915 list_move_tail(&obj
->mm_list
, &dev_priv
->mm
.inactive_list
);
1917 list_del_init(&obj
->ring_list
);
1920 obj
->last_read_seqno
= 0;
1921 obj
->last_write_seqno
= 0;
1922 obj
->base
.write_domain
= 0;
1924 obj
->last_fenced_seqno
= 0;
1925 obj
->fenced_gpu_access
= false;
1928 drm_gem_object_unreference(&obj
->base
);
1930 WARN_ON(i915_verify_lists(dev
));
1934 i915_gem_init_seqno(struct drm_device
*dev
, u32 seqno
)
1936 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1937 struct intel_ring_buffer
*ring
;
1940 /* Carefully retire all requests without writing to the rings */
1941 for_each_ring(ring
, dev_priv
, i
) {
1942 ret
= intel_ring_idle(ring
);
1946 i915_gem_retire_requests(dev
);
1948 /* Finally reset hw state */
1949 for_each_ring(ring
, dev_priv
, i
) {
1950 intel_ring_init_seqno(ring
, seqno
);
1952 for (j
= 0; j
< ARRAY_SIZE(ring
->sync_seqno
); j
++)
1953 ring
->sync_seqno
[j
] = 0;
1959 int i915_gem_set_seqno(struct drm_device
*dev
, u32 seqno
)
1961 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1967 /* HWS page needs to be set less than what we
1968 * will inject to ring
1970 ret
= i915_gem_init_seqno(dev
, seqno
- 1);
1974 /* Carefully set the last_seqno value so that wrap
1975 * detection still works
1977 dev_priv
->next_seqno
= seqno
;
1978 dev_priv
->last_seqno
= seqno
- 1;
1979 if (dev_priv
->last_seqno
== 0)
1980 dev_priv
->last_seqno
--;
1986 i915_gem_get_seqno(struct drm_device
*dev
, u32
*seqno
)
1988 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1990 /* reserve 0 for non-seqno */
1991 if (dev_priv
->next_seqno
== 0) {
1992 int ret
= i915_gem_init_seqno(dev
, 0);
1996 dev_priv
->next_seqno
= 1;
1999 *seqno
= dev_priv
->last_seqno
= dev_priv
->next_seqno
++;
2003 int __i915_add_request(struct intel_ring_buffer
*ring
,
2004 struct drm_file
*file
,
2005 struct drm_i915_gem_object
*obj
,
2008 drm_i915_private_t
*dev_priv
= ring
->dev
->dev_private
;
2009 struct drm_i915_gem_request
*request
;
2010 u32 request_ring_position
, request_start
;
2014 request_start
= intel_ring_get_tail(ring
);
2016 * Emit any outstanding flushes - execbuf can fail to emit the flush
2017 * after having emitted the batchbuffer command. Hence we need to fix
2018 * things up similar to emitting the lazy request. The difference here
2019 * is that the flush _must_ happen before the next request, no matter
2022 ret
= intel_ring_flush_all_caches(ring
);
2026 request
= kmalloc(sizeof(*request
), GFP_KERNEL
);
2027 if (request
== NULL
)
2031 /* Record the position of the start of the request so that
2032 * should we detect the updated seqno part-way through the
2033 * GPU processing the request, we never over-estimate the
2034 * position of the head.
2036 request_ring_position
= intel_ring_get_tail(ring
);
2038 ret
= ring
->add_request(ring
);
2044 request
->seqno
= intel_ring_get_seqno(ring
);
2045 request
->ring
= ring
;
2046 request
->head
= request_start
;
2047 request
->tail
= request_ring_position
;
2048 request
->ctx
= ring
->last_context
;
2049 request
->batch_obj
= obj
;
2051 /* Whilst this request exists, batch_obj will be on the
2052 * active_list, and so will hold the active reference. Only when this
2053 * request is retired will the the batch_obj be moved onto the
2054 * inactive_list and lose its active reference. Hence we do not need
2055 * to explicitly hold another reference here.
2059 i915_gem_context_reference(request
->ctx
);
2061 request
->emitted_jiffies
= jiffies
;
2062 was_empty
= list_empty(&ring
->request_list
);
2063 list_add_tail(&request
->list
, &ring
->request_list
);
2064 request
->file_priv
= NULL
;
2067 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
2069 spin_lock(&file_priv
->mm
.lock
);
2070 request
->file_priv
= file_priv
;
2071 list_add_tail(&request
->client_list
,
2072 &file_priv
->mm
.request_list
);
2073 spin_unlock(&file_priv
->mm
.lock
);
2076 trace_i915_gem_request_add(ring
, request
->seqno
);
2077 ring
->outstanding_lazy_request
= 0;
2079 if (!dev_priv
->mm
.suspended
) {
2080 if (i915_enable_hangcheck
) {
2081 mod_timer(&dev_priv
->gpu_error
.hangcheck_timer
,
2082 round_jiffies_up(jiffies
+ DRM_I915_HANGCHECK_JIFFIES
));
2085 queue_delayed_work(dev_priv
->wq
,
2086 &dev_priv
->mm
.retire_work
,
2087 round_jiffies_up_relative(HZ
));
2088 intel_mark_busy(dev_priv
->dev
);
2093 *out_seqno
= request
->seqno
;
2098 i915_gem_request_remove_from_client(struct drm_i915_gem_request
*request
)
2100 struct drm_i915_file_private
*file_priv
= request
->file_priv
;
2105 spin_lock(&file_priv
->mm
.lock
);
2106 if (request
->file_priv
) {
2107 list_del(&request
->client_list
);
2108 request
->file_priv
= NULL
;
2110 spin_unlock(&file_priv
->mm
.lock
);
2113 static bool i915_head_inside_object(u32 acthd
, struct drm_i915_gem_object
*obj
)
2115 if (acthd
>= obj
->gtt_offset
&&
2116 acthd
< obj
->gtt_offset
+ obj
->base
.size
)
2122 static bool i915_head_inside_request(const u32 acthd_unmasked
,
2123 const u32 request_start
,
2124 const u32 request_end
)
2126 const u32 acthd
= acthd_unmasked
& HEAD_ADDR
;
2128 if (request_start
< request_end
) {
2129 if (acthd
>= request_start
&& acthd
< request_end
)
2131 } else if (request_start
> request_end
) {
2132 if (acthd
>= request_start
|| acthd
< request_end
)
2139 static bool i915_request_guilty(struct drm_i915_gem_request
*request
,
2140 const u32 acthd
, bool *inside
)
2142 /* There is a possibility that unmasked head address
2143 * pointing inside the ring, matches the batch_obj address range.
2144 * However this is extremely unlikely.
2147 if (request
->batch_obj
) {
2148 if (i915_head_inside_object(acthd
, request
->batch_obj
)) {
2154 if (i915_head_inside_request(acthd
, request
->head
, request
->tail
)) {
2162 static void i915_set_reset_status(struct intel_ring_buffer
*ring
,
2163 struct drm_i915_gem_request
*request
,
2166 struct i915_ctx_hang_stats
*hs
= NULL
;
2167 bool inside
, guilty
;
2169 /* Innocent until proven guilty */
2172 if (ring
->hangcheck
.action
!= wait
&&
2173 i915_request_guilty(request
, acthd
, &inside
)) {
2174 DRM_ERROR("%s hung %s bo (0x%x ctx %d) at 0x%x\n",
2176 inside
? "inside" : "flushing",
2177 request
->batch_obj
?
2178 request
->batch_obj
->gtt_offset
: 0,
2179 request
->ctx
? request
->ctx
->id
: 0,
2185 /* If contexts are disabled or this is the default context, use
2186 * file_priv->reset_state
2188 if (request
->ctx
&& request
->ctx
->id
!= DEFAULT_CONTEXT_ID
)
2189 hs
= &request
->ctx
->hang_stats
;
2190 else if (request
->file_priv
)
2191 hs
= &request
->file_priv
->hang_stats
;
2197 hs
->batch_pending
++;
2201 static void i915_gem_free_request(struct drm_i915_gem_request
*request
)
2203 list_del(&request
->list
);
2204 i915_gem_request_remove_from_client(request
);
2207 i915_gem_context_unreference(request
->ctx
);
2212 static void i915_gem_reset_ring_lists(struct drm_i915_private
*dev_priv
,
2213 struct intel_ring_buffer
*ring
)
2215 u32 completed_seqno
;
2218 acthd
= intel_ring_get_active_head(ring
);
2219 completed_seqno
= ring
->get_seqno(ring
, false);
2221 while (!list_empty(&ring
->request_list
)) {
2222 struct drm_i915_gem_request
*request
;
2224 request
= list_first_entry(&ring
->request_list
,
2225 struct drm_i915_gem_request
,
2228 if (request
->seqno
> completed_seqno
)
2229 i915_set_reset_status(ring
, request
, acthd
);
2231 i915_gem_free_request(request
);
2234 while (!list_empty(&ring
->active_list
)) {
2235 struct drm_i915_gem_object
*obj
;
2237 obj
= list_first_entry(&ring
->active_list
,
2238 struct drm_i915_gem_object
,
2241 i915_gem_object_move_to_inactive(obj
);
2245 static void i915_gem_reset_fences(struct drm_device
*dev
)
2247 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2250 for (i
= 0; i
< dev_priv
->num_fence_regs
; i
++) {
2251 struct drm_i915_fence_reg
*reg
= &dev_priv
->fence_regs
[i
];
2254 i915_gem_object_fence_lost(reg
->obj
);
2256 i915_gem_write_fence(dev
, i
, NULL
);
2260 INIT_LIST_HEAD(®
->lru_list
);
2263 INIT_LIST_HEAD(&dev_priv
->mm
.fence_list
);
2266 void i915_gem_reset(struct drm_device
*dev
)
2268 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2269 struct drm_i915_gem_object
*obj
;
2270 struct intel_ring_buffer
*ring
;
2273 for_each_ring(ring
, dev_priv
, i
)
2274 i915_gem_reset_ring_lists(dev_priv
, ring
);
2276 /* Move everything out of the GPU domains to ensure we do any
2277 * necessary invalidation upon reuse.
2279 list_for_each_entry(obj
,
2280 &dev_priv
->mm
.inactive_list
,
2283 obj
->base
.read_domains
&= ~I915_GEM_GPU_DOMAINS
;
2286 /* The fence registers are invalidated so clear them out */
2287 i915_gem_reset_fences(dev
);
2291 * This function clears the request list as sequence numbers are passed.
2294 i915_gem_retire_requests_ring(struct intel_ring_buffer
*ring
)
2298 if (list_empty(&ring
->request_list
))
2301 WARN_ON(i915_verify_lists(ring
->dev
));
2303 seqno
= ring
->get_seqno(ring
, true);
2305 while (!list_empty(&ring
->request_list
)) {
2306 struct drm_i915_gem_request
*request
;
2308 request
= list_first_entry(&ring
->request_list
,
2309 struct drm_i915_gem_request
,
2312 if (!i915_seqno_passed(seqno
, request
->seqno
))
2315 trace_i915_gem_request_retire(ring
, request
->seqno
);
2316 /* We know the GPU must have read the request to have
2317 * sent us the seqno + interrupt, so use the position
2318 * of tail of the request to update the last known position
2321 ring
->last_retired_head
= request
->tail
;
2323 i915_gem_free_request(request
);
2326 /* Move any buffers on the active list that are no longer referenced
2327 * by the ringbuffer to the flushing/inactive lists as appropriate.
2329 while (!list_empty(&ring
->active_list
)) {
2330 struct drm_i915_gem_object
*obj
;
2332 obj
= list_first_entry(&ring
->active_list
,
2333 struct drm_i915_gem_object
,
2336 if (!i915_seqno_passed(seqno
, obj
->last_read_seqno
))
2339 i915_gem_object_move_to_inactive(obj
);
2342 if (unlikely(ring
->trace_irq_seqno
&&
2343 i915_seqno_passed(seqno
, ring
->trace_irq_seqno
))) {
2344 ring
->irq_put(ring
);
2345 ring
->trace_irq_seqno
= 0;
2348 WARN_ON(i915_verify_lists(ring
->dev
));
2352 i915_gem_retire_requests(struct drm_device
*dev
)
2354 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
2355 struct intel_ring_buffer
*ring
;
2358 for_each_ring(ring
, dev_priv
, i
)
2359 i915_gem_retire_requests_ring(ring
);
2363 i915_gem_retire_work_handler(struct work_struct
*work
)
2365 drm_i915_private_t
*dev_priv
;
2366 struct drm_device
*dev
;
2367 struct intel_ring_buffer
*ring
;
2371 dev_priv
= container_of(work
, drm_i915_private_t
,
2372 mm
.retire_work
.work
);
2373 dev
= dev_priv
->dev
;
2375 /* Come back later if the device is busy... */
2376 if (!mutex_trylock(&dev
->struct_mutex
)) {
2377 queue_delayed_work(dev_priv
->wq
, &dev_priv
->mm
.retire_work
,
2378 round_jiffies_up_relative(HZ
));
2382 i915_gem_retire_requests(dev
);
2384 /* Send a periodic flush down the ring so we don't hold onto GEM
2385 * objects indefinitely.
2388 for_each_ring(ring
, dev_priv
, i
) {
2389 if (ring
->gpu_caches_dirty
)
2390 i915_add_request(ring
, NULL
);
2392 idle
&= list_empty(&ring
->request_list
);
2395 if (!dev_priv
->mm
.suspended
&& !idle
)
2396 queue_delayed_work(dev_priv
->wq
, &dev_priv
->mm
.retire_work
,
2397 round_jiffies_up_relative(HZ
));
2399 intel_mark_idle(dev
);
2401 mutex_unlock(&dev
->struct_mutex
);
2405 * Ensures that an object will eventually get non-busy by flushing any required
2406 * write domains, emitting any outstanding lazy request and retiring and
2407 * completed requests.
2410 i915_gem_object_flush_active(struct drm_i915_gem_object
*obj
)
2415 ret
= i915_gem_check_olr(obj
->ring
, obj
->last_read_seqno
);
2419 i915_gem_retire_requests_ring(obj
->ring
);
2426 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
2427 * @DRM_IOCTL_ARGS: standard ioctl arguments
2429 * Returns 0 if successful, else an error is returned with the remaining time in
2430 * the timeout parameter.
2431 * -ETIME: object is still busy after timeout
2432 * -ERESTARTSYS: signal interrupted the wait
2433 * -ENONENT: object doesn't exist
2434 * Also possible, but rare:
2435 * -EAGAIN: GPU wedged
2437 * -ENODEV: Internal IRQ fail
2438 * -E?: The add request failed
2440 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
2441 * non-zero timeout parameter the wait ioctl will wait for the given number of
2442 * nanoseconds on an object becoming unbusy. Since the wait itself does so
2443 * without holding struct_mutex the object may become re-busied before this
2444 * function completes. A similar but shorter * race condition exists in the busy
2448 i915_gem_wait_ioctl(struct drm_device
*dev
, void *data
, struct drm_file
*file
)
2450 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
2451 struct drm_i915_gem_wait
*args
= data
;
2452 struct drm_i915_gem_object
*obj
;
2453 struct intel_ring_buffer
*ring
= NULL
;
2454 struct timespec timeout_stack
, *timeout
= NULL
;
2455 unsigned reset_counter
;
2459 if (args
->timeout_ns
>= 0) {
2460 timeout_stack
= ns_to_timespec(args
->timeout_ns
);
2461 timeout
= &timeout_stack
;
2464 ret
= i915_mutex_lock_interruptible(dev
);
2468 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->bo_handle
));
2469 if (&obj
->base
== NULL
) {
2470 mutex_unlock(&dev
->struct_mutex
);
2474 /* Need to make sure the object gets inactive eventually. */
2475 ret
= i915_gem_object_flush_active(obj
);
2480 seqno
= obj
->last_read_seqno
;
2487 /* Do this after OLR check to make sure we make forward progress polling
2488 * on this IOCTL with a 0 timeout (like busy ioctl)
2490 if (!args
->timeout_ns
) {
2495 drm_gem_object_unreference(&obj
->base
);
2496 reset_counter
= atomic_read(&dev_priv
->gpu_error
.reset_counter
);
2497 mutex_unlock(&dev
->struct_mutex
);
2499 ret
= __wait_seqno(ring
, seqno
, reset_counter
, true, timeout
);
2501 args
->timeout_ns
= timespec_to_ns(timeout
);
2505 drm_gem_object_unreference(&obj
->base
);
2506 mutex_unlock(&dev
->struct_mutex
);
2511 * i915_gem_object_sync - sync an object to a ring.
2513 * @obj: object which may be in use on another ring.
2514 * @to: ring we wish to use the object on. May be NULL.
2516 * This code is meant to abstract object synchronization with the GPU.
2517 * Calling with NULL implies synchronizing the object with the CPU
2518 * rather than a particular GPU ring.
2520 * Returns 0 if successful, else propagates up the lower layer error.
2523 i915_gem_object_sync(struct drm_i915_gem_object
*obj
,
2524 struct intel_ring_buffer
*to
)
2526 struct intel_ring_buffer
*from
= obj
->ring
;
2530 if (from
== NULL
|| to
== from
)
2533 if (to
== NULL
|| !i915_semaphore_is_enabled(obj
->base
.dev
))
2534 return i915_gem_object_wait_rendering(obj
, false);
2536 idx
= intel_ring_sync_index(from
, to
);
2538 seqno
= obj
->last_read_seqno
;
2539 if (seqno
<= from
->sync_seqno
[idx
])
2542 ret
= i915_gem_check_olr(obj
->ring
, seqno
);
2546 ret
= to
->sync_to(to
, from
, seqno
);
2548 /* We use last_read_seqno because sync_to()
2549 * might have just caused seqno wrap under
2552 from
->sync_seqno
[idx
] = obj
->last_read_seqno
;
2557 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object
*obj
)
2559 u32 old_write_domain
, old_read_domains
;
2561 /* Force a pagefault for domain tracking on next user access */
2562 i915_gem_release_mmap(obj
);
2564 if ((obj
->base
.read_domains
& I915_GEM_DOMAIN_GTT
) == 0)
2567 /* Wait for any direct GTT access to complete */
2570 old_read_domains
= obj
->base
.read_domains
;
2571 old_write_domain
= obj
->base
.write_domain
;
2573 obj
->base
.read_domains
&= ~I915_GEM_DOMAIN_GTT
;
2574 obj
->base
.write_domain
&= ~I915_GEM_DOMAIN_GTT
;
2576 trace_i915_gem_object_change_domain(obj
,
2582 * Unbinds an object from the GTT aperture.
2585 i915_gem_object_unbind(struct drm_i915_gem_object
*obj
)
2587 drm_i915_private_t
*dev_priv
= obj
->base
.dev
->dev_private
;
2590 if (obj
->gtt_space
== NULL
)
2596 BUG_ON(obj
->pages
== NULL
);
2598 ret
= i915_gem_object_finish_gpu(obj
);
2601 /* Continue on if we fail due to EIO, the GPU is hung so we
2602 * should be safe and we need to cleanup or else we might
2603 * cause memory corruption through use-after-free.
2606 i915_gem_object_finish_gtt(obj
);
2608 /* release the fence reg _after_ flushing */
2609 ret
= i915_gem_object_put_fence(obj
);
2613 trace_i915_gem_object_unbind(obj
);
2615 if (obj
->has_global_gtt_mapping
)
2616 i915_gem_gtt_unbind_object(obj
);
2617 if (obj
->has_aliasing_ppgtt_mapping
) {
2618 i915_ppgtt_unbind_object(dev_priv
->mm
.aliasing_ppgtt
, obj
);
2619 obj
->has_aliasing_ppgtt_mapping
= 0;
2621 i915_gem_gtt_finish_object(obj
);
2622 i915_gem_object_unpin_pages(obj
);
2624 list_del(&obj
->mm_list
);
2625 list_move_tail(&obj
->global_list
, &dev_priv
->mm
.unbound_list
);
2626 /* Avoid an unnecessary call to unbind on rebind. */
2627 obj
->map_and_fenceable
= true;
2629 drm_mm_put_block(obj
->gtt_space
);
2630 obj
->gtt_space
= NULL
;
2631 obj
->gtt_offset
= 0;
2636 int i915_gpu_idle(struct drm_device
*dev
)
2638 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
2639 struct intel_ring_buffer
*ring
;
2642 /* Flush everything onto the inactive list. */
2643 for_each_ring(ring
, dev_priv
, i
) {
2644 ret
= i915_switch_context(ring
, NULL
, DEFAULT_CONTEXT_ID
);
2648 ret
= intel_ring_idle(ring
);
2656 static void i965_write_fence_reg(struct drm_device
*dev
, int reg
,
2657 struct drm_i915_gem_object
*obj
)
2659 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
2661 int fence_pitch_shift
;
2664 if (INTEL_INFO(dev
)->gen
>= 6) {
2665 fence_reg
= FENCE_REG_SANDYBRIDGE_0
;
2666 fence_pitch_shift
= SANDYBRIDGE_FENCE_PITCH_SHIFT
;
2668 fence_reg
= FENCE_REG_965_0
;
2669 fence_pitch_shift
= I965_FENCE_PITCH_SHIFT
;
2673 u32 size
= obj
->gtt_space
->size
;
2675 val
= (uint64_t)((obj
->gtt_offset
+ size
- 4096) &
2677 val
|= obj
->gtt_offset
& 0xfffff000;
2678 val
|= (uint64_t)((obj
->stride
/ 128) - 1) << fence_pitch_shift
;
2679 if (obj
->tiling_mode
== I915_TILING_Y
)
2680 val
|= 1 << I965_FENCE_TILING_Y_SHIFT
;
2681 val
|= I965_FENCE_REG_VALID
;
2685 fence_reg
+= reg
* 8;
2686 I915_WRITE64(fence_reg
, val
);
2687 POSTING_READ(fence_reg
);
2690 static void i915_write_fence_reg(struct drm_device
*dev
, int reg
,
2691 struct drm_i915_gem_object
*obj
)
2693 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
2697 u32 size
= obj
->gtt_space
->size
;
2701 WARN((obj
->gtt_offset
& ~I915_FENCE_START_MASK
) ||
2702 (size
& -size
) != size
||
2703 (obj
->gtt_offset
& (size
- 1)),
2704 "object 0x%08x [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
2705 obj
->gtt_offset
, obj
->map_and_fenceable
, size
);
2707 if (obj
->tiling_mode
== I915_TILING_Y
&& HAS_128_BYTE_Y_TILING(dev
))
2712 /* Note: pitch better be a power of two tile widths */
2713 pitch_val
= obj
->stride
/ tile_width
;
2714 pitch_val
= ffs(pitch_val
) - 1;
2716 val
= obj
->gtt_offset
;
2717 if (obj
->tiling_mode
== I915_TILING_Y
)
2718 val
|= 1 << I830_FENCE_TILING_Y_SHIFT
;
2719 val
|= I915_FENCE_SIZE_BITS(size
);
2720 val
|= pitch_val
<< I830_FENCE_PITCH_SHIFT
;
2721 val
|= I830_FENCE_REG_VALID
;
2726 reg
= FENCE_REG_830_0
+ reg
* 4;
2728 reg
= FENCE_REG_945_8
+ (reg
- 8) * 4;
2730 I915_WRITE(reg
, val
);
2734 static void i830_write_fence_reg(struct drm_device
*dev
, int reg
,
2735 struct drm_i915_gem_object
*obj
)
2737 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
2741 u32 size
= obj
->gtt_space
->size
;
2744 WARN((obj
->gtt_offset
& ~I830_FENCE_START_MASK
) ||
2745 (size
& -size
) != size
||
2746 (obj
->gtt_offset
& (size
- 1)),
2747 "object 0x%08x not 512K or pot-size 0x%08x aligned\n",
2748 obj
->gtt_offset
, size
);
2750 pitch_val
= obj
->stride
/ 128;
2751 pitch_val
= ffs(pitch_val
) - 1;
2753 val
= obj
->gtt_offset
;
2754 if (obj
->tiling_mode
== I915_TILING_Y
)
2755 val
|= 1 << I830_FENCE_TILING_Y_SHIFT
;
2756 val
|= I830_FENCE_SIZE_BITS(size
);
2757 val
|= pitch_val
<< I830_FENCE_PITCH_SHIFT
;
2758 val
|= I830_FENCE_REG_VALID
;
2762 I915_WRITE(FENCE_REG_830_0
+ reg
* 4, val
);
2763 POSTING_READ(FENCE_REG_830_0
+ reg
* 4);
2766 inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object
*obj
)
2768 return obj
&& obj
->base
.read_domains
& I915_GEM_DOMAIN_GTT
;
2771 static void i915_gem_write_fence(struct drm_device
*dev
, int reg
,
2772 struct drm_i915_gem_object
*obj
)
2774 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2776 /* Ensure that all CPU reads are completed before installing a fence
2777 * and all writes before removing the fence.
2779 if (i915_gem_object_needs_mb(dev_priv
->fence_regs
[reg
].obj
))
2782 switch (INTEL_INFO(dev
)->gen
) {
2786 case 4: i965_write_fence_reg(dev
, reg
, obj
); break;
2787 case 3: i915_write_fence_reg(dev
, reg
, obj
); break;
2788 case 2: i830_write_fence_reg(dev
, reg
, obj
); break;
2792 /* And similarly be paranoid that no direct access to this region
2793 * is reordered to before the fence is installed.
2795 if (i915_gem_object_needs_mb(obj
))
2799 static inline int fence_number(struct drm_i915_private
*dev_priv
,
2800 struct drm_i915_fence_reg
*fence
)
2802 return fence
- dev_priv
->fence_regs
;
2805 struct write_fence
{
2806 struct drm_device
*dev
;
2807 struct drm_i915_gem_object
*obj
;
2811 static void i915_gem_write_fence__ipi(void *data
)
2813 struct write_fence
*args
= data
;
2815 /* Required for SNB+ with LLC */
2818 /* Required for VLV */
2819 i915_gem_write_fence(args
->dev
, args
->fence
, args
->obj
);
2822 static void i915_gem_object_update_fence(struct drm_i915_gem_object
*obj
,
2823 struct drm_i915_fence_reg
*fence
,
2826 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
2827 struct write_fence args
= {
2828 .dev
= obj
->base
.dev
,
2829 .fence
= fence_number(dev_priv
, fence
),
2830 .obj
= enable
? obj
: NULL
,
2833 /* In order to fully serialize access to the fenced region and
2834 * the update to the fence register we need to take extreme
2835 * measures on SNB+. In theory, the write to the fence register
2836 * flushes all memory transactions before, and coupled with the
2837 * mb() placed around the register write we serialise all memory
2838 * operations with respect to the changes in the tiler. Yet, on
2839 * SNB+ we need to take a step further and emit an explicit wbinvd()
2840 * on each processor in order to manually flush all memory
2841 * transactions before updating the fence register.
2843 * However, Valleyview complicates matter. There the wbinvd is
2844 * insufficient and unlike SNB/IVB requires the serialising
2845 * register write. (Note that that register write by itself is
2846 * conversely not sufficient for SNB+.) To compromise, we do both.
2848 if (INTEL_INFO(args
.dev
)->gen
>= 6)
2849 on_each_cpu(i915_gem_write_fence__ipi
, &args
, 1);
2851 i915_gem_write_fence(args
.dev
, args
.fence
, args
.obj
);
2854 obj
->fence_reg
= args
.fence
;
2856 list_move_tail(&fence
->lru_list
, &dev_priv
->mm
.fence_list
);
2858 obj
->fence_reg
= I915_FENCE_REG_NONE
;
2860 list_del_init(&fence
->lru_list
);
2865 i915_gem_object_wait_fence(struct drm_i915_gem_object
*obj
)
2867 if (obj
->last_fenced_seqno
) {
2868 int ret
= i915_wait_seqno(obj
->ring
, obj
->last_fenced_seqno
);
2872 obj
->last_fenced_seqno
= 0;
2875 obj
->fenced_gpu_access
= false;
2880 i915_gem_object_put_fence(struct drm_i915_gem_object
*obj
)
2882 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
2883 struct drm_i915_fence_reg
*fence
;
2886 ret
= i915_gem_object_wait_fence(obj
);
2890 if (obj
->fence_reg
== I915_FENCE_REG_NONE
)
2893 fence
= &dev_priv
->fence_regs
[obj
->fence_reg
];
2895 i915_gem_object_fence_lost(obj
);
2896 i915_gem_object_update_fence(obj
, fence
, false);
2901 static struct drm_i915_fence_reg
*
2902 i915_find_fence_reg(struct drm_device
*dev
)
2904 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2905 struct drm_i915_fence_reg
*reg
, *avail
;
2908 /* First try to find a free reg */
2910 for (i
= dev_priv
->fence_reg_start
; i
< dev_priv
->num_fence_regs
; i
++) {
2911 reg
= &dev_priv
->fence_regs
[i
];
2915 if (!reg
->pin_count
)
2922 /* None available, try to steal one or wait for a user to finish */
2923 list_for_each_entry(reg
, &dev_priv
->mm
.fence_list
, lru_list
) {
2934 * i915_gem_object_get_fence - set up fencing for an object
2935 * @obj: object to map through a fence reg
2937 * When mapping objects through the GTT, userspace wants to be able to write
2938 * to them without having to worry about swizzling if the object is tiled.
2939 * This function walks the fence regs looking for a free one for @obj,
2940 * stealing one if it can't find any.
2942 * It then sets up the reg based on the object's properties: address, pitch
2943 * and tiling format.
2945 * For an untiled surface, this removes any existing fence.
2948 i915_gem_object_get_fence(struct drm_i915_gem_object
*obj
)
2950 struct drm_device
*dev
= obj
->base
.dev
;
2951 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2952 bool enable
= obj
->tiling_mode
!= I915_TILING_NONE
;
2953 struct drm_i915_fence_reg
*reg
;
2956 /* Have we updated the tiling parameters upon the object and so
2957 * will need to serialise the write to the associated fence register?
2959 if (obj
->fence_dirty
) {
2960 ret
= i915_gem_object_wait_fence(obj
);
2965 /* Just update our place in the LRU if our fence is getting reused. */
2966 if (obj
->fence_reg
!= I915_FENCE_REG_NONE
) {
2967 reg
= &dev_priv
->fence_regs
[obj
->fence_reg
];
2968 if (!obj
->fence_dirty
) {
2969 list_move_tail(®
->lru_list
,
2970 &dev_priv
->mm
.fence_list
);
2973 } else if (enable
) {
2974 reg
= i915_find_fence_reg(dev
);
2979 struct drm_i915_gem_object
*old
= reg
->obj
;
2981 ret
= i915_gem_object_wait_fence(old
);
2985 i915_gem_object_fence_lost(old
);
2990 i915_gem_object_update_fence(obj
, reg
, enable
);
2991 obj
->fence_dirty
= false;
2996 static bool i915_gem_valid_gtt_space(struct drm_device
*dev
,
2997 struct drm_mm_node
*gtt_space
,
2998 unsigned long cache_level
)
3000 struct drm_mm_node
*other
;
3002 /* On non-LLC machines we have to be careful when putting differing
3003 * types of snoopable memory together to avoid the prefetcher
3004 * crossing memory domains and dying.
3009 if (gtt_space
== NULL
)
3012 if (list_empty(>t_space
->node_list
))
3015 other
= list_entry(gtt_space
->node_list
.prev
, struct drm_mm_node
, node_list
);
3016 if (other
->allocated
&& !other
->hole_follows
&& other
->color
!= cache_level
)
3019 other
= list_entry(gtt_space
->node_list
.next
, struct drm_mm_node
, node_list
);
3020 if (other
->allocated
&& !gtt_space
->hole_follows
&& other
->color
!= cache_level
)
3026 static void i915_gem_verify_gtt(struct drm_device
*dev
)
3029 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3030 struct drm_i915_gem_object
*obj
;
3033 list_for_each_entry(obj
, &dev_priv
->mm
.gtt_list
, global_list
) {
3034 if (obj
->gtt_space
== NULL
) {
3035 printk(KERN_ERR
"object found on GTT list with no space reserved\n");
3040 if (obj
->cache_level
!= obj
->gtt_space
->color
) {
3041 printk(KERN_ERR
"object reserved space [%08lx, %08lx] with wrong color, cache_level=%x, color=%lx\n",
3042 obj
->gtt_space
->start
,
3043 obj
->gtt_space
->start
+ obj
->gtt_space
->size
,
3045 obj
->gtt_space
->color
);
3050 if (!i915_gem_valid_gtt_space(dev
,
3052 obj
->cache_level
)) {
3053 printk(KERN_ERR
"invalid GTT space found at [%08lx, %08lx] - color=%x\n",
3054 obj
->gtt_space
->start
,
3055 obj
->gtt_space
->start
+ obj
->gtt_space
->size
,
3067 * Finds free space in the GTT aperture and binds the object there.
3070 i915_gem_object_bind_to_gtt(struct drm_i915_gem_object
*obj
,
3072 bool map_and_fenceable
,
3075 struct drm_device
*dev
= obj
->base
.dev
;
3076 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
3077 struct drm_mm_node
*node
;
3078 u32 size
, fence_size
, fence_alignment
, unfenced_alignment
;
3079 bool mappable
, fenceable
;
3080 size_t gtt_max
= map_and_fenceable
?
3081 dev_priv
->gtt
.mappable_end
: dev_priv
->gtt
.total
;
3084 fence_size
= i915_gem_get_gtt_size(dev
,
3087 fence_alignment
= i915_gem_get_gtt_alignment(dev
,
3089 obj
->tiling_mode
, true);
3090 unfenced_alignment
=
3091 i915_gem_get_gtt_alignment(dev
,
3093 obj
->tiling_mode
, false);
3096 alignment
= map_and_fenceable
? fence_alignment
:
3098 if (map_and_fenceable
&& alignment
& (fence_alignment
- 1)) {
3099 DRM_ERROR("Invalid object alignment requested %u\n", alignment
);
3103 size
= map_and_fenceable
? fence_size
: obj
->base
.size
;
3105 /* If the object is bigger than the entire aperture, reject it early
3106 * before evicting everything in a vain attempt to find space.
3108 if (obj
->base
.size
> gtt_max
) {
3109 DRM_ERROR("Attempting to bind an object larger than the aperture: object=%zd > %s aperture=%ld\n",
3111 map_and_fenceable
? "mappable" : "total",
3116 ret
= i915_gem_object_get_pages(obj
);
3120 i915_gem_object_pin_pages(obj
);
3122 node
= kzalloc(sizeof(*node
), GFP_KERNEL
);
3124 i915_gem_object_unpin_pages(obj
);
3129 ret
= drm_mm_insert_node_in_range_generic(&dev_priv
->mm
.gtt_space
, node
,
3131 obj
->cache_level
, 0, gtt_max
);
3133 ret
= i915_gem_evict_something(dev
, size
, alignment
,
3140 i915_gem_object_unpin_pages(obj
);
3144 if (WARN_ON(!i915_gem_valid_gtt_space(dev
, node
, obj
->cache_level
))) {
3145 i915_gem_object_unpin_pages(obj
);
3146 drm_mm_put_block(node
);
3150 ret
= i915_gem_gtt_prepare_object(obj
);
3152 i915_gem_object_unpin_pages(obj
);
3153 drm_mm_put_block(node
);
3157 list_move_tail(&obj
->global_list
, &dev_priv
->mm
.bound_list
);
3158 list_add_tail(&obj
->mm_list
, &dev_priv
->mm
.inactive_list
);
3160 obj
->gtt_space
= node
;
3161 obj
->gtt_offset
= node
->start
;
3164 node
->size
== fence_size
&&
3165 (node
->start
& (fence_alignment
- 1)) == 0;
3168 obj
->gtt_offset
+ obj
->base
.size
<= dev_priv
->gtt
.mappable_end
;
3170 obj
->map_and_fenceable
= mappable
&& fenceable
;
3172 trace_i915_gem_object_bind(obj
, map_and_fenceable
);
3173 i915_gem_verify_gtt(dev
);
3178 i915_gem_clflush_object(struct drm_i915_gem_object
*obj
)
3180 /* If we don't have a page list set up, then we're not pinned
3181 * to GPU, and we can ignore the cache flush because it'll happen
3182 * again at bind time.
3184 if (obj
->pages
== NULL
)
3188 * Stolen memory is always coherent with the GPU as it is explicitly
3189 * marked as wc by the system, or the system is cache-coherent.
3194 /* If the GPU is snooping the contents of the CPU cache,
3195 * we do not need to manually clear the CPU cache lines. However,
3196 * the caches are only snooped when the render cache is
3197 * flushed/invalidated. As we always have to emit invalidations
3198 * and flushes when moving into and out of the RENDER domain, correct
3199 * snooping behaviour occurs naturally as the result of our domain
3202 if (obj
->cache_level
!= I915_CACHE_NONE
)
3205 trace_i915_gem_object_clflush(obj
);
3207 drm_clflush_sg(obj
->pages
);
3210 /** Flushes the GTT write domain for the object if it's dirty. */
3212 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object
*obj
)
3214 uint32_t old_write_domain
;
3216 if (obj
->base
.write_domain
!= I915_GEM_DOMAIN_GTT
)
3219 /* No actual flushing is required for the GTT write domain. Writes
3220 * to it immediately go to main memory as far as we know, so there's
3221 * no chipset flush. It also doesn't land in render cache.
3223 * However, we do have to enforce the order so that all writes through
3224 * the GTT land before any writes to the device, such as updates to
3229 old_write_domain
= obj
->base
.write_domain
;
3230 obj
->base
.write_domain
= 0;
3232 trace_i915_gem_object_change_domain(obj
,
3233 obj
->base
.read_domains
,
3237 /** Flushes the CPU write domain for the object if it's dirty. */
3239 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object
*obj
)
3241 uint32_t old_write_domain
;
3243 if (obj
->base
.write_domain
!= I915_GEM_DOMAIN_CPU
)
3246 i915_gem_clflush_object(obj
);
3247 i915_gem_chipset_flush(obj
->base
.dev
);
3248 old_write_domain
= obj
->base
.write_domain
;
3249 obj
->base
.write_domain
= 0;
3251 trace_i915_gem_object_change_domain(obj
,
3252 obj
->base
.read_domains
,
3257 * Moves a single object to the GTT read, and possibly write domain.
3259 * This function returns when the move is complete, including waiting on
3263 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object
*obj
, bool write
)
3265 drm_i915_private_t
*dev_priv
= obj
->base
.dev
->dev_private
;
3266 uint32_t old_write_domain
, old_read_domains
;
3269 /* Not valid to be called on unbound objects. */
3270 if (obj
->gtt_space
== NULL
)
3273 if (obj
->base
.write_domain
== I915_GEM_DOMAIN_GTT
)
3276 ret
= i915_gem_object_wait_rendering(obj
, !write
);
3280 i915_gem_object_flush_cpu_write_domain(obj
);
3282 /* Serialise direct access to this object with the barriers for
3283 * coherent writes from the GPU, by effectively invalidating the
3284 * GTT domain upon first access.
3286 if ((obj
->base
.read_domains
& I915_GEM_DOMAIN_GTT
) == 0)
3289 old_write_domain
= obj
->base
.write_domain
;
3290 old_read_domains
= obj
->base
.read_domains
;
3292 /* It should now be out of any other write domains, and we can update
3293 * the domain values for our changes.
3295 BUG_ON((obj
->base
.write_domain
& ~I915_GEM_DOMAIN_GTT
) != 0);
3296 obj
->base
.read_domains
|= I915_GEM_DOMAIN_GTT
;
3298 obj
->base
.read_domains
= I915_GEM_DOMAIN_GTT
;
3299 obj
->base
.write_domain
= I915_GEM_DOMAIN_GTT
;
3303 trace_i915_gem_object_change_domain(obj
,
3307 /* And bump the LRU for this access */
3308 if (i915_gem_object_is_inactive(obj
))
3309 list_move_tail(&obj
->mm_list
, &dev_priv
->mm
.inactive_list
);
3314 int i915_gem_object_set_cache_level(struct drm_i915_gem_object
*obj
,
3315 enum i915_cache_level cache_level
)
3317 struct drm_device
*dev
= obj
->base
.dev
;
3318 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
3321 if (obj
->cache_level
== cache_level
)
3324 if (obj
->pin_count
) {
3325 DRM_DEBUG("can not change the cache level of pinned objects\n");
3329 if (!i915_gem_valid_gtt_space(dev
, obj
->gtt_space
, cache_level
)) {
3330 ret
= i915_gem_object_unbind(obj
);
3335 if (obj
->gtt_space
) {
3336 ret
= i915_gem_object_finish_gpu(obj
);
3340 i915_gem_object_finish_gtt(obj
);
3342 /* Before SandyBridge, you could not use tiling or fence
3343 * registers with snooped memory, so relinquish any fences
3344 * currently pointing to our region in the aperture.
3346 if (INTEL_INFO(dev
)->gen
< 6) {
3347 ret
= i915_gem_object_put_fence(obj
);
3352 if (obj
->has_global_gtt_mapping
)
3353 i915_gem_gtt_bind_object(obj
, cache_level
);
3354 if (obj
->has_aliasing_ppgtt_mapping
)
3355 i915_ppgtt_bind_object(dev_priv
->mm
.aliasing_ppgtt
,
3358 obj
->gtt_space
->color
= cache_level
;
3361 if (cache_level
== I915_CACHE_NONE
) {
3362 u32 old_read_domains
, old_write_domain
;
3364 /* If we're coming from LLC cached, then we haven't
3365 * actually been tracking whether the data is in the
3366 * CPU cache or not, since we only allow one bit set
3367 * in obj->write_domain and have been skipping the clflushes.
3368 * Just set it to the CPU cache for now.
3370 WARN_ON(obj
->base
.write_domain
& ~I915_GEM_DOMAIN_CPU
);
3371 WARN_ON(obj
->base
.read_domains
& ~I915_GEM_DOMAIN_CPU
);
3373 old_read_domains
= obj
->base
.read_domains
;
3374 old_write_domain
= obj
->base
.write_domain
;
3376 obj
->base
.read_domains
= I915_GEM_DOMAIN_CPU
;
3377 obj
->base
.write_domain
= I915_GEM_DOMAIN_CPU
;
3379 trace_i915_gem_object_change_domain(obj
,
3384 obj
->cache_level
= cache_level
;
3385 i915_gem_verify_gtt(dev
);
3389 int i915_gem_get_caching_ioctl(struct drm_device
*dev
, void *data
,
3390 struct drm_file
*file
)
3392 struct drm_i915_gem_caching
*args
= data
;
3393 struct drm_i915_gem_object
*obj
;
3396 ret
= i915_mutex_lock_interruptible(dev
);
3400 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
3401 if (&obj
->base
== NULL
) {
3406 args
->caching
= obj
->cache_level
!= I915_CACHE_NONE
;
3408 drm_gem_object_unreference(&obj
->base
);
3410 mutex_unlock(&dev
->struct_mutex
);
3414 int i915_gem_set_caching_ioctl(struct drm_device
*dev
, void *data
,
3415 struct drm_file
*file
)
3417 struct drm_i915_gem_caching
*args
= data
;
3418 struct drm_i915_gem_object
*obj
;
3419 enum i915_cache_level level
;
3422 switch (args
->caching
) {
3423 case I915_CACHING_NONE
:
3424 level
= I915_CACHE_NONE
;
3426 case I915_CACHING_CACHED
:
3427 level
= I915_CACHE_LLC
;
3433 ret
= i915_mutex_lock_interruptible(dev
);
3437 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
3438 if (&obj
->base
== NULL
) {
3443 ret
= i915_gem_object_set_cache_level(obj
, level
);
3445 drm_gem_object_unreference(&obj
->base
);
3447 mutex_unlock(&dev
->struct_mutex
);
3452 * Prepare buffer for display plane (scanout, cursors, etc).
3453 * Can be called from an uninterruptible phase (modesetting) and allows
3454 * any flushes to be pipelined (for pageflips).
3457 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object
*obj
,
3459 struct intel_ring_buffer
*pipelined
)
3461 u32 old_read_domains
, old_write_domain
;
3464 if (pipelined
!= obj
->ring
) {
3465 ret
= i915_gem_object_sync(obj
, pipelined
);
3470 /* The display engine is not coherent with the LLC cache on gen6. As
3471 * a result, we make sure that the pinning that is about to occur is
3472 * done with uncached PTEs. This is lowest common denominator for all
3475 * However for gen6+, we could do better by using the GFDT bit instead
3476 * of uncaching, which would allow us to flush all the LLC-cached data
3477 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3479 ret
= i915_gem_object_set_cache_level(obj
, I915_CACHE_NONE
);
3483 /* As the user may map the buffer once pinned in the display plane
3484 * (e.g. libkms for the bootup splash), we have to ensure that we
3485 * always use map_and_fenceable for all scanout buffers.
3487 ret
= i915_gem_object_pin(obj
, alignment
, true, false);
3491 i915_gem_object_flush_cpu_write_domain(obj
);
3493 old_write_domain
= obj
->base
.write_domain
;
3494 old_read_domains
= obj
->base
.read_domains
;
3496 /* It should now be out of any other write domains, and we can update
3497 * the domain values for our changes.
3499 obj
->base
.write_domain
= 0;
3500 obj
->base
.read_domains
|= I915_GEM_DOMAIN_GTT
;
3502 trace_i915_gem_object_change_domain(obj
,
3510 i915_gem_object_finish_gpu(struct drm_i915_gem_object
*obj
)
3514 if ((obj
->base
.read_domains
& I915_GEM_GPU_DOMAINS
) == 0)
3517 ret
= i915_gem_object_wait_rendering(obj
, false);
3521 /* Ensure that we invalidate the GPU's caches and TLBs. */
3522 obj
->base
.read_domains
&= ~I915_GEM_GPU_DOMAINS
;
3527 * Moves a single object to the CPU read, and possibly write domain.
3529 * This function returns when the move is complete, including waiting on
3533 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object
*obj
, bool write
)
3535 uint32_t old_write_domain
, old_read_domains
;
3538 if (obj
->base
.write_domain
== I915_GEM_DOMAIN_CPU
)
3541 ret
= i915_gem_object_wait_rendering(obj
, !write
);
3545 i915_gem_object_flush_gtt_write_domain(obj
);
3547 old_write_domain
= obj
->base
.write_domain
;
3548 old_read_domains
= obj
->base
.read_domains
;
3550 /* Flush the CPU cache if it's still invalid. */
3551 if ((obj
->base
.read_domains
& I915_GEM_DOMAIN_CPU
) == 0) {
3552 i915_gem_clflush_object(obj
);
3554 obj
->base
.read_domains
|= I915_GEM_DOMAIN_CPU
;
3557 /* It should now be out of any other write domains, and we can update
3558 * the domain values for our changes.
3560 BUG_ON((obj
->base
.write_domain
& ~I915_GEM_DOMAIN_CPU
) != 0);
3562 /* If we're writing through the CPU, then the GPU read domains will
3563 * need to be invalidated at next use.
3566 obj
->base
.read_domains
= I915_GEM_DOMAIN_CPU
;
3567 obj
->base
.write_domain
= I915_GEM_DOMAIN_CPU
;
3570 trace_i915_gem_object_change_domain(obj
,
3577 /* Throttle our rendering by waiting until the ring has completed our requests
3578 * emitted over 20 msec ago.
3580 * Note that if we were to use the current jiffies each time around the loop,
3581 * we wouldn't escape the function with any frames outstanding if the time to
3582 * render a frame was over 20ms.
3584 * This should get us reasonable parallelism between CPU and GPU but also
3585 * relatively low latency when blocking on a particular request to finish.
3588 i915_gem_ring_throttle(struct drm_device
*dev
, struct drm_file
*file
)
3590 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3591 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
3592 unsigned long recent_enough
= jiffies
- msecs_to_jiffies(20);
3593 struct drm_i915_gem_request
*request
;
3594 struct intel_ring_buffer
*ring
= NULL
;
3595 unsigned reset_counter
;
3599 ret
= i915_gem_wait_for_error(&dev_priv
->gpu_error
);
3603 ret
= i915_gem_check_wedge(&dev_priv
->gpu_error
, false);
3607 spin_lock(&file_priv
->mm
.lock
);
3608 list_for_each_entry(request
, &file_priv
->mm
.request_list
, client_list
) {
3609 if (time_after_eq(request
->emitted_jiffies
, recent_enough
))
3612 ring
= request
->ring
;
3613 seqno
= request
->seqno
;
3615 reset_counter
= atomic_read(&dev_priv
->gpu_error
.reset_counter
);
3616 spin_unlock(&file_priv
->mm
.lock
);
3621 ret
= __wait_seqno(ring
, seqno
, reset_counter
, true, NULL
);
3623 queue_delayed_work(dev_priv
->wq
, &dev_priv
->mm
.retire_work
, 0);
3629 i915_gem_object_pin(struct drm_i915_gem_object
*obj
,
3631 bool map_and_fenceable
,
3636 if (WARN_ON(obj
->pin_count
== DRM_I915_GEM_OBJECT_MAX_PIN_COUNT
))
3639 if (obj
->gtt_space
!= NULL
) {
3640 if ((alignment
&& obj
->gtt_offset
& (alignment
- 1)) ||
3641 (map_and_fenceable
&& !obj
->map_and_fenceable
)) {
3642 WARN(obj
->pin_count
,
3643 "bo is already pinned with incorrect alignment:"
3644 " offset=%x, req.alignment=%x, req.map_and_fenceable=%d,"
3645 " obj->map_and_fenceable=%d\n",
3646 obj
->gtt_offset
, alignment
,
3648 obj
->map_and_fenceable
);
3649 ret
= i915_gem_object_unbind(obj
);
3655 if (obj
->gtt_space
== NULL
) {
3656 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
3658 ret
= i915_gem_object_bind_to_gtt(obj
, alignment
,
3664 if (!dev_priv
->mm
.aliasing_ppgtt
)
3665 i915_gem_gtt_bind_object(obj
, obj
->cache_level
);
3668 if (!obj
->has_global_gtt_mapping
&& map_and_fenceable
)
3669 i915_gem_gtt_bind_object(obj
, obj
->cache_level
);
3672 obj
->pin_mappable
|= map_and_fenceable
;
3678 i915_gem_object_unpin(struct drm_i915_gem_object
*obj
)
3680 BUG_ON(obj
->pin_count
== 0);
3681 BUG_ON(obj
->gtt_space
== NULL
);
3683 if (--obj
->pin_count
== 0)
3684 obj
->pin_mappable
= false;
3688 i915_gem_pin_ioctl(struct drm_device
*dev
, void *data
,
3689 struct drm_file
*file
)
3691 struct drm_i915_gem_pin
*args
= data
;
3692 struct drm_i915_gem_object
*obj
;
3695 ret
= i915_mutex_lock_interruptible(dev
);
3699 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
3700 if (&obj
->base
== NULL
) {
3705 if (obj
->madv
!= I915_MADV_WILLNEED
) {
3706 DRM_ERROR("Attempting to pin a purgeable buffer\n");
3711 if (obj
->pin_filp
!= NULL
&& obj
->pin_filp
!= file
) {
3712 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
3718 if (obj
->user_pin_count
== 0) {
3719 ret
= i915_gem_object_pin(obj
, args
->alignment
, true, false);
3724 obj
->user_pin_count
++;
3725 obj
->pin_filp
= file
;
3727 /* XXX - flush the CPU caches for pinned objects
3728 * as the X server doesn't manage domains yet
3730 i915_gem_object_flush_cpu_write_domain(obj
);
3731 args
->offset
= obj
->gtt_offset
;
3733 drm_gem_object_unreference(&obj
->base
);
3735 mutex_unlock(&dev
->struct_mutex
);
3740 i915_gem_unpin_ioctl(struct drm_device
*dev
, void *data
,
3741 struct drm_file
*file
)
3743 struct drm_i915_gem_pin
*args
= data
;
3744 struct drm_i915_gem_object
*obj
;
3747 ret
= i915_mutex_lock_interruptible(dev
);
3751 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
3752 if (&obj
->base
== NULL
) {
3757 if (obj
->pin_filp
!= file
) {
3758 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
3763 obj
->user_pin_count
--;
3764 if (obj
->user_pin_count
== 0) {
3765 obj
->pin_filp
= NULL
;
3766 i915_gem_object_unpin(obj
);
3770 drm_gem_object_unreference(&obj
->base
);
3772 mutex_unlock(&dev
->struct_mutex
);
3777 i915_gem_busy_ioctl(struct drm_device
*dev
, void *data
,
3778 struct drm_file
*file
)
3780 struct drm_i915_gem_busy
*args
= data
;
3781 struct drm_i915_gem_object
*obj
;
3784 ret
= i915_mutex_lock_interruptible(dev
);
3788 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
3789 if (&obj
->base
== NULL
) {
3794 /* Count all active objects as busy, even if they are currently not used
3795 * by the gpu. Users of this interface expect objects to eventually
3796 * become non-busy without any further actions, therefore emit any
3797 * necessary flushes here.
3799 ret
= i915_gem_object_flush_active(obj
);
3801 args
->busy
= obj
->active
;
3803 BUILD_BUG_ON(I915_NUM_RINGS
> 16);
3804 args
->busy
|= intel_ring_flag(obj
->ring
) << 16;
3807 drm_gem_object_unreference(&obj
->base
);
3809 mutex_unlock(&dev
->struct_mutex
);
3814 i915_gem_throttle_ioctl(struct drm_device
*dev
, void *data
,
3815 struct drm_file
*file_priv
)
3817 return i915_gem_ring_throttle(dev
, file_priv
);
3821 i915_gem_madvise_ioctl(struct drm_device
*dev
, void *data
,
3822 struct drm_file
*file_priv
)
3824 struct drm_i915_gem_madvise
*args
= data
;
3825 struct drm_i915_gem_object
*obj
;
3828 switch (args
->madv
) {
3829 case I915_MADV_DONTNEED
:
3830 case I915_MADV_WILLNEED
:
3836 ret
= i915_mutex_lock_interruptible(dev
);
3840 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file_priv
, args
->handle
));
3841 if (&obj
->base
== NULL
) {
3846 if (obj
->pin_count
) {
3851 if (obj
->madv
!= __I915_MADV_PURGED
)
3852 obj
->madv
= args
->madv
;
3854 /* if the object is no longer attached, discard its backing storage */
3855 if (i915_gem_object_is_purgeable(obj
) && obj
->pages
== NULL
)
3856 i915_gem_object_truncate(obj
);
3858 args
->retained
= obj
->madv
!= __I915_MADV_PURGED
;
3861 drm_gem_object_unreference(&obj
->base
);
3863 mutex_unlock(&dev
->struct_mutex
);
3867 void i915_gem_object_init(struct drm_i915_gem_object
*obj
,
3868 const struct drm_i915_gem_object_ops
*ops
)
3870 INIT_LIST_HEAD(&obj
->mm_list
);
3871 INIT_LIST_HEAD(&obj
->global_list
);
3872 INIT_LIST_HEAD(&obj
->ring_list
);
3873 INIT_LIST_HEAD(&obj
->exec_list
);
3877 obj
->fence_reg
= I915_FENCE_REG_NONE
;
3878 obj
->madv
= I915_MADV_WILLNEED
;
3879 /* Avoid an unnecessary call to unbind on the first bind. */
3880 obj
->map_and_fenceable
= true;
3882 i915_gem_info_add_obj(obj
->base
.dev
->dev_private
, obj
->base
.size
);
3885 static const struct drm_i915_gem_object_ops i915_gem_object_ops
= {
3886 .get_pages
= i915_gem_object_get_pages_gtt
,
3887 .put_pages
= i915_gem_object_put_pages_gtt
,
3890 struct drm_i915_gem_object
*i915_gem_alloc_object(struct drm_device
*dev
,
3893 struct drm_i915_gem_object
*obj
;
3894 struct address_space
*mapping
;
3897 obj
= i915_gem_object_alloc(dev
);
3901 if (drm_gem_object_init(dev
, &obj
->base
, size
) != 0) {
3902 i915_gem_object_free(obj
);
3906 mask
= GFP_HIGHUSER
| __GFP_RECLAIMABLE
;
3907 if (IS_CRESTLINE(dev
) || IS_BROADWATER(dev
)) {
3908 /* 965gm cannot relocate objects above 4GiB. */
3909 mask
&= ~__GFP_HIGHMEM
;
3910 mask
|= __GFP_DMA32
;
3913 mapping
= file_inode(obj
->base
.filp
)->i_mapping
;
3914 mapping_set_gfp_mask(mapping
, mask
);
3916 i915_gem_object_init(obj
, &i915_gem_object_ops
);
3918 obj
->base
.write_domain
= I915_GEM_DOMAIN_CPU
;
3919 obj
->base
.read_domains
= I915_GEM_DOMAIN_CPU
;
3922 /* On some devices, we can have the GPU use the LLC (the CPU
3923 * cache) for about a 10% performance improvement
3924 * compared to uncached. Graphics requests other than
3925 * display scanout are coherent with the CPU in
3926 * accessing this cache. This means in this mode we
3927 * don't need to clflush on the CPU side, and on the
3928 * GPU side we only need to flush internal caches to
3929 * get data visible to the CPU.
3931 * However, we maintain the display planes as UC, and so
3932 * need to rebind when first used as such.
3934 obj
->cache_level
= I915_CACHE_LLC
;
3936 obj
->cache_level
= I915_CACHE_NONE
;
3941 int i915_gem_init_object(struct drm_gem_object
*obj
)
3948 void i915_gem_free_object(struct drm_gem_object
*gem_obj
)
3950 struct drm_i915_gem_object
*obj
= to_intel_bo(gem_obj
);
3951 struct drm_device
*dev
= obj
->base
.dev
;
3952 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
3954 trace_i915_gem_object_destroy(obj
);
3957 i915_gem_detach_phys_object(dev
, obj
);
3960 if (WARN_ON(i915_gem_object_unbind(obj
) == -ERESTARTSYS
)) {
3961 bool was_interruptible
;
3963 was_interruptible
= dev_priv
->mm
.interruptible
;
3964 dev_priv
->mm
.interruptible
= false;
3966 WARN_ON(i915_gem_object_unbind(obj
));
3968 dev_priv
->mm
.interruptible
= was_interruptible
;
3971 /* Stolen objects don't hold a ref, but do hold pin count. Fix that up
3972 * before progressing. */
3974 i915_gem_object_unpin_pages(obj
);
3976 if (WARN_ON(obj
->pages_pin_count
))
3977 obj
->pages_pin_count
= 0;
3978 i915_gem_object_put_pages(obj
);
3979 i915_gem_object_free_mmap_offset(obj
);
3980 i915_gem_object_release_stolen(obj
);
3984 if (obj
->base
.import_attach
)
3985 drm_prime_gem_destroy(&obj
->base
, NULL
);
3987 drm_gem_object_release(&obj
->base
);
3988 i915_gem_info_remove_obj(dev_priv
, obj
->base
.size
);
3991 i915_gem_object_free(obj
);
3995 i915_gem_idle(struct drm_device
*dev
)
3997 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
4000 mutex_lock(&dev
->struct_mutex
);
4002 if (dev_priv
->mm
.suspended
) {
4003 mutex_unlock(&dev
->struct_mutex
);
4007 ret
= i915_gpu_idle(dev
);
4009 mutex_unlock(&dev
->struct_mutex
);
4012 i915_gem_retire_requests(dev
);
4014 /* Under UMS, be paranoid and evict. */
4015 if (!drm_core_check_feature(dev
, DRIVER_MODESET
))
4016 i915_gem_evict_everything(dev
);
4018 i915_gem_reset_fences(dev
);
4020 /* Hack! Don't let anybody do execbuf while we don't control the chip.
4021 * We need to replace this with a semaphore, or something.
4022 * And not confound mm.suspended!
4024 dev_priv
->mm
.suspended
= 1;
4025 del_timer_sync(&dev_priv
->gpu_error
.hangcheck_timer
);
4027 i915_kernel_lost_context(dev
);
4028 i915_gem_cleanup_ringbuffer(dev
);
4030 mutex_unlock(&dev
->struct_mutex
);
4032 /* Cancel the retire work handler, which should be idle now. */
4033 cancel_delayed_work_sync(&dev_priv
->mm
.retire_work
);
4038 void i915_gem_l3_remap(struct drm_device
*dev
)
4040 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
4044 if (!HAS_L3_GPU_CACHE(dev
))
4047 if (!dev_priv
->l3_parity
.remap_info
)
4050 misccpctl
= I915_READ(GEN7_MISCCPCTL
);
4051 I915_WRITE(GEN7_MISCCPCTL
, misccpctl
& ~GEN7_DOP_CLOCK_GATE_ENABLE
);
4052 POSTING_READ(GEN7_MISCCPCTL
);
4054 for (i
= 0; i
< GEN7_L3LOG_SIZE
; i
+= 4) {
4055 u32 remap
= I915_READ(GEN7_L3LOG_BASE
+ i
);
4056 if (remap
&& remap
!= dev_priv
->l3_parity
.remap_info
[i
/4])
4057 DRM_DEBUG("0x%x was already programmed to %x\n",
4058 GEN7_L3LOG_BASE
+ i
, remap
);
4059 if (remap
&& !dev_priv
->l3_parity
.remap_info
[i
/4])
4060 DRM_DEBUG_DRIVER("Clearing remapped register\n");
4061 I915_WRITE(GEN7_L3LOG_BASE
+ i
, dev_priv
->l3_parity
.remap_info
[i
/4]);
4064 /* Make sure all the writes land before disabling dop clock gating */
4065 POSTING_READ(GEN7_L3LOG_BASE
);
4067 I915_WRITE(GEN7_MISCCPCTL
, misccpctl
);
4070 void i915_gem_init_swizzling(struct drm_device
*dev
)
4072 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
4074 if (INTEL_INFO(dev
)->gen
< 5 ||
4075 dev_priv
->mm
.bit_6_swizzle_x
== I915_BIT_6_SWIZZLE_NONE
)
4078 I915_WRITE(DISP_ARB_CTL
, I915_READ(DISP_ARB_CTL
) |
4079 DISP_TILE_SURFACE_SWIZZLING
);
4084 I915_WRITE(TILECTL
, I915_READ(TILECTL
) | TILECTL_SWZCTL
);
4086 I915_WRITE(ARB_MODE
, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB
));
4087 else if (IS_GEN7(dev
))
4088 I915_WRITE(ARB_MODE
, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB
));
4094 intel_enable_blt(struct drm_device
*dev
)
4099 /* The blitter was dysfunctional on early prototypes */
4100 if (IS_GEN6(dev
) && dev
->pdev
->revision
< 8) {
4101 DRM_INFO("BLT not supported on this pre-production hardware;"
4102 " graphics performance will be degraded.\n");
4109 static int i915_gem_init_rings(struct drm_device
*dev
)
4111 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4114 ret
= intel_init_render_ring_buffer(dev
);
4119 ret
= intel_init_bsd_ring_buffer(dev
);
4121 goto cleanup_render_ring
;
4124 if (intel_enable_blt(dev
)) {
4125 ret
= intel_init_blt_ring_buffer(dev
);
4127 goto cleanup_bsd_ring
;
4130 if (HAS_VEBOX(dev
)) {
4131 ret
= intel_init_vebox_ring_buffer(dev
);
4133 goto cleanup_blt_ring
;
4137 ret
= i915_gem_set_seqno(dev
, ((u32
)~0 - 0x1000));
4139 goto cleanup_vebox_ring
;
4144 intel_cleanup_ring_buffer(&dev_priv
->ring
[VECS
]);
4146 intel_cleanup_ring_buffer(&dev_priv
->ring
[BCS
]);
4148 intel_cleanup_ring_buffer(&dev_priv
->ring
[VCS
]);
4149 cleanup_render_ring
:
4150 intel_cleanup_ring_buffer(&dev_priv
->ring
[RCS
]);
4156 i915_gem_init_hw(struct drm_device
*dev
)
4158 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
4161 if (INTEL_INFO(dev
)->gen
< 6 && !intel_enable_gtt())
4164 if (IS_HASWELL(dev
) && (I915_READ(0x120010) == 1))
4165 I915_WRITE(0x9008, I915_READ(0x9008) | 0xf0000);
4167 if (HAS_PCH_NOP(dev
)) {
4168 u32 temp
= I915_READ(GEN7_MSG_CTL
);
4169 temp
&= ~(WAIT_FOR_PCH_FLR_ACK
| WAIT_FOR_PCH_RESET_ACK
);
4170 I915_WRITE(GEN7_MSG_CTL
, temp
);
4173 i915_gem_l3_remap(dev
);
4175 i915_gem_init_swizzling(dev
);
4177 ret
= i915_gem_init_rings(dev
);
4182 * XXX: There was some w/a described somewhere suggesting loading
4183 * contexts before PPGTT.
4185 i915_gem_context_init(dev
);
4186 if (dev_priv
->mm
.aliasing_ppgtt
) {
4187 ret
= dev_priv
->mm
.aliasing_ppgtt
->enable(dev
);
4189 i915_gem_cleanup_aliasing_ppgtt(dev
);
4190 DRM_INFO("PPGTT enable failed. This is not fatal, but unexpected\n");
4197 int i915_gem_init(struct drm_device
*dev
)
4199 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4202 mutex_lock(&dev
->struct_mutex
);
4204 if (IS_VALLEYVIEW(dev
)) {
4205 /* VLVA0 (potential hack), BIOS isn't actually waking us */
4206 I915_WRITE(VLV_GTLC_WAKE_CTRL
, 1);
4207 if (wait_for((I915_READ(VLV_GTLC_PW_STATUS
) & 1) == 1, 10))
4208 DRM_DEBUG_DRIVER("allow wake ack timed out\n");
4211 i915_gem_init_global_gtt(dev
);
4213 ret
= i915_gem_init_hw(dev
);
4214 mutex_unlock(&dev
->struct_mutex
);
4216 i915_gem_cleanup_aliasing_ppgtt(dev
);
4220 /* Allow hardware batchbuffers unless told otherwise, but not for KMS. */
4221 if (!drm_core_check_feature(dev
, DRIVER_MODESET
))
4222 dev_priv
->dri1
.allow_batchbuffer
= 1;
4227 i915_gem_cleanup_ringbuffer(struct drm_device
*dev
)
4229 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
4230 struct intel_ring_buffer
*ring
;
4233 for_each_ring(ring
, dev_priv
, i
)
4234 intel_cleanup_ring_buffer(ring
);
4238 i915_gem_entervt_ioctl(struct drm_device
*dev
, void *data
,
4239 struct drm_file
*file_priv
)
4241 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
4244 if (drm_core_check_feature(dev
, DRIVER_MODESET
))
4247 if (i915_reset_in_progress(&dev_priv
->gpu_error
)) {
4248 DRM_ERROR("Reenabling wedged hardware, good luck\n");
4249 atomic_set(&dev_priv
->gpu_error
.reset_counter
, 0);
4252 mutex_lock(&dev
->struct_mutex
);
4253 dev_priv
->mm
.suspended
= 0;
4255 ret
= i915_gem_init_hw(dev
);
4257 mutex_unlock(&dev
->struct_mutex
);
4261 BUG_ON(!list_empty(&dev_priv
->mm
.active_list
));
4262 mutex_unlock(&dev
->struct_mutex
);
4264 ret
= drm_irq_install(dev
);
4266 goto cleanup_ringbuffer
;
4271 mutex_lock(&dev
->struct_mutex
);
4272 i915_gem_cleanup_ringbuffer(dev
);
4273 dev_priv
->mm
.suspended
= 1;
4274 mutex_unlock(&dev
->struct_mutex
);
4280 i915_gem_leavevt_ioctl(struct drm_device
*dev
, void *data
,
4281 struct drm_file
*file_priv
)
4283 if (drm_core_check_feature(dev
, DRIVER_MODESET
))
4286 drm_irq_uninstall(dev
);
4287 return i915_gem_idle(dev
);
4291 i915_gem_lastclose(struct drm_device
*dev
)
4295 if (drm_core_check_feature(dev
, DRIVER_MODESET
))
4298 ret
= i915_gem_idle(dev
);
4300 DRM_ERROR("failed to idle hardware: %d\n", ret
);
4304 init_ring_lists(struct intel_ring_buffer
*ring
)
4306 INIT_LIST_HEAD(&ring
->active_list
);
4307 INIT_LIST_HEAD(&ring
->request_list
);
4311 i915_gem_load(struct drm_device
*dev
)
4313 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
4317 kmem_cache_create("i915_gem_object",
4318 sizeof(struct drm_i915_gem_object
), 0,
4322 INIT_LIST_HEAD(&dev_priv
->mm
.active_list
);
4323 INIT_LIST_HEAD(&dev_priv
->mm
.inactive_list
);
4324 INIT_LIST_HEAD(&dev_priv
->mm
.unbound_list
);
4325 INIT_LIST_HEAD(&dev_priv
->mm
.bound_list
);
4326 INIT_LIST_HEAD(&dev_priv
->mm
.fence_list
);
4327 for (i
= 0; i
< I915_NUM_RINGS
; i
++)
4328 init_ring_lists(&dev_priv
->ring
[i
]);
4329 for (i
= 0; i
< I915_MAX_NUM_FENCES
; i
++)
4330 INIT_LIST_HEAD(&dev_priv
->fence_regs
[i
].lru_list
);
4331 INIT_DELAYED_WORK(&dev_priv
->mm
.retire_work
,
4332 i915_gem_retire_work_handler
);
4333 init_waitqueue_head(&dev_priv
->gpu_error
.reset_queue
);
4335 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
4337 I915_WRITE(MI_ARB_STATE
,
4338 _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE
));
4341 dev_priv
->relative_constants_mode
= I915_EXEC_CONSTANTS_REL_GENERAL
;
4343 /* Old X drivers will take 0-2 for front, back, depth buffers */
4344 if (!drm_core_check_feature(dev
, DRIVER_MODESET
))
4345 dev_priv
->fence_reg_start
= 3;
4347 if (INTEL_INFO(dev
)->gen
>= 7 && !IS_VALLEYVIEW(dev
))
4348 dev_priv
->num_fence_regs
= 32;
4349 else if (INTEL_INFO(dev
)->gen
>= 4 || IS_I945G(dev
) || IS_I945GM(dev
) || IS_G33(dev
))
4350 dev_priv
->num_fence_regs
= 16;
4352 dev_priv
->num_fence_regs
= 8;
4354 /* Initialize fence registers to zero */
4355 i915_gem_reset_fences(dev
);
4357 i915_gem_detect_bit_6_swizzle(dev
);
4358 init_waitqueue_head(&dev_priv
->pending_flip_queue
);
4360 dev_priv
->mm
.interruptible
= true;
4362 dev_priv
->mm
.inactive_shrinker
.shrink
= i915_gem_inactive_shrink
;
4363 dev_priv
->mm
.inactive_shrinker
.seeks
= DEFAULT_SEEKS
;
4364 register_shrinker(&dev_priv
->mm
.inactive_shrinker
);
4368 * Create a physically contiguous memory object for this object
4369 * e.g. for cursor + overlay regs
4371 static int i915_gem_init_phys_object(struct drm_device
*dev
,
4372 int id
, int size
, int align
)
4374 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
4375 struct drm_i915_gem_phys_object
*phys_obj
;
4378 if (dev_priv
->mm
.phys_objs
[id
- 1] || !size
)
4381 phys_obj
= kzalloc(sizeof(struct drm_i915_gem_phys_object
), GFP_KERNEL
);
4387 phys_obj
->handle
= drm_pci_alloc(dev
, size
, align
);
4388 if (!phys_obj
->handle
) {
4393 set_memory_wc((unsigned long)phys_obj
->handle
->vaddr
, phys_obj
->handle
->size
/ PAGE_SIZE
);
4396 dev_priv
->mm
.phys_objs
[id
- 1] = phys_obj
;
4404 static void i915_gem_free_phys_object(struct drm_device
*dev
, int id
)
4406 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
4407 struct drm_i915_gem_phys_object
*phys_obj
;
4409 if (!dev_priv
->mm
.phys_objs
[id
- 1])
4412 phys_obj
= dev_priv
->mm
.phys_objs
[id
- 1];
4413 if (phys_obj
->cur_obj
) {
4414 i915_gem_detach_phys_object(dev
, phys_obj
->cur_obj
);
4418 set_memory_wb((unsigned long)phys_obj
->handle
->vaddr
, phys_obj
->handle
->size
/ PAGE_SIZE
);
4420 drm_pci_free(dev
, phys_obj
->handle
);
4422 dev_priv
->mm
.phys_objs
[id
- 1] = NULL
;
4425 void i915_gem_free_all_phys_object(struct drm_device
*dev
)
4429 for (i
= I915_GEM_PHYS_CURSOR_0
; i
<= I915_MAX_PHYS_OBJECT
; i
++)
4430 i915_gem_free_phys_object(dev
, i
);
4433 void i915_gem_detach_phys_object(struct drm_device
*dev
,
4434 struct drm_i915_gem_object
*obj
)
4436 struct address_space
*mapping
= file_inode(obj
->base
.filp
)->i_mapping
;
4443 vaddr
= obj
->phys_obj
->handle
->vaddr
;
4445 page_count
= obj
->base
.size
/ PAGE_SIZE
;
4446 for (i
= 0; i
< page_count
; i
++) {
4447 struct page
*page
= shmem_read_mapping_page(mapping
, i
);
4448 if (!IS_ERR(page
)) {
4449 char *dst
= kmap_atomic(page
);
4450 memcpy(dst
, vaddr
+ i
*PAGE_SIZE
, PAGE_SIZE
);
4453 drm_clflush_pages(&page
, 1);
4455 set_page_dirty(page
);
4456 mark_page_accessed(page
);
4457 page_cache_release(page
);
4460 i915_gem_chipset_flush(dev
);
4462 obj
->phys_obj
->cur_obj
= NULL
;
4463 obj
->phys_obj
= NULL
;
4467 i915_gem_attach_phys_object(struct drm_device
*dev
,
4468 struct drm_i915_gem_object
*obj
,
4472 struct address_space
*mapping
= file_inode(obj
->base
.filp
)->i_mapping
;
4473 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
4478 if (id
> I915_MAX_PHYS_OBJECT
)
4481 if (obj
->phys_obj
) {
4482 if (obj
->phys_obj
->id
== id
)
4484 i915_gem_detach_phys_object(dev
, obj
);
4487 /* create a new object */
4488 if (!dev_priv
->mm
.phys_objs
[id
- 1]) {
4489 ret
= i915_gem_init_phys_object(dev
, id
,
4490 obj
->base
.size
, align
);
4492 DRM_ERROR("failed to init phys object %d size: %zu\n",
4493 id
, obj
->base
.size
);
4498 /* bind to the object */
4499 obj
->phys_obj
= dev_priv
->mm
.phys_objs
[id
- 1];
4500 obj
->phys_obj
->cur_obj
= obj
;
4502 page_count
= obj
->base
.size
/ PAGE_SIZE
;
4504 for (i
= 0; i
< page_count
; i
++) {
4508 page
= shmem_read_mapping_page(mapping
, i
);
4510 return PTR_ERR(page
);
4512 src
= kmap_atomic(page
);
4513 dst
= obj
->phys_obj
->handle
->vaddr
+ (i
* PAGE_SIZE
);
4514 memcpy(dst
, src
, PAGE_SIZE
);
4517 mark_page_accessed(page
);
4518 page_cache_release(page
);
4525 i915_gem_phys_pwrite(struct drm_device
*dev
,
4526 struct drm_i915_gem_object
*obj
,
4527 struct drm_i915_gem_pwrite
*args
,
4528 struct drm_file
*file_priv
)
4530 void *vaddr
= obj
->phys_obj
->handle
->vaddr
+ args
->offset
;
4531 char __user
*user_data
= to_user_ptr(args
->data_ptr
);
4533 if (__copy_from_user_inatomic_nocache(vaddr
, user_data
, args
->size
)) {
4534 unsigned long unwritten
;
4536 /* The physical object once assigned is fixed for the lifetime
4537 * of the obj, so we can safely drop the lock and continue
4540 mutex_unlock(&dev
->struct_mutex
);
4541 unwritten
= copy_from_user(vaddr
, user_data
, args
->size
);
4542 mutex_lock(&dev
->struct_mutex
);
4547 i915_gem_chipset_flush(dev
);
4551 void i915_gem_release(struct drm_device
*dev
, struct drm_file
*file
)
4553 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
4555 /* Clean up our request list when the client is going away, so that
4556 * later retire_requests won't dereference our soon-to-be-gone
4559 spin_lock(&file_priv
->mm
.lock
);
4560 while (!list_empty(&file_priv
->mm
.request_list
)) {
4561 struct drm_i915_gem_request
*request
;
4563 request
= list_first_entry(&file_priv
->mm
.request_list
,
4564 struct drm_i915_gem_request
,
4566 list_del(&request
->client_list
);
4567 request
->file_priv
= NULL
;
4569 spin_unlock(&file_priv
->mm
.lock
);
4572 static bool mutex_is_locked_by(struct mutex
*mutex
, struct task_struct
*task
)
4574 if (!mutex_is_locked(mutex
))
4577 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES)
4578 return mutex
->owner
== task
;
4580 /* Since UP may be pre-empted, we cannot assume that we own the lock */
4586 i915_gem_inactive_shrink(struct shrinker
*shrinker
, struct shrink_control
*sc
)
4588 struct drm_i915_private
*dev_priv
=
4589 container_of(shrinker
,
4590 struct drm_i915_private
,
4591 mm
.inactive_shrinker
);
4592 struct drm_device
*dev
= dev_priv
->dev
;
4593 struct drm_i915_gem_object
*obj
;
4594 int nr_to_scan
= sc
->nr_to_scan
;
4598 if (!mutex_trylock(&dev
->struct_mutex
)) {
4599 if (!mutex_is_locked_by(&dev
->struct_mutex
, current
))
4602 if (dev_priv
->mm
.shrinker_no_lock_stealing
)
4609 nr_to_scan
-= i915_gem_purge(dev_priv
, nr_to_scan
);
4611 nr_to_scan
-= __i915_gem_shrink(dev_priv
, nr_to_scan
,
4614 i915_gem_shrink_all(dev_priv
);
4618 list_for_each_entry(obj
, &dev_priv
->mm
.unbound_list
, global_list
)
4619 if (obj
->pages_pin_count
== 0)
4620 cnt
+= obj
->base
.size
>> PAGE_SHIFT
;
4621 list_for_each_entry(obj
, &dev_priv
->mm
.inactive_list
, global_list
)
4622 if (obj
->pin_count
== 0 && obj
->pages_pin_count
== 0)
4623 cnt
+= obj
->base
.size
>> PAGE_SHIFT
;
4626 mutex_unlock(&dev
->struct_mutex
);