2 * Copyright © 2008 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eric Anholt <eric@anholt.net>
32 #include "i915_trace.h"
33 #include "intel_drv.h"
34 #include <linux/shmem_fs.h>
35 #include <linux/slab.h>
36 #include <linux/swap.h>
37 #include <linux/pci.h>
39 static __must_check
int i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object
*obj
);
40 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object
*obj
);
41 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object
*obj
);
42 static __must_check
int i915_gem_object_bind_to_gtt(struct drm_i915_gem_object
*obj
,
44 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 void i915_gem_object_truncate(struct drm_i915_gem_object
*obj
);
60 static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object
*obj
)
63 i915_gem_release_mmap(obj
);
65 /* As we do not have an associated fence register, we will force
66 * a tiling change if we ever need to acquire one.
68 obj
->fence_dirty
= false;
69 obj
->fence_reg
= I915_FENCE_REG_NONE
;
72 /* some bookkeeping */
73 static void i915_gem_info_add_obj(struct drm_i915_private
*dev_priv
,
76 dev_priv
->mm
.object_count
++;
77 dev_priv
->mm
.object_memory
+= size
;
80 static void i915_gem_info_remove_obj(struct drm_i915_private
*dev_priv
,
83 dev_priv
->mm
.object_count
--;
84 dev_priv
->mm
.object_memory
-= size
;
88 i915_gem_wait_for_error(struct drm_device
*dev
)
90 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
91 struct completion
*x
= &dev_priv
->error_completion
;
95 if (!atomic_read(&dev_priv
->mm
.wedged
))
98 ret
= wait_for_completion_interruptible(x
);
102 if (atomic_read(&dev_priv
->mm
.wedged
)) {
103 /* GPU is hung, bump the completion count to account for
104 * the token we just consumed so that we never hit zero and
105 * end up waiting upon a subsequent completion event that
108 spin_lock_irqsave(&x
->wait
.lock
, flags
);
110 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
115 int i915_mutex_lock_interruptible(struct drm_device
*dev
)
119 ret
= i915_gem_wait_for_error(dev
);
123 ret
= mutex_lock_interruptible(&dev
->struct_mutex
);
127 WARN_ON(i915_verify_lists(dev
));
132 i915_gem_object_is_inactive(struct drm_i915_gem_object
*obj
)
138 i915_gem_init_ioctl(struct drm_device
*dev
, void *data
,
139 struct drm_file
*file
)
141 struct drm_i915_gem_init
*args
= data
;
143 if (drm_core_check_feature(dev
, DRIVER_MODESET
))
146 if (args
->gtt_start
>= args
->gtt_end
||
147 (args
->gtt_end
| args
->gtt_start
) & (PAGE_SIZE
- 1))
150 /* GEM with user mode setting was never supported on ilk and later. */
151 if (INTEL_INFO(dev
)->gen
>= 5)
154 mutex_lock(&dev
->struct_mutex
);
155 i915_gem_init_global_gtt(dev
, args
->gtt_start
,
156 args
->gtt_end
, args
->gtt_end
);
157 mutex_unlock(&dev
->struct_mutex
);
163 i915_gem_get_aperture_ioctl(struct drm_device
*dev
, void *data
,
164 struct drm_file
*file
)
166 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
167 struct drm_i915_gem_get_aperture
*args
= data
;
168 struct drm_i915_gem_object
*obj
;
172 mutex_lock(&dev
->struct_mutex
);
173 list_for_each_entry(obj
, &dev_priv
->mm
.gtt_list
, gtt_list
)
175 pinned
+= obj
->gtt_space
->size
;
176 mutex_unlock(&dev
->struct_mutex
);
178 args
->aper_size
= dev_priv
->mm
.gtt_total
;
179 args
->aper_available_size
= args
->aper_size
- pinned
;
185 i915_gem_create(struct drm_file
*file
,
186 struct drm_device
*dev
,
190 struct drm_i915_gem_object
*obj
;
194 size
= roundup(size
, PAGE_SIZE
);
198 /* Allocate the new object */
199 obj
= i915_gem_alloc_object(dev
, size
);
203 ret
= drm_gem_handle_create(file
, &obj
->base
, &handle
);
205 drm_gem_object_release(&obj
->base
);
206 i915_gem_info_remove_obj(dev
->dev_private
, obj
->base
.size
);
211 /* drop reference from allocate - handle holds it now */
212 drm_gem_object_unreference(&obj
->base
);
213 trace_i915_gem_object_create(obj
);
220 i915_gem_dumb_create(struct drm_file
*file
,
221 struct drm_device
*dev
,
222 struct drm_mode_create_dumb
*args
)
224 /* have to work out size/pitch and return them */
225 args
->pitch
= ALIGN(args
->width
* ((args
->bpp
+ 7) / 8), 64);
226 args
->size
= args
->pitch
* args
->height
;
227 return i915_gem_create(file
, dev
,
228 args
->size
, &args
->handle
);
231 int i915_gem_dumb_destroy(struct drm_file
*file
,
232 struct drm_device
*dev
,
235 return drm_gem_handle_delete(file
, handle
);
239 * Creates a new mm object and returns a handle to it.
242 i915_gem_create_ioctl(struct drm_device
*dev
, void *data
,
243 struct drm_file
*file
)
245 struct drm_i915_gem_create
*args
= data
;
247 return i915_gem_create(file
, dev
,
248 args
->size
, &args
->handle
);
251 static int i915_gem_object_needs_bit17_swizzle(struct drm_i915_gem_object
*obj
)
253 drm_i915_private_t
*dev_priv
= obj
->base
.dev
->dev_private
;
255 return dev_priv
->mm
.bit_6_swizzle_x
== I915_BIT_6_SWIZZLE_9_10_17
&&
256 obj
->tiling_mode
!= I915_TILING_NONE
;
260 __copy_to_user_swizzled(char __user
*cpu_vaddr
,
261 const char *gpu_vaddr
, int gpu_offset
,
264 int ret
, cpu_offset
= 0;
267 int cacheline_end
= ALIGN(gpu_offset
+ 1, 64);
268 int this_length
= min(cacheline_end
- gpu_offset
, length
);
269 int swizzled_gpu_offset
= gpu_offset
^ 64;
271 ret
= __copy_to_user(cpu_vaddr
+ cpu_offset
,
272 gpu_vaddr
+ swizzled_gpu_offset
,
277 cpu_offset
+= this_length
;
278 gpu_offset
+= this_length
;
279 length
-= this_length
;
286 __copy_from_user_swizzled(char *gpu_vaddr
, int gpu_offset
,
287 const char __user
*cpu_vaddr
,
290 int ret
, cpu_offset
= 0;
293 int cacheline_end
= ALIGN(gpu_offset
+ 1, 64);
294 int this_length
= min(cacheline_end
- gpu_offset
, length
);
295 int swizzled_gpu_offset
= gpu_offset
^ 64;
297 ret
= __copy_from_user(gpu_vaddr
+ swizzled_gpu_offset
,
298 cpu_vaddr
+ cpu_offset
,
303 cpu_offset
+= this_length
;
304 gpu_offset
+= this_length
;
305 length
-= this_length
;
311 /* Per-page copy function for the shmem pread fastpath.
312 * Flushes invalid cachelines before reading the target if
313 * needs_clflush is set. */
315 shmem_pread_fast(struct page
*page
, int shmem_page_offset
, int page_length
,
316 char __user
*user_data
,
317 bool page_do_bit17_swizzling
, bool needs_clflush
)
322 if (unlikely(page_do_bit17_swizzling
))
325 vaddr
= kmap_atomic(page
);
327 drm_clflush_virt_range(vaddr
+ shmem_page_offset
,
329 ret
= __copy_to_user_inatomic(user_data
,
330 vaddr
+ shmem_page_offset
,
332 kunmap_atomic(vaddr
);
338 shmem_clflush_swizzled_range(char *addr
, unsigned long length
,
341 if (unlikely(swizzled
)) {
342 unsigned long start
= (unsigned long) addr
;
343 unsigned long end
= (unsigned long) addr
+ length
;
345 /* For swizzling simply ensure that we always flush both
346 * channels. Lame, but simple and it works. Swizzled
347 * pwrite/pread is far from a hotpath - current userspace
348 * doesn't use it at all. */
349 start
= round_down(start
, 128);
350 end
= round_up(end
, 128);
352 drm_clflush_virt_range((void *)start
, end
- start
);
354 drm_clflush_virt_range(addr
, length
);
359 /* Only difference to the fast-path function is that this can handle bit17
360 * and uses non-atomic copy and kmap functions. */
362 shmem_pread_slow(struct page
*page
, int shmem_page_offset
, int page_length
,
363 char __user
*user_data
,
364 bool page_do_bit17_swizzling
, bool needs_clflush
)
371 shmem_clflush_swizzled_range(vaddr
+ shmem_page_offset
,
373 page_do_bit17_swizzling
);
375 if (page_do_bit17_swizzling
)
376 ret
= __copy_to_user_swizzled(user_data
,
377 vaddr
, shmem_page_offset
,
380 ret
= __copy_to_user(user_data
,
381 vaddr
+ shmem_page_offset
,
389 i915_gem_shmem_pread(struct drm_device
*dev
,
390 struct drm_i915_gem_object
*obj
,
391 struct drm_i915_gem_pread
*args
,
392 struct drm_file
*file
)
394 struct address_space
*mapping
= obj
->base
.filp
->f_path
.dentry
->d_inode
->i_mapping
;
395 char __user
*user_data
;
398 int shmem_page_offset
, page_length
, ret
= 0;
399 int obj_do_bit17_swizzling
, page_do_bit17_swizzling
;
400 int hit_slowpath
= 0;
402 int needs_clflush
= 0;
405 user_data
= (char __user
*) (uintptr_t) args
->data_ptr
;
408 obj_do_bit17_swizzling
= i915_gem_object_needs_bit17_swizzle(obj
);
410 if (!(obj
->base
.read_domains
& I915_GEM_DOMAIN_CPU
)) {
411 /* If we're not in the cpu read domain, set ourself into the gtt
412 * read domain and manually flush cachelines (if required). This
413 * optimizes for the case when the gpu will dirty the data
414 * anyway again before the next pread happens. */
415 if (obj
->cache_level
== I915_CACHE_NONE
)
417 ret
= i915_gem_object_set_to_gtt_domain(obj
, false);
422 offset
= args
->offset
;
427 /* Operation in this page
429 * shmem_page_offset = offset within page in shmem file
430 * page_length = bytes to copy for this page
432 shmem_page_offset
= offset_in_page(offset
);
433 page_length
= remain
;
434 if ((shmem_page_offset
+ page_length
) > PAGE_SIZE
)
435 page_length
= PAGE_SIZE
- shmem_page_offset
;
438 page
= obj
->pages
[offset
>> PAGE_SHIFT
];
441 page
= shmem_read_mapping_page(mapping
, offset
>> PAGE_SHIFT
);
449 page_do_bit17_swizzling
= obj_do_bit17_swizzling
&&
450 (page_to_phys(page
) & (1 << 17)) != 0;
452 ret
= shmem_pread_fast(page
, shmem_page_offset
, page_length
,
453 user_data
, page_do_bit17_swizzling
,
459 page_cache_get(page
);
460 mutex_unlock(&dev
->struct_mutex
);
463 ret
= fault_in_multipages_writeable(user_data
, remain
);
464 /* Userspace is tricking us, but we've already clobbered
465 * its pages with the prefault and promised to write the
466 * data up to the first fault. Hence ignore any errors
467 * and just continue. */
472 ret
= shmem_pread_slow(page
, shmem_page_offset
, page_length
,
473 user_data
, page_do_bit17_swizzling
,
476 mutex_lock(&dev
->struct_mutex
);
477 page_cache_release(page
);
479 mark_page_accessed(page
);
481 page_cache_release(page
);
488 remain
-= page_length
;
489 user_data
+= page_length
;
490 offset
+= page_length
;
495 /* Fixup: Kill any reinstated backing storage pages */
496 if (obj
->madv
== __I915_MADV_PURGED
)
497 i915_gem_object_truncate(obj
);
504 * Reads data from the object referenced by handle.
506 * On error, the contents of *data are undefined.
509 i915_gem_pread_ioctl(struct drm_device
*dev
, void *data
,
510 struct drm_file
*file
)
512 struct drm_i915_gem_pread
*args
= data
;
513 struct drm_i915_gem_object
*obj
;
519 if (!access_ok(VERIFY_WRITE
,
520 (char __user
*)(uintptr_t)args
->data_ptr
,
524 ret
= i915_mutex_lock_interruptible(dev
);
528 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
529 if (&obj
->base
== NULL
) {
534 /* Bounds check source. */
535 if (args
->offset
> obj
->base
.size
||
536 args
->size
> obj
->base
.size
- args
->offset
) {
541 trace_i915_gem_object_pread(obj
, args
->offset
, args
->size
);
543 ret
= i915_gem_shmem_pread(dev
, obj
, args
, file
);
546 drm_gem_object_unreference(&obj
->base
);
548 mutex_unlock(&dev
->struct_mutex
);
552 /* This is the fast write path which cannot handle
553 * page faults in the source data
557 fast_user_write(struct io_mapping
*mapping
,
558 loff_t page_base
, int page_offset
,
559 char __user
*user_data
,
562 void __iomem
*vaddr_atomic
;
564 unsigned long unwritten
;
566 vaddr_atomic
= io_mapping_map_atomic_wc(mapping
, page_base
);
567 /* We can use the cpu mem copy function because this is X86. */
568 vaddr
= (void __force
*)vaddr_atomic
+ page_offset
;
569 unwritten
= __copy_from_user_inatomic_nocache(vaddr
,
571 io_mapping_unmap_atomic(vaddr_atomic
);
576 * This is the fast pwrite path, where we copy the data directly from the
577 * user into the GTT, uncached.
580 i915_gem_gtt_pwrite_fast(struct drm_device
*dev
,
581 struct drm_i915_gem_object
*obj
,
582 struct drm_i915_gem_pwrite
*args
,
583 struct drm_file
*file
)
585 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
587 loff_t offset
, page_base
;
588 char __user
*user_data
;
589 int page_offset
, page_length
, ret
;
591 ret
= i915_gem_object_pin(obj
, 0, true);
595 ret
= i915_gem_object_set_to_gtt_domain(obj
, true);
599 ret
= i915_gem_object_put_fence(obj
);
603 user_data
= (char __user
*) (uintptr_t) args
->data_ptr
;
606 offset
= obj
->gtt_offset
+ args
->offset
;
609 /* Operation in this page
611 * page_base = page offset within aperture
612 * page_offset = offset within page
613 * page_length = bytes to copy for this page
615 page_base
= offset
& PAGE_MASK
;
616 page_offset
= offset_in_page(offset
);
617 page_length
= remain
;
618 if ((page_offset
+ remain
) > PAGE_SIZE
)
619 page_length
= PAGE_SIZE
- page_offset
;
621 /* If we get a fault while copying data, then (presumably) our
622 * source page isn't available. Return the error and we'll
623 * retry in the slow path.
625 if (fast_user_write(dev_priv
->mm
.gtt_mapping
, page_base
,
626 page_offset
, user_data
, page_length
)) {
631 remain
-= page_length
;
632 user_data
+= page_length
;
633 offset
+= page_length
;
637 i915_gem_object_unpin(obj
);
642 /* Per-page copy function for the shmem pwrite fastpath.
643 * Flushes invalid cachelines before writing to the target if
644 * needs_clflush_before is set and flushes out any written cachelines after
645 * writing if needs_clflush is set. */
647 shmem_pwrite_fast(struct page
*page
, int shmem_page_offset
, int page_length
,
648 char __user
*user_data
,
649 bool page_do_bit17_swizzling
,
650 bool needs_clflush_before
,
651 bool needs_clflush_after
)
656 if (unlikely(page_do_bit17_swizzling
))
659 vaddr
= kmap_atomic(page
);
660 if (needs_clflush_before
)
661 drm_clflush_virt_range(vaddr
+ shmem_page_offset
,
663 ret
= __copy_from_user_inatomic_nocache(vaddr
+ shmem_page_offset
,
666 if (needs_clflush_after
)
667 drm_clflush_virt_range(vaddr
+ shmem_page_offset
,
669 kunmap_atomic(vaddr
);
674 /* Only difference to the fast-path function is that this can handle bit17
675 * and uses non-atomic copy and kmap functions. */
677 shmem_pwrite_slow(struct page
*page
, int shmem_page_offset
, int page_length
,
678 char __user
*user_data
,
679 bool page_do_bit17_swizzling
,
680 bool needs_clflush_before
,
681 bool needs_clflush_after
)
687 if (unlikely(needs_clflush_before
|| page_do_bit17_swizzling
))
688 shmem_clflush_swizzled_range(vaddr
+ shmem_page_offset
,
690 page_do_bit17_swizzling
);
691 if (page_do_bit17_swizzling
)
692 ret
= __copy_from_user_swizzled(vaddr
, shmem_page_offset
,
696 ret
= __copy_from_user(vaddr
+ shmem_page_offset
,
699 if (needs_clflush_after
)
700 shmem_clflush_swizzled_range(vaddr
+ shmem_page_offset
,
702 page_do_bit17_swizzling
);
709 i915_gem_shmem_pwrite(struct drm_device
*dev
,
710 struct drm_i915_gem_object
*obj
,
711 struct drm_i915_gem_pwrite
*args
,
712 struct drm_file
*file
)
714 struct address_space
*mapping
= obj
->base
.filp
->f_path
.dentry
->d_inode
->i_mapping
;
717 char __user
*user_data
;
718 int shmem_page_offset
, page_length
, ret
= 0;
719 int obj_do_bit17_swizzling
, page_do_bit17_swizzling
;
720 int hit_slowpath
= 0;
721 int needs_clflush_after
= 0;
722 int needs_clflush_before
= 0;
725 user_data
= (char __user
*) (uintptr_t) args
->data_ptr
;
728 obj_do_bit17_swizzling
= i915_gem_object_needs_bit17_swizzle(obj
);
730 if (obj
->base
.write_domain
!= I915_GEM_DOMAIN_CPU
) {
731 /* If we're not in the cpu write domain, set ourself into the gtt
732 * write domain and manually flush cachelines (if required). This
733 * optimizes for the case when the gpu will use the data
734 * right away and we therefore have to clflush anyway. */
735 if (obj
->cache_level
== I915_CACHE_NONE
)
736 needs_clflush_after
= 1;
737 ret
= i915_gem_object_set_to_gtt_domain(obj
, true);
741 /* Same trick applies for invalidate partially written cachelines before
743 if (!(obj
->base
.read_domains
& I915_GEM_DOMAIN_CPU
)
744 && obj
->cache_level
== I915_CACHE_NONE
)
745 needs_clflush_before
= 1;
747 offset
= args
->offset
;
752 int partial_cacheline_write
;
754 /* Operation in this page
756 * shmem_page_offset = offset within page in shmem file
757 * page_length = bytes to copy for this page
759 shmem_page_offset
= offset_in_page(offset
);
761 page_length
= remain
;
762 if ((shmem_page_offset
+ page_length
) > PAGE_SIZE
)
763 page_length
= PAGE_SIZE
- shmem_page_offset
;
765 /* If we don't overwrite a cacheline completely we need to be
766 * careful to have up-to-date data by first clflushing. Don't
767 * overcomplicate things and flush the entire patch. */
768 partial_cacheline_write
= needs_clflush_before
&&
769 ((shmem_page_offset
| page_length
)
770 & (boot_cpu_data
.x86_clflush_size
- 1));
773 page
= obj
->pages
[offset
>> PAGE_SHIFT
];
776 page
= shmem_read_mapping_page(mapping
, offset
>> PAGE_SHIFT
);
784 page_do_bit17_swizzling
= obj_do_bit17_swizzling
&&
785 (page_to_phys(page
) & (1 << 17)) != 0;
787 ret
= shmem_pwrite_fast(page
, shmem_page_offset
, page_length
,
788 user_data
, page_do_bit17_swizzling
,
789 partial_cacheline_write
,
790 needs_clflush_after
);
795 page_cache_get(page
);
796 mutex_unlock(&dev
->struct_mutex
);
798 ret
= shmem_pwrite_slow(page
, shmem_page_offset
, page_length
,
799 user_data
, page_do_bit17_swizzling
,
800 partial_cacheline_write
,
801 needs_clflush_after
);
803 mutex_lock(&dev
->struct_mutex
);
804 page_cache_release(page
);
806 set_page_dirty(page
);
807 mark_page_accessed(page
);
809 page_cache_release(page
);
816 remain
-= page_length
;
817 user_data
+= page_length
;
818 offset
+= page_length
;
823 /* Fixup: Kill any reinstated backing storage pages */
824 if (obj
->madv
== __I915_MADV_PURGED
)
825 i915_gem_object_truncate(obj
);
826 /* and flush dirty cachelines in case the object isn't in the cpu write
828 if (obj
->base
.write_domain
!= I915_GEM_DOMAIN_CPU
) {
829 i915_gem_clflush_object(obj
);
830 intel_gtt_chipset_flush();
834 if (needs_clflush_after
)
835 intel_gtt_chipset_flush();
841 * Writes data to the object referenced by handle.
843 * On error, the contents of the buffer that were to be modified are undefined.
846 i915_gem_pwrite_ioctl(struct drm_device
*dev
, void *data
,
847 struct drm_file
*file
)
849 struct drm_i915_gem_pwrite
*args
= data
;
850 struct drm_i915_gem_object
*obj
;
856 if (!access_ok(VERIFY_READ
,
857 (char __user
*)(uintptr_t)args
->data_ptr
,
861 ret
= fault_in_multipages_readable((char __user
*)(uintptr_t)args
->data_ptr
,
866 ret
= i915_mutex_lock_interruptible(dev
);
870 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
871 if (&obj
->base
== NULL
) {
876 /* Bounds check destination. */
877 if (args
->offset
> obj
->base
.size
||
878 args
->size
> obj
->base
.size
- args
->offset
) {
883 trace_i915_gem_object_pwrite(obj
, args
->offset
, args
->size
);
886 /* We can only do the GTT pwrite on untiled buffers, as otherwise
887 * it would end up going through the fenced access, and we'll get
888 * different detiling behavior between reading and writing.
889 * pread/pwrite currently are reading and writing from the CPU
890 * perspective, requiring manual detiling by the client.
893 ret
= i915_gem_phys_pwrite(dev
, obj
, args
, file
);
897 if (obj
->gtt_space
&&
898 obj
->cache_level
== I915_CACHE_NONE
&&
899 obj
->tiling_mode
== I915_TILING_NONE
&&
900 obj
->map_and_fenceable
&&
901 obj
->base
.write_domain
!= I915_GEM_DOMAIN_CPU
) {
902 ret
= i915_gem_gtt_pwrite_fast(dev
, obj
, args
, file
);
903 /* Note that the gtt paths might fail with non-page-backed user
904 * pointers (e.g. gtt mappings when moving data between
905 * textures). Fallback to the shmem path in that case. */
909 ret
= i915_gem_shmem_pwrite(dev
, obj
, args
, file
);
912 drm_gem_object_unreference(&obj
->base
);
914 mutex_unlock(&dev
->struct_mutex
);
919 * Called when user space prepares to use an object with the CPU, either
920 * through the mmap ioctl's mapping or a GTT mapping.
923 i915_gem_set_domain_ioctl(struct drm_device
*dev
, void *data
,
924 struct drm_file
*file
)
926 struct drm_i915_gem_set_domain
*args
= data
;
927 struct drm_i915_gem_object
*obj
;
928 uint32_t read_domains
= args
->read_domains
;
929 uint32_t write_domain
= args
->write_domain
;
932 /* Only handle setting domains to types used by the CPU. */
933 if (write_domain
& I915_GEM_GPU_DOMAINS
)
936 if (read_domains
& I915_GEM_GPU_DOMAINS
)
939 /* Having something in the write domain implies it's in the read
940 * domain, and only that read domain. Enforce that in the request.
942 if (write_domain
!= 0 && read_domains
!= write_domain
)
945 ret
= i915_mutex_lock_interruptible(dev
);
949 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
950 if (&obj
->base
== NULL
) {
955 if (read_domains
& I915_GEM_DOMAIN_GTT
) {
956 ret
= i915_gem_object_set_to_gtt_domain(obj
, write_domain
!= 0);
958 /* Silently promote "you're not bound, there was nothing to do"
959 * to success, since the client was just asking us to
960 * make sure everything was done.
965 ret
= i915_gem_object_set_to_cpu_domain(obj
, write_domain
!= 0);
968 drm_gem_object_unreference(&obj
->base
);
970 mutex_unlock(&dev
->struct_mutex
);
975 * Called when user space has done writes to this buffer
978 i915_gem_sw_finish_ioctl(struct drm_device
*dev
, void *data
,
979 struct drm_file
*file
)
981 struct drm_i915_gem_sw_finish
*args
= data
;
982 struct drm_i915_gem_object
*obj
;
985 ret
= i915_mutex_lock_interruptible(dev
);
989 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
990 if (&obj
->base
== NULL
) {
995 /* Pinned buffers may be scanout, so flush the cache */
997 i915_gem_object_flush_cpu_write_domain(obj
);
999 drm_gem_object_unreference(&obj
->base
);
1001 mutex_unlock(&dev
->struct_mutex
);
1006 * Maps the contents of an object, returning the address it is mapped
1009 * While the mapping holds a reference on the contents of the object, it doesn't
1010 * imply a ref on the object itself.
1013 i915_gem_mmap_ioctl(struct drm_device
*dev
, void *data
,
1014 struct drm_file
*file
)
1016 struct drm_i915_gem_mmap
*args
= data
;
1017 struct drm_gem_object
*obj
;
1020 obj
= drm_gem_object_lookup(dev
, file
, args
->handle
);
1024 down_write(¤t
->mm
->mmap_sem
);
1025 addr
= do_mmap(obj
->filp
, 0, args
->size
,
1026 PROT_READ
| PROT_WRITE
, MAP_SHARED
,
1028 up_write(¤t
->mm
->mmap_sem
);
1029 drm_gem_object_unreference_unlocked(obj
);
1030 if (IS_ERR((void *)addr
))
1033 args
->addr_ptr
= (uint64_t) addr
;
1039 * i915_gem_fault - fault a page into the GTT
1040 * vma: VMA in question
1043 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1044 * from userspace. The fault handler takes care of binding the object to
1045 * the GTT (if needed), allocating and programming a fence register (again,
1046 * only if needed based on whether the old reg is still valid or the object
1047 * is tiled) and inserting a new PTE into the faulting process.
1049 * Note that the faulting process may involve evicting existing objects
1050 * from the GTT and/or fence registers to make room. So performance may
1051 * suffer if the GTT working set is large or there are few fence registers
1054 int i915_gem_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1056 struct drm_i915_gem_object
*obj
= to_intel_bo(vma
->vm_private_data
);
1057 struct drm_device
*dev
= obj
->base
.dev
;
1058 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
1059 pgoff_t page_offset
;
1062 bool write
= !!(vmf
->flags
& FAULT_FLAG_WRITE
);
1064 /* We don't use vmf->pgoff since that has the fake offset */
1065 page_offset
= ((unsigned long)vmf
->virtual_address
- vma
->vm_start
) >>
1068 ret
= i915_mutex_lock_interruptible(dev
);
1072 trace_i915_gem_object_fault(obj
, page_offset
, true, write
);
1074 /* Now bind it into the GTT if needed */
1075 if (!obj
->map_and_fenceable
) {
1076 ret
= i915_gem_object_unbind(obj
);
1080 if (!obj
->gtt_space
) {
1081 ret
= i915_gem_object_bind_to_gtt(obj
, 0, true);
1085 ret
= i915_gem_object_set_to_gtt_domain(obj
, write
);
1090 if (!obj
->has_global_gtt_mapping
)
1091 i915_gem_gtt_bind_object(obj
, obj
->cache_level
);
1093 ret
= i915_gem_object_get_fence(obj
);
1097 if (i915_gem_object_is_inactive(obj
))
1098 list_move_tail(&obj
->mm_list
, &dev_priv
->mm
.inactive_list
);
1100 obj
->fault_mappable
= true;
1102 pfn
= ((dev
->agp
->base
+ obj
->gtt_offset
) >> PAGE_SHIFT
) +
1105 /* Finally, remap it using the new GTT offset */
1106 ret
= vm_insert_pfn(vma
, (unsigned long)vmf
->virtual_address
, pfn
);
1108 mutex_unlock(&dev
->struct_mutex
);
1113 /* Give the error handler a chance to run and move the
1114 * objects off the GPU active list. Next time we service the
1115 * fault, we should be able to transition the page into the
1116 * GTT without touching the GPU (and so avoid further
1117 * EIO/EGAIN). If the GPU is wedged, then there is no issue
1118 * with coherency, just lost writes.
1124 return VM_FAULT_NOPAGE
;
1126 return VM_FAULT_OOM
;
1128 return VM_FAULT_SIGBUS
;
1133 * i915_gem_release_mmap - remove physical page mappings
1134 * @obj: obj in question
1136 * Preserve the reservation of the mmapping with the DRM core code, but
1137 * relinquish ownership of the pages back to the system.
1139 * It is vital that we remove the page mapping if we have mapped a tiled
1140 * object through the GTT and then lose the fence register due to
1141 * resource pressure. Similarly if the object has been moved out of the
1142 * aperture, than pages mapped into userspace must be revoked. Removing the
1143 * mapping will then trigger a page fault on the next user access, allowing
1144 * fixup by i915_gem_fault().
1147 i915_gem_release_mmap(struct drm_i915_gem_object
*obj
)
1149 if (!obj
->fault_mappable
)
1152 if (obj
->base
.dev
->dev_mapping
)
1153 unmap_mapping_range(obj
->base
.dev
->dev_mapping
,
1154 (loff_t
)obj
->base
.map_list
.hash
.key
<<PAGE_SHIFT
,
1157 obj
->fault_mappable
= false;
1161 i915_gem_get_gtt_size(struct drm_device
*dev
, uint32_t size
, int tiling_mode
)
1165 if (INTEL_INFO(dev
)->gen
>= 4 ||
1166 tiling_mode
== I915_TILING_NONE
)
1169 /* Previous chips need a power-of-two fence region when tiling */
1170 if (INTEL_INFO(dev
)->gen
== 3)
1171 gtt_size
= 1024*1024;
1173 gtt_size
= 512*1024;
1175 while (gtt_size
< size
)
1182 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1183 * @obj: object to check
1185 * Return the required GTT alignment for an object, taking into account
1186 * potential fence register mapping.
1189 i915_gem_get_gtt_alignment(struct drm_device
*dev
,
1194 * Minimum alignment is 4k (GTT page size), but might be greater
1195 * if a fence register is needed for the object.
1197 if (INTEL_INFO(dev
)->gen
>= 4 ||
1198 tiling_mode
== I915_TILING_NONE
)
1202 * Previous chips need to be aligned to the size of the smallest
1203 * fence register that can contain the object.
1205 return i915_gem_get_gtt_size(dev
, size
, tiling_mode
);
1209 * i915_gem_get_unfenced_gtt_alignment - return required GTT alignment for an
1212 * @size: size of the object
1213 * @tiling_mode: tiling mode of the object
1215 * Return the required GTT alignment for an object, only taking into account
1216 * unfenced tiled surface requirements.
1219 i915_gem_get_unfenced_gtt_alignment(struct drm_device
*dev
,
1224 * Minimum alignment is 4k (GTT page size) for sane hw.
1226 if (INTEL_INFO(dev
)->gen
>= 4 || IS_G33(dev
) ||
1227 tiling_mode
== I915_TILING_NONE
)
1230 /* Previous hardware however needs to be aligned to a power-of-two
1231 * tile height. The simplest method for determining this is to reuse
1232 * the power-of-tile object size.
1234 return i915_gem_get_gtt_size(dev
, size
, tiling_mode
);
1238 i915_gem_mmap_gtt(struct drm_file
*file
,
1239 struct drm_device
*dev
,
1243 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1244 struct drm_i915_gem_object
*obj
;
1247 ret
= i915_mutex_lock_interruptible(dev
);
1251 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, handle
));
1252 if (&obj
->base
== NULL
) {
1257 if (obj
->base
.size
> dev_priv
->mm
.gtt_mappable_end
) {
1262 if (obj
->madv
!= I915_MADV_WILLNEED
) {
1263 DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1268 if (!obj
->base
.map_list
.map
) {
1269 ret
= drm_gem_create_mmap_offset(&obj
->base
);
1274 *offset
= (u64
)obj
->base
.map_list
.hash
.key
<< PAGE_SHIFT
;
1277 drm_gem_object_unreference(&obj
->base
);
1279 mutex_unlock(&dev
->struct_mutex
);
1284 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1286 * @data: GTT mapping ioctl data
1287 * @file: GEM object info
1289 * Simply returns the fake offset to userspace so it can mmap it.
1290 * The mmap call will end up in drm_gem_mmap(), which will set things
1291 * up so we can get faults in the handler above.
1293 * The fault handler will take care of binding the object into the GTT
1294 * (since it may have been evicted to make room for something), allocating
1295 * a fence register, and mapping the appropriate aperture address into
1299 i915_gem_mmap_gtt_ioctl(struct drm_device
*dev
, void *data
,
1300 struct drm_file
*file
)
1302 struct drm_i915_gem_mmap_gtt
*args
= data
;
1304 return i915_gem_mmap_gtt(file
, dev
, args
->handle
, &args
->offset
);
1309 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object
*obj
,
1313 struct address_space
*mapping
;
1314 struct inode
*inode
;
1317 /* Get the list of pages out of our struct file. They'll be pinned
1318 * at this point until we release them.
1320 page_count
= obj
->base
.size
/ PAGE_SIZE
;
1321 BUG_ON(obj
->pages
!= NULL
);
1322 obj
->pages
= drm_malloc_ab(page_count
, sizeof(struct page
*));
1323 if (obj
->pages
== NULL
)
1326 inode
= obj
->base
.filp
->f_path
.dentry
->d_inode
;
1327 mapping
= inode
->i_mapping
;
1328 gfpmask
|= mapping_gfp_mask(mapping
);
1330 for (i
= 0; i
< page_count
; i
++) {
1331 page
= shmem_read_mapping_page_gfp(mapping
, i
, gfpmask
);
1335 obj
->pages
[i
] = page
;
1338 if (i915_gem_object_needs_bit17_swizzle(obj
))
1339 i915_gem_object_do_bit_17_swizzle(obj
);
1345 page_cache_release(obj
->pages
[i
]);
1347 drm_free_large(obj
->pages
);
1349 return PTR_ERR(page
);
1353 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object
*obj
)
1355 int page_count
= obj
->base
.size
/ PAGE_SIZE
;
1358 BUG_ON(obj
->madv
== __I915_MADV_PURGED
);
1360 if (i915_gem_object_needs_bit17_swizzle(obj
))
1361 i915_gem_object_save_bit_17_swizzle(obj
);
1363 if (obj
->madv
== I915_MADV_DONTNEED
)
1366 for (i
= 0; i
< page_count
; i
++) {
1368 set_page_dirty(obj
->pages
[i
]);
1370 if (obj
->madv
== I915_MADV_WILLNEED
)
1371 mark_page_accessed(obj
->pages
[i
]);
1373 page_cache_release(obj
->pages
[i
]);
1377 drm_free_large(obj
->pages
);
1382 i915_gem_object_move_to_active(struct drm_i915_gem_object
*obj
,
1383 struct intel_ring_buffer
*ring
,
1386 struct drm_device
*dev
= obj
->base
.dev
;
1387 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1389 BUG_ON(ring
== NULL
);
1392 /* Add a reference if we're newly entering the active list. */
1394 drm_gem_object_reference(&obj
->base
);
1398 /* Move from whatever list we were on to the tail of execution. */
1399 list_move_tail(&obj
->mm_list
, &dev_priv
->mm
.active_list
);
1400 list_move_tail(&obj
->ring_list
, &ring
->active_list
);
1402 obj
->last_rendering_seqno
= seqno
;
1404 if (obj
->fenced_gpu_access
) {
1405 obj
->last_fenced_seqno
= seqno
;
1407 /* Bump MRU to take account of the delayed flush */
1408 if (obj
->fence_reg
!= I915_FENCE_REG_NONE
) {
1409 struct drm_i915_fence_reg
*reg
;
1411 reg
= &dev_priv
->fence_regs
[obj
->fence_reg
];
1412 list_move_tail(®
->lru_list
,
1413 &dev_priv
->mm
.fence_list
);
1419 i915_gem_object_move_off_active(struct drm_i915_gem_object
*obj
)
1421 list_del_init(&obj
->ring_list
);
1422 obj
->last_rendering_seqno
= 0;
1423 obj
->last_fenced_seqno
= 0;
1427 i915_gem_object_move_to_flushing(struct drm_i915_gem_object
*obj
)
1429 struct drm_device
*dev
= obj
->base
.dev
;
1430 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
1432 BUG_ON(!obj
->active
);
1433 list_move_tail(&obj
->mm_list
, &dev_priv
->mm
.flushing_list
);
1435 i915_gem_object_move_off_active(obj
);
1439 i915_gem_object_move_to_inactive(struct drm_i915_gem_object
*obj
)
1441 struct drm_device
*dev
= obj
->base
.dev
;
1442 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1444 list_move_tail(&obj
->mm_list
, &dev_priv
->mm
.inactive_list
);
1446 BUG_ON(!list_empty(&obj
->gpu_write_list
));
1447 BUG_ON(!obj
->active
);
1450 i915_gem_object_move_off_active(obj
);
1451 obj
->fenced_gpu_access
= false;
1454 obj
->pending_gpu_write
= false;
1455 drm_gem_object_unreference(&obj
->base
);
1457 WARN_ON(i915_verify_lists(dev
));
1460 /* Immediately discard the backing storage */
1462 i915_gem_object_truncate(struct drm_i915_gem_object
*obj
)
1464 struct inode
*inode
;
1466 /* Our goal here is to return as much of the memory as
1467 * is possible back to the system as we are called from OOM.
1468 * To do this we must instruct the shmfs to drop all of its
1469 * backing pages, *now*.
1471 inode
= obj
->base
.filp
->f_path
.dentry
->d_inode
;
1472 shmem_truncate_range(inode
, 0, (loff_t
)-1);
1474 if (obj
->base
.map_list
.map
)
1475 drm_gem_free_mmap_offset(&obj
->base
);
1477 obj
->madv
= __I915_MADV_PURGED
;
1481 i915_gem_object_is_purgeable(struct drm_i915_gem_object
*obj
)
1483 return obj
->madv
== I915_MADV_DONTNEED
;
1487 i915_gem_process_flushing_list(struct intel_ring_buffer
*ring
,
1488 uint32_t flush_domains
)
1490 struct drm_i915_gem_object
*obj
, *next
;
1492 list_for_each_entry_safe(obj
, next
,
1493 &ring
->gpu_write_list
,
1495 if (obj
->base
.write_domain
& flush_domains
) {
1496 uint32_t old_write_domain
= obj
->base
.write_domain
;
1498 obj
->base
.write_domain
= 0;
1499 list_del_init(&obj
->gpu_write_list
);
1500 i915_gem_object_move_to_active(obj
, ring
,
1501 i915_gem_next_request_seqno(ring
));
1503 trace_i915_gem_object_change_domain(obj
,
1504 obj
->base
.read_domains
,
1511 i915_gem_get_seqno(struct drm_device
*dev
)
1513 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
1514 u32 seqno
= dev_priv
->next_seqno
;
1516 /* reserve 0 for non-seqno */
1517 if (++dev_priv
->next_seqno
== 0)
1518 dev_priv
->next_seqno
= 1;
1524 i915_gem_next_request_seqno(struct intel_ring_buffer
*ring
)
1526 if (ring
->outstanding_lazy_request
== 0)
1527 ring
->outstanding_lazy_request
= i915_gem_get_seqno(ring
->dev
);
1529 return ring
->outstanding_lazy_request
;
1533 i915_add_request(struct intel_ring_buffer
*ring
,
1534 struct drm_file
*file
,
1535 struct drm_i915_gem_request
*request
)
1537 drm_i915_private_t
*dev_priv
= ring
->dev
->dev_private
;
1539 u32 request_ring_position
;
1543 BUG_ON(request
== NULL
);
1544 seqno
= i915_gem_next_request_seqno(ring
);
1546 /* Record the position of the start of the request so that
1547 * should we detect the updated seqno part-way through the
1548 * GPU processing the request, we never over-estimate the
1549 * position of the head.
1551 request_ring_position
= intel_ring_get_tail(ring
);
1553 ret
= ring
->add_request(ring
, &seqno
);
1557 trace_i915_gem_request_add(ring
, seqno
);
1559 request
->seqno
= seqno
;
1560 request
->ring
= ring
;
1561 request
->tail
= request_ring_position
;
1562 request
->emitted_jiffies
= jiffies
;
1563 was_empty
= list_empty(&ring
->request_list
);
1564 list_add_tail(&request
->list
, &ring
->request_list
);
1567 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
1569 spin_lock(&file_priv
->mm
.lock
);
1570 request
->file_priv
= file_priv
;
1571 list_add_tail(&request
->client_list
,
1572 &file_priv
->mm
.request_list
);
1573 spin_unlock(&file_priv
->mm
.lock
);
1576 ring
->outstanding_lazy_request
= 0;
1578 if (!dev_priv
->mm
.suspended
) {
1579 if (i915_enable_hangcheck
) {
1580 mod_timer(&dev_priv
->hangcheck_timer
,
1582 msecs_to_jiffies(DRM_I915_HANGCHECK_PERIOD
));
1585 queue_delayed_work(dev_priv
->wq
,
1586 &dev_priv
->mm
.retire_work
, HZ
);
1592 i915_gem_request_remove_from_client(struct drm_i915_gem_request
*request
)
1594 struct drm_i915_file_private
*file_priv
= request
->file_priv
;
1599 spin_lock(&file_priv
->mm
.lock
);
1600 if (request
->file_priv
) {
1601 list_del(&request
->client_list
);
1602 request
->file_priv
= NULL
;
1604 spin_unlock(&file_priv
->mm
.lock
);
1607 static void i915_gem_reset_ring_lists(struct drm_i915_private
*dev_priv
,
1608 struct intel_ring_buffer
*ring
)
1610 while (!list_empty(&ring
->request_list
)) {
1611 struct drm_i915_gem_request
*request
;
1613 request
= list_first_entry(&ring
->request_list
,
1614 struct drm_i915_gem_request
,
1617 list_del(&request
->list
);
1618 i915_gem_request_remove_from_client(request
);
1622 while (!list_empty(&ring
->active_list
)) {
1623 struct drm_i915_gem_object
*obj
;
1625 obj
= list_first_entry(&ring
->active_list
,
1626 struct drm_i915_gem_object
,
1629 obj
->base
.write_domain
= 0;
1630 list_del_init(&obj
->gpu_write_list
);
1631 i915_gem_object_move_to_inactive(obj
);
1635 static void i915_gem_reset_fences(struct drm_device
*dev
)
1637 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1640 for (i
= 0; i
< dev_priv
->num_fence_regs
; i
++) {
1641 struct drm_i915_fence_reg
*reg
= &dev_priv
->fence_regs
[i
];
1643 i915_gem_write_fence(dev
, i
, NULL
);
1646 i915_gem_object_fence_lost(reg
->obj
);
1650 INIT_LIST_HEAD(®
->lru_list
);
1653 INIT_LIST_HEAD(&dev_priv
->mm
.fence_list
);
1656 void i915_gem_reset(struct drm_device
*dev
)
1658 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1659 struct drm_i915_gem_object
*obj
;
1662 for (i
= 0; i
< I915_NUM_RINGS
; i
++)
1663 i915_gem_reset_ring_lists(dev_priv
, &dev_priv
->ring
[i
]);
1665 /* Remove anything from the flushing lists. The GPU cache is likely
1666 * to be lost on reset along with the data, so simply move the
1667 * lost bo to the inactive list.
1669 while (!list_empty(&dev_priv
->mm
.flushing_list
)) {
1670 obj
= list_first_entry(&dev_priv
->mm
.flushing_list
,
1671 struct drm_i915_gem_object
,
1674 obj
->base
.write_domain
= 0;
1675 list_del_init(&obj
->gpu_write_list
);
1676 i915_gem_object_move_to_inactive(obj
);
1679 /* Move everything out of the GPU domains to ensure we do any
1680 * necessary invalidation upon reuse.
1682 list_for_each_entry(obj
,
1683 &dev_priv
->mm
.inactive_list
,
1686 obj
->base
.read_domains
&= ~I915_GEM_GPU_DOMAINS
;
1689 /* The fence registers are invalidated so clear them out */
1690 i915_gem_reset_fences(dev
);
1694 * This function clears the request list as sequence numbers are passed.
1697 i915_gem_retire_requests_ring(struct intel_ring_buffer
*ring
)
1702 if (list_empty(&ring
->request_list
))
1705 WARN_ON(i915_verify_lists(ring
->dev
));
1707 seqno
= ring
->get_seqno(ring
);
1709 for (i
= 0; i
< ARRAY_SIZE(ring
->sync_seqno
); i
++)
1710 if (seqno
>= ring
->sync_seqno
[i
])
1711 ring
->sync_seqno
[i
] = 0;
1713 while (!list_empty(&ring
->request_list
)) {
1714 struct drm_i915_gem_request
*request
;
1716 request
= list_first_entry(&ring
->request_list
,
1717 struct drm_i915_gem_request
,
1720 if (!i915_seqno_passed(seqno
, request
->seqno
))
1723 trace_i915_gem_request_retire(ring
, request
->seqno
);
1724 /* We know the GPU must have read the request to have
1725 * sent us the seqno + interrupt, so use the position
1726 * of tail of the request to update the last known position
1729 ring
->last_retired_head
= request
->tail
;
1731 list_del(&request
->list
);
1732 i915_gem_request_remove_from_client(request
);
1736 /* Move any buffers on the active list that are no longer referenced
1737 * by the ringbuffer to the flushing/inactive lists as appropriate.
1739 while (!list_empty(&ring
->active_list
)) {
1740 struct drm_i915_gem_object
*obj
;
1742 obj
= list_first_entry(&ring
->active_list
,
1743 struct drm_i915_gem_object
,
1746 if (!i915_seqno_passed(seqno
, obj
->last_rendering_seqno
))
1749 if (obj
->base
.write_domain
!= 0)
1750 i915_gem_object_move_to_flushing(obj
);
1752 i915_gem_object_move_to_inactive(obj
);
1755 if (unlikely(ring
->trace_irq_seqno
&&
1756 i915_seqno_passed(seqno
, ring
->trace_irq_seqno
))) {
1757 ring
->irq_put(ring
);
1758 ring
->trace_irq_seqno
= 0;
1761 WARN_ON(i915_verify_lists(ring
->dev
));
1765 i915_gem_retire_requests(struct drm_device
*dev
)
1767 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
1770 for (i
= 0; i
< I915_NUM_RINGS
; i
++)
1771 i915_gem_retire_requests_ring(&dev_priv
->ring
[i
]);
1775 i915_gem_retire_work_handler(struct work_struct
*work
)
1777 drm_i915_private_t
*dev_priv
;
1778 struct drm_device
*dev
;
1782 dev_priv
= container_of(work
, drm_i915_private_t
,
1783 mm
.retire_work
.work
);
1784 dev
= dev_priv
->dev
;
1786 /* Come back later if the device is busy... */
1787 if (!mutex_trylock(&dev
->struct_mutex
)) {
1788 queue_delayed_work(dev_priv
->wq
, &dev_priv
->mm
.retire_work
, HZ
);
1792 i915_gem_retire_requests(dev
);
1794 /* Send a periodic flush down the ring so we don't hold onto GEM
1795 * objects indefinitely.
1798 for (i
= 0; i
< I915_NUM_RINGS
; i
++) {
1799 struct intel_ring_buffer
*ring
= &dev_priv
->ring
[i
];
1801 if (!list_empty(&ring
->gpu_write_list
)) {
1802 struct drm_i915_gem_request
*request
;
1805 ret
= i915_gem_flush_ring(ring
,
1806 0, I915_GEM_GPU_DOMAINS
);
1807 request
= kzalloc(sizeof(*request
), GFP_KERNEL
);
1808 if (ret
|| request
== NULL
||
1809 i915_add_request(ring
, NULL
, request
))
1813 idle
&= list_empty(&ring
->request_list
);
1816 if (!dev_priv
->mm
.suspended
&& !idle
)
1817 queue_delayed_work(dev_priv
->wq
, &dev_priv
->mm
.retire_work
, HZ
);
1819 mutex_unlock(&dev
->struct_mutex
);
1823 * Waits for a sequence number to be signaled, and cleans up the
1824 * request and object lists appropriately for that event.
1827 i915_wait_request(struct intel_ring_buffer
*ring
,
1831 drm_i915_private_t
*dev_priv
= ring
->dev
->dev_private
;
1837 if (atomic_read(&dev_priv
->mm
.wedged
)) {
1838 struct completion
*x
= &dev_priv
->error_completion
;
1839 bool recovery_complete
;
1840 unsigned long flags
;
1842 /* Give the error handler a chance to run. */
1843 spin_lock_irqsave(&x
->wait
.lock
, flags
);
1844 recovery_complete
= x
->done
> 0;
1845 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
1847 return recovery_complete
? -EIO
: -EAGAIN
;
1850 if (seqno
== ring
->outstanding_lazy_request
) {
1851 struct drm_i915_gem_request
*request
;
1853 request
= kzalloc(sizeof(*request
), GFP_KERNEL
);
1854 if (request
== NULL
)
1857 ret
= i915_add_request(ring
, NULL
, request
);
1863 seqno
= request
->seqno
;
1866 if (!i915_seqno_passed(ring
->get_seqno(ring
), seqno
)) {
1867 if (HAS_PCH_SPLIT(ring
->dev
))
1868 ier
= I915_READ(DEIER
) | I915_READ(GTIER
);
1869 else if (IS_VALLEYVIEW(ring
->dev
))
1870 ier
= I915_READ(GTIER
) | I915_READ(VLV_IER
);
1872 ier
= I915_READ(IER
);
1874 DRM_ERROR("something (likely vbetool) disabled "
1875 "interrupts, re-enabling\n");
1876 ring
->dev
->driver
->irq_preinstall(ring
->dev
);
1877 ring
->dev
->driver
->irq_postinstall(ring
->dev
);
1880 trace_i915_gem_request_wait_begin(ring
, seqno
);
1882 ring
->waiting_seqno
= seqno
;
1883 if (ring
->irq_get(ring
)) {
1884 if (dev_priv
->mm
.interruptible
)
1885 ret
= wait_event_interruptible(ring
->irq_queue
,
1886 i915_seqno_passed(ring
->get_seqno(ring
), seqno
)
1887 || atomic_read(&dev_priv
->mm
.wedged
));
1889 wait_event(ring
->irq_queue
,
1890 i915_seqno_passed(ring
->get_seqno(ring
), seqno
)
1891 || atomic_read(&dev_priv
->mm
.wedged
));
1893 ring
->irq_put(ring
);
1894 } else if (wait_for_atomic(i915_seqno_passed(ring
->get_seqno(ring
),
1896 atomic_read(&dev_priv
->mm
.wedged
), 3000))
1898 ring
->waiting_seqno
= 0;
1900 trace_i915_gem_request_wait_end(ring
, seqno
);
1902 if (atomic_read(&dev_priv
->mm
.wedged
))
1905 /* Directly dispatch request retiring. While we have the work queue
1906 * to handle this, the waiter on a request often wants an associated
1907 * buffer to have made it to the inactive list, and we would need
1908 * a separate wait queue to handle that.
1910 if (ret
== 0 && do_retire
)
1911 i915_gem_retire_requests_ring(ring
);
1917 * Ensures that all rendering to the object has completed and the object is
1918 * safe to unbind from the GTT or access from the CPU.
1921 i915_gem_object_wait_rendering(struct drm_i915_gem_object
*obj
)
1925 /* This function only exists to support waiting for existing rendering,
1926 * not for emitting required flushes.
1928 BUG_ON((obj
->base
.write_domain
& I915_GEM_GPU_DOMAINS
) != 0);
1930 /* If there is rendering queued on the buffer being evicted, wait for
1934 ret
= i915_wait_request(obj
->ring
, obj
->last_rendering_seqno
,
1944 * i915_gem_object_sync - sync an object to a ring.
1946 * @obj: object which may be in use on another ring.
1947 * @to: ring we wish to use the object on. May be NULL.
1949 * This code is meant to abstract object synchronization with the GPU.
1950 * Calling with NULL implies synchronizing the object with the CPU
1951 * rather than a particular GPU ring.
1953 * Returns 0 if successful, else propagates up the lower layer error.
1956 i915_gem_object_sync(struct drm_i915_gem_object
*obj
,
1957 struct intel_ring_buffer
*to
)
1959 struct intel_ring_buffer
*from
= obj
->ring
;
1963 if (from
== NULL
|| to
== from
)
1966 if (to
== NULL
|| !i915_semaphore_is_enabled(obj
->base
.dev
))
1967 return i915_gem_object_wait_rendering(obj
);
1969 idx
= intel_ring_sync_index(from
, to
);
1971 seqno
= obj
->last_rendering_seqno
;
1972 if (seqno
<= from
->sync_seqno
[idx
])
1975 if (seqno
== from
->outstanding_lazy_request
) {
1976 struct drm_i915_gem_request
*request
;
1978 request
= kzalloc(sizeof(*request
), GFP_KERNEL
);
1979 if (request
== NULL
)
1982 ret
= i915_add_request(from
, NULL
, request
);
1988 seqno
= request
->seqno
;
1992 ret
= to
->sync_to(to
, from
, seqno
);
1994 from
->sync_seqno
[idx
] = seqno
;
1999 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object
*obj
)
2001 u32 old_write_domain
, old_read_domains
;
2003 /* Act a barrier for all accesses through the GTT */
2006 /* Force a pagefault for domain tracking on next user access */
2007 i915_gem_release_mmap(obj
);
2009 if ((obj
->base
.read_domains
& I915_GEM_DOMAIN_GTT
) == 0)
2012 old_read_domains
= obj
->base
.read_domains
;
2013 old_write_domain
= obj
->base
.write_domain
;
2015 obj
->base
.read_domains
&= ~I915_GEM_DOMAIN_GTT
;
2016 obj
->base
.write_domain
&= ~I915_GEM_DOMAIN_GTT
;
2018 trace_i915_gem_object_change_domain(obj
,
2024 * Unbinds an object from the GTT aperture.
2027 i915_gem_object_unbind(struct drm_i915_gem_object
*obj
)
2029 drm_i915_private_t
*dev_priv
= obj
->base
.dev
->dev_private
;
2032 if (obj
->gtt_space
== NULL
)
2035 if (obj
->pin_count
!= 0) {
2036 DRM_ERROR("Attempting to unbind pinned buffer\n");
2040 ret
= i915_gem_object_finish_gpu(obj
);
2043 /* Continue on if we fail due to EIO, the GPU is hung so we
2044 * should be safe and we need to cleanup or else we might
2045 * cause memory corruption through use-after-free.
2048 i915_gem_object_finish_gtt(obj
);
2050 /* Move the object to the CPU domain to ensure that
2051 * any possible CPU writes while it's not in the GTT
2052 * are flushed when we go to remap it.
2055 ret
= i915_gem_object_set_to_cpu_domain(obj
, 1);
2056 if (ret
== -ERESTARTSYS
)
2059 /* In the event of a disaster, abandon all caches and
2060 * hope for the best.
2062 i915_gem_clflush_object(obj
);
2063 obj
->base
.read_domains
= obj
->base
.write_domain
= I915_GEM_DOMAIN_CPU
;
2066 /* release the fence reg _after_ flushing */
2067 ret
= i915_gem_object_put_fence(obj
);
2071 trace_i915_gem_object_unbind(obj
);
2073 if (obj
->has_global_gtt_mapping
)
2074 i915_gem_gtt_unbind_object(obj
);
2075 if (obj
->has_aliasing_ppgtt_mapping
) {
2076 i915_ppgtt_unbind_object(dev_priv
->mm
.aliasing_ppgtt
, obj
);
2077 obj
->has_aliasing_ppgtt_mapping
= 0;
2079 i915_gem_gtt_finish_object(obj
);
2081 i915_gem_object_put_pages_gtt(obj
);
2083 list_del_init(&obj
->gtt_list
);
2084 list_del_init(&obj
->mm_list
);
2085 /* Avoid an unnecessary call to unbind on rebind. */
2086 obj
->map_and_fenceable
= true;
2088 drm_mm_put_block(obj
->gtt_space
);
2089 obj
->gtt_space
= NULL
;
2090 obj
->gtt_offset
= 0;
2092 if (i915_gem_object_is_purgeable(obj
))
2093 i915_gem_object_truncate(obj
);
2099 i915_gem_flush_ring(struct intel_ring_buffer
*ring
,
2100 uint32_t invalidate_domains
,
2101 uint32_t flush_domains
)
2105 if (((invalidate_domains
| flush_domains
) & I915_GEM_GPU_DOMAINS
) == 0)
2108 trace_i915_gem_ring_flush(ring
, invalidate_domains
, flush_domains
);
2110 ret
= ring
->flush(ring
, invalidate_domains
, flush_domains
);
2114 if (flush_domains
& I915_GEM_GPU_DOMAINS
)
2115 i915_gem_process_flushing_list(ring
, flush_domains
);
2120 static int i915_ring_idle(struct intel_ring_buffer
*ring
, bool do_retire
)
2124 if (list_empty(&ring
->gpu_write_list
) && list_empty(&ring
->active_list
))
2127 if (!list_empty(&ring
->gpu_write_list
)) {
2128 ret
= i915_gem_flush_ring(ring
,
2129 I915_GEM_GPU_DOMAINS
, I915_GEM_GPU_DOMAINS
);
2134 return i915_wait_request(ring
, i915_gem_next_request_seqno(ring
),
2138 int i915_gpu_idle(struct drm_device
*dev
, bool do_retire
)
2140 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
2143 /* Flush everything onto the inactive list. */
2144 for (i
= 0; i
< I915_NUM_RINGS
; i
++) {
2145 ret
= i915_ring_idle(&dev_priv
->ring
[i
], do_retire
);
2153 static void sandybridge_write_fence_reg(struct drm_device
*dev
, int reg
,
2154 struct drm_i915_gem_object
*obj
)
2156 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
2160 u32 size
= obj
->gtt_space
->size
;
2162 val
= (uint64_t)((obj
->gtt_offset
+ size
- 4096) &
2164 val
|= obj
->gtt_offset
& 0xfffff000;
2165 val
|= (uint64_t)((obj
->stride
/ 128) - 1) <<
2166 SANDYBRIDGE_FENCE_PITCH_SHIFT
;
2168 if (obj
->tiling_mode
== I915_TILING_Y
)
2169 val
|= 1 << I965_FENCE_TILING_Y_SHIFT
;
2170 val
|= I965_FENCE_REG_VALID
;
2174 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0
+ reg
* 8, val
);
2175 POSTING_READ(FENCE_REG_SANDYBRIDGE_0
+ reg
* 8);
2178 static void i965_write_fence_reg(struct drm_device
*dev
, int reg
,
2179 struct drm_i915_gem_object
*obj
)
2181 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
2185 u32 size
= obj
->gtt_space
->size
;
2187 val
= (uint64_t)((obj
->gtt_offset
+ size
- 4096) &
2189 val
|= obj
->gtt_offset
& 0xfffff000;
2190 val
|= ((obj
->stride
/ 128) - 1) << I965_FENCE_PITCH_SHIFT
;
2191 if (obj
->tiling_mode
== I915_TILING_Y
)
2192 val
|= 1 << I965_FENCE_TILING_Y_SHIFT
;
2193 val
|= I965_FENCE_REG_VALID
;
2197 I915_WRITE64(FENCE_REG_965_0
+ reg
* 8, val
);
2198 POSTING_READ(FENCE_REG_965_0
+ reg
* 8);
2201 static void i915_write_fence_reg(struct drm_device
*dev
, int reg
,
2202 struct drm_i915_gem_object
*obj
)
2204 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
2208 u32 size
= obj
->gtt_space
->size
;
2212 WARN((obj
->gtt_offset
& ~I915_FENCE_START_MASK
) ||
2213 (size
& -size
) != size
||
2214 (obj
->gtt_offset
& (size
- 1)),
2215 "object 0x%08x [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
2216 obj
->gtt_offset
, obj
->map_and_fenceable
, size
);
2218 if (obj
->tiling_mode
== I915_TILING_Y
&& HAS_128_BYTE_Y_TILING(dev
))
2223 /* Note: pitch better be a power of two tile widths */
2224 pitch_val
= obj
->stride
/ tile_width
;
2225 pitch_val
= ffs(pitch_val
) - 1;
2227 val
= obj
->gtt_offset
;
2228 if (obj
->tiling_mode
== I915_TILING_Y
)
2229 val
|= 1 << I830_FENCE_TILING_Y_SHIFT
;
2230 val
|= I915_FENCE_SIZE_BITS(size
);
2231 val
|= pitch_val
<< I830_FENCE_PITCH_SHIFT
;
2232 val
|= I830_FENCE_REG_VALID
;
2237 reg
= FENCE_REG_830_0
+ reg
* 4;
2239 reg
= FENCE_REG_945_8
+ (reg
- 8) * 4;
2241 I915_WRITE(reg
, val
);
2245 static void i830_write_fence_reg(struct drm_device
*dev
, int reg
,
2246 struct drm_i915_gem_object
*obj
)
2248 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
2252 u32 size
= obj
->gtt_space
->size
;
2255 WARN((obj
->gtt_offset
& ~I830_FENCE_START_MASK
) ||
2256 (size
& -size
) != size
||
2257 (obj
->gtt_offset
& (size
- 1)),
2258 "object 0x%08x not 512K or pot-size 0x%08x aligned\n",
2259 obj
->gtt_offset
, size
);
2261 pitch_val
= obj
->stride
/ 128;
2262 pitch_val
= ffs(pitch_val
) - 1;
2264 val
= obj
->gtt_offset
;
2265 if (obj
->tiling_mode
== I915_TILING_Y
)
2266 val
|= 1 << I830_FENCE_TILING_Y_SHIFT
;
2267 val
|= I830_FENCE_SIZE_BITS(size
);
2268 val
|= pitch_val
<< I830_FENCE_PITCH_SHIFT
;
2269 val
|= I830_FENCE_REG_VALID
;
2273 I915_WRITE(FENCE_REG_830_0
+ reg
* 4, val
);
2274 POSTING_READ(FENCE_REG_830_0
+ reg
* 4);
2277 static void i915_gem_write_fence(struct drm_device
*dev
, int reg
,
2278 struct drm_i915_gem_object
*obj
)
2280 switch (INTEL_INFO(dev
)->gen
) {
2282 case 6: sandybridge_write_fence_reg(dev
, reg
, obj
); break;
2284 case 4: i965_write_fence_reg(dev
, reg
, obj
); break;
2285 case 3: i915_write_fence_reg(dev
, reg
, obj
); break;
2286 case 2: i830_write_fence_reg(dev
, reg
, obj
); break;
2291 static inline int fence_number(struct drm_i915_private
*dev_priv
,
2292 struct drm_i915_fence_reg
*fence
)
2294 return fence
- dev_priv
->fence_regs
;
2297 static void i915_gem_object_update_fence(struct drm_i915_gem_object
*obj
,
2298 struct drm_i915_fence_reg
*fence
,
2301 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
2302 int reg
= fence_number(dev_priv
, fence
);
2304 i915_gem_write_fence(obj
->base
.dev
, reg
, enable
? obj
: NULL
);
2307 obj
->fence_reg
= reg
;
2309 list_move_tail(&fence
->lru_list
, &dev_priv
->mm
.fence_list
);
2311 obj
->fence_reg
= I915_FENCE_REG_NONE
;
2313 list_del_init(&fence
->lru_list
);
2318 i915_gem_object_flush_fence(struct drm_i915_gem_object
*obj
)
2322 if (obj
->fenced_gpu_access
) {
2323 if (obj
->base
.write_domain
& I915_GEM_GPU_DOMAINS
) {
2324 ret
= i915_gem_flush_ring(obj
->ring
,
2325 0, obj
->base
.write_domain
);
2330 obj
->fenced_gpu_access
= false;
2333 if (obj
->last_fenced_seqno
) {
2334 ret
= i915_wait_request(obj
->ring
,
2335 obj
->last_fenced_seqno
,
2340 obj
->last_fenced_seqno
= 0;
2343 /* Ensure that all CPU reads are completed before installing a fence
2344 * and all writes before removing the fence.
2346 if (obj
->base
.read_domains
& I915_GEM_DOMAIN_GTT
)
2353 i915_gem_object_put_fence(struct drm_i915_gem_object
*obj
)
2355 struct drm_i915_private
*dev_priv
= obj
->base
.dev
->dev_private
;
2358 ret
= i915_gem_object_flush_fence(obj
);
2362 if (obj
->fence_reg
== I915_FENCE_REG_NONE
)
2365 i915_gem_object_update_fence(obj
,
2366 &dev_priv
->fence_regs
[obj
->fence_reg
],
2368 i915_gem_object_fence_lost(obj
);
2373 static struct drm_i915_fence_reg
*
2374 i915_find_fence_reg(struct drm_device
*dev
)
2376 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2377 struct drm_i915_fence_reg
*reg
, *avail
;
2380 /* First try to find a free reg */
2382 for (i
= dev_priv
->fence_reg_start
; i
< dev_priv
->num_fence_regs
; i
++) {
2383 reg
= &dev_priv
->fence_regs
[i
];
2387 if (!reg
->pin_count
)
2394 /* None available, try to steal one or wait for a user to finish */
2395 list_for_each_entry(reg
, &dev_priv
->mm
.fence_list
, lru_list
) {
2406 * i915_gem_object_get_fence - set up fencing for an object
2407 * @obj: object to map through a fence reg
2409 * When mapping objects through the GTT, userspace wants to be able to write
2410 * to them without having to worry about swizzling if the object is tiled.
2411 * This function walks the fence regs looking for a free one for @obj,
2412 * stealing one if it can't find any.
2414 * It then sets up the reg based on the object's properties: address, pitch
2415 * and tiling format.
2417 * For an untiled surface, this removes any existing fence.
2420 i915_gem_object_get_fence(struct drm_i915_gem_object
*obj
)
2422 struct drm_device
*dev
= obj
->base
.dev
;
2423 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2424 bool enable
= obj
->tiling_mode
!= I915_TILING_NONE
;
2425 struct drm_i915_fence_reg
*reg
;
2428 /* Have we updated the tiling parameters upon the object and so
2429 * will need to serialise the write to the associated fence register?
2431 if (obj
->fence_dirty
) {
2432 ret
= i915_gem_object_flush_fence(obj
);
2437 /* Just update our place in the LRU if our fence is getting reused. */
2438 if (obj
->fence_reg
!= I915_FENCE_REG_NONE
) {
2439 reg
= &dev_priv
->fence_regs
[obj
->fence_reg
];
2440 if (!obj
->fence_dirty
) {
2441 list_move_tail(®
->lru_list
,
2442 &dev_priv
->mm
.fence_list
);
2445 } else if (enable
) {
2446 reg
= i915_find_fence_reg(dev
);
2451 struct drm_i915_gem_object
*old
= reg
->obj
;
2453 ret
= i915_gem_object_flush_fence(old
);
2457 i915_gem_object_fence_lost(old
);
2462 i915_gem_object_update_fence(obj
, reg
, enable
);
2463 obj
->fence_dirty
= false;
2469 * Finds free space in the GTT aperture and binds the object there.
2472 i915_gem_object_bind_to_gtt(struct drm_i915_gem_object
*obj
,
2474 bool map_and_fenceable
)
2476 struct drm_device
*dev
= obj
->base
.dev
;
2477 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
2478 struct drm_mm_node
*free_space
;
2479 gfp_t gfpmask
= __GFP_NORETRY
| __GFP_NOWARN
;
2480 u32 size
, fence_size
, fence_alignment
, unfenced_alignment
;
2481 bool mappable
, fenceable
;
2484 if (obj
->madv
!= I915_MADV_WILLNEED
) {
2485 DRM_ERROR("Attempting to bind a purgeable object\n");
2489 fence_size
= i915_gem_get_gtt_size(dev
,
2492 fence_alignment
= i915_gem_get_gtt_alignment(dev
,
2495 unfenced_alignment
=
2496 i915_gem_get_unfenced_gtt_alignment(dev
,
2501 alignment
= map_and_fenceable
? fence_alignment
:
2503 if (map_and_fenceable
&& alignment
& (fence_alignment
- 1)) {
2504 DRM_ERROR("Invalid object alignment requested %u\n", alignment
);
2508 size
= map_and_fenceable
? fence_size
: obj
->base
.size
;
2510 /* If the object is bigger than the entire aperture, reject it early
2511 * before evicting everything in a vain attempt to find space.
2513 if (obj
->base
.size
>
2514 (map_and_fenceable
? dev_priv
->mm
.gtt_mappable_end
: dev_priv
->mm
.gtt_total
)) {
2515 DRM_ERROR("Attempting to bind an object larger than the aperture\n");
2520 if (map_and_fenceable
)
2522 drm_mm_search_free_in_range(&dev_priv
->mm
.gtt_space
,
2524 dev_priv
->mm
.gtt_mappable_end
,
2527 free_space
= drm_mm_search_free(&dev_priv
->mm
.gtt_space
,
2528 size
, alignment
, 0);
2530 if (free_space
!= NULL
) {
2531 if (map_and_fenceable
)
2533 drm_mm_get_block_range_generic(free_space
,
2535 dev_priv
->mm
.gtt_mappable_end
,
2539 drm_mm_get_block(free_space
, size
, alignment
);
2541 if (obj
->gtt_space
== NULL
) {
2542 /* If the gtt is empty and we're still having trouble
2543 * fitting our object in, we're out of memory.
2545 ret
= i915_gem_evict_something(dev
, size
, alignment
,
2553 ret
= i915_gem_object_get_pages_gtt(obj
, gfpmask
);
2555 drm_mm_put_block(obj
->gtt_space
);
2556 obj
->gtt_space
= NULL
;
2558 if (ret
== -ENOMEM
) {
2559 /* first try to reclaim some memory by clearing the GTT */
2560 ret
= i915_gem_evict_everything(dev
, false);
2562 /* now try to shrink everyone else */
2577 ret
= i915_gem_gtt_prepare_object(obj
);
2579 i915_gem_object_put_pages_gtt(obj
);
2580 drm_mm_put_block(obj
->gtt_space
);
2581 obj
->gtt_space
= NULL
;
2583 if (i915_gem_evict_everything(dev
, false))
2589 if (!dev_priv
->mm
.aliasing_ppgtt
)
2590 i915_gem_gtt_bind_object(obj
, obj
->cache_level
);
2592 list_add_tail(&obj
->gtt_list
, &dev_priv
->mm
.gtt_list
);
2593 list_add_tail(&obj
->mm_list
, &dev_priv
->mm
.inactive_list
);
2595 /* Assert that the object is not currently in any GPU domain. As it
2596 * wasn't in the GTT, there shouldn't be any way it could have been in
2599 BUG_ON(obj
->base
.read_domains
& I915_GEM_GPU_DOMAINS
);
2600 BUG_ON(obj
->base
.write_domain
& I915_GEM_GPU_DOMAINS
);
2602 obj
->gtt_offset
= obj
->gtt_space
->start
;
2605 obj
->gtt_space
->size
== fence_size
&&
2606 (obj
->gtt_space
->start
& (fence_alignment
- 1)) == 0;
2609 obj
->gtt_offset
+ obj
->base
.size
<= dev_priv
->mm
.gtt_mappable_end
;
2611 obj
->map_and_fenceable
= mappable
&& fenceable
;
2613 trace_i915_gem_object_bind(obj
, map_and_fenceable
);
2618 i915_gem_clflush_object(struct drm_i915_gem_object
*obj
)
2620 /* If we don't have a page list set up, then we're not pinned
2621 * to GPU, and we can ignore the cache flush because it'll happen
2622 * again at bind time.
2624 if (obj
->pages
== NULL
)
2627 /* If the GPU is snooping the contents of the CPU cache,
2628 * we do not need to manually clear the CPU cache lines. However,
2629 * the caches are only snooped when the render cache is
2630 * flushed/invalidated. As we always have to emit invalidations
2631 * and flushes when moving into and out of the RENDER domain, correct
2632 * snooping behaviour occurs naturally as the result of our domain
2635 if (obj
->cache_level
!= I915_CACHE_NONE
)
2638 trace_i915_gem_object_clflush(obj
);
2640 drm_clflush_pages(obj
->pages
, obj
->base
.size
/ PAGE_SIZE
);
2643 /** Flushes any GPU write domain for the object if it's dirty. */
2645 i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object
*obj
)
2647 if ((obj
->base
.write_domain
& I915_GEM_GPU_DOMAINS
) == 0)
2650 /* Queue the GPU write cache flushing we need. */
2651 return i915_gem_flush_ring(obj
->ring
, 0, obj
->base
.write_domain
);
2654 /** Flushes the GTT write domain for the object if it's dirty. */
2656 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object
*obj
)
2658 uint32_t old_write_domain
;
2660 if (obj
->base
.write_domain
!= I915_GEM_DOMAIN_GTT
)
2663 /* No actual flushing is required for the GTT write domain. Writes
2664 * to it immediately go to main memory as far as we know, so there's
2665 * no chipset flush. It also doesn't land in render cache.
2667 * However, we do have to enforce the order so that all writes through
2668 * the GTT land before any writes to the device, such as updates to
2673 old_write_domain
= obj
->base
.write_domain
;
2674 obj
->base
.write_domain
= 0;
2676 trace_i915_gem_object_change_domain(obj
,
2677 obj
->base
.read_domains
,
2681 /** Flushes the CPU write domain for the object if it's dirty. */
2683 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object
*obj
)
2685 uint32_t old_write_domain
;
2687 if (obj
->base
.write_domain
!= I915_GEM_DOMAIN_CPU
)
2690 i915_gem_clflush_object(obj
);
2691 intel_gtt_chipset_flush();
2692 old_write_domain
= obj
->base
.write_domain
;
2693 obj
->base
.write_domain
= 0;
2695 trace_i915_gem_object_change_domain(obj
,
2696 obj
->base
.read_domains
,
2701 * Moves a single object to the GTT read, and possibly write domain.
2703 * This function returns when the move is complete, including waiting on
2707 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object
*obj
, bool write
)
2709 drm_i915_private_t
*dev_priv
= obj
->base
.dev
->dev_private
;
2710 uint32_t old_write_domain
, old_read_domains
;
2713 /* Not valid to be called on unbound objects. */
2714 if (obj
->gtt_space
== NULL
)
2717 if (obj
->base
.write_domain
== I915_GEM_DOMAIN_GTT
)
2720 ret
= i915_gem_object_flush_gpu_write_domain(obj
);
2724 if (obj
->pending_gpu_write
|| write
) {
2725 ret
= i915_gem_object_wait_rendering(obj
);
2730 i915_gem_object_flush_cpu_write_domain(obj
);
2732 old_write_domain
= obj
->base
.write_domain
;
2733 old_read_domains
= obj
->base
.read_domains
;
2735 /* It should now be out of any other write domains, and we can update
2736 * the domain values for our changes.
2738 BUG_ON((obj
->base
.write_domain
& ~I915_GEM_DOMAIN_GTT
) != 0);
2739 obj
->base
.read_domains
|= I915_GEM_DOMAIN_GTT
;
2741 obj
->base
.read_domains
= I915_GEM_DOMAIN_GTT
;
2742 obj
->base
.write_domain
= I915_GEM_DOMAIN_GTT
;
2746 trace_i915_gem_object_change_domain(obj
,
2750 /* And bump the LRU for this access */
2751 if (i915_gem_object_is_inactive(obj
))
2752 list_move_tail(&obj
->mm_list
, &dev_priv
->mm
.inactive_list
);
2757 int i915_gem_object_set_cache_level(struct drm_i915_gem_object
*obj
,
2758 enum i915_cache_level cache_level
)
2760 struct drm_device
*dev
= obj
->base
.dev
;
2761 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
2764 if (obj
->cache_level
== cache_level
)
2767 if (obj
->pin_count
) {
2768 DRM_DEBUG("can not change the cache level of pinned objects\n");
2772 if (obj
->gtt_space
) {
2773 ret
= i915_gem_object_finish_gpu(obj
);
2777 i915_gem_object_finish_gtt(obj
);
2779 /* Before SandyBridge, you could not use tiling or fence
2780 * registers with snooped memory, so relinquish any fences
2781 * currently pointing to our region in the aperture.
2783 if (INTEL_INFO(obj
->base
.dev
)->gen
< 6) {
2784 ret
= i915_gem_object_put_fence(obj
);
2789 if (obj
->has_global_gtt_mapping
)
2790 i915_gem_gtt_bind_object(obj
, cache_level
);
2791 if (obj
->has_aliasing_ppgtt_mapping
)
2792 i915_ppgtt_bind_object(dev_priv
->mm
.aliasing_ppgtt
,
2796 if (cache_level
== I915_CACHE_NONE
) {
2797 u32 old_read_domains
, old_write_domain
;
2799 /* If we're coming from LLC cached, then we haven't
2800 * actually been tracking whether the data is in the
2801 * CPU cache or not, since we only allow one bit set
2802 * in obj->write_domain and have been skipping the clflushes.
2803 * Just set it to the CPU cache for now.
2805 WARN_ON(obj
->base
.write_domain
& ~I915_GEM_DOMAIN_CPU
);
2806 WARN_ON(obj
->base
.read_domains
& ~I915_GEM_DOMAIN_CPU
);
2808 old_read_domains
= obj
->base
.read_domains
;
2809 old_write_domain
= obj
->base
.write_domain
;
2811 obj
->base
.read_domains
= I915_GEM_DOMAIN_CPU
;
2812 obj
->base
.write_domain
= I915_GEM_DOMAIN_CPU
;
2814 trace_i915_gem_object_change_domain(obj
,
2819 obj
->cache_level
= cache_level
;
2824 * Prepare buffer for display plane (scanout, cursors, etc).
2825 * Can be called from an uninterruptible phase (modesetting) and allows
2826 * any flushes to be pipelined (for pageflips).
2829 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object
*obj
,
2831 struct intel_ring_buffer
*pipelined
)
2833 u32 old_read_domains
, old_write_domain
;
2836 ret
= i915_gem_object_flush_gpu_write_domain(obj
);
2840 if (pipelined
!= obj
->ring
) {
2841 ret
= i915_gem_object_sync(obj
, pipelined
);
2846 /* The display engine is not coherent with the LLC cache on gen6. As
2847 * a result, we make sure that the pinning that is about to occur is
2848 * done with uncached PTEs. This is lowest common denominator for all
2851 * However for gen6+, we could do better by using the GFDT bit instead
2852 * of uncaching, which would allow us to flush all the LLC-cached data
2853 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
2855 ret
= i915_gem_object_set_cache_level(obj
, I915_CACHE_NONE
);
2859 /* As the user may map the buffer once pinned in the display plane
2860 * (e.g. libkms for the bootup splash), we have to ensure that we
2861 * always use map_and_fenceable for all scanout buffers.
2863 ret
= i915_gem_object_pin(obj
, alignment
, true);
2867 i915_gem_object_flush_cpu_write_domain(obj
);
2869 old_write_domain
= obj
->base
.write_domain
;
2870 old_read_domains
= obj
->base
.read_domains
;
2872 /* It should now be out of any other write domains, and we can update
2873 * the domain values for our changes.
2875 BUG_ON((obj
->base
.write_domain
& ~I915_GEM_DOMAIN_GTT
) != 0);
2876 obj
->base
.read_domains
|= I915_GEM_DOMAIN_GTT
;
2878 trace_i915_gem_object_change_domain(obj
,
2886 i915_gem_object_finish_gpu(struct drm_i915_gem_object
*obj
)
2890 if ((obj
->base
.read_domains
& I915_GEM_GPU_DOMAINS
) == 0)
2893 if (obj
->base
.write_domain
& I915_GEM_GPU_DOMAINS
) {
2894 ret
= i915_gem_flush_ring(obj
->ring
, 0, obj
->base
.write_domain
);
2899 ret
= i915_gem_object_wait_rendering(obj
);
2903 /* Ensure that we invalidate the GPU's caches and TLBs. */
2904 obj
->base
.read_domains
&= ~I915_GEM_GPU_DOMAINS
;
2909 * Moves a single object to the CPU read, and possibly write domain.
2911 * This function returns when the move is complete, including waiting on
2915 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object
*obj
, bool write
)
2917 uint32_t old_write_domain
, old_read_domains
;
2920 if (obj
->base
.write_domain
== I915_GEM_DOMAIN_CPU
)
2923 ret
= i915_gem_object_flush_gpu_write_domain(obj
);
2927 if (write
|| obj
->pending_gpu_write
) {
2928 ret
= i915_gem_object_wait_rendering(obj
);
2933 i915_gem_object_flush_gtt_write_domain(obj
);
2935 old_write_domain
= obj
->base
.write_domain
;
2936 old_read_domains
= obj
->base
.read_domains
;
2938 /* Flush the CPU cache if it's still invalid. */
2939 if ((obj
->base
.read_domains
& I915_GEM_DOMAIN_CPU
) == 0) {
2940 i915_gem_clflush_object(obj
);
2942 obj
->base
.read_domains
|= I915_GEM_DOMAIN_CPU
;
2945 /* It should now be out of any other write domains, and we can update
2946 * the domain values for our changes.
2948 BUG_ON((obj
->base
.write_domain
& ~I915_GEM_DOMAIN_CPU
) != 0);
2950 /* If we're writing through the CPU, then the GPU read domains will
2951 * need to be invalidated at next use.
2954 obj
->base
.read_domains
= I915_GEM_DOMAIN_CPU
;
2955 obj
->base
.write_domain
= I915_GEM_DOMAIN_CPU
;
2958 trace_i915_gem_object_change_domain(obj
,
2965 /* Throttle our rendering by waiting until the ring has completed our requests
2966 * emitted over 20 msec ago.
2968 * Note that if we were to use the current jiffies each time around the loop,
2969 * we wouldn't escape the function with any frames outstanding if the time to
2970 * render a frame was over 20ms.
2972 * This should get us reasonable parallelism between CPU and GPU but also
2973 * relatively low latency when blocking on a particular request to finish.
2976 i915_gem_ring_throttle(struct drm_device
*dev
, struct drm_file
*file
)
2978 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2979 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
2980 unsigned long recent_enough
= jiffies
- msecs_to_jiffies(20);
2981 struct drm_i915_gem_request
*request
;
2982 struct intel_ring_buffer
*ring
= NULL
;
2986 if (atomic_read(&dev_priv
->mm
.wedged
))
2989 spin_lock(&file_priv
->mm
.lock
);
2990 list_for_each_entry(request
, &file_priv
->mm
.request_list
, client_list
) {
2991 if (time_after_eq(request
->emitted_jiffies
, recent_enough
))
2994 ring
= request
->ring
;
2995 seqno
= request
->seqno
;
2997 spin_unlock(&file_priv
->mm
.lock
);
3003 if (!i915_seqno_passed(ring
->get_seqno(ring
), seqno
)) {
3004 /* And wait for the seqno passing without holding any locks and
3005 * causing extra latency for others. This is safe as the irq
3006 * generation is designed to be run atomically and so is
3009 if (ring
->irq_get(ring
)) {
3010 ret
= wait_event_interruptible(ring
->irq_queue
,
3011 i915_seqno_passed(ring
->get_seqno(ring
), seqno
)
3012 || atomic_read(&dev_priv
->mm
.wedged
));
3013 ring
->irq_put(ring
);
3015 if (ret
== 0 && atomic_read(&dev_priv
->mm
.wedged
))
3017 } else if (wait_for_atomic(i915_seqno_passed(ring
->get_seqno(ring
),
3019 atomic_read(&dev_priv
->mm
.wedged
), 3000)) {
3025 queue_delayed_work(dev_priv
->wq
, &dev_priv
->mm
.retire_work
, 0);
3031 i915_gem_object_pin(struct drm_i915_gem_object
*obj
,
3033 bool map_and_fenceable
)
3037 BUG_ON(obj
->pin_count
== DRM_I915_GEM_OBJECT_MAX_PIN_COUNT
);
3039 if (obj
->gtt_space
!= NULL
) {
3040 if ((alignment
&& obj
->gtt_offset
& (alignment
- 1)) ||
3041 (map_and_fenceable
&& !obj
->map_and_fenceable
)) {
3042 WARN(obj
->pin_count
,
3043 "bo is already pinned with incorrect alignment:"
3044 " offset=%x, req.alignment=%x, req.map_and_fenceable=%d,"
3045 " obj->map_and_fenceable=%d\n",
3046 obj
->gtt_offset
, alignment
,
3048 obj
->map_and_fenceable
);
3049 ret
= i915_gem_object_unbind(obj
);
3055 if (obj
->gtt_space
== NULL
) {
3056 ret
= i915_gem_object_bind_to_gtt(obj
, alignment
,
3062 if (!obj
->has_global_gtt_mapping
&& map_and_fenceable
)
3063 i915_gem_gtt_bind_object(obj
, obj
->cache_level
);
3066 obj
->pin_mappable
|= map_and_fenceable
;
3072 i915_gem_object_unpin(struct drm_i915_gem_object
*obj
)
3074 BUG_ON(obj
->pin_count
== 0);
3075 BUG_ON(obj
->gtt_space
== NULL
);
3077 if (--obj
->pin_count
== 0)
3078 obj
->pin_mappable
= false;
3082 i915_gem_pin_ioctl(struct drm_device
*dev
, void *data
,
3083 struct drm_file
*file
)
3085 struct drm_i915_gem_pin
*args
= data
;
3086 struct drm_i915_gem_object
*obj
;
3089 ret
= i915_mutex_lock_interruptible(dev
);
3093 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
3094 if (&obj
->base
== NULL
) {
3099 if (obj
->madv
!= I915_MADV_WILLNEED
) {
3100 DRM_ERROR("Attempting to pin a purgeable buffer\n");
3105 if (obj
->pin_filp
!= NULL
&& obj
->pin_filp
!= file
) {
3106 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
3112 obj
->user_pin_count
++;
3113 obj
->pin_filp
= file
;
3114 if (obj
->user_pin_count
== 1) {
3115 ret
= i915_gem_object_pin(obj
, args
->alignment
, true);
3120 /* XXX - flush the CPU caches for pinned objects
3121 * as the X server doesn't manage domains yet
3123 i915_gem_object_flush_cpu_write_domain(obj
);
3124 args
->offset
= obj
->gtt_offset
;
3126 drm_gem_object_unreference(&obj
->base
);
3128 mutex_unlock(&dev
->struct_mutex
);
3133 i915_gem_unpin_ioctl(struct drm_device
*dev
, void *data
,
3134 struct drm_file
*file
)
3136 struct drm_i915_gem_pin
*args
= data
;
3137 struct drm_i915_gem_object
*obj
;
3140 ret
= i915_mutex_lock_interruptible(dev
);
3144 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
3145 if (&obj
->base
== NULL
) {
3150 if (obj
->pin_filp
!= file
) {
3151 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
3156 obj
->user_pin_count
--;
3157 if (obj
->user_pin_count
== 0) {
3158 obj
->pin_filp
= NULL
;
3159 i915_gem_object_unpin(obj
);
3163 drm_gem_object_unreference(&obj
->base
);
3165 mutex_unlock(&dev
->struct_mutex
);
3170 i915_gem_busy_ioctl(struct drm_device
*dev
, void *data
,
3171 struct drm_file
*file
)
3173 struct drm_i915_gem_busy
*args
= data
;
3174 struct drm_i915_gem_object
*obj
;
3177 ret
= i915_mutex_lock_interruptible(dev
);
3181 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file
, args
->handle
));
3182 if (&obj
->base
== NULL
) {
3187 /* Count all active objects as busy, even if they are currently not used
3188 * by the gpu. Users of this interface expect objects to eventually
3189 * become non-busy without any further actions, therefore emit any
3190 * necessary flushes here.
3192 args
->busy
= obj
->active
;
3194 /* Unconditionally flush objects, even when the gpu still uses this
3195 * object. Userspace calling this function indicates that it wants to
3196 * use this buffer rather sooner than later, so issuing the required
3197 * flush earlier is beneficial.
3199 if (obj
->base
.write_domain
& I915_GEM_GPU_DOMAINS
) {
3200 ret
= i915_gem_flush_ring(obj
->ring
,
3201 0, obj
->base
.write_domain
);
3202 } else if (obj
->ring
->outstanding_lazy_request
==
3203 obj
->last_rendering_seqno
) {
3204 struct drm_i915_gem_request
*request
;
3206 /* This ring is not being cleared by active usage,
3207 * so emit a request to do so.
3209 request
= kzalloc(sizeof(*request
), GFP_KERNEL
);
3211 ret
= i915_add_request(obj
->ring
, NULL
, request
);
3218 /* Update the active list for the hardware's current position.
3219 * Otherwise this only updates on a delayed timer or when irqs
3220 * are actually unmasked, and our working set ends up being
3221 * larger than required.
3223 i915_gem_retire_requests_ring(obj
->ring
);
3225 args
->busy
= obj
->active
;
3228 drm_gem_object_unreference(&obj
->base
);
3230 mutex_unlock(&dev
->struct_mutex
);
3235 i915_gem_throttle_ioctl(struct drm_device
*dev
, void *data
,
3236 struct drm_file
*file_priv
)
3238 return i915_gem_ring_throttle(dev
, file_priv
);
3242 i915_gem_madvise_ioctl(struct drm_device
*dev
, void *data
,
3243 struct drm_file
*file_priv
)
3245 struct drm_i915_gem_madvise
*args
= data
;
3246 struct drm_i915_gem_object
*obj
;
3249 switch (args
->madv
) {
3250 case I915_MADV_DONTNEED
:
3251 case I915_MADV_WILLNEED
:
3257 ret
= i915_mutex_lock_interruptible(dev
);
3261 obj
= to_intel_bo(drm_gem_object_lookup(dev
, file_priv
, args
->handle
));
3262 if (&obj
->base
== NULL
) {
3267 if (obj
->pin_count
) {
3272 if (obj
->madv
!= __I915_MADV_PURGED
)
3273 obj
->madv
= args
->madv
;
3275 /* if the object is no longer bound, discard its backing storage */
3276 if (i915_gem_object_is_purgeable(obj
) &&
3277 obj
->gtt_space
== NULL
)
3278 i915_gem_object_truncate(obj
);
3280 args
->retained
= obj
->madv
!= __I915_MADV_PURGED
;
3283 drm_gem_object_unreference(&obj
->base
);
3285 mutex_unlock(&dev
->struct_mutex
);
3289 struct drm_i915_gem_object
*i915_gem_alloc_object(struct drm_device
*dev
,
3292 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3293 struct drm_i915_gem_object
*obj
;
3294 struct address_space
*mapping
;
3296 obj
= kzalloc(sizeof(*obj
), GFP_KERNEL
);
3300 if (drm_gem_object_init(dev
, &obj
->base
, size
) != 0) {
3305 mapping
= obj
->base
.filp
->f_path
.dentry
->d_inode
->i_mapping
;
3306 mapping_set_gfp_mask(mapping
, GFP_HIGHUSER
| __GFP_RECLAIMABLE
);
3308 i915_gem_info_add_obj(dev_priv
, size
);
3310 obj
->base
.write_domain
= I915_GEM_DOMAIN_CPU
;
3311 obj
->base
.read_domains
= I915_GEM_DOMAIN_CPU
;
3314 /* On some devices, we can have the GPU use the LLC (the CPU
3315 * cache) for about a 10% performance improvement
3316 * compared to uncached. Graphics requests other than
3317 * display scanout are coherent with the CPU in
3318 * accessing this cache. This means in this mode we
3319 * don't need to clflush on the CPU side, and on the
3320 * GPU side we only need to flush internal caches to
3321 * get data visible to the CPU.
3323 * However, we maintain the display planes as UC, and so
3324 * need to rebind when first used as such.
3326 obj
->cache_level
= I915_CACHE_LLC
;
3328 obj
->cache_level
= I915_CACHE_NONE
;
3330 obj
->base
.driver_private
= NULL
;
3331 obj
->fence_reg
= I915_FENCE_REG_NONE
;
3332 INIT_LIST_HEAD(&obj
->mm_list
);
3333 INIT_LIST_HEAD(&obj
->gtt_list
);
3334 INIT_LIST_HEAD(&obj
->ring_list
);
3335 INIT_LIST_HEAD(&obj
->exec_list
);
3336 INIT_LIST_HEAD(&obj
->gpu_write_list
);
3337 obj
->madv
= I915_MADV_WILLNEED
;
3338 /* Avoid an unnecessary call to unbind on the first bind. */
3339 obj
->map_and_fenceable
= true;
3344 int i915_gem_init_object(struct drm_gem_object
*obj
)
3351 void i915_gem_free_object(struct drm_gem_object
*gem_obj
)
3353 struct drm_i915_gem_object
*obj
= to_intel_bo(gem_obj
);
3354 struct drm_device
*dev
= obj
->base
.dev
;
3355 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
3357 trace_i915_gem_object_destroy(obj
);
3360 i915_gem_detach_phys_object(dev
, obj
);
3363 if (WARN_ON(i915_gem_object_unbind(obj
) == -ERESTARTSYS
)) {
3364 bool was_interruptible
;
3366 was_interruptible
= dev_priv
->mm
.interruptible
;
3367 dev_priv
->mm
.interruptible
= false;
3369 WARN_ON(i915_gem_object_unbind(obj
));
3371 dev_priv
->mm
.interruptible
= was_interruptible
;
3374 if (obj
->base
.map_list
.map
)
3375 drm_gem_free_mmap_offset(&obj
->base
);
3377 drm_gem_object_release(&obj
->base
);
3378 i915_gem_info_remove_obj(dev_priv
, obj
->base
.size
);
3385 i915_gem_idle(struct drm_device
*dev
)
3387 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
3390 mutex_lock(&dev
->struct_mutex
);
3392 if (dev_priv
->mm
.suspended
) {
3393 mutex_unlock(&dev
->struct_mutex
);
3397 ret
= i915_gpu_idle(dev
, true);
3399 mutex_unlock(&dev
->struct_mutex
);
3403 /* Under UMS, be paranoid and evict. */
3404 if (!drm_core_check_feature(dev
, DRIVER_MODESET
))
3405 i915_gem_evict_everything(dev
, false);
3407 i915_gem_reset_fences(dev
);
3409 /* Hack! Don't let anybody do execbuf while we don't control the chip.
3410 * We need to replace this with a semaphore, or something.
3411 * And not confound mm.suspended!
3413 dev_priv
->mm
.suspended
= 1;
3414 del_timer_sync(&dev_priv
->hangcheck_timer
);
3416 i915_kernel_lost_context(dev
);
3417 i915_gem_cleanup_ringbuffer(dev
);
3419 mutex_unlock(&dev
->struct_mutex
);
3421 /* Cancel the retire work handler, which should be idle now. */
3422 cancel_delayed_work_sync(&dev_priv
->mm
.retire_work
);
3427 void i915_gem_init_swizzling(struct drm_device
*dev
)
3429 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
3431 if (INTEL_INFO(dev
)->gen
< 5 ||
3432 dev_priv
->mm
.bit_6_swizzle_x
== I915_BIT_6_SWIZZLE_NONE
)
3435 I915_WRITE(DISP_ARB_CTL
, I915_READ(DISP_ARB_CTL
) |
3436 DISP_TILE_SURFACE_SWIZZLING
);
3441 I915_WRITE(TILECTL
, I915_READ(TILECTL
) | TILECTL_SWZCTL
);
3443 I915_WRITE(ARB_MODE
, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB
));
3445 I915_WRITE(ARB_MODE
, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB
));
3448 void i915_gem_init_ppgtt(struct drm_device
*dev
)
3450 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
3452 struct intel_ring_buffer
*ring
;
3453 struct i915_hw_ppgtt
*ppgtt
= dev_priv
->mm
.aliasing_ppgtt
;
3454 uint32_t __iomem
*pd_addr
;
3458 if (!dev_priv
->mm
.aliasing_ppgtt
)
3462 pd_addr
= dev_priv
->mm
.gtt
->gtt
+ ppgtt
->pd_offset
/sizeof(uint32_t);
3463 for (i
= 0; i
< ppgtt
->num_pd_entries
; i
++) {
3466 if (dev_priv
->mm
.gtt
->needs_dmar
)
3467 pt_addr
= ppgtt
->pt_dma_addr
[i
];
3469 pt_addr
= page_to_phys(ppgtt
->pt_pages
[i
]);
3471 pd_entry
= GEN6_PDE_ADDR_ENCODE(pt_addr
);
3472 pd_entry
|= GEN6_PDE_VALID
;
3474 writel(pd_entry
, pd_addr
+ i
);
3478 pd_offset
= ppgtt
->pd_offset
;
3479 pd_offset
/= 64; /* in cachelines, */
3482 if (INTEL_INFO(dev
)->gen
== 6) {
3483 uint32_t ecochk
, gab_ctl
, ecobits
;
3485 ecobits
= I915_READ(GAC_ECO_BITS
);
3486 I915_WRITE(GAC_ECO_BITS
, ecobits
| ECOBITS_PPGTT_CACHE64B
);
3488 gab_ctl
= I915_READ(GAB_CTL
);
3489 I915_WRITE(GAB_CTL
, gab_ctl
| GAB_CTL_CONT_AFTER_PAGEFAULT
);
3491 ecochk
= I915_READ(GAM_ECOCHK
);
3492 I915_WRITE(GAM_ECOCHK
, ecochk
| ECOCHK_SNB_BIT
|
3493 ECOCHK_PPGTT_CACHE64B
);
3494 I915_WRITE(GFX_MODE
, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE
));
3495 } else if (INTEL_INFO(dev
)->gen
>= 7) {
3496 I915_WRITE(GAM_ECOCHK
, ECOCHK_PPGTT_CACHE64B
);
3497 /* GFX_MODE is per-ring on gen7+ */
3500 for (i
= 0; i
< I915_NUM_RINGS
; i
++) {
3501 ring
= &dev_priv
->ring
[i
];
3503 if (INTEL_INFO(dev
)->gen
>= 7)
3504 I915_WRITE(RING_MODE_GEN7(ring
),
3505 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE
));
3507 I915_WRITE(RING_PP_DIR_DCLV(ring
), PP_DIR_DCLV_2G
);
3508 I915_WRITE(RING_PP_DIR_BASE(ring
), pd_offset
);
3513 i915_gem_init_hw(struct drm_device
*dev
)
3515 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
3518 i915_gem_init_swizzling(dev
);
3520 ret
= intel_init_render_ring_buffer(dev
);
3525 ret
= intel_init_bsd_ring_buffer(dev
);
3527 goto cleanup_render_ring
;
3531 ret
= intel_init_blt_ring_buffer(dev
);
3533 goto cleanup_bsd_ring
;
3536 dev_priv
->next_seqno
= 1;
3538 i915_gem_init_ppgtt(dev
);
3543 intel_cleanup_ring_buffer(&dev_priv
->ring
[VCS
]);
3544 cleanup_render_ring
:
3545 intel_cleanup_ring_buffer(&dev_priv
->ring
[RCS
]);
3550 intel_enable_ppgtt(struct drm_device
*dev
)
3552 if (i915_enable_ppgtt
>= 0)
3553 return i915_enable_ppgtt
;
3555 #ifdef CONFIG_INTEL_IOMMU
3556 /* Disable ppgtt on SNB if VT-d is on. */
3557 if (INTEL_INFO(dev
)->gen
== 6 && intel_iommu_gfx_mapped
)
3564 int i915_gem_init(struct drm_device
*dev
)
3566 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3567 unsigned long gtt_size
, mappable_size
;
3570 gtt_size
= dev_priv
->mm
.gtt
->gtt_total_entries
<< PAGE_SHIFT
;
3571 mappable_size
= dev_priv
->mm
.gtt
->gtt_mappable_entries
<< PAGE_SHIFT
;
3573 mutex_lock(&dev
->struct_mutex
);
3574 if (intel_enable_ppgtt(dev
) && HAS_ALIASING_PPGTT(dev
)) {
3575 /* PPGTT pdes are stolen from global gtt ptes, so shrink the
3576 * aperture accordingly when using aliasing ppgtt. */
3577 gtt_size
-= I915_PPGTT_PD_ENTRIES
*PAGE_SIZE
;
3579 i915_gem_init_global_gtt(dev
, 0, mappable_size
, gtt_size
);
3581 ret
= i915_gem_init_aliasing_ppgtt(dev
);
3583 mutex_unlock(&dev
->struct_mutex
);
3587 /* Let GEM Manage all of the aperture.
3589 * However, leave one page at the end still bound to the scratch
3590 * page. There are a number of places where the hardware
3591 * apparently prefetches past the end of the object, and we've
3592 * seen multiple hangs with the GPU head pointer stuck in a
3593 * batchbuffer bound at the last page of the aperture. One page
3594 * should be enough to keep any prefetching inside of the
3597 i915_gem_init_global_gtt(dev
, 0, mappable_size
,
3601 ret
= i915_gem_init_hw(dev
);
3602 mutex_unlock(&dev
->struct_mutex
);
3604 i915_gem_cleanup_aliasing_ppgtt(dev
);
3608 /* Allow hardware batchbuffers unless told otherwise. */
3609 dev_priv
->allow_batchbuffer
= 1;
3614 i915_gem_cleanup_ringbuffer(struct drm_device
*dev
)
3616 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
3619 for (i
= 0; i
< I915_NUM_RINGS
; i
++)
3620 intel_cleanup_ring_buffer(&dev_priv
->ring
[i
]);
3624 i915_gem_entervt_ioctl(struct drm_device
*dev
, void *data
,
3625 struct drm_file
*file_priv
)
3627 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
3630 if (drm_core_check_feature(dev
, DRIVER_MODESET
))
3633 if (atomic_read(&dev_priv
->mm
.wedged
)) {
3634 DRM_ERROR("Reenabling wedged hardware, good luck\n");
3635 atomic_set(&dev_priv
->mm
.wedged
, 0);
3638 mutex_lock(&dev
->struct_mutex
);
3639 dev_priv
->mm
.suspended
= 0;
3641 ret
= i915_gem_init_hw(dev
);
3643 mutex_unlock(&dev
->struct_mutex
);
3647 BUG_ON(!list_empty(&dev_priv
->mm
.active_list
));
3648 BUG_ON(!list_empty(&dev_priv
->mm
.flushing_list
));
3649 BUG_ON(!list_empty(&dev_priv
->mm
.inactive_list
));
3650 for (i
= 0; i
< I915_NUM_RINGS
; i
++) {
3651 BUG_ON(!list_empty(&dev_priv
->ring
[i
].active_list
));
3652 BUG_ON(!list_empty(&dev_priv
->ring
[i
].request_list
));
3654 mutex_unlock(&dev
->struct_mutex
);
3656 ret
= drm_irq_install(dev
);
3658 goto cleanup_ringbuffer
;
3663 mutex_lock(&dev
->struct_mutex
);
3664 i915_gem_cleanup_ringbuffer(dev
);
3665 dev_priv
->mm
.suspended
= 1;
3666 mutex_unlock(&dev
->struct_mutex
);
3672 i915_gem_leavevt_ioctl(struct drm_device
*dev
, void *data
,
3673 struct drm_file
*file_priv
)
3675 if (drm_core_check_feature(dev
, DRIVER_MODESET
))
3678 drm_irq_uninstall(dev
);
3679 return i915_gem_idle(dev
);
3683 i915_gem_lastclose(struct drm_device
*dev
)
3687 if (drm_core_check_feature(dev
, DRIVER_MODESET
))
3690 ret
= i915_gem_idle(dev
);
3692 DRM_ERROR("failed to idle hardware: %d\n", ret
);
3696 init_ring_lists(struct intel_ring_buffer
*ring
)
3698 INIT_LIST_HEAD(&ring
->active_list
);
3699 INIT_LIST_HEAD(&ring
->request_list
);
3700 INIT_LIST_HEAD(&ring
->gpu_write_list
);
3704 i915_gem_load(struct drm_device
*dev
)
3707 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
3709 INIT_LIST_HEAD(&dev_priv
->mm
.active_list
);
3710 INIT_LIST_HEAD(&dev_priv
->mm
.flushing_list
);
3711 INIT_LIST_HEAD(&dev_priv
->mm
.inactive_list
);
3712 INIT_LIST_HEAD(&dev_priv
->mm
.fence_list
);
3713 INIT_LIST_HEAD(&dev_priv
->mm
.gtt_list
);
3714 for (i
= 0; i
< I915_NUM_RINGS
; i
++)
3715 init_ring_lists(&dev_priv
->ring
[i
]);
3716 for (i
= 0; i
< I915_MAX_NUM_FENCES
; i
++)
3717 INIT_LIST_HEAD(&dev_priv
->fence_regs
[i
].lru_list
);
3718 INIT_DELAYED_WORK(&dev_priv
->mm
.retire_work
,
3719 i915_gem_retire_work_handler
);
3720 init_completion(&dev_priv
->error_completion
);
3722 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
3724 I915_WRITE(MI_ARB_STATE
,
3725 _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE
));
3728 dev_priv
->relative_constants_mode
= I915_EXEC_CONSTANTS_REL_GENERAL
;
3730 /* Old X drivers will take 0-2 for front, back, depth buffers */
3731 if (!drm_core_check_feature(dev
, DRIVER_MODESET
))
3732 dev_priv
->fence_reg_start
= 3;
3734 if (INTEL_INFO(dev
)->gen
>= 4 || IS_I945G(dev
) || IS_I945GM(dev
) || IS_G33(dev
))
3735 dev_priv
->num_fence_regs
= 16;
3737 dev_priv
->num_fence_regs
= 8;
3739 /* Initialize fence registers to zero */
3740 i915_gem_reset_fences(dev
);
3742 i915_gem_detect_bit_6_swizzle(dev
);
3743 init_waitqueue_head(&dev_priv
->pending_flip_queue
);
3745 dev_priv
->mm
.interruptible
= true;
3747 dev_priv
->mm
.inactive_shrinker
.shrink
= i915_gem_inactive_shrink
;
3748 dev_priv
->mm
.inactive_shrinker
.seeks
= DEFAULT_SEEKS
;
3749 register_shrinker(&dev_priv
->mm
.inactive_shrinker
);
3753 * Create a physically contiguous memory object for this object
3754 * e.g. for cursor + overlay regs
3756 static int i915_gem_init_phys_object(struct drm_device
*dev
,
3757 int id
, int size
, int align
)
3759 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
3760 struct drm_i915_gem_phys_object
*phys_obj
;
3763 if (dev_priv
->mm
.phys_objs
[id
- 1] || !size
)
3766 phys_obj
= kzalloc(sizeof(struct drm_i915_gem_phys_object
), GFP_KERNEL
);
3772 phys_obj
->handle
= drm_pci_alloc(dev
, size
, align
);
3773 if (!phys_obj
->handle
) {
3778 set_memory_wc((unsigned long)phys_obj
->handle
->vaddr
, phys_obj
->handle
->size
/ PAGE_SIZE
);
3781 dev_priv
->mm
.phys_objs
[id
- 1] = phys_obj
;
3789 static void i915_gem_free_phys_object(struct drm_device
*dev
, int id
)
3791 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
3792 struct drm_i915_gem_phys_object
*phys_obj
;
3794 if (!dev_priv
->mm
.phys_objs
[id
- 1])
3797 phys_obj
= dev_priv
->mm
.phys_objs
[id
- 1];
3798 if (phys_obj
->cur_obj
) {
3799 i915_gem_detach_phys_object(dev
, phys_obj
->cur_obj
);
3803 set_memory_wb((unsigned long)phys_obj
->handle
->vaddr
, phys_obj
->handle
->size
/ PAGE_SIZE
);
3805 drm_pci_free(dev
, phys_obj
->handle
);
3807 dev_priv
->mm
.phys_objs
[id
- 1] = NULL
;
3810 void i915_gem_free_all_phys_object(struct drm_device
*dev
)
3814 for (i
= I915_GEM_PHYS_CURSOR_0
; i
<= I915_MAX_PHYS_OBJECT
; i
++)
3815 i915_gem_free_phys_object(dev
, i
);
3818 void i915_gem_detach_phys_object(struct drm_device
*dev
,
3819 struct drm_i915_gem_object
*obj
)
3821 struct address_space
*mapping
= obj
->base
.filp
->f_path
.dentry
->d_inode
->i_mapping
;
3828 vaddr
= obj
->phys_obj
->handle
->vaddr
;
3830 page_count
= obj
->base
.size
/ PAGE_SIZE
;
3831 for (i
= 0; i
< page_count
; i
++) {
3832 struct page
*page
= shmem_read_mapping_page(mapping
, i
);
3833 if (!IS_ERR(page
)) {
3834 char *dst
= kmap_atomic(page
);
3835 memcpy(dst
, vaddr
+ i
*PAGE_SIZE
, PAGE_SIZE
);
3838 drm_clflush_pages(&page
, 1);
3840 set_page_dirty(page
);
3841 mark_page_accessed(page
);
3842 page_cache_release(page
);
3845 intel_gtt_chipset_flush();
3847 obj
->phys_obj
->cur_obj
= NULL
;
3848 obj
->phys_obj
= NULL
;
3852 i915_gem_attach_phys_object(struct drm_device
*dev
,
3853 struct drm_i915_gem_object
*obj
,
3857 struct address_space
*mapping
= obj
->base
.filp
->f_path
.dentry
->d_inode
->i_mapping
;
3858 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
3863 if (id
> I915_MAX_PHYS_OBJECT
)
3866 if (obj
->phys_obj
) {
3867 if (obj
->phys_obj
->id
== id
)
3869 i915_gem_detach_phys_object(dev
, obj
);
3872 /* create a new object */
3873 if (!dev_priv
->mm
.phys_objs
[id
- 1]) {
3874 ret
= i915_gem_init_phys_object(dev
, id
,
3875 obj
->base
.size
, align
);
3877 DRM_ERROR("failed to init phys object %d size: %zu\n",
3878 id
, obj
->base
.size
);
3883 /* bind to the object */
3884 obj
->phys_obj
= dev_priv
->mm
.phys_objs
[id
- 1];
3885 obj
->phys_obj
->cur_obj
= obj
;
3887 page_count
= obj
->base
.size
/ PAGE_SIZE
;
3889 for (i
= 0; i
< page_count
; i
++) {
3893 page
= shmem_read_mapping_page(mapping
, i
);
3895 return PTR_ERR(page
);
3897 src
= kmap_atomic(page
);
3898 dst
= obj
->phys_obj
->handle
->vaddr
+ (i
* PAGE_SIZE
);
3899 memcpy(dst
, src
, PAGE_SIZE
);
3902 mark_page_accessed(page
);
3903 page_cache_release(page
);
3910 i915_gem_phys_pwrite(struct drm_device
*dev
,
3911 struct drm_i915_gem_object
*obj
,
3912 struct drm_i915_gem_pwrite
*args
,
3913 struct drm_file
*file_priv
)
3915 void *vaddr
= obj
->phys_obj
->handle
->vaddr
+ args
->offset
;
3916 char __user
*user_data
= (char __user
*) (uintptr_t) args
->data_ptr
;
3918 if (__copy_from_user_inatomic_nocache(vaddr
, user_data
, args
->size
)) {
3919 unsigned long unwritten
;
3921 /* The physical object once assigned is fixed for the lifetime
3922 * of the obj, so we can safely drop the lock and continue
3925 mutex_unlock(&dev
->struct_mutex
);
3926 unwritten
= copy_from_user(vaddr
, user_data
, args
->size
);
3927 mutex_lock(&dev
->struct_mutex
);
3932 intel_gtt_chipset_flush();
3936 void i915_gem_release(struct drm_device
*dev
, struct drm_file
*file
)
3938 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
3940 /* Clean up our request list when the client is going away, so that
3941 * later retire_requests won't dereference our soon-to-be-gone
3944 spin_lock(&file_priv
->mm
.lock
);
3945 while (!list_empty(&file_priv
->mm
.request_list
)) {
3946 struct drm_i915_gem_request
*request
;
3948 request
= list_first_entry(&file_priv
->mm
.request_list
,
3949 struct drm_i915_gem_request
,
3951 list_del(&request
->client_list
);
3952 request
->file_priv
= NULL
;
3954 spin_unlock(&file_priv
->mm
.lock
);
3958 i915_gpu_is_active(struct drm_device
*dev
)
3960 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
3963 lists_empty
= list_empty(&dev_priv
->mm
.flushing_list
) &&
3964 list_empty(&dev_priv
->mm
.active_list
);
3966 return !lists_empty
;
3970 i915_gem_inactive_shrink(struct shrinker
*shrinker
, struct shrink_control
*sc
)
3972 struct drm_i915_private
*dev_priv
=
3973 container_of(shrinker
,
3974 struct drm_i915_private
,
3975 mm
.inactive_shrinker
);
3976 struct drm_device
*dev
= dev_priv
->dev
;
3977 struct drm_i915_gem_object
*obj
, *next
;
3978 int nr_to_scan
= sc
->nr_to_scan
;
3981 if (!mutex_trylock(&dev
->struct_mutex
))
3984 /* "fast-path" to count number of available objects */
3985 if (nr_to_scan
== 0) {
3987 list_for_each_entry(obj
,
3988 &dev_priv
->mm
.inactive_list
,
3991 mutex_unlock(&dev
->struct_mutex
);
3992 return cnt
/ 100 * sysctl_vfs_cache_pressure
;
3996 /* first scan for clean buffers */
3997 i915_gem_retire_requests(dev
);
3999 list_for_each_entry_safe(obj
, next
,
4000 &dev_priv
->mm
.inactive_list
,
4002 if (i915_gem_object_is_purgeable(obj
)) {
4003 if (i915_gem_object_unbind(obj
) == 0 &&
4009 /* second pass, evict/count anything still on the inactive list */
4011 list_for_each_entry_safe(obj
, next
,
4012 &dev_priv
->mm
.inactive_list
,
4015 i915_gem_object_unbind(obj
) == 0)
4021 if (nr_to_scan
&& i915_gpu_is_active(dev
)) {
4023 * We are desperate for pages, so as a last resort, wait
4024 * for the GPU to finish and discard whatever we can.
4025 * This has a dramatic impact to reduce the number of
4026 * OOM-killer events whilst running the GPU aggressively.
4028 if (i915_gpu_idle(dev
, true) == 0)
4031 mutex_unlock(&dev
->struct_mutex
);
4032 return cnt
/ 100 * sysctl_vfs_cache_pressure
;