2 * Copyright 2009 Jerome Glisse.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
28 * Jerome Glisse <glisse@freedesktop.org>
29 * Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
32 #include <linux/list.h>
33 #include <linux/slab.h>
35 #include <drm/radeon_drm.h>
37 #include "radeon_trace.h"
40 int radeon_ttm_init(struct radeon_device
*rdev
);
41 void radeon_ttm_fini(struct radeon_device
*rdev
);
42 static void radeon_bo_clear_surface_reg(struct radeon_bo
*bo
);
45 * To exclude mutual BO access we rely on bo_reserve exclusion, as all
46 * function are calling it.
49 static void radeon_bo_clear_va(struct radeon_bo
*bo
)
51 struct radeon_bo_va
*bo_va
, *tmp
;
53 list_for_each_entry_safe(bo_va
, tmp
, &bo
->va
, bo_list
) {
54 /* remove from all vm address space */
55 radeon_vm_bo_rmv(bo
->rdev
, bo_va
);
59 static void radeon_update_memory_usage(struct radeon_bo
*bo
,
60 unsigned mem_type
, int sign
)
62 struct radeon_device
*rdev
= bo
->rdev
;
63 u64 size
= (u64
)bo
->tbo
.num_pages
<< PAGE_SHIFT
;
68 atomic64_add(size
, &rdev
->gtt_usage
);
70 atomic64_sub(size
, &rdev
->gtt_usage
);
74 atomic64_add(size
, &rdev
->vram_usage
);
76 atomic64_sub(size
, &rdev
->vram_usage
);
81 static void radeon_ttm_bo_destroy(struct ttm_buffer_object
*tbo
)
85 bo
= container_of(tbo
, struct radeon_bo
, tbo
);
87 radeon_update_memory_usage(bo
, bo
->tbo
.mem
.mem_type
, -1);
89 mutex_lock(&bo
->rdev
->gem
.mutex
);
90 list_del_init(&bo
->list
);
91 mutex_unlock(&bo
->rdev
->gem
.mutex
);
92 radeon_bo_clear_surface_reg(bo
);
93 radeon_bo_clear_va(bo
);
94 drm_gem_object_release(&bo
->gem_base
);
98 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object
*bo
)
100 if (bo
->destroy
== &radeon_ttm_bo_destroy
)
105 void radeon_ttm_placement_from_domain(struct radeon_bo
*rbo
, u32 domain
)
109 rbo
->placement
.fpfn
= 0;
110 rbo
->placement
.lpfn
= 0;
111 rbo
->placement
.placement
= rbo
->placements
;
112 rbo
->placement
.busy_placement
= rbo
->placements
;
113 if (domain
& RADEON_GEM_DOMAIN_VRAM
)
114 rbo
->placements
[c
++] = TTM_PL_FLAG_WC
| TTM_PL_FLAG_UNCACHED
|
116 if (domain
& RADEON_GEM_DOMAIN_GTT
) {
117 if (rbo
->rdev
->flags
& RADEON_IS_AGP
) {
118 rbo
->placements
[c
++] = TTM_PL_FLAG_WC
| TTM_PL_FLAG_TT
;
120 rbo
->placements
[c
++] = TTM_PL_FLAG_CACHED
| TTM_PL_FLAG_TT
;
123 if (domain
& RADEON_GEM_DOMAIN_CPU
) {
124 if (rbo
->rdev
->flags
& RADEON_IS_AGP
) {
125 rbo
->placements
[c
++] = TTM_PL_FLAG_WC
| TTM_PL_FLAG_SYSTEM
;
127 rbo
->placements
[c
++] = TTM_PL_FLAG_CACHED
| TTM_PL_FLAG_SYSTEM
;
131 rbo
->placements
[c
++] = TTM_PL_MASK_CACHING
| TTM_PL_FLAG_SYSTEM
;
132 rbo
->placement
.num_placement
= c
;
133 rbo
->placement
.num_busy_placement
= c
;
136 int radeon_bo_create(struct radeon_device
*rdev
,
137 unsigned long size
, int byte_align
, bool kernel
, u32 domain
,
138 struct sg_table
*sg
, struct radeon_bo
**bo_ptr
)
140 struct radeon_bo
*bo
;
141 enum ttm_bo_type type
;
142 unsigned long page_align
= roundup(byte_align
, PAGE_SIZE
) >> PAGE_SHIFT
;
146 size
= ALIGN(size
, PAGE_SIZE
);
149 type
= ttm_bo_type_kernel
;
151 type
= ttm_bo_type_sg
;
153 type
= ttm_bo_type_device
;
157 acc_size
= ttm_bo_dma_acc_size(&rdev
->mman
.bdev
, size
,
158 sizeof(struct radeon_bo
));
160 bo
= kzalloc(sizeof(struct radeon_bo
), GFP_KERNEL
);
163 r
= drm_gem_object_init(rdev
->ddev
, &bo
->gem_base
, size
);
169 bo
->surface_reg
= -1;
170 INIT_LIST_HEAD(&bo
->list
);
171 INIT_LIST_HEAD(&bo
->va
);
172 bo
->initial_domain
= domain
& (RADEON_GEM_DOMAIN_VRAM
|
173 RADEON_GEM_DOMAIN_GTT
|
174 RADEON_GEM_DOMAIN_CPU
);
175 radeon_ttm_placement_from_domain(bo
, domain
);
176 /* Kernel allocation are uninterruptible */
177 down_read(&rdev
->pm
.mclk_lock
);
178 r
= ttm_bo_init(&rdev
->mman
.bdev
, &bo
->tbo
, size
, type
,
179 &bo
->placement
, page_align
, !kernel
, NULL
,
180 acc_size
, sg
, &radeon_ttm_bo_destroy
);
181 up_read(&rdev
->pm
.mclk_lock
);
182 if (unlikely(r
!= 0)) {
187 trace_radeon_bo_create(bo
);
192 int radeon_bo_kmap(struct radeon_bo
*bo
, void **ptr
)
203 r
= ttm_bo_kmap(&bo
->tbo
, 0, bo
->tbo
.num_pages
, &bo
->kmap
);
207 bo
->kptr
= ttm_kmap_obj_virtual(&bo
->kmap
, &is_iomem
);
211 radeon_bo_check_tiling(bo
, 0, 0);
215 void radeon_bo_kunmap(struct radeon_bo
*bo
)
217 if (bo
->kptr
== NULL
)
220 radeon_bo_check_tiling(bo
, 0, 0);
221 ttm_bo_kunmap(&bo
->kmap
);
224 void radeon_bo_unref(struct radeon_bo
**bo
)
226 struct ttm_buffer_object
*tbo
;
227 struct radeon_device
*rdev
;
233 down_read(&rdev
->pm
.mclk_lock
);
235 up_read(&rdev
->pm
.mclk_lock
);
240 int radeon_bo_pin_restricted(struct radeon_bo
*bo
, u32 domain
, u64 max_offset
,
248 *gpu_addr
= radeon_bo_gpu_offset(bo
);
250 if (max_offset
!= 0) {
253 if (domain
== RADEON_GEM_DOMAIN_VRAM
)
254 domain_start
= bo
->rdev
->mc
.vram_start
;
256 domain_start
= bo
->rdev
->mc
.gtt_start
;
257 WARN_ON_ONCE(max_offset
<
258 (radeon_bo_gpu_offset(bo
) - domain_start
));
263 radeon_ttm_placement_from_domain(bo
, domain
);
264 if (domain
== RADEON_GEM_DOMAIN_VRAM
) {
265 /* force to pin into visible video ram */
266 bo
->placement
.lpfn
= bo
->rdev
->mc
.visible_vram_size
>> PAGE_SHIFT
;
269 u64 lpfn
= max_offset
>> PAGE_SHIFT
;
271 if (!bo
->placement
.lpfn
)
272 bo
->placement
.lpfn
= bo
->rdev
->mc
.gtt_size
>> PAGE_SHIFT
;
274 if (lpfn
< bo
->placement
.lpfn
)
275 bo
->placement
.lpfn
= lpfn
;
277 for (i
= 0; i
< bo
->placement
.num_placement
; i
++)
278 bo
->placements
[i
] |= TTM_PL_FLAG_NO_EVICT
;
279 r
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, false, false);
280 if (likely(r
== 0)) {
282 if (gpu_addr
!= NULL
)
283 *gpu_addr
= radeon_bo_gpu_offset(bo
);
285 if (unlikely(r
!= 0))
286 dev_err(bo
->rdev
->dev
, "%p pin failed\n", bo
);
290 int radeon_bo_pin(struct radeon_bo
*bo
, u32 domain
, u64
*gpu_addr
)
292 return radeon_bo_pin_restricted(bo
, domain
, 0, gpu_addr
);
295 int radeon_bo_unpin(struct radeon_bo
*bo
)
299 if (!bo
->pin_count
) {
300 dev_warn(bo
->rdev
->dev
, "%p unpin not necessary\n", bo
);
306 for (i
= 0; i
< bo
->placement
.num_placement
; i
++)
307 bo
->placements
[i
] &= ~TTM_PL_FLAG_NO_EVICT
;
308 r
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, false, false);
309 if (unlikely(r
!= 0))
310 dev_err(bo
->rdev
->dev
, "%p validate failed for unpin\n", bo
);
314 int radeon_bo_evict_vram(struct radeon_device
*rdev
)
316 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
317 if (0 && (rdev
->flags
& RADEON_IS_IGP
)) {
318 if (rdev
->mc
.igp_sideport_enabled
== false)
319 /* Useless to evict on IGP chips */
322 return ttm_bo_evict_mm(&rdev
->mman
.bdev
, TTM_PL_VRAM
);
325 void radeon_bo_force_delete(struct radeon_device
*rdev
)
327 struct radeon_bo
*bo
, *n
;
329 if (list_empty(&rdev
->gem
.objects
)) {
332 dev_err(rdev
->dev
, "Userspace still has active objects !\n");
333 list_for_each_entry_safe(bo
, n
, &rdev
->gem
.objects
, list
) {
334 mutex_lock(&rdev
->ddev
->struct_mutex
);
335 dev_err(rdev
->dev
, "%p %p %lu %lu force free\n",
336 &bo
->gem_base
, bo
, (unsigned long)bo
->gem_base
.size
,
337 *((unsigned long *)&bo
->gem_base
.refcount
));
338 mutex_lock(&bo
->rdev
->gem
.mutex
);
339 list_del_init(&bo
->list
);
340 mutex_unlock(&bo
->rdev
->gem
.mutex
);
341 /* this should unref the ttm bo */
342 drm_gem_object_unreference(&bo
->gem_base
);
343 mutex_unlock(&rdev
->ddev
->struct_mutex
);
347 int radeon_bo_init(struct radeon_device
*rdev
)
349 /* Add an MTRR for the VRAM */
350 if (!rdev
->fastfb_working
) {
351 rdev
->mc
.vram_mtrr
= arch_phys_wc_add(rdev
->mc
.aper_base
,
354 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
355 rdev
->mc
.mc_vram_size
>> 20,
356 (unsigned long long)rdev
->mc
.aper_size
>> 20);
357 DRM_INFO("RAM width %dbits %cDR\n",
358 rdev
->mc
.vram_width
, rdev
->mc
.vram_is_ddr
? 'D' : 'S');
359 return radeon_ttm_init(rdev
);
362 void radeon_bo_fini(struct radeon_device
*rdev
)
364 radeon_ttm_fini(rdev
);
365 arch_phys_wc_del(rdev
->mc
.vram_mtrr
);
368 /* Returns how many bytes TTM can move per IB.
370 static u64
radeon_bo_get_threshold_for_moves(struct radeon_device
*rdev
)
372 u64 real_vram_size
= rdev
->mc
.real_vram_size
;
373 u64 vram_usage
= atomic64_read(&rdev
->vram_usage
);
375 /* This function is based on the current VRAM usage.
377 * - If all of VRAM is free, allow relocating the number of bytes that
378 * is equal to 1/4 of the size of VRAM for this IB.
380 * - If more than one half of VRAM is occupied, only allow relocating
381 * 1 MB of data for this IB.
383 * - From 0 to one half of used VRAM, the threshold decreases
398 * Note: It's a threshold, not a limit. The threshold must be crossed
399 * for buffer relocations to stop, so any buffer of an arbitrary size
400 * can be moved as long as the threshold isn't crossed before
401 * the relocation takes place. We don't want to disable buffer
402 * relocations completely.
404 * The idea is that buffers should be placed in VRAM at creation time
405 * and TTM should only do a minimum number of relocations during
406 * command submission. In practice, you need to submit at least
407 * a dozen IBs to move all buffers to VRAM if they are in GTT.
409 * Also, things can get pretty crazy under memory pressure and actual
410 * VRAM usage can change a lot, so playing safe even at 50% does
411 * consistently increase performance.
414 u64 half_vram
= real_vram_size
>> 1;
415 u64 half_free_vram
= vram_usage
>= half_vram
? 0 : half_vram
- vram_usage
;
416 u64 bytes_moved_threshold
= half_free_vram
>> 1;
417 return max(bytes_moved_threshold
, 1024*1024ull);
420 int radeon_bo_list_validate(struct radeon_device
*rdev
,
421 struct ww_acquire_ctx
*ticket
,
422 struct list_head
*head
, int ring
)
424 struct radeon_cs_reloc
*lobj
;
425 struct radeon_bo
*bo
;
427 u64 bytes_moved
= 0, initial_bytes_moved
;
428 u64 bytes_moved_threshold
= radeon_bo_get_threshold_for_moves(rdev
);
430 r
= ttm_eu_reserve_buffers(ticket
, head
);
431 if (unlikely(r
!= 0)) {
435 list_for_each_entry(lobj
, head
, tv
.head
) {
437 if (!bo
->pin_count
) {
438 u32 domain
= lobj
->domain
;
440 radeon_mem_type_to_domain(bo
->tbo
.mem
.mem_type
);
442 /* Check if this buffer will be moved and don't move it
443 * if we have moved too many buffers for this IB already.
445 * Note that this allows moving at least one buffer of
446 * any size, because it doesn't take the current "bo"
447 * into account. We don't want to disallow buffer moves
450 if (current_domain
!= RADEON_GEM_DOMAIN_CPU
&&
451 (domain
& current_domain
) == 0 && /* will be moved */
452 bytes_moved
> bytes_moved_threshold
) {
454 domain
= current_domain
;
458 radeon_ttm_placement_from_domain(bo
, domain
);
459 if (ring
== R600_RING_TYPE_UVD_INDEX
)
460 radeon_uvd_force_into_uvd_segment(bo
);
462 initial_bytes_moved
= atomic64_read(&rdev
->num_bytes_moved
);
463 r
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, true, false);
464 bytes_moved
+= atomic64_read(&rdev
->num_bytes_moved
) -
468 if (r
!= -ERESTARTSYS
&& domain
!= lobj
->alt_domain
) {
469 domain
= lobj
->alt_domain
;
472 ttm_eu_backoff_reservation(ticket
, head
);
476 lobj
->gpu_offset
= radeon_bo_gpu_offset(bo
);
477 lobj
->tiling_flags
= bo
->tiling_flags
;
482 int radeon_bo_fbdev_mmap(struct radeon_bo
*bo
,
483 struct vm_area_struct
*vma
)
485 return ttm_fbdev_mmap(vma
, &bo
->tbo
);
488 int radeon_bo_get_surface_reg(struct radeon_bo
*bo
)
490 struct radeon_device
*rdev
= bo
->rdev
;
491 struct radeon_surface_reg
*reg
;
492 struct radeon_bo
*old_object
;
496 lockdep_assert_held(&bo
->tbo
.resv
->lock
.base
);
498 if (!bo
->tiling_flags
)
501 if (bo
->surface_reg
>= 0) {
502 reg
= &rdev
->surface_regs
[bo
->surface_reg
];
508 for (i
= 0; i
< RADEON_GEM_MAX_SURFACES
; i
++) {
510 reg
= &rdev
->surface_regs
[i
];
514 old_object
= reg
->bo
;
515 if (old_object
->pin_count
== 0)
519 /* if we are all out */
520 if (i
== RADEON_GEM_MAX_SURFACES
) {
523 /* find someone with a surface reg and nuke their BO */
524 reg
= &rdev
->surface_regs
[steal
];
525 old_object
= reg
->bo
;
526 /* blow away the mapping */
527 DRM_DEBUG("stealing surface reg %d from %p\n", steal
, old_object
);
528 ttm_bo_unmap_virtual(&old_object
->tbo
);
529 old_object
->surface_reg
= -1;
537 radeon_set_surface_reg(rdev
, i
, bo
->tiling_flags
, bo
->pitch
,
538 bo
->tbo
.mem
.start
<< PAGE_SHIFT
,
539 bo
->tbo
.num_pages
<< PAGE_SHIFT
);
543 static void radeon_bo_clear_surface_reg(struct radeon_bo
*bo
)
545 struct radeon_device
*rdev
= bo
->rdev
;
546 struct radeon_surface_reg
*reg
;
548 if (bo
->surface_reg
== -1)
551 reg
= &rdev
->surface_regs
[bo
->surface_reg
];
552 radeon_clear_surface_reg(rdev
, bo
->surface_reg
);
555 bo
->surface_reg
= -1;
558 int radeon_bo_set_tiling_flags(struct radeon_bo
*bo
,
559 uint32_t tiling_flags
, uint32_t pitch
)
561 struct radeon_device
*rdev
= bo
->rdev
;
564 if (rdev
->family
>= CHIP_CEDAR
) {
565 unsigned bankw
, bankh
, mtaspect
, tilesplit
, stilesplit
;
567 bankw
= (tiling_flags
>> RADEON_TILING_EG_BANKW_SHIFT
) & RADEON_TILING_EG_BANKW_MASK
;
568 bankh
= (tiling_flags
>> RADEON_TILING_EG_BANKH_SHIFT
) & RADEON_TILING_EG_BANKH_MASK
;
569 mtaspect
= (tiling_flags
>> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT
) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK
;
570 tilesplit
= (tiling_flags
>> RADEON_TILING_EG_TILE_SPLIT_SHIFT
) & RADEON_TILING_EG_TILE_SPLIT_MASK
;
571 stilesplit
= (tiling_flags
>> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT
) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK
;
605 if (stilesplit
> 6) {
609 r
= radeon_bo_reserve(bo
, false);
610 if (unlikely(r
!= 0))
612 bo
->tiling_flags
= tiling_flags
;
614 radeon_bo_unreserve(bo
);
618 void radeon_bo_get_tiling_flags(struct radeon_bo
*bo
,
619 uint32_t *tiling_flags
,
622 lockdep_assert_held(&bo
->tbo
.resv
->lock
.base
);
625 *tiling_flags
= bo
->tiling_flags
;
630 int radeon_bo_check_tiling(struct radeon_bo
*bo
, bool has_moved
,
634 lockdep_assert_held(&bo
->tbo
.resv
->lock
.base
);
636 if (!(bo
->tiling_flags
& RADEON_TILING_SURFACE
))
640 radeon_bo_clear_surface_reg(bo
);
644 if (bo
->tbo
.mem
.mem_type
!= TTM_PL_VRAM
) {
648 if (bo
->surface_reg
>= 0)
649 radeon_bo_clear_surface_reg(bo
);
653 if ((bo
->surface_reg
>= 0) && !has_moved
)
656 return radeon_bo_get_surface_reg(bo
);
659 void radeon_bo_move_notify(struct ttm_buffer_object
*bo
,
660 struct ttm_mem_reg
*new_mem
)
662 struct radeon_bo
*rbo
;
664 if (!radeon_ttm_bo_is_radeon_bo(bo
))
667 rbo
= container_of(bo
, struct radeon_bo
, tbo
);
668 radeon_bo_check_tiling(rbo
, 0, 1);
669 radeon_vm_bo_invalidate(rbo
->rdev
, rbo
);
671 /* update statistics */
675 radeon_update_memory_usage(rbo
, bo
->mem
.mem_type
, -1);
676 radeon_update_memory_usage(rbo
, new_mem
->mem_type
, 1);
679 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object
*bo
)
681 struct radeon_device
*rdev
;
682 struct radeon_bo
*rbo
;
683 unsigned long offset
, size
;
686 if (!radeon_ttm_bo_is_radeon_bo(bo
))
688 rbo
= container_of(bo
, struct radeon_bo
, tbo
);
689 radeon_bo_check_tiling(rbo
, 0, 0);
691 if (bo
->mem
.mem_type
== TTM_PL_VRAM
) {
692 size
= bo
->mem
.num_pages
<< PAGE_SHIFT
;
693 offset
= bo
->mem
.start
<< PAGE_SHIFT
;
694 if ((offset
+ size
) > rdev
->mc
.visible_vram_size
) {
695 /* hurrah the memory is not visible ! */
696 radeon_ttm_placement_from_domain(rbo
, RADEON_GEM_DOMAIN_VRAM
);
697 rbo
->placement
.lpfn
= rdev
->mc
.visible_vram_size
>> PAGE_SHIFT
;
698 r
= ttm_bo_validate(bo
, &rbo
->placement
, false, false);
699 if (unlikely(r
!= 0))
701 offset
= bo
->mem
.start
<< PAGE_SHIFT
;
702 /* this should not happen */
703 if ((offset
+ size
) > rdev
->mc
.visible_vram_size
)
710 int radeon_bo_wait(struct radeon_bo
*bo
, u32
*mem_type
, bool no_wait
)
714 r
= ttm_bo_reserve(&bo
->tbo
, true, no_wait
, false, 0);
715 if (unlikely(r
!= 0))
717 spin_lock(&bo
->tbo
.bdev
->fence_lock
);
719 *mem_type
= bo
->tbo
.mem
.mem_type
;
720 if (bo
->tbo
.sync_obj
)
721 r
= ttm_bo_wait(&bo
->tbo
, true, true, no_wait
);
722 spin_unlock(&bo
->tbo
.bdev
->fence_lock
);
723 ttm_bo_unreserve(&bo
->tbo
);