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
);
148 rdev
->mman
.bdev
.dev_mapping
= rdev
->ddev
->dev_mapping
;
150 type
= ttm_bo_type_kernel
;
152 type
= ttm_bo_type_sg
;
154 type
= ttm_bo_type_device
;
158 acc_size
= ttm_bo_dma_acc_size(&rdev
->mman
.bdev
, size
,
159 sizeof(struct radeon_bo
));
161 bo
= kzalloc(sizeof(struct radeon_bo
), GFP_KERNEL
);
164 r
= drm_gem_object_init(rdev
->ddev
, &bo
->gem_base
, size
);
170 bo
->surface_reg
= -1;
171 INIT_LIST_HEAD(&bo
->list
);
172 INIT_LIST_HEAD(&bo
->va
);
173 bo
->initial_domain
= domain
& (RADEON_GEM_DOMAIN_VRAM
|
174 RADEON_GEM_DOMAIN_GTT
|
175 RADEON_GEM_DOMAIN_CPU
);
176 radeon_ttm_placement_from_domain(bo
, domain
);
177 /* Kernel allocation are uninterruptible */
178 down_read(&rdev
->pm
.mclk_lock
);
179 r
= ttm_bo_init(&rdev
->mman
.bdev
, &bo
->tbo
, size
, type
,
180 &bo
->placement
, page_align
, !kernel
, NULL
,
181 acc_size
, sg
, &radeon_ttm_bo_destroy
);
182 up_read(&rdev
->pm
.mclk_lock
);
183 if (unlikely(r
!= 0)) {
188 trace_radeon_bo_create(bo
);
193 int radeon_bo_kmap(struct radeon_bo
*bo
, void **ptr
)
204 r
= ttm_bo_kmap(&bo
->tbo
, 0, bo
->tbo
.num_pages
, &bo
->kmap
);
208 bo
->kptr
= ttm_kmap_obj_virtual(&bo
->kmap
, &is_iomem
);
212 radeon_bo_check_tiling(bo
, 0, 0);
216 void radeon_bo_kunmap(struct radeon_bo
*bo
)
218 if (bo
->kptr
== NULL
)
221 radeon_bo_check_tiling(bo
, 0, 0);
222 ttm_bo_kunmap(&bo
->kmap
);
225 void radeon_bo_unref(struct radeon_bo
**bo
)
227 struct ttm_buffer_object
*tbo
;
228 struct radeon_device
*rdev
;
234 down_read(&rdev
->pm
.mclk_lock
);
236 up_read(&rdev
->pm
.mclk_lock
);
241 int radeon_bo_pin_restricted(struct radeon_bo
*bo
, u32 domain
, u64 max_offset
,
249 *gpu_addr
= radeon_bo_gpu_offset(bo
);
251 if (max_offset
!= 0) {
254 if (domain
== RADEON_GEM_DOMAIN_VRAM
)
255 domain_start
= bo
->rdev
->mc
.vram_start
;
257 domain_start
= bo
->rdev
->mc
.gtt_start
;
258 WARN_ON_ONCE(max_offset
<
259 (radeon_bo_gpu_offset(bo
) - domain_start
));
264 radeon_ttm_placement_from_domain(bo
, domain
);
265 if (domain
== RADEON_GEM_DOMAIN_VRAM
) {
266 /* force to pin into visible video ram */
267 bo
->placement
.lpfn
= bo
->rdev
->mc
.visible_vram_size
>> PAGE_SHIFT
;
270 u64 lpfn
= max_offset
>> PAGE_SHIFT
;
272 if (!bo
->placement
.lpfn
)
273 bo
->placement
.lpfn
= bo
->rdev
->mc
.gtt_size
>> PAGE_SHIFT
;
275 if (lpfn
< bo
->placement
.lpfn
)
276 bo
->placement
.lpfn
= lpfn
;
278 for (i
= 0; i
< bo
->placement
.num_placement
; i
++)
279 bo
->placements
[i
] |= TTM_PL_FLAG_NO_EVICT
;
280 r
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, false, false);
281 if (likely(r
== 0)) {
283 if (gpu_addr
!= NULL
)
284 *gpu_addr
= radeon_bo_gpu_offset(bo
);
286 if (unlikely(r
!= 0))
287 dev_err(bo
->rdev
->dev
, "%p pin failed\n", bo
);
291 int radeon_bo_pin(struct radeon_bo
*bo
, u32 domain
, u64
*gpu_addr
)
293 return radeon_bo_pin_restricted(bo
, domain
, 0, gpu_addr
);
296 int radeon_bo_unpin(struct radeon_bo
*bo
)
300 if (!bo
->pin_count
) {
301 dev_warn(bo
->rdev
->dev
, "%p unpin not necessary\n", bo
);
307 for (i
= 0; i
< bo
->placement
.num_placement
; i
++)
308 bo
->placements
[i
] &= ~TTM_PL_FLAG_NO_EVICT
;
309 r
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, false, false);
310 if (unlikely(r
!= 0))
311 dev_err(bo
->rdev
->dev
, "%p validate failed for unpin\n", bo
);
315 int radeon_bo_evict_vram(struct radeon_device
*rdev
)
317 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
318 if (0 && (rdev
->flags
& RADEON_IS_IGP
)) {
319 if (rdev
->mc
.igp_sideport_enabled
== false)
320 /* Useless to evict on IGP chips */
323 return ttm_bo_evict_mm(&rdev
->mman
.bdev
, TTM_PL_VRAM
);
326 void radeon_bo_force_delete(struct radeon_device
*rdev
)
328 struct radeon_bo
*bo
, *n
;
330 if (list_empty(&rdev
->gem
.objects
)) {
333 dev_err(rdev
->dev
, "Userspace still has active objects !\n");
334 list_for_each_entry_safe(bo
, n
, &rdev
->gem
.objects
, list
) {
335 mutex_lock(&rdev
->ddev
->struct_mutex
);
336 dev_err(rdev
->dev
, "%p %p %lu %lu force free\n",
337 &bo
->gem_base
, bo
, (unsigned long)bo
->gem_base
.size
,
338 *((unsigned long *)&bo
->gem_base
.refcount
));
339 mutex_lock(&bo
->rdev
->gem
.mutex
);
340 list_del_init(&bo
->list
);
341 mutex_unlock(&bo
->rdev
->gem
.mutex
);
342 /* this should unref the ttm bo */
343 drm_gem_object_unreference(&bo
->gem_base
);
344 mutex_unlock(&rdev
->ddev
->struct_mutex
);
348 int radeon_bo_init(struct radeon_device
*rdev
)
350 /* Add an MTRR for the VRAM */
351 if (!rdev
->fastfb_working
) {
352 rdev
->mc
.vram_mtrr
= arch_phys_wc_add(rdev
->mc
.aper_base
,
355 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
356 rdev
->mc
.mc_vram_size
>> 20,
357 (unsigned long long)rdev
->mc
.aper_size
>> 20);
358 DRM_INFO("RAM width %dbits %cDR\n",
359 rdev
->mc
.vram_width
, rdev
->mc
.vram_is_ddr
? 'D' : 'S');
360 return radeon_ttm_init(rdev
);
363 void radeon_bo_fini(struct radeon_device
*rdev
)
365 radeon_ttm_fini(rdev
);
366 arch_phys_wc_del(rdev
->mc
.vram_mtrr
);
369 /* Returns how many bytes TTM can move per IB.
371 static u64
radeon_bo_get_threshold_for_moves(struct radeon_device
*rdev
)
373 u64 real_vram_size
= rdev
->mc
.real_vram_size
;
374 u64 vram_usage
= atomic64_read(&rdev
->vram_usage
);
376 /* This function is based on the current VRAM usage.
378 * - If all of VRAM is free, allow relocating the number of bytes that
379 * is equal to 1/4 of the size of VRAM for this IB.
381 * - If more than one half of VRAM is occupied, only allow relocating
382 * 1 MB of data for this IB.
384 * - From 0 to one half of used VRAM, the threshold decreases
399 * Note: It's a threshold, not a limit. The threshold must be crossed
400 * for buffer relocations to stop, so any buffer of an arbitrary size
401 * can be moved as long as the threshold isn't crossed before
402 * the relocation takes place. We don't want to disable buffer
403 * relocations completely.
405 * The idea is that buffers should be placed in VRAM at creation time
406 * and TTM should only do a minimum number of relocations during
407 * command submission. In practice, you need to submit at least
408 * a dozen IBs to move all buffers to VRAM if they are in GTT.
410 * Also, things can get pretty crazy under memory pressure and actual
411 * VRAM usage can change a lot, so playing safe even at 50% does
412 * consistently increase performance.
415 u64 half_vram
= real_vram_size
>> 1;
416 u64 half_free_vram
= vram_usage
>= half_vram
? 0 : half_vram
- vram_usage
;
417 u64 bytes_moved_threshold
= half_free_vram
>> 1;
418 return max(bytes_moved_threshold
, 1024*1024ull);
421 int radeon_bo_list_validate(struct radeon_device
*rdev
,
422 struct ww_acquire_ctx
*ticket
,
423 struct list_head
*head
, int ring
)
425 struct radeon_bo_list
*lobj
;
426 struct radeon_bo
*bo
;
428 u64 bytes_moved
= 0, initial_bytes_moved
;
429 u64 bytes_moved_threshold
= radeon_bo_get_threshold_for_moves(rdev
);
431 r
= ttm_eu_reserve_buffers(ticket
, head
);
432 if (unlikely(r
!= 0)) {
436 list_for_each_entry(lobj
, head
, tv
.head
) {
438 if (!bo
->pin_count
) {
439 u32 domain
= lobj
->domain
;
441 radeon_mem_type_to_domain(bo
->tbo
.mem
.mem_type
);
443 /* Check if this buffer will be moved and don't move it
444 * if we have moved too many buffers for this IB already.
446 * Note that this allows moving at least one buffer of
447 * any size, because it doesn't take the current "bo"
448 * into account. We don't want to disallow buffer moves
451 if (current_domain
!= RADEON_GEM_DOMAIN_CPU
&&
452 (domain
& current_domain
) == 0 && /* will be moved */
453 bytes_moved
> bytes_moved_threshold
) {
455 domain
= current_domain
;
459 radeon_ttm_placement_from_domain(bo
, domain
);
460 if (ring
== R600_RING_TYPE_UVD_INDEX
)
461 radeon_uvd_force_into_uvd_segment(bo
);
463 initial_bytes_moved
= atomic64_read(&rdev
->num_bytes_moved
);
464 r
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, true, false);
465 bytes_moved
+= atomic64_read(&rdev
->num_bytes_moved
) -
469 if (r
!= -ERESTARTSYS
&& domain
!= lobj
->alt_domain
) {
470 domain
= lobj
->alt_domain
;
473 ttm_eu_backoff_reservation(ticket
, head
);
477 lobj
->gpu_offset
= radeon_bo_gpu_offset(bo
);
478 lobj
->tiling_flags
= bo
->tiling_flags
;
483 int radeon_bo_fbdev_mmap(struct radeon_bo
*bo
,
484 struct vm_area_struct
*vma
)
486 return ttm_fbdev_mmap(vma
, &bo
->tbo
);
489 int radeon_bo_get_surface_reg(struct radeon_bo
*bo
)
491 struct radeon_device
*rdev
= bo
->rdev
;
492 struct radeon_surface_reg
*reg
;
493 struct radeon_bo
*old_object
;
497 lockdep_assert_held(&bo
->tbo
.resv
->lock
.base
);
499 if (!bo
->tiling_flags
)
502 if (bo
->surface_reg
>= 0) {
503 reg
= &rdev
->surface_regs
[bo
->surface_reg
];
509 for (i
= 0; i
< RADEON_GEM_MAX_SURFACES
; i
++) {
511 reg
= &rdev
->surface_regs
[i
];
515 old_object
= reg
->bo
;
516 if (old_object
->pin_count
== 0)
520 /* if we are all out */
521 if (i
== RADEON_GEM_MAX_SURFACES
) {
524 /* find someone with a surface reg and nuke their BO */
525 reg
= &rdev
->surface_regs
[steal
];
526 old_object
= reg
->bo
;
527 /* blow away the mapping */
528 DRM_DEBUG("stealing surface reg %d from %p\n", steal
, old_object
);
529 ttm_bo_unmap_virtual(&old_object
->tbo
);
530 old_object
->surface_reg
= -1;
538 radeon_set_surface_reg(rdev
, i
, bo
->tiling_flags
, bo
->pitch
,
539 bo
->tbo
.mem
.start
<< PAGE_SHIFT
,
540 bo
->tbo
.num_pages
<< PAGE_SHIFT
);
544 static void radeon_bo_clear_surface_reg(struct radeon_bo
*bo
)
546 struct radeon_device
*rdev
= bo
->rdev
;
547 struct radeon_surface_reg
*reg
;
549 if (bo
->surface_reg
== -1)
552 reg
= &rdev
->surface_regs
[bo
->surface_reg
];
553 radeon_clear_surface_reg(rdev
, bo
->surface_reg
);
556 bo
->surface_reg
= -1;
559 int radeon_bo_set_tiling_flags(struct radeon_bo
*bo
,
560 uint32_t tiling_flags
, uint32_t pitch
)
562 struct radeon_device
*rdev
= bo
->rdev
;
565 if (rdev
->family
>= CHIP_CEDAR
) {
566 unsigned bankw
, bankh
, mtaspect
, tilesplit
, stilesplit
;
568 bankw
= (tiling_flags
>> RADEON_TILING_EG_BANKW_SHIFT
) & RADEON_TILING_EG_BANKW_MASK
;
569 bankh
= (tiling_flags
>> RADEON_TILING_EG_BANKH_SHIFT
) & RADEON_TILING_EG_BANKH_MASK
;
570 mtaspect
= (tiling_flags
>> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT
) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK
;
571 tilesplit
= (tiling_flags
>> RADEON_TILING_EG_TILE_SPLIT_SHIFT
) & RADEON_TILING_EG_TILE_SPLIT_MASK
;
572 stilesplit
= (tiling_flags
>> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT
) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK
;
606 if (stilesplit
> 6) {
610 r
= radeon_bo_reserve(bo
, false);
611 if (unlikely(r
!= 0))
613 bo
->tiling_flags
= tiling_flags
;
615 radeon_bo_unreserve(bo
);
619 void radeon_bo_get_tiling_flags(struct radeon_bo
*bo
,
620 uint32_t *tiling_flags
,
623 lockdep_assert_held(&bo
->tbo
.resv
->lock
.base
);
626 *tiling_flags
= bo
->tiling_flags
;
631 int radeon_bo_check_tiling(struct radeon_bo
*bo
, bool has_moved
,
635 lockdep_assert_held(&bo
->tbo
.resv
->lock
.base
);
637 if (!(bo
->tiling_flags
& RADEON_TILING_SURFACE
))
641 radeon_bo_clear_surface_reg(bo
);
645 if (bo
->tbo
.mem
.mem_type
!= TTM_PL_VRAM
) {
649 if (bo
->surface_reg
>= 0)
650 radeon_bo_clear_surface_reg(bo
);
654 if ((bo
->surface_reg
>= 0) && !has_moved
)
657 return radeon_bo_get_surface_reg(bo
);
660 void radeon_bo_move_notify(struct ttm_buffer_object
*bo
,
661 struct ttm_mem_reg
*new_mem
)
663 struct radeon_bo
*rbo
;
665 if (!radeon_ttm_bo_is_radeon_bo(bo
))
668 rbo
= container_of(bo
, struct radeon_bo
, tbo
);
669 radeon_bo_check_tiling(rbo
, 0, 1);
670 radeon_vm_bo_invalidate(rbo
->rdev
, rbo
);
672 /* update statistics */
676 radeon_update_memory_usage(rbo
, bo
->mem
.mem_type
, -1);
677 radeon_update_memory_usage(rbo
, new_mem
->mem_type
, 1);
680 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object
*bo
)
682 struct radeon_device
*rdev
;
683 struct radeon_bo
*rbo
;
684 unsigned long offset
, size
;
687 if (!radeon_ttm_bo_is_radeon_bo(bo
))
689 rbo
= container_of(bo
, struct radeon_bo
, tbo
);
690 radeon_bo_check_tiling(rbo
, 0, 0);
692 if (bo
->mem
.mem_type
== TTM_PL_VRAM
) {
693 size
= bo
->mem
.num_pages
<< PAGE_SHIFT
;
694 offset
= bo
->mem
.start
<< PAGE_SHIFT
;
695 if ((offset
+ size
) > rdev
->mc
.visible_vram_size
) {
696 /* hurrah the memory is not visible ! */
697 radeon_ttm_placement_from_domain(rbo
, RADEON_GEM_DOMAIN_VRAM
);
698 rbo
->placement
.lpfn
= rdev
->mc
.visible_vram_size
>> PAGE_SHIFT
;
699 r
= ttm_bo_validate(bo
, &rbo
->placement
, false, false);
700 if (unlikely(r
!= 0))
702 offset
= bo
->mem
.start
<< PAGE_SHIFT
;
703 /* this should not happen */
704 if ((offset
+ size
) > rdev
->mc
.visible_vram_size
)
711 int radeon_bo_wait(struct radeon_bo
*bo
, u32
*mem_type
, bool no_wait
)
715 r
= ttm_bo_reserve(&bo
->tbo
, true, no_wait
, false, 0);
716 if (unlikely(r
!= 0))
718 spin_lock(&bo
->tbo
.bdev
->fence_lock
);
720 *mem_type
= bo
->tbo
.mem
.mem_type
;
721 if (bo
->tbo
.sync_obj
)
722 r
= ttm_bo_wait(&bo
->tbo
, true, true, no_wait
);
723 spin_unlock(&bo
->tbo
.bdev
->fence_lock
);
724 ttm_bo_unreserve(&bo
->tbo
);