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>
33 #include <uapi_drm/radeon_drm.h>
36 #include "radeon_trace.h"
41 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
44 * To exclude mutual BO access we rely on bo_reserve exclusion, as all
45 * function are calling it.
48 static void radeon_update_memory_usage(struct radeon_bo *bo,
49 unsigned mem_type, int sign)
51 struct radeon_device *rdev = bo->rdev;
52 u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT;
57 atomic64_add(size, &rdev->gtt_usage);
59 atomic64_sub(size, &rdev->gtt_usage);
63 atomic64_add(size, &rdev->vram_usage);
65 atomic64_sub(size, &rdev->vram_usage);
70 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
74 bo = container_of(tbo, struct radeon_bo, tbo);
76 radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1);
78 mutex_lock(&bo->rdev->gem.mutex);
79 list_del_init(&bo->list);
80 mutex_unlock(&bo->rdev->gem.mutex);
81 radeon_bo_clear_surface_reg(bo);
82 WARN_ON(!list_empty(&bo->va));
83 drm_gem_object_release(&bo->gem_base);
87 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
89 if (bo->destroy == &radeon_ttm_bo_destroy)
94 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
98 rbo->placement.placement = rbo->placements;
99 rbo->placement.busy_placement = rbo->placements;
100 if (domain & RADEON_GEM_DOMAIN_VRAM) {
101 /* Try placing BOs which don't need CPU access outside of the
102 * CPU accessible part of VRAM
104 if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
105 rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
106 rbo->placements[c].fpfn =
107 rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
108 rbo->placements[c++].flags = TTM_PL_FLAG_WC |
109 TTM_PL_FLAG_UNCACHED |
113 rbo->placements[c].fpfn = 0;
114 rbo->placements[c++].flags = TTM_PL_FLAG_WC |
115 TTM_PL_FLAG_UNCACHED |
119 if (domain & RADEON_GEM_DOMAIN_GTT) {
120 if (rbo->flags & RADEON_GEM_GTT_UC) {
121 rbo->placements[c].fpfn = 0;
122 rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
125 } else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
126 (rbo->rdev->flags & RADEON_IS_AGP)) {
127 rbo->placements[c].fpfn = 0;
128 rbo->placements[c++].flags = TTM_PL_FLAG_WC |
129 TTM_PL_FLAG_UNCACHED |
132 rbo->placements[c].fpfn = 0;
133 rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
138 if (domain & RADEON_GEM_DOMAIN_CPU) {
139 if (rbo->flags & RADEON_GEM_GTT_UC) {
140 rbo->placements[c].fpfn = 0;
141 rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
144 } else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
145 rbo->rdev->flags & RADEON_IS_AGP) {
146 rbo->placements[c].fpfn = 0;
147 rbo->placements[c++].flags = TTM_PL_FLAG_WC |
148 TTM_PL_FLAG_UNCACHED |
151 rbo->placements[c].fpfn = 0;
152 rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
157 rbo->placements[c].fpfn = 0;
158 rbo->placements[c++].flags = TTM_PL_MASK_CACHING |
162 rbo->placement.num_placement = c;
163 rbo->placement.num_busy_placement = c;
165 for (i = 0; i < c; ++i) {
166 if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
167 (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
168 !rbo->placements[i].fpfn)
169 rbo->placements[i].lpfn =
170 rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
172 rbo->placements[i].lpfn = 0;
177 * Use two-ended allocation depending on the buffer size to
178 * improve fragmentation quality.
179 * 512kb was measured as the most optimal number.
181 if (!((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
182 (rbo->placements[i].flags & TTM_PL_FLAG_VRAM)) &&
183 rbo->tbo.mem.size > 512 * 1024) {
184 for (i = 0; i < c; i++) {
185 rbo->placements[i].flags |= TTM_PL_FLAG_TOPDOWN;
191 int radeon_bo_create(struct radeon_device *rdev,
192 unsigned long size, int byte_align, bool kernel, u32 domain,
193 u32 flags, struct sg_table *sg, struct radeon_bo **bo_ptr)
195 struct radeon_bo *bo;
196 enum ttm_bo_type type;
197 unsigned long page_align = roundup2(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
201 size = ALIGN(size, PAGE_SIZE);
204 type = ttm_bo_type_kernel;
206 type = ttm_bo_type_sg;
208 type = ttm_bo_type_device;
212 acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
213 sizeof(struct radeon_bo));
215 bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
218 r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size);
224 bo->surface_reg = -1;
225 INIT_LIST_HEAD(&bo->list);
226 INIT_LIST_HEAD(&bo->va);
227 bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
228 RADEON_GEM_DOMAIN_GTT |
229 RADEON_GEM_DOMAIN_CPU);
232 /* PCI GART is always snooped */
233 if (!(rdev->flags & RADEON_IS_PCIE))
234 bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
236 /* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx
237 * See https://bugs.freedesktop.org/show_bug.cgi?id=91268
239 if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635)
240 bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
242 /* DragonFly only supported on __x86_64__ and supports PAT */
243 #if !defined (__DragonFly__)
245 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
246 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
248 bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
249 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
250 /* Don't try to enable write-combining when it can't work, or things
252 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
255 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
256 thanks to write-combining
258 if (bo->flags & RADEON_GEM_GTT_WC)
259 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
260 "better performance thanks to write-combining\n");
261 bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
263 /* For architectures that don't support WC memory,
264 * mask out the WC flag from the BO
266 if (!drm_arch_can_wc_memory())
267 bo->flags &= ~RADEON_GEM_GTT_WC;
269 #endif /* __DragonFly__*/
271 radeon_ttm_placement_from_domain(bo, domain);
272 /* Kernel allocation are uninterruptible */
273 lockmgr(&rdev->pm.mclk_lock, LK_SHARED);
274 r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
275 &bo->placement, page_align, !kernel, NULL,
276 acc_size, sg, &radeon_ttm_bo_destroy);
277 lockmgr(&rdev->pm.mclk_lock, LK_RELEASE);
278 if (unlikely(r != 0)) {
284 trace_radeon_bo_create(bo);
290 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
301 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
305 bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
309 radeon_bo_check_tiling(bo, 0, 0);
313 void radeon_bo_kunmap(struct radeon_bo *bo)
315 if (bo->kptr == NULL)
318 radeon_bo_check_tiling(bo, 0, 0);
319 ttm_bo_kunmap(&bo->kmap);
322 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
327 ttm_bo_reference(&bo->tbo);
331 void radeon_bo_unref(struct radeon_bo **bo)
333 struct ttm_buffer_object *tbo;
334 struct radeon_device *rdev;
335 struct radeon_bo *rbo;
337 if ((rbo = *bo) == NULL)
345 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
353 *gpu_addr = radeon_bo_gpu_offset(bo);
355 if (max_offset != 0) {
358 if (domain == RADEON_GEM_DOMAIN_VRAM)
359 domain_start = bo->rdev->mc.vram_start;
361 domain_start = bo->rdev->mc.gtt_start;
362 if (max_offset < (radeon_bo_gpu_offset(bo) - domain_start)) {
363 DRM_ERROR("radeon_bo_pin_restricted: "
365 "(radeon_bo_gpu_offset(%ju) - "
367 (uintmax_t)max_offset, (uintmax_t)radeon_bo_gpu_offset(bo),
368 (uintmax_t)domain_start);
374 radeon_ttm_placement_from_domain(bo, domain);
375 for (i = 0; i < bo->placement.num_placement; i++) {
376 /* force to pin into visible video ram */
377 if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
378 !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
379 (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
380 bo->placements[i].lpfn =
381 bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
383 bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
385 bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
388 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
389 if (likely(r == 0)) {
391 if (gpu_addr != NULL)
392 *gpu_addr = radeon_bo_gpu_offset(bo);
393 if (domain == RADEON_GEM_DOMAIN_VRAM)
394 bo->rdev->vram_pin_size += radeon_bo_size(bo);
396 bo->rdev->gart_pin_size += radeon_bo_size(bo);
398 dev_err(bo->rdev->dev, "%p pin failed\n", bo);
403 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
405 return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
408 int radeon_bo_unpin(struct radeon_bo *bo)
412 if (!bo->pin_count) {
413 dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
419 for (i = 0; i < bo->placement.num_placement; i++) {
420 bo->placements[i].lpfn = 0;
421 bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
423 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
424 if (likely(r == 0)) {
425 if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
426 bo->rdev->vram_pin_size -= radeon_bo_size(bo);
428 bo->rdev->gart_pin_size -= radeon_bo_size(bo);
430 dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
435 int radeon_bo_evict_vram(struct radeon_device *rdev)
437 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
438 if (0 && (rdev->flags & RADEON_IS_IGP)) {
439 if (rdev->mc.igp_sideport_enabled == false)
440 /* Useless to evict on IGP chips */
443 return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
446 void radeon_bo_force_delete(struct radeon_device *rdev)
448 struct radeon_bo *bo, *n;
450 if (list_empty(&rdev->gem.objects)) {
453 dev_err(rdev->dev, "Userspace still has active objects !\n");
454 list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
455 dev_err(rdev->dev, "%p %p %lu %lu force free\n",
456 &bo->gem_base, bo, (unsigned long)bo->gem_base.size,
457 *((unsigned long *)&bo->gem_base.refcount));
458 mutex_lock(&bo->rdev->gem.mutex);
459 list_del_init(&bo->list);
460 mutex_unlock(&bo->rdev->gem.mutex);
461 /* this should unref the ttm bo */
462 drm_gem_object_unreference_unlocked(&bo->gem_base);
466 int radeon_bo_init(struct radeon_device *rdev)
468 /* Add an MTRR for the VRAM */
469 if (!rdev->fastfb_working) {
470 rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
473 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
474 rdev->mc.mc_vram_size >> 20,
475 (unsigned long long)rdev->mc.aper_size >> 20);
476 DRM_INFO("RAM width %dbits %cDR\n",
477 rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
478 return radeon_ttm_init(rdev);
481 void radeon_bo_fini(struct radeon_device *rdev)
483 radeon_ttm_fini(rdev);
484 arch_phys_wc_del(rdev->mc.vram_mtrr);
487 /* Returns how many bytes TTM can move per IB.
489 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
491 u64 real_vram_size = rdev->mc.real_vram_size;
492 u64 vram_usage = atomic64_read(&rdev->vram_usage);
494 /* This function is based on the current VRAM usage.
496 * - If all of VRAM is free, allow relocating the number of bytes that
497 * is equal to 1/4 of the size of VRAM for this IB.
499 * - If more than one half of VRAM is occupied, only allow relocating
500 * 1 MB of data for this IB.
502 * - From 0 to one half of used VRAM, the threshold decreases
517 * Note: It's a threshold, not a limit. The threshold must be crossed
518 * for buffer relocations to stop, so any buffer of an arbitrary size
519 * can be moved as long as the threshold isn't crossed before
520 * the relocation takes place. We don't want to disable buffer
521 * relocations completely.
523 * The idea is that buffers should be placed in VRAM at creation time
524 * and TTM should only do a minimum number of relocations during
525 * command submission. In practice, you need to submit at least
526 * a dozen IBs to move all buffers to VRAM if they are in GTT.
528 * Also, things can get pretty crazy under memory pressure and actual
529 * VRAM usage can change a lot, so playing safe even at 50% does
530 * consistently increase performance.
533 u64 half_vram = real_vram_size >> 1;
534 u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
535 u64 bytes_moved_threshold = half_free_vram >> 1;
536 return max(bytes_moved_threshold, 1024*1024ull);
539 int radeon_bo_list_validate(struct radeon_device *rdev,
540 struct ww_acquire_ctx *ticket,
541 struct list_head *head, int ring)
543 struct radeon_bo_list *lobj;
544 struct radeon_bo *bo;
546 u64 bytes_moved = 0, initial_bytes_moved;
547 u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
549 r = ttm_eu_reserve_buffers(ticket, head);
550 if (unlikely(r != 0)) {
554 list_for_each_entry(lobj, head, tv.head) {
556 if (!bo->pin_count) {
557 u32 domain = lobj->prefered_domains;
558 u32 allowed = lobj->allowed_domains;
560 radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
562 /* Check if this buffer will be moved and don't move it
563 * if we have moved too many buffers for this IB already.
565 * Note that this allows moving at least one buffer of
566 * any size, because it doesn't take the current "bo"
567 * into account. We don't want to disallow buffer moves
570 if ((allowed & current_domain) != 0 &&
571 (domain & current_domain) == 0 && /* will be moved */
572 bytes_moved > bytes_moved_threshold) {
574 domain = current_domain;
578 radeon_ttm_placement_from_domain(bo, domain);
579 if (ring == R600_RING_TYPE_UVD_INDEX)
580 radeon_uvd_force_into_uvd_segment(bo, allowed);
582 initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
583 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
584 bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
588 if (r != -ERESTARTSYS &&
589 domain != lobj->allowed_domains) {
590 domain = lobj->allowed_domains;
593 ttm_eu_backoff_reservation(ticket, head);
597 lobj->gpu_offset = radeon_bo_gpu_offset(bo);
598 lobj->tiling_flags = bo->tiling_flags;
603 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
605 struct radeon_device *rdev = bo->rdev;
606 struct radeon_surface_reg *reg;
607 struct radeon_bo *old_object;
611 KASSERT(radeon_bo_is_reserved(bo),
612 ("radeon_bo_get_surface_reg: radeon_bo is not reserved"));
614 if (!bo->tiling_flags)
617 if (bo->surface_reg >= 0) {
618 reg = &rdev->surface_regs[bo->surface_reg];
624 for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
626 reg = &rdev->surface_regs[i];
630 old_object = reg->bo;
631 if (old_object->pin_count == 0)
635 /* if we are all out */
636 if (i == RADEON_GEM_MAX_SURFACES) {
639 /* find someone with a surface reg and nuke their BO */
640 reg = &rdev->surface_regs[steal];
641 old_object = reg->bo;
642 /* blow away the mapping */
643 DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
644 ttm_bo_unmap_virtual(&old_object->tbo);
645 old_object->surface_reg = -1;
653 radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
654 bo->tbo.mem.start << PAGE_SHIFT,
655 bo->tbo.num_pages << PAGE_SHIFT);
659 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
661 struct radeon_device *rdev = bo->rdev;
662 struct radeon_surface_reg *reg;
664 if (bo->surface_reg == -1)
667 reg = &rdev->surface_regs[bo->surface_reg];
668 radeon_clear_surface_reg(rdev, bo->surface_reg);
671 bo->surface_reg = -1;
674 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
675 uint32_t tiling_flags, uint32_t pitch)
677 struct radeon_device *rdev = bo->rdev;
680 if (rdev->family >= CHIP_CEDAR) {
681 unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
683 bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
684 bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
685 mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
686 tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
687 stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
721 if (stilesplit > 6) {
725 r = radeon_bo_reserve(bo, false);
726 if (unlikely(r != 0))
728 bo->tiling_flags = tiling_flags;
730 radeon_bo_unreserve(bo);
734 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
735 uint32_t *tiling_flags,
738 KASSERT(radeon_bo_is_reserved(bo),
739 ("radeon_bo_get_tiling_flags: radeon_bo is not reserved"));
741 *tiling_flags = bo->tiling_flags;
746 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
749 KASSERT((radeon_bo_is_reserved(bo) || force_drop),
750 ("radeon_bo_check_tiling: radeon_bo is not reserved && !force_drop"));
752 if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
756 radeon_bo_clear_surface_reg(bo);
760 if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
764 if (bo->surface_reg >= 0)
765 radeon_bo_clear_surface_reg(bo);
769 if ((bo->surface_reg >= 0) && !has_moved)
772 return radeon_bo_get_surface_reg(bo);
775 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
776 struct ttm_mem_reg *new_mem)
778 struct radeon_bo *rbo;
780 if (!radeon_ttm_bo_is_radeon_bo(bo))
783 rbo = container_of(bo, struct radeon_bo, tbo);
784 radeon_bo_check_tiling(rbo, 0, 1);
785 radeon_vm_bo_invalidate(rbo->rdev, rbo);
787 /* update statistics */
791 radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
792 radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
795 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
797 struct radeon_device *rdev;
798 struct radeon_bo *rbo;
799 unsigned long offset, size, lpfn;
802 if (!radeon_ttm_bo_is_radeon_bo(bo))
804 rbo = container_of(bo, struct radeon_bo, tbo);
805 radeon_bo_check_tiling(rbo, 0, 0);
807 if (bo->mem.mem_type != TTM_PL_VRAM)
810 size = bo->mem.num_pages << PAGE_SHIFT;
811 offset = bo->mem.start << PAGE_SHIFT;
812 if ((offset + size) <= rdev->mc.visible_vram_size)
815 /* Can't move a pinned BO to visible VRAM */
816 if (rbo->pin_count > 0)
819 /* hurrah the memory is not visible ! */
820 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
821 lpfn = rdev->mc.visible_vram_size >> PAGE_SHIFT;
822 for (i = 0; i < rbo->placement.num_placement; i++) {
823 /* Force into visible VRAM */
824 if ((rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
825 (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
826 rbo->placements[i].lpfn = lpfn;
828 r = ttm_bo_validate(bo, &rbo->placement, false, false);
829 if (unlikely(r == -ENOMEM)) {
830 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
831 return ttm_bo_validate(bo, &rbo->placement, false, false);
832 } else if (unlikely(r != 0)) {
836 offset = bo->mem.start << PAGE_SHIFT;
837 /* this should never happen */
838 if ((offset + size) > rdev->mc.visible_vram_size)
844 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
848 r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, NULL);
849 if (unlikely(r != 0))
851 lockmgr(&bo->tbo.bdev->fence_lock, LK_EXCLUSIVE);
853 *mem_type = bo->tbo.mem.mem_type;
854 if (bo->tbo.sync_obj)
855 r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
856 lockmgr(&bo->tbo.bdev->fence_lock, LK_RELEASE);
857 ttm_bo_unreserve(&bo->tbo);