2 * Copyright 2007 Dave Airlied
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * VA LINUX SYSTEMS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
25 * Authors: Dave Airlied <airlied@linux.ie>
26 * Ben Skeggs <darktama@iinet.net.au>
27 * Jeremy Kolb <jkolb@brandeis.edu>
30 #include <linux/dma-mapping.h>
31 #include <linux/swiotlb.h>
33 #include "nouveau_drv.h"
34 #include "nouveau_chan.h"
35 #include "nouveau_fence.h"
37 #include "nouveau_bo.h"
38 #include "nouveau_ttm.h"
39 #include "nouveau_gem.h"
40 #include "nouveau_mem.h"
41 #include "nouveau_vmm.h"
43 #include <nvif/class.h>
44 #include <nvif/if500b.h>
45 #include <nvif/if900b.h>
48 * NV10-NV40 tiling helpers
52 nv10_bo_update_tile_region(struct drm_device *dev, struct nouveau_drm_tile *reg,
53 u32 addr, u32 size, u32 pitch, u32 flags)
55 struct nouveau_drm *drm = nouveau_drm(dev);
56 int i = reg - drm->tile.reg;
57 struct nvkm_fb *fb = nvxx_fb(&drm->client.device);
58 struct nvkm_fb_tile *tile = &fb->tile.region[i];
60 nouveau_fence_unref(®->fence);
63 nvkm_fb_tile_fini(fb, i, tile);
66 nvkm_fb_tile_init(fb, i, addr, size, pitch, flags, tile);
68 nvkm_fb_tile_prog(fb, i, tile);
71 static struct nouveau_drm_tile *
72 nv10_bo_get_tile_region(struct drm_device *dev, int i)
74 struct nouveau_drm *drm = nouveau_drm(dev);
75 struct nouveau_drm_tile *tile = &drm->tile.reg[i];
77 spin_lock(&drm->tile.lock);
80 (!tile->fence || nouveau_fence_done(tile->fence)))
85 spin_unlock(&drm->tile.lock);
90 nv10_bo_put_tile_region(struct drm_device *dev, struct nouveau_drm_tile *tile,
91 struct dma_fence *fence)
93 struct nouveau_drm *drm = nouveau_drm(dev);
96 spin_lock(&drm->tile.lock);
97 tile->fence = (struct nouveau_fence *)dma_fence_get(fence);
99 spin_unlock(&drm->tile.lock);
103 static struct nouveau_drm_tile *
104 nv10_bo_set_tiling(struct drm_device *dev, u32 addr,
105 u32 size, u32 pitch, u32 zeta)
107 struct nouveau_drm *drm = nouveau_drm(dev);
108 struct nvkm_fb *fb = nvxx_fb(&drm->client.device);
109 struct nouveau_drm_tile *tile, *found = NULL;
112 for (i = 0; i < fb->tile.regions; i++) {
113 tile = nv10_bo_get_tile_region(dev, i);
115 if (pitch && !found) {
119 } else if (tile && fb->tile.region[i].pitch) {
120 /* Kill an unused tile region. */
121 nv10_bo_update_tile_region(dev, tile, 0, 0, 0, 0);
124 nv10_bo_put_tile_region(dev, tile, NULL);
128 nv10_bo_update_tile_region(dev, found, addr, size, pitch, zeta);
133 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
135 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
136 struct drm_device *dev = drm->dev;
137 struct nouveau_bo *nvbo = nouveau_bo(bo);
139 WARN_ON(nvbo->pin_refcnt > 0);
140 nv10_bo_put_tile_region(dev, nvbo->tile, NULL);
143 * If nouveau_bo_new() allocated this buffer, the GEM object was never
144 * initialized, so don't attempt to release it.
147 drm_gem_object_release(&bo->base);
153 roundup_64(u64 x, u32 y)
161 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags,
162 int *align, u64 *size)
164 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
165 struct nvif_device *device = &drm->client.device;
167 if (device->info.family < NV_DEVICE_INFO_V0_TESLA) {
169 if (device->info.chipset >= 0x40) {
171 *size = roundup_64(*size, 64 * nvbo->mode);
173 } else if (device->info.chipset >= 0x30) {
175 *size = roundup_64(*size, 64 * nvbo->mode);
177 } else if (device->info.chipset >= 0x20) {
179 *size = roundup_64(*size, 64 * nvbo->mode);
181 } else if (device->info.chipset >= 0x10) {
183 *size = roundup_64(*size, 32 * nvbo->mode);
187 *size = roundup_64(*size, (1 << nvbo->page));
188 *align = max((1 << nvbo->page), *align);
191 *size = roundup_64(*size, PAGE_SIZE);
195 nouveau_bo_alloc(struct nouveau_cli *cli, u64 *size, int *align, u32 flags,
196 u32 tile_mode, u32 tile_flags)
198 struct nouveau_drm *drm = cli->drm;
199 struct nouveau_bo *nvbo;
200 struct nvif_mmu *mmu = &cli->mmu;
201 struct nvif_vmm *vmm = cli->svm.cli ? &cli->svm.vmm : &cli->vmm.vmm;
205 NV_WARN(drm, "skipped size %016llx\n", *size);
206 return ERR_PTR(-EINVAL);
209 nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
211 return ERR_PTR(-ENOMEM);
212 INIT_LIST_HEAD(&nvbo->head);
213 INIT_LIST_HEAD(&nvbo->entry);
214 INIT_LIST_HEAD(&nvbo->vma_list);
215 nvbo->bo.bdev = &drm->ttm.bdev;
217 /* This is confusing, and doesn't actually mean we want an uncached
218 * mapping, but is what NOUVEAU_GEM_DOMAIN_COHERENT gets translated
219 * into in nouveau_gem_new().
221 if (flags & TTM_PL_FLAG_UNCACHED) {
222 /* Determine if we can get a cache-coherent map, forcing
223 * uncached mapping if we can't.
225 if (!nouveau_drm_use_coherent_gpu_mapping(drm))
226 nvbo->force_coherent = true;
229 if (cli->device.info.family >= NV_DEVICE_INFO_V0_FERMI) {
230 nvbo->kind = (tile_flags & 0x0000ff00) >> 8;
231 if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
233 return ERR_PTR(-EINVAL);
236 nvbo->comp = mmu->kind[nvbo->kind] != nvbo->kind;
238 if (cli->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
239 nvbo->kind = (tile_flags & 0x00007f00) >> 8;
240 nvbo->comp = (tile_flags & 0x00030000) >> 16;
241 if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
243 return ERR_PTR(-EINVAL);
246 nvbo->zeta = (tile_flags & 0x00000007);
248 nvbo->mode = tile_mode;
249 nvbo->contig = !(tile_flags & NOUVEAU_GEM_TILE_NONCONTIG);
251 /* Determine the desirable target GPU page size for the buffer. */
252 for (i = 0; i < vmm->page_nr; i++) {
253 /* Because we cannot currently allow VMM maps to fail
254 * during buffer migration, we need to determine page
255 * size for the buffer up-front, and pre-allocate its
258 * Skip page sizes that can't support needed domains.
260 if (cli->device.info.family > NV_DEVICE_INFO_V0_CURIE &&
261 (flags & TTM_PL_FLAG_VRAM) && !vmm->page[i].vram)
263 if ((flags & TTM_PL_FLAG_TT) &&
264 (!vmm->page[i].host || vmm->page[i].shift > PAGE_SHIFT))
267 /* Select this page size if it's the first that supports
268 * the potential memory domains, or when it's compatible
269 * with the requested compression settings.
271 if (pi < 0 || !nvbo->comp || vmm->page[i].comp)
274 /* Stop once the buffer is larger than the current page size. */
275 if (*size >= 1ULL << vmm->page[i].shift)
280 return ERR_PTR(-EINVAL);
282 /* Disable compression if suitable settings couldn't be found. */
283 if (nvbo->comp && !vmm->page[pi].comp) {
284 if (mmu->object.oclass >= NVIF_CLASS_MMU_GF100)
285 nvbo->kind = mmu->kind[nvbo->kind];
288 nvbo->page = vmm->page[pi].shift;
290 nouveau_bo_fixup_align(nvbo, flags, align, size);
296 nouveau_bo_init(struct nouveau_bo *nvbo, u64 size, int align, u32 flags,
297 struct sg_table *sg, struct dma_resv *robj)
299 int type = sg ? ttm_bo_type_sg : ttm_bo_type_device;
303 acc_size = ttm_bo_dma_acc_size(nvbo->bo.bdev, size, sizeof(*nvbo));
305 nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
306 nouveau_bo_placement_set(nvbo, flags, 0);
308 ret = ttm_bo_init(nvbo->bo.bdev, &nvbo->bo, size, type,
309 &nvbo->placement, align >> PAGE_SHIFT, false,
310 acc_size, sg, robj, nouveau_bo_del_ttm);
312 /* ttm will call nouveau_bo_del_ttm if it fails.. */
320 nouveau_bo_new(struct nouveau_cli *cli, u64 size, int align,
321 uint32_t flags, uint32_t tile_mode, uint32_t tile_flags,
322 struct sg_table *sg, struct dma_resv *robj,
323 struct nouveau_bo **pnvbo)
325 struct nouveau_bo *nvbo;
328 nvbo = nouveau_bo_alloc(cli, &size, &align, flags, tile_mode,
331 return PTR_ERR(nvbo);
333 ret = nouveau_bo_init(nvbo, size, align, flags, sg, robj);
342 set_placement_list(struct ttm_place *pl, unsigned *n, uint32_t type, uint32_t flags)
346 if (type & TTM_PL_FLAG_VRAM)
347 pl[(*n)++].flags = TTM_PL_FLAG_VRAM | flags;
348 if (type & TTM_PL_FLAG_TT)
349 pl[(*n)++].flags = TTM_PL_FLAG_TT | flags;
350 if (type & TTM_PL_FLAG_SYSTEM)
351 pl[(*n)++].flags = TTM_PL_FLAG_SYSTEM | flags;
355 set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
357 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
358 u32 vram_pages = drm->client.device.info.ram_size >> PAGE_SHIFT;
359 unsigned i, fpfn, lpfn;
361 if (drm->client.device.info.family == NV_DEVICE_INFO_V0_CELSIUS &&
362 nvbo->mode && (type & TTM_PL_FLAG_VRAM) &&
363 nvbo->bo.mem.num_pages < vram_pages / 4) {
365 * Make sure that the color and depth buffers are handled
366 * by independent memory controller units. Up to a 9x
367 * speed up when alpha-blending and depth-test are enabled
371 fpfn = vram_pages / 2;
375 lpfn = vram_pages / 2;
377 for (i = 0; i < nvbo->placement.num_placement; ++i) {
378 nvbo->placements[i].fpfn = fpfn;
379 nvbo->placements[i].lpfn = lpfn;
381 for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
382 nvbo->busy_placements[i].fpfn = fpfn;
383 nvbo->busy_placements[i].lpfn = lpfn;
389 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
391 struct ttm_placement *pl = &nvbo->placement;
392 uint32_t flags = (nvbo->force_coherent ? TTM_PL_FLAG_UNCACHED :
393 TTM_PL_MASK_CACHING) |
394 (nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
396 pl->placement = nvbo->placements;
397 set_placement_list(nvbo->placements, &pl->num_placement,
400 pl->busy_placement = nvbo->busy_placements;
401 set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
404 set_placement_range(nvbo, type);
408 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype, bool contig)
410 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
411 struct ttm_buffer_object *bo = &nvbo->bo;
412 bool force = false, evict = false;
415 ret = ttm_bo_reserve(bo, false, false, NULL);
419 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
420 memtype == TTM_PL_FLAG_VRAM && contig) {
428 if (nvbo->pin_refcnt) {
429 if (!(memtype & (1 << bo->mem.mem_type)) || evict) {
430 NV_ERROR(drm, "bo %p pinned elsewhere: "
431 "0x%08x vs 0x%08x\n", bo,
432 1 << bo->mem.mem_type, memtype);
440 nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT, 0);
441 ret = nouveau_bo_validate(nvbo, false, false);
447 nouveau_bo_placement_set(nvbo, memtype, 0);
449 /* drop pin_refcnt temporarily, so we don't trip the assertion
450 * in nouveau_bo_move() that makes sure we're not trying to
451 * move a pinned buffer
454 ret = nouveau_bo_validate(nvbo, false, false);
459 switch (bo->mem.mem_type) {
461 drm->gem.vram_available -= bo->mem.size;
464 drm->gem.gart_available -= bo->mem.size;
472 nvbo->contig = false;
473 ttm_bo_unreserve(bo);
478 nouveau_bo_unpin(struct nouveau_bo *nvbo)
480 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
481 struct ttm_buffer_object *bo = &nvbo->bo;
484 ret = ttm_bo_reserve(bo, false, false, NULL);
488 ref = --nvbo->pin_refcnt;
489 WARN_ON_ONCE(ref < 0);
493 nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
495 ret = nouveau_bo_validate(nvbo, false, false);
497 switch (bo->mem.mem_type) {
499 drm->gem.vram_available += bo->mem.size;
502 drm->gem.gart_available += bo->mem.size;
510 ttm_bo_unreserve(bo);
515 nouveau_bo_map(struct nouveau_bo *nvbo)
519 ret = ttm_bo_reserve(&nvbo->bo, false, false, NULL);
523 ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
525 ttm_bo_unreserve(&nvbo->bo);
530 nouveau_bo_unmap(struct nouveau_bo *nvbo)
535 ttm_bo_kunmap(&nvbo->kmap);
539 nouveau_bo_sync_for_device(struct nouveau_bo *nvbo)
541 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
542 struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
548 /* Don't waste time looping if the object is coherent */
549 if (nvbo->force_coherent)
552 for (i = 0; i < ttm_dma->ttm.num_pages; i++)
553 dma_sync_single_for_device(drm->dev->dev,
554 ttm_dma->dma_address[i],
555 PAGE_SIZE, DMA_TO_DEVICE);
559 nouveau_bo_sync_for_cpu(struct nouveau_bo *nvbo)
561 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
562 struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
568 /* Don't waste time looping if the object is coherent */
569 if (nvbo->force_coherent)
572 for (i = 0; i < ttm_dma->ttm.num_pages; i++)
573 dma_sync_single_for_cpu(drm->dev->dev, ttm_dma->dma_address[i],
574 PAGE_SIZE, DMA_FROM_DEVICE);
578 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
581 struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
584 ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement, &ctx);
588 nouveau_bo_sync_for_device(nvbo);
594 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
597 u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
602 iowrite16_native(val, (void __force __iomem *)mem);
608 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
611 u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
616 return ioread32_native((void __force __iomem *)mem);
622 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
625 u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
630 iowrite32_native(val, (void __force __iomem *)mem);
635 static struct ttm_tt *
636 nouveau_ttm_tt_create(struct ttm_buffer_object *bo, uint32_t page_flags)
638 #if IS_ENABLED(CONFIG_AGP)
639 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
641 if (drm->agp.bridge) {
642 return ttm_agp_tt_create(bo, drm->agp.bridge, page_flags);
646 return nouveau_sgdma_create_ttm(bo, page_flags);
650 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
652 struct nouveau_bo *nvbo = nouveau_bo(bo);
654 switch (bo->mem.mem_type) {
656 nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
660 nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
664 *pl = nvbo->placement;
668 nouveau_bo_move_prep(struct nouveau_drm *drm, struct ttm_buffer_object *bo,
669 struct ttm_resource *reg)
671 struct nouveau_mem *old_mem = nouveau_mem(&bo->mem);
672 struct nouveau_mem *new_mem = nouveau_mem(reg);
673 struct nvif_vmm *vmm = &drm->client.vmm.vmm;
676 ret = nvif_vmm_get(vmm, LAZY, false, old_mem->mem.page, 0,
677 old_mem->mem.size, &old_mem->vma[0]);
681 ret = nvif_vmm_get(vmm, LAZY, false, new_mem->mem.page, 0,
682 new_mem->mem.size, &old_mem->vma[1]);
686 ret = nouveau_mem_map(old_mem, vmm, &old_mem->vma[0]);
690 ret = nouveau_mem_map(new_mem, vmm, &old_mem->vma[1]);
693 nvif_vmm_put(vmm, &old_mem->vma[1]);
694 nvif_vmm_put(vmm, &old_mem->vma[0]);
700 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
701 bool no_wait_gpu, struct ttm_resource *new_reg)
703 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
704 struct nouveau_channel *chan = drm->ttm.chan;
705 struct nouveau_cli *cli = (void *)chan->user.client;
706 struct nouveau_fence *fence;
709 /* create temporary vmas for the transfer and attach them to the
710 * old nvkm_mem node, these will get cleaned up after ttm has
711 * destroyed the ttm_resource
713 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
714 ret = nouveau_bo_move_prep(drm, bo, new_reg);
719 mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
720 ret = nouveau_fence_sync(nouveau_bo(bo), chan, true, intr);
722 ret = drm->ttm.move(chan, bo, &bo->mem, new_reg);
724 ret = nouveau_fence_new(chan, false, &fence);
726 ret = ttm_bo_move_accel_cleanup(bo,
730 nouveau_fence_unref(&fence);
734 mutex_unlock(&cli->mutex);
739 nouveau_bo_move_init(struct nouveau_drm *drm)
741 static const struct _method_table {
745 int (*exec)(struct nouveau_channel *,
746 struct ttm_buffer_object *,
747 struct ttm_resource *, struct ttm_resource *);
748 int (*init)(struct nouveau_channel *, u32 handle);
750 { "COPY", 4, 0xc5b5, nve0_bo_move_copy, nve0_bo_move_init },
751 { "GRCE", 0, 0xc5b5, nve0_bo_move_copy, nvc0_bo_move_init },
752 { "COPY", 4, 0xc3b5, nve0_bo_move_copy, nve0_bo_move_init },
753 { "GRCE", 0, 0xc3b5, nve0_bo_move_copy, nvc0_bo_move_init },
754 { "COPY", 4, 0xc1b5, nve0_bo_move_copy, nve0_bo_move_init },
755 { "GRCE", 0, 0xc1b5, nve0_bo_move_copy, nvc0_bo_move_init },
756 { "COPY", 4, 0xc0b5, nve0_bo_move_copy, nve0_bo_move_init },
757 { "GRCE", 0, 0xc0b5, nve0_bo_move_copy, nvc0_bo_move_init },
758 { "COPY", 4, 0xb0b5, nve0_bo_move_copy, nve0_bo_move_init },
759 { "GRCE", 0, 0xb0b5, nve0_bo_move_copy, nvc0_bo_move_init },
760 { "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
761 { "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
762 { "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
763 { "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
764 { "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
765 { "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
766 { "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
767 { "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
768 { "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
771 const struct _method_table *mthd = _methods;
772 const char *name = "CPU";
776 struct nouveau_channel *chan;
785 ret = nvif_object_ctor(&chan->user, "ttmBoMove",
786 mthd->oclass | (mthd->engine << 16),
787 mthd->oclass, NULL, 0,
790 ret = mthd->init(chan, drm->ttm.copy.handle);
792 nvif_object_dtor(&drm->ttm.copy);
796 drm->ttm.move = mthd->exec;
797 drm->ttm.chan = chan;
801 } while ((++mthd)->exec);
803 NV_INFO(drm, "MM: using %s for buffer copies\n", name);
807 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
808 bool no_wait_gpu, struct ttm_resource *new_reg)
810 struct ttm_operation_ctx ctx = { intr, no_wait_gpu };
811 struct ttm_place placement_memtype = {
814 .flags = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING
816 struct ttm_placement placement;
817 struct ttm_resource tmp_reg;
820 placement.num_placement = placement.num_busy_placement = 1;
821 placement.placement = placement.busy_placement = &placement_memtype;
824 tmp_reg.mm_node = NULL;
825 ret = ttm_bo_mem_space(bo, &placement, &tmp_reg, &ctx);
829 ret = ttm_tt_bind(bo->ttm, &tmp_reg, &ctx);
833 ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, &tmp_reg);
837 ret = ttm_bo_move_ttm(bo, &ctx, new_reg);
839 ttm_resource_free(bo, &tmp_reg);
844 nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
845 bool no_wait_gpu, struct ttm_resource *new_reg)
847 struct ttm_operation_ctx ctx = { intr, no_wait_gpu };
848 struct ttm_place placement_memtype = {
851 .flags = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING
853 struct ttm_placement placement;
854 struct ttm_resource tmp_reg;
857 placement.num_placement = placement.num_busy_placement = 1;
858 placement.placement = placement.busy_placement = &placement_memtype;
861 tmp_reg.mm_node = NULL;
862 ret = ttm_bo_mem_space(bo, &placement, &tmp_reg, &ctx);
866 ret = ttm_bo_move_ttm(bo, &ctx, &tmp_reg);
870 ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, new_reg);
875 ttm_resource_free(bo, &tmp_reg);
880 nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, bool evict,
881 struct ttm_resource *new_reg)
883 struct nouveau_mem *mem = new_reg ? nouveau_mem(new_reg) : NULL;
884 struct nouveau_bo *nvbo = nouveau_bo(bo);
885 struct nouveau_vma *vma;
887 /* ttm can now (stupidly) pass the driver bos it didn't create... */
888 if (bo->destroy != nouveau_bo_del_ttm)
891 if (mem && new_reg->mem_type != TTM_PL_SYSTEM &&
892 mem->mem.page == nvbo->page) {
893 list_for_each_entry(vma, &nvbo->vma_list, head) {
894 nouveau_vma_map(vma, mem);
897 list_for_each_entry(vma, &nvbo->vma_list, head) {
898 WARN_ON(ttm_bo_wait(bo, false, false));
899 nouveau_vma_unmap(vma);
904 if (new_reg->mm_node)
905 nvbo->offset = (new_reg->start << PAGE_SHIFT);
913 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_resource *new_reg,
914 struct nouveau_drm_tile **new_tile)
916 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
917 struct drm_device *dev = drm->dev;
918 struct nouveau_bo *nvbo = nouveau_bo(bo);
919 u64 offset = new_reg->start << PAGE_SHIFT;
922 if (new_reg->mem_type != TTM_PL_VRAM)
925 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_CELSIUS) {
926 *new_tile = nv10_bo_set_tiling(dev, offset, new_reg->size,
927 nvbo->mode, nvbo->zeta);
934 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
935 struct nouveau_drm_tile *new_tile,
936 struct nouveau_drm_tile **old_tile)
938 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
939 struct drm_device *dev = drm->dev;
940 struct dma_fence *fence = dma_resv_get_excl(bo->base.resv);
942 nv10_bo_put_tile_region(dev, *old_tile, fence);
943 *old_tile = new_tile;
947 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict,
948 struct ttm_operation_ctx *ctx,
949 struct ttm_resource *new_reg)
951 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
952 struct nouveau_bo *nvbo = nouveau_bo(bo);
953 struct ttm_resource *old_reg = &bo->mem;
954 struct nouveau_drm_tile *new_tile = NULL;
957 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
961 if (nvbo->pin_refcnt)
962 NV_WARN(drm, "Moving pinned object %p!\n", nvbo);
964 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
965 ret = nouveau_bo_vm_bind(bo, new_reg, &new_tile);
971 if (old_reg->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
972 BUG_ON(bo->mem.mm_node != NULL);
974 new_reg->mm_node = NULL;
978 /* Hardware assisted copy. */
980 if (new_reg->mem_type == TTM_PL_SYSTEM)
981 ret = nouveau_bo_move_flipd(bo, evict,
983 ctx->no_wait_gpu, new_reg);
984 else if (old_reg->mem_type == TTM_PL_SYSTEM)
985 ret = nouveau_bo_move_flips(bo, evict,
987 ctx->no_wait_gpu, new_reg);
989 ret = nouveau_bo_move_m2mf(bo, evict,
991 ctx->no_wait_gpu, new_reg);
996 /* Fallback to software copy. */
997 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
999 ret = ttm_bo_move_memcpy(bo, ctx, new_reg);
1002 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1004 nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1006 nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1013 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
1015 struct nouveau_bo *nvbo = nouveau_bo(bo);
1017 return drm_vma_node_verify_access(&nvbo->bo.base.vma_node,
1018 filp->private_data);
1022 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_resource *reg)
1024 struct nouveau_drm *drm = nouveau_bdev(bdev);
1025 struct nvkm_device *device = nvxx_device(&drm->client.device);
1026 struct nouveau_mem *mem = nouveau_mem(reg);
1028 switch (reg->mem_type) {
1033 #if IS_ENABLED(CONFIG_AGP)
1034 if (drm->agp.bridge) {
1035 reg->bus.offset = reg->start << PAGE_SHIFT;
1036 reg->bus.base = drm->agp.base;
1037 reg->bus.is_iomem = !drm->agp.cma;
1040 if (drm->client.mem->oclass < NVIF_CLASS_MEM_NV50 || !mem->kind)
1043 fallthrough; /* tiled memory */
1045 reg->bus.offset = reg->start << PAGE_SHIFT;
1046 reg->bus.base = device->func->resource_addr(device, 1);
1047 reg->bus.is_iomem = true;
1048 if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1050 struct nv50_mem_map_v0 nv50;
1051 struct gf100_mem_map_v0 gf100;
1057 switch (mem->mem.object.oclass) {
1058 case NVIF_CLASS_MEM_NV50:
1059 args.nv50.version = 0;
1061 args.nv50.kind = mem->kind;
1062 args.nv50.comp = mem->comp;
1063 argc = sizeof(args.nv50);
1065 case NVIF_CLASS_MEM_GF100:
1066 args.gf100.version = 0;
1068 args.gf100.kind = mem->kind;
1069 argc = sizeof(args.gf100);
1076 ret = nvif_object_map_handle(&mem->mem.object,
1080 if (WARN_ON(ret == 0))
1086 reg->bus.offset = handle;
1096 nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_resource *reg)
1098 struct nouveau_drm *drm = nouveau_bdev(bdev);
1099 struct nouveau_mem *mem = nouveau_mem(reg);
1101 if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1102 switch (reg->mem_type) {
1105 nvif_object_unmap_handle(&mem->mem.object);
1108 nvif_object_unmap_handle(&mem->mem.object);
1117 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1119 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1120 struct nouveau_bo *nvbo = nouveau_bo(bo);
1121 struct nvkm_device *device = nvxx_device(&drm->client.device);
1122 u32 mappable = device->func->resource_size(device, 1) >> PAGE_SHIFT;
1125 /* as long as the bo isn't in vram, and isn't tiled, we've got
1126 * nothing to do here.
1128 if (bo->mem.mem_type != TTM_PL_VRAM) {
1129 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA ||
1133 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
1134 nouveau_bo_placement_set(nvbo, TTM_PL_TT, 0);
1136 ret = nouveau_bo_validate(nvbo, false, false);
1143 /* make sure bo is in mappable vram */
1144 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA ||
1145 bo->mem.start + bo->mem.num_pages < mappable)
1148 for (i = 0; i < nvbo->placement.num_placement; ++i) {
1149 nvbo->placements[i].fpfn = 0;
1150 nvbo->placements[i].lpfn = mappable;
1153 for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
1154 nvbo->busy_placements[i].fpfn = 0;
1155 nvbo->busy_placements[i].lpfn = mappable;
1158 nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_VRAM, 0);
1159 return nouveau_bo_validate(nvbo, false, false);
1163 nouveau_ttm_tt_populate(struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
1165 struct ttm_dma_tt *ttm_dma = (void *)ttm;
1166 struct nouveau_drm *drm;
1168 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1170 if (ttm->state != tt_unpopulated)
1173 if (slave && ttm->sg) {
1174 /* make userspace faulting work */
1175 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1176 ttm_dma->dma_address, ttm->num_pages);
1177 ttm->state = tt_unbound;
1181 drm = nouveau_bdev(ttm->bdev);
1182 dev = drm->dev->dev;
1184 #if IS_ENABLED(CONFIG_AGP)
1185 if (drm->agp.bridge) {
1186 return ttm_agp_tt_populate(ttm, ctx);
1190 #if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1191 if (swiotlb_nr_tbl()) {
1192 return ttm_dma_populate((void *)ttm, dev, ctx);
1195 return ttm_populate_and_map_pages(dev, ttm_dma, ctx);
1199 nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
1201 struct ttm_dma_tt *ttm_dma = (void *)ttm;
1202 struct nouveau_drm *drm;
1204 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1209 drm = nouveau_bdev(ttm->bdev);
1210 dev = drm->dev->dev;
1212 #if IS_ENABLED(CONFIG_AGP)
1213 if (drm->agp.bridge) {
1214 ttm_agp_tt_unpopulate(ttm);
1219 #if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1220 if (swiotlb_nr_tbl()) {
1221 ttm_dma_unpopulate((void *)ttm, dev);
1226 ttm_unmap_and_unpopulate_pages(dev, ttm_dma);
1230 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence, bool exclusive)
1232 struct dma_resv *resv = nvbo->bo.base.resv;
1235 dma_resv_add_excl_fence(resv, &fence->base);
1237 dma_resv_add_shared_fence(resv, &fence->base);
1240 struct ttm_bo_driver nouveau_bo_driver = {
1241 .ttm_tt_create = &nouveau_ttm_tt_create,
1242 .ttm_tt_populate = &nouveau_ttm_tt_populate,
1243 .ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1244 .eviction_valuable = ttm_bo_eviction_valuable,
1245 .evict_flags = nouveau_bo_evict_flags,
1246 .move_notify = nouveau_bo_move_ntfy,
1247 .move = nouveau_bo_move,
1248 .verify_access = nouveau_bo_verify_access,
1249 .fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1250 .io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1251 .io_mem_free = &nouveau_ttm_io_mem_free,