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>
32 #include "nouveau_drv.h"
33 #include "nouveau_chan.h"
34 #include "nouveau_fence.h"
36 #include "nouveau_bo.h"
37 #include "nouveau_ttm.h"
38 #include "nouveau_gem.h"
39 #include "nouveau_mem.h"
40 #include "nouveau_vmm.h"
42 #include <nvif/class.h>
43 #include <nvif/if500b.h>
44 #include <nvif/if900b.h>
46 static int nouveau_ttm_tt_bind(struct ttm_device *bdev, struct ttm_tt *ttm,
47 struct ttm_resource *reg);
48 static void nouveau_ttm_tt_unbind(struct ttm_device *bdev, struct ttm_tt *ttm);
51 * NV10-NV40 tiling helpers
55 nv10_bo_update_tile_region(struct drm_device *dev, struct nouveau_drm_tile *reg,
56 u32 addr, u32 size, u32 pitch, u32 flags)
58 struct nouveau_drm *drm = nouveau_drm(dev);
59 int i = reg - drm->tile.reg;
60 struct nvkm_fb *fb = nvxx_fb(&drm->client.device);
61 struct nvkm_fb_tile *tile = &fb->tile.region[i];
63 nouveau_fence_unref(®->fence);
66 nvkm_fb_tile_fini(fb, i, tile);
69 nvkm_fb_tile_init(fb, i, addr, size, pitch, flags, tile);
71 nvkm_fb_tile_prog(fb, i, tile);
74 static struct nouveau_drm_tile *
75 nv10_bo_get_tile_region(struct drm_device *dev, int i)
77 struct nouveau_drm *drm = nouveau_drm(dev);
78 struct nouveau_drm_tile *tile = &drm->tile.reg[i];
80 spin_lock(&drm->tile.lock);
83 (!tile->fence || nouveau_fence_done(tile->fence)))
88 spin_unlock(&drm->tile.lock);
93 nv10_bo_put_tile_region(struct drm_device *dev, struct nouveau_drm_tile *tile,
94 struct dma_fence *fence)
96 struct nouveau_drm *drm = nouveau_drm(dev);
99 spin_lock(&drm->tile.lock);
100 tile->fence = (struct nouveau_fence *)dma_fence_get(fence);
102 spin_unlock(&drm->tile.lock);
106 static struct nouveau_drm_tile *
107 nv10_bo_set_tiling(struct drm_device *dev, u32 addr,
108 u32 size, u32 pitch, u32 zeta)
110 struct nouveau_drm *drm = nouveau_drm(dev);
111 struct nvkm_fb *fb = nvxx_fb(&drm->client.device);
112 struct nouveau_drm_tile *tile, *found = NULL;
115 for (i = 0; i < fb->tile.regions; i++) {
116 tile = nv10_bo_get_tile_region(dev, i);
118 if (pitch && !found) {
122 } else if (tile && fb->tile.region[i].pitch) {
123 /* Kill an unused tile region. */
124 nv10_bo_update_tile_region(dev, tile, 0, 0, 0, 0);
127 nv10_bo_put_tile_region(dev, tile, NULL);
131 nv10_bo_update_tile_region(dev, found, addr, size, pitch, zeta);
136 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
138 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
139 struct drm_device *dev = drm->dev;
140 struct nouveau_bo *nvbo = nouveau_bo(bo);
142 WARN_ON(nvbo->bo.pin_count > 0);
143 nouveau_bo_del_io_reserve_lru(bo);
144 nv10_bo_put_tile_region(dev, nvbo->tile, NULL);
147 * If nouveau_bo_new() allocated this buffer, the GEM object was never
148 * initialized, so don't attempt to release it.
151 drm_gem_object_release(&bo->base);
153 dma_resv_fini(&bo->base._resv);
159 roundup_64(u64 x, u32 y)
167 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, int *align, u64 *size)
169 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
170 struct nvif_device *device = &drm->client.device;
172 if (device->info.family < NV_DEVICE_INFO_V0_TESLA) {
174 if (device->info.chipset >= 0x40) {
176 *size = roundup_64(*size, 64 * nvbo->mode);
178 } else if (device->info.chipset >= 0x30) {
180 *size = roundup_64(*size, 64 * nvbo->mode);
182 } else if (device->info.chipset >= 0x20) {
184 *size = roundup_64(*size, 64 * nvbo->mode);
186 } else if (device->info.chipset >= 0x10) {
188 *size = roundup_64(*size, 32 * nvbo->mode);
192 *size = roundup_64(*size, (1 << nvbo->page));
193 *align = max((1 << nvbo->page), *align);
196 *size = roundup_64(*size, PAGE_SIZE);
200 nouveau_bo_alloc(struct nouveau_cli *cli, u64 *size, int *align, u32 domain,
201 u32 tile_mode, u32 tile_flags)
203 struct nouveau_drm *drm = cli->drm;
204 struct nouveau_bo *nvbo;
205 struct nvif_mmu *mmu = &cli->mmu;
206 struct nvif_vmm *vmm = cli->svm.cli ? &cli->svm.vmm : &cli->vmm.vmm;
210 NV_WARN(drm, "skipped size %016llx\n", *size);
211 return ERR_PTR(-EINVAL);
214 nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
216 return ERR_PTR(-ENOMEM);
217 INIT_LIST_HEAD(&nvbo->head);
218 INIT_LIST_HEAD(&nvbo->entry);
219 INIT_LIST_HEAD(&nvbo->vma_list);
220 nvbo->bo.bdev = &drm->ttm.bdev;
222 /* This is confusing, and doesn't actually mean we want an uncached
223 * mapping, but is what NOUVEAU_GEM_DOMAIN_COHERENT gets translated
224 * into in nouveau_gem_new().
226 if (domain & NOUVEAU_GEM_DOMAIN_COHERENT) {
227 /* Determine if we can get a cache-coherent map, forcing
228 * uncached mapping if we can't.
230 if (!nouveau_drm_use_coherent_gpu_mapping(drm))
231 nvbo->force_coherent = true;
234 if (cli->device.info.family >= NV_DEVICE_INFO_V0_FERMI) {
235 nvbo->kind = (tile_flags & 0x0000ff00) >> 8;
236 if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
238 return ERR_PTR(-EINVAL);
241 nvbo->comp = mmu->kind[nvbo->kind] != nvbo->kind;
243 if (cli->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
244 nvbo->kind = (tile_flags & 0x00007f00) >> 8;
245 nvbo->comp = (tile_flags & 0x00030000) >> 16;
246 if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
248 return ERR_PTR(-EINVAL);
251 nvbo->zeta = (tile_flags & 0x00000007);
253 nvbo->mode = tile_mode;
254 nvbo->contig = !(tile_flags & NOUVEAU_GEM_TILE_NONCONTIG);
256 /* Determine the desirable target GPU page size for the buffer. */
257 for (i = 0; i < vmm->page_nr; i++) {
258 /* Because we cannot currently allow VMM maps to fail
259 * during buffer migration, we need to determine page
260 * size for the buffer up-front, and pre-allocate its
263 * Skip page sizes that can't support needed domains.
265 if (cli->device.info.family > NV_DEVICE_INFO_V0_CURIE &&
266 (domain & NOUVEAU_GEM_DOMAIN_VRAM) && !vmm->page[i].vram)
268 if ((domain & NOUVEAU_GEM_DOMAIN_GART) &&
269 (!vmm->page[i].host || vmm->page[i].shift > PAGE_SHIFT))
272 /* Select this page size if it's the first that supports
273 * the potential memory domains, or when it's compatible
274 * with the requested compression settings.
276 if (pi < 0 || !nvbo->comp || vmm->page[i].comp)
279 /* Stop once the buffer is larger than the current page size. */
280 if (*size >= 1ULL << vmm->page[i].shift)
284 if (WARN_ON(pi < 0)) {
286 return ERR_PTR(-EINVAL);
289 /* Disable compression if suitable settings couldn't be found. */
290 if (nvbo->comp && !vmm->page[pi].comp) {
291 if (mmu->object.oclass >= NVIF_CLASS_MMU_GF100)
292 nvbo->kind = mmu->kind[nvbo->kind];
295 nvbo->page = vmm->page[pi].shift;
297 nouveau_bo_fixup_align(nvbo, align, size);
303 nouveau_bo_init(struct nouveau_bo *nvbo, u64 size, int align, u32 domain,
304 struct sg_table *sg, struct dma_resv *robj)
306 int type = sg ? ttm_bo_type_sg : ttm_bo_type_device;
309 nouveau_bo_placement_set(nvbo, domain, 0);
310 INIT_LIST_HEAD(&nvbo->io_reserve_lru);
312 ret = ttm_bo_init_validate(nvbo->bo.bdev, &nvbo->bo, type,
313 &nvbo->placement, align >> PAGE_SHIFT, false,
314 sg, robj, nouveau_bo_del_ttm);
316 /* ttm will call nouveau_bo_del_ttm if it fails.. */
324 nouveau_bo_new(struct nouveau_cli *cli, u64 size, int align,
325 uint32_t domain, uint32_t tile_mode, uint32_t tile_flags,
326 struct sg_table *sg, struct dma_resv *robj,
327 struct nouveau_bo **pnvbo)
329 struct nouveau_bo *nvbo;
332 nvbo = nouveau_bo_alloc(cli, &size, &align, domain, tile_mode,
335 return PTR_ERR(nvbo);
337 nvbo->bo.base.size = size;
338 dma_resv_init(&nvbo->bo.base._resv);
339 drm_vma_node_reset(&nvbo->bo.base.vma_node);
341 ret = nouveau_bo_init(nvbo, size, align, domain, sg, robj);
350 set_placement_list(struct ttm_place *pl, unsigned *n, uint32_t domain)
354 if (domain & NOUVEAU_GEM_DOMAIN_VRAM) {
355 pl[*n].mem_type = TTM_PL_VRAM;
359 if (domain & NOUVEAU_GEM_DOMAIN_GART) {
360 pl[*n].mem_type = TTM_PL_TT;
364 if (domain & NOUVEAU_GEM_DOMAIN_CPU) {
365 pl[*n].mem_type = TTM_PL_SYSTEM;
366 pl[(*n)++].flags = 0;
371 set_placement_range(struct nouveau_bo *nvbo, uint32_t domain)
373 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
374 u64 vram_size = drm->client.device.info.ram_size;
375 unsigned i, fpfn, lpfn;
377 if (drm->client.device.info.family == NV_DEVICE_INFO_V0_CELSIUS &&
378 nvbo->mode && (domain & NOUVEAU_GEM_DOMAIN_VRAM) &&
379 nvbo->bo.base.size < vram_size / 4) {
381 * Make sure that the color and depth buffers are handled
382 * by independent memory controller units. Up to a 9x
383 * speed up when alpha-blending and depth-test are enabled
387 fpfn = (vram_size / 2) >> PAGE_SHIFT;
391 lpfn = (vram_size / 2) >> PAGE_SHIFT;
393 for (i = 0; i < nvbo->placement.num_placement; ++i) {
394 nvbo->placements[i].fpfn = fpfn;
395 nvbo->placements[i].lpfn = lpfn;
397 for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
398 nvbo->busy_placements[i].fpfn = fpfn;
399 nvbo->busy_placements[i].lpfn = lpfn;
405 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t domain,
408 struct ttm_placement *pl = &nvbo->placement;
410 pl->placement = nvbo->placements;
411 set_placement_list(nvbo->placements, &pl->num_placement, domain);
413 pl->busy_placement = nvbo->busy_placements;
414 set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
417 set_placement_range(nvbo, domain);
421 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t domain, bool contig)
423 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
424 struct ttm_buffer_object *bo = &nvbo->bo;
425 bool force = false, evict = false;
428 ret = ttm_bo_reserve(bo, false, false, NULL);
432 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
433 domain == NOUVEAU_GEM_DOMAIN_VRAM && contig) {
441 if (nvbo->bo.pin_count) {
444 switch (bo->resource->mem_type) {
446 error |= !(domain & NOUVEAU_GEM_DOMAIN_VRAM);
449 error |= !(domain & NOUVEAU_GEM_DOMAIN_GART);
456 NV_ERROR(drm, "bo %p pinned elsewhere: "
457 "0x%08x vs 0x%08x\n", bo,
458 bo->resource->mem_type, domain);
461 ttm_bo_pin(&nvbo->bo);
466 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART, 0);
467 ret = nouveau_bo_validate(nvbo, false, false);
472 nouveau_bo_placement_set(nvbo, domain, 0);
473 ret = nouveau_bo_validate(nvbo, false, false);
477 ttm_bo_pin(&nvbo->bo);
479 switch (bo->resource->mem_type) {
481 drm->gem.vram_available -= bo->base.size;
484 drm->gem.gart_available -= bo->base.size;
492 nvbo->contig = false;
493 ttm_bo_unreserve(bo);
498 nouveau_bo_unpin(struct nouveau_bo *nvbo)
500 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
501 struct ttm_buffer_object *bo = &nvbo->bo;
504 ret = ttm_bo_reserve(bo, false, false, NULL);
508 ttm_bo_unpin(&nvbo->bo);
509 if (!nvbo->bo.pin_count) {
510 switch (bo->resource->mem_type) {
512 drm->gem.vram_available += bo->base.size;
515 drm->gem.gart_available += bo->base.size;
522 ttm_bo_unreserve(bo);
527 nouveau_bo_map(struct nouveau_bo *nvbo)
531 ret = ttm_bo_reserve(&nvbo->bo, false, false, NULL);
535 ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.resource->num_pages, &nvbo->kmap);
537 ttm_bo_unreserve(&nvbo->bo);
542 nouveau_bo_unmap(struct nouveau_bo *nvbo)
547 ttm_bo_kunmap(&nvbo->kmap);
551 nouveau_bo_sync_for_device(struct nouveau_bo *nvbo)
553 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
554 struct ttm_tt *ttm_dma = (struct ttm_tt *)nvbo->bo.ttm;
557 if (!ttm_dma || !ttm_dma->dma_address)
559 if (!ttm_dma->pages) {
560 NV_DEBUG(drm, "ttm_dma 0x%p: pages NULL\n", ttm_dma);
564 /* Don't waste time looping if the object is coherent */
565 if (nvbo->force_coherent)
569 while (i < ttm_dma->num_pages) {
570 struct page *p = ttm_dma->pages[i];
571 size_t num_pages = 1;
573 for (j = i + 1; j < ttm_dma->num_pages; ++j) {
574 if (++p != ttm_dma->pages[j])
579 dma_sync_single_for_device(drm->dev->dev,
580 ttm_dma->dma_address[i],
581 num_pages * PAGE_SIZE, DMA_TO_DEVICE);
587 nouveau_bo_sync_for_cpu(struct nouveau_bo *nvbo)
589 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
590 struct ttm_tt *ttm_dma = (struct ttm_tt *)nvbo->bo.ttm;
593 if (!ttm_dma || !ttm_dma->dma_address)
595 if (!ttm_dma->pages) {
596 NV_DEBUG(drm, "ttm_dma 0x%p: pages NULL\n", ttm_dma);
600 /* Don't waste time looping if the object is coherent */
601 if (nvbo->force_coherent)
605 while (i < ttm_dma->num_pages) {
606 struct page *p = ttm_dma->pages[i];
607 size_t num_pages = 1;
609 for (j = i + 1; j < ttm_dma->num_pages; ++j) {
610 if (++p != ttm_dma->pages[j])
616 dma_sync_single_for_cpu(drm->dev->dev, ttm_dma->dma_address[i],
617 num_pages * PAGE_SIZE, DMA_FROM_DEVICE);
622 void nouveau_bo_add_io_reserve_lru(struct ttm_buffer_object *bo)
624 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
625 struct nouveau_bo *nvbo = nouveau_bo(bo);
627 mutex_lock(&drm->ttm.io_reserve_mutex);
628 list_move_tail(&nvbo->io_reserve_lru, &drm->ttm.io_reserve_lru);
629 mutex_unlock(&drm->ttm.io_reserve_mutex);
632 void nouveau_bo_del_io_reserve_lru(struct ttm_buffer_object *bo)
634 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
635 struct nouveau_bo *nvbo = nouveau_bo(bo);
637 mutex_lock(&drm->ttm.io_reserve_mutex);
638 list_del_init(&nvbo->io_reserve_lru);
639 mutex_unlock(&drm->ttm.io_reserve_mutex);
643 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
646 struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
649 ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement, &ctx);
653 nouveau_bo_sync_for_device(nvbo);
659 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
662 u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
667 iowrite16_native(val, (void __force __iomem *)mem);
673 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
676 u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
681 return ioread32_native((void __force __iomem *)mem);
687 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
690 u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
695 iowrite32_native(val, (void __force __iomem *)mem);
700 static struct ttm_tt *
701 nouveau_ttm_tt_create(struct ttm_buffer_object *bo, uint32_t page_flags)
703 #if IS_ENABLED(CONFIG_AGP)
704 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
706 if (drm->agp.bridge) {
707 return ttm_agp_tt_create(bo, drm->agp.bridge, page_flags);
711 return nouveau_sgdma_create_ttm(bo, page_flags);
715 nouveau_ttm_tt_bind(struct ttm_device *bdev, struct ttm_tt *ttm,
716 struct ttm_resource *reg)
718 #if IS_ENABLED(CONFIG_AGP)
719 struct nouveau_drm *drm = nouveau_bdev(bdev);
723 #if IS_ENABLED(CONFIG_AGP)
725 return ttm_agp_bind(ttm, reg);
727 return nouveau_sgdma_bind(bdev, ttm, reg);
731 nouveau_ttm_tt_unbind(struct ttm_device *bdev, struct ttm_tt *ttm)
733 #if IS_ENABLED(CONFIG_AGP)
734 struct nouveau_drm *drm = nouveau_bdev(bdev);
736 if (drm->agp.bridge) {
741 nouveau_sgdma_unbind(bdev, ttm);
745 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
747 struct nouveau_bo *nvbo = nouveau_bo(bo);
749 switch (bo->resource->mem_type) {
751 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART,
752 NOUVEAU_GEM_DOMAIN_CPU);
755 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_CPU, 0);
759 *pl = nvbo->placement;
763 nouveau_bo_move_prep(struct nouveau_drm *drm, struct ttm_buffer_object *bo,
764 struct ttm_resource *reg)
766 struct nouveau_mem *old_mem = nouveau_mem(bo->resource);
767 struct nouveau_mem *new_mem = nouveau_mem(reg);
768 struct nvif_vmm *vmm = &drm->client.vmm.vmm;
771 ret = nvif_vmm_get(vmm, LAZY, false, old_mem->mem.page, 0,
772 old_mem->mem.size, &old_mem->vma[0]);
776 ret = nvif_vmm_get(vmm, LAZY, false, new_mem->mem.page, 0,
777 new_mem->mem.size, &old_mem->vma[1]);
781 ret = nouveau_mem_map(old_mem, vmm, &old_mem->vma[0]);
785 ret = nouveau_mem_map(new_mem, vmm, &old_mem->vma[1]);
788 nvif_vmm_put(vmm, &old_mem->vma[1]);
789 nvif_vmm_put(vmm, &old_mem->vma[0]);
795 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict,
796 struct ttm_operation_ctx *ctx,
797 struct ttm_resource *new_reg)
799 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
800 struct nouveau_channel *chan = drm->ttm.chan;
801 struct nouveau_cli *cli = (void *)chan->user.client;
802 struct nouveau_fence *fence;
805 /* create temporary vmas for the transfer and attach them to the
806 * old nvkm_mem node, these will get cleaned up after ttm has
807 * destroyed the ttm_resource
809 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
810 ret = nouveau_bo_move_prep(drm, bo, new_reg);
815 if (drm_drv_uses_atomic_modeset(drm->dev))
816 mutex_lock(&cli->mutex);
818 mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
819 ret = nouveau_fence_sync(nouveau_bo(bo), chan, true, ctx->interruptible);
821 ret = drm->ttm.move(chan, bo, bo->resource, new_reg);
823 ret = nouveau_fence_new(chan, false, &fence);
825 /* TODO: figure out a better solution here
827 * wait on the fence here explicitly as going through
828 * ttm_bo_move_accel_cleanup somehow doesn't seem to do it.
830 * Without this the operation can timeout and we'll fallback to a
831 * software copy, which might take several minutes to finish.
833 nouveau_fence_wait(fence, false, false);
834 ret = ttm_bo_move_accel_cleanup(bo,
838 nouveau_fence_unref(&fence);
842 mutex_unlock(&cli->mutex);
847 nouveau_bo_move_init(struct nouveau_drm *drm)
849 static const struct _method_table {
853 int (*exec)(struct nouveau_channel *,
854 struct ttm_buffer_object *,
855 struct ttm_resource *, struct ttm_resource *);
856 int (*init)(struct nouveau_channel *, u32 handle);
858 { "COPY", 4, 0xc7b5, nve0_bo_move_copy, nve0_bo_move_init },
859 { "COPY", 4, 0xc5b5, nve0_bo_move_copy, nve0_bo_move_init },
860 { "GRCE", 0, 0xc5b5, nve0_bo_move_copy, nvc0_bo_move_init },
861 { "COPY", 4, 0xc3b5, nve0_bo_move_copy, nve0_bo_move_init },
862 { "GRCE", 0, 0xc3b5, nve0_bo_move_copy, nvc0_bo_move_init },
863 { "COPY", 4, 0xc1b5, nve0_bo_move_copy, nve0_bo_move_init },
864 { "GRCE", 0, 0xc1b5, nve0_bo_move_copy, nvc0_bo_move_init },
865 { "COPY", 4, 0xc0b5, nve0_bo_move_copy, nve0_bo_move_init },
866 { "GRCE", 0, 0xc0b5, nve0_bo_move_copy, nvc0_bo_move_init },
867 { "COPY", 4, 0xb0b5, nve0_bo_move_copy, nve0_bo_move_init },
868 { "GRCE", 0, 0xb0b5, nve0_bo_move_copy, nvc0_bo_move_init },
869 { "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
870 { "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
871 { "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
872 { "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
873 { "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
874 { "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
875 { "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
876 { "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
877 { "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
880 const struct _method_table *mthd = _methods;
881 const char *name = "CPU";
885 struct nouveau_channel *chan;
894 ret = nvif_object_ctor(&chan->user, "ttmBoMove",
895 mthd->oclass | (mthd->engine << 16),
896 mthd->oclass, NULL, 0,
899 ret = mthd->init(chan, drm->ttm.copy.handle);
901 nvif_object_dtor(&drm->ttm.copy);
905 drm->ttm.move = mthd->exec;
906 drm->ttm.chan = chan;
910 } while ((++mthd)->exec);
912 NV_INFO(drm, "MM: using %s for buffer copies\n", name);
915 static void nouveau_bo_move_ntfy(struct ttm_buffer_object *bo,
916 struct ttm_resource *new_reg)
918 struct nouveau_mem *mem = new_reg ? nouveau_mem(new_reg) : NULL;
919 struct nouveau_bo *nvbo = nouveau_bo(bo);
920 struct nouveau_vma *vma;
922 /* ttm can now (stupidly) pass the driver bos it didn't create... */
923 if (bo->destroy != nouveau_bo_del_ttm)
926 nouveau_bo_del_io_reserve_lru(bo);
928 if (mem && new_reg->mem_type != TTM_PL_SYSTEM &&
929 mem->mem.page == nvbo->page) {
930 list_for_each_entry(vma, &nvbo->vma_list, head) {
931 nouveau_vma_map(vma, mem);
934 list_for_each_entry(vma, &nvbo->vma_list, head) {
935 WARN_ON(ttm_bo_wait(bo, false, false));
936 nouveau_vma_unmap(vma);
941 nvbo->offset = (new_reg->start << PAGE_SHIFT);
946 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_resource *new_reg,
947 struct nouveau_drm_tile **new_tile)
949 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
950 struct drm_device *dev = drm->dev;
951 struct nouveau_bo *nvbo = nouveau_bo(bo);
952 u64 offset = new_reg->start << PAGE_SHIFT;
955 if (new_reg->mem_type != TTM_PL_VRAM)
958 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_CELSIUS) {
959 *new_tile = nv10_bo_set_tiling(dev, offset, bo->base.size,
960 nvbo->mode, nvbo->zeta);
967 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
968 struct nouveau_drm_tile *new_tile,
969 struct nouveau_drm_tile **old_tile)
971 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
972 struct drm_device *dev = drm->dev;
973 struct dma_fence *fence;
976 ret = dma_resv_get_singleton(bo->base.resv, DMA_RESV_USAGE_WRITE,
979 dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_WRITE,
980 false, MAX_SCHEDULE_TIMEOUT);
982 nv10_bo_put_tile_region(dev, *old_tile, fence);
983 *old_tile = new_tile;
987 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict,
988 struct ttm_operation_ctx *ctx,
989 struct ttm_resource *new_reg,
990 struct ttm_place *hop)
992 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
993 struct nouveau_bo *nvbo = nouveau_bo(bo);
994 struct ttm_resource *old_reg = bo->resource;
995 struct nouveau_drm_tile *new_tile = NULL;
999 if (new_reg->mem_type == TTM_PL_TT) {
1000 ret = nouveau_ttm_tt_bind(bo->bdev, bo->ttm, new_reg);
1005 nouveau_bo_move_ntfy(bo, new_reg);
1006 ret = ttm_bo_wait_ctx(bo, ctx);
1010 if (nvbo->bo.pin_count)
1011 NV_WARN(drm, "Moving pinned object %p!\n", nvbo);
1013 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1014 ret = nouveau_bo_vm_bind(bo, new_reg, &new_tile);
1020 if (!old_reg || (old_reg->mem_type == TTM_PL_SYSTEM &&
1022 ttm_bo_move_null(bo, new_reg);
1026 if (old_reg->mem_type == TTM_PL_SYSTEM &&
1027 new_reg->mem_type == TTM_PL_TT) {
1028 ttm_bo_move_null(bo, new_reg);
1032 if (old_reg->mem_type == TTM_PL_TT &&
1033 new_reg->mem_type == TTM_PL_SYSTEM) {
1034 nouveau_ttm_tt_unbind(bo->bdev, bo->ttm);
1035 ttm_resource_free(bo, &bo->resource);
1036 ttm_bo_assign_mem(bo, new_reg);
1040 /* Hardware assisted copy. */
1041 if (drm->ttm.move) {
1042 if ((old_reg->mem_type == TTM_PL_SYSTEM &&
1043 new_reg->mem_type == TTM_PL_VRAM) ||
1044 (old_reg->mem_type == TTM_PL_VRAM &&
1045 new_reg->mem_type == TTM_PL_SYSTEM)) {
1048 hop->mem_type = TTM_PL_TT;
1052 ret = nouveau_bo_move_m2mf(bo, evict, ctx,
1058 /* Fallback to software copy. */
1059 ret = ttm_bo_move_memcpy(bo, ctx, new_reg);
1063 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1065 nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1067 nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1071 nouveau_bo_move_ntfy(bo, bo->resource);
1077 nouveau_ttm_io_mem_free_locked(struct nouveau_drm *drm,
1078 struct ttm_resource *reg)
1080 struct nouveau_mem *mem = nouveau_mem(reg);
1082 if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1083 switch (reg->mem_type) {
1086 nvif_object_unmap_handle(&mem->mem.object);
1089 nvif_object_unmap_handle(&mem->mem.object);
1098 nouveau_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *reg)
1100 struct nouveau_drm *drm = nouveau_bdev(bdev);
1101 struct nvkm_device *device = nvxx_device(&drm->client.device);
1102 struct nouveau_mem *mem = nouveau_mem(reg);
1103 struct nvif_mmu *mmu = &drm->client.mmu;
1106 mutex_lock(&drm->ttm.io_reserve_mutex);
1108 switch (reg->mem_type) {
1114 #if IS_ENABLED(CONFIG_AGP)
1115 if (drm->agp.bridge) {
1116 reg->bus.offset = (reg->start << PAGE_SHIFT) +
1118 reg->bus.is_iomem = !drm->agp.cma;
1119 reg->bus.caching = ttm_write_combined;
1122 if (drm->client.mem->oclass < NVIF_CLASS_MEM_NV50 ||
1128 fallthrough; /* tiled memory */
1130 reg->bus.offset = (reg->start << PAGE_SHIFT) +
1131 device->func->resource_addr(device, 1);
1132 reg->bus.is_iomem = true;
1134 /* Some BARs do not support being ioremapped WC */
1135 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
1136 mmu->type[drm->ttm.type_vram].type & NVIF_MEM_UNCACHED)
1137 reg->bus.caching = ttm_uncached;
1139 reg->bus.caching = ttm_write_combined;
1141 if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1143 struct nv50_mem_map_v0 nv50;
1144 struct gf100_mem_map_v0 gf100;
1149 switch (mem->mem.object.oclass) {
1150 case NVIF_CLASS_MEM_NV50:
1151 args.nv50.version = 0;
1153 args.nv50.kind = mem->kind;
1154 args.nv50.comp = mem->comp;
1155 argc = sizeof(args.nv50);
1157 case NVIF_CLASS_MEM_GF100:
1158 args.gf100.version = 0;
1160 args.gf100.kind = mem->kind;
1161 argc = sizeof(args.gf100);
1168 ret = nvif_object_map_handle(&mem->mem.object,
1172 if (WARN_ON(ret == 0))
1177 reg->bus.offset = handle;
1186 if (ret == -ENOSPC) {
1187 struct nouveau_bo *nvbo;
1189 nvbo = list_first_entry_or_null(&drm->ttm.io_reserve_lru,
1193 list_del_init(&nvbo->io_reserve_lru);
1194 drm_vma_node_unmap(&nvbo->bo.base.vma_node,
1196 nouveau_ttm_io_mem_free_locked(drm, nvbo->bo.resource);
1201 mutex_unlock(&drm->ttm.io_reserve_mutex);
1206 nouveau_ttm_io_mem_free(struct ttm_device *bdev, struct ttm_resource *reg)
1208 struct nouveau_drm *drm = nouveau_bdev(bdev);
1210 mutex_lock(&drm->ttm.io_reserve_mutex);
1211 nouveau_ttm_io_mem_free_locked(drm, reg);
1212 mutex_unlock(&drm->ttm.io_reserve_mutex);
1215 vm_fault_t nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1217 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1218 struct nouveau_bo *nvbo = nouveau_bo(bo);
1219 struct nvkm_device *device = nvxx_device(&drm->client.device);
1220 u32 mappable = device->func->resource_size(device, 1) >> PAGE_SHIFT;
1223 /* as long as the bo isn't in vram, and isn't tiled, we've got
1224 * nothing to do here.
1226 if (bo->resource->mem_type != TTM_PL_VRAM) {
1227 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA ||
1231 if (bo->resource->mem_type != TTM_PL_SYSTEM)
1234 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART, 0);
1237 /* make sure bo is in mappable vram */
1238 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA ||
1239 bo->resource->start + bo->resource->num_pages < mappable)
1242 for (i = 0; i < nvbo->placement.num_placement; ++i) {
1243 nvbo->placements[i].fpfn = 0;
1244 nvbo->placements[i].lpfn = mappable;
1247 for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
1248 nvbo->busy_placements[i].fpfn = 0;
1249 nvbo->busy_placements[i].lpfn = mappable;
1252 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_VRAM, 0);
1255 ret = nouveau_bo_validate(nvbo, false, false);
1256 if (unlikely(ret == -EBUSY || ret == -ERESTARTSYS))
1257 return VM_FAULT_NOPAGE;
1258 else if (unlikely(ret))
1259 return VM_FAULT_SIGBUS;
1261 ttm_bo_move_to_lru_tail_unlocked(bo);
1266 nouveau_ttm_tt_populate(struct ttm_device *bdev,
1267 struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
1269 struct ttm_tt *ttm_dma = (void *)ttm;
1270 struct nouveau_drm *drm;
1271 bool slave = !!(ttm->page_flags & TTM_TT_FLAG_EXTERNAL);
1273 if (ttm_tt_is_populated(ttm))
1276 if (slave && ttm->sg) {
1277 drm_prime_sg_to_dma_addr_array(ttm->sg, ttm_dma->dma_address,
1282 drm = nouveau_bdev(bdev);
1284 return ttm_pool_alloc(&drm->ttm.bdev.pool, ttm, ctx);
1288 nouveau_ttm_tt_unpopulate(struct ttm_device *bdev,
1291 struct nouveau_drm *drm;
1292 bool slave = !!(ttm->page_flags & TTM_TT_FLAG_EXTERNAL);
1297 nouveau_ttm_tt_unbind(bdev, ttm);
1299 drm = nouveau_bdev(bdev);
1301 return ttm_pool_free(&drm->ttm.bdev.pool, ttm);
1305 nouveau_ttm_tt_destroy(struct ttm_device *bdev,
1308 #if IS_ENABLED(CONFIG_AGP)
1309 struct nouveau_drm *drm = nouveau_bdev(bdev);
1310 if (drm->agp.bridge) {
1311 ttm_agp_destroy(ttm);
1315 nouveau_sgdma_destroy(bdev, ttm);
1319 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence, bool exclusive)
1321 struct dma_resv *resv = nvbo->bo.base.resv;
1326 dma_resv_add_fence(resv, &fence->base, exclusive ?
1327 DMA_RESV_USAGE_WRITE : DMA_RESV_USAGE_READ);
1331 nouveau_bo_delete_mem_notify(struct ttm_buffer_object *bo)
1333 nouveau_bo_move_ntfy(bo, NULL);
1336 struct ttm_device_funcs nouveau_bo_driver = {
1337 .ttm_tt_create = &nouveau_ttm_tt_create,
1338 .ttm_tt_populate = &nouveau_ttm_tt_populate,
1339 .ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1340 .ttm_tt_destroy = &nouveau_ttm_tt_destroy,
1341 .eviction_valuable = ttm_bo_eviction_valuable,
1342 .evict_flags = nouveau_bo_evict_flags,
1343 .delete_mem_notify = nouveau_bo_delete_mem_notify,
1344 .move = nouveau_bo_move,
1345 .io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1346 .io_mem_free = &nouveau_ttm_io_mem_free,