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 nouveau_bo_del_io_reserve_lru(bo);
141 nv10_bo_put_tile_region(dev, nvbo->tile, NULL);
144 * If nouveau_bo_new() allocated this buffer, the GEM object was never
145 * initialized, so don't attempt to release it.
148 drm_gem_object_release(&bo->base);
154 roundup_64(u64 x, u32 y)
162 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, 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 domain,
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 (domain & NOUVEAU_GEM_DOMAIN_COHERENT) {
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 (domain & NOUVEAU_GEM_DOMAIN_VRAM) && !vmm->page[i].vram)
263 if ((domain & NOUVEAU_GEM_DOMAIN_GART) &&
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, align, size);
296 nouveau_bo_init(struct nouveau_bo *nvbo, u64 size, int align, u32 domain,
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, domain, 0);
307 INIT_LIST_HEAD(&nvbo->io_reserve_lru);
309 ret = ttm_bo_init(nvbo->bo.bdev, &nvbo->bo, size, type,
310 &nvbo->placement, align >> PAGE_SHIFT, false,
311 acc_size, sg, robj, nouveau_bo_del_ttm);
313 /* ttm will call nouveau_bo_del_ttm if it fails.. */
321 nouveau_bo_new(struct nouveau_cli *cli, u64 size, int align,
322 uint32_t domain, uint32_t tile_mode, uint32_t tile_flags,
323 struct sg_table *sg, struct dma_resv *robj,
324 struct nouveau_bo **pnvbo)
326 struct nouveau_bo *nvbo;
329 nvbo = nouveau_bo_alloc(cli, &size, &align, domain, tile_mode,
332 return PTR_ERR(nvbo);
334 ret = nouveau_bo_init(nvbo, size, align, domain, sg, robj);
343 set_placement_list(struct nouveau_drm *drm, struct ttm_place *pl, unsigned *n,
344 uint32_t domain, uint32_t flags)
348 if (domain & NOUVEAU_GEM_DOMAIN_VRAM) {
349 struct nvif_mmu *mmu = &drm->client.mmu;
350 const u8 type = mmu->type[drm->ttm.type_vram].type;
352 pl[*n].mem_type = TTM_PL_VRAM;
353 pl[*n].flags = flags & ~TTM_PL_FLAG_CACHED;
355 /* Some BARs do not support being ioremapped WC */
356 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
357 type & NVIF_MEM_UNCACHED)
358 pl[*n].flags &= ~TTM_PL_FLAG_WC;
362 if (domain & NOUVEAU_GEM_DOMAIN_GART) {
363 pl[*n].mem_type = TTM_PL_TT;
364 pl[*n].flags = flags;
367 pl[*n].flags &= ~TTM_PL_FLAG_CACHED;
371 if (domain & NOUVEAU_GEM_DOMAIN_CPU) {
372 pl[*n].mem_type = TTM_PL_SYSTEM;
373 pl[(*n)++].flags = flags;
378 set_placement_range(struct nouveau_bo *nvbo, uint32_t domain)
380 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
381 u32 vram_pages = drm->client.device.info.ram_size >> PAGE_SHIFT;
382 unsigned i, fpfn, lpfn;
384 if (drm->client.device.info.family == NV_DEVICE_INFO_V0_CELSIUS &&
385 nvbo->mode && (domain & NOUVEAU_GEM_DOMAIN_VRAM) &&
386 nvbo->bo.mem.num_pages < vram_pages / 4) {
388 * Make sure that the color and depth buffers are handled
389 * by independent memory controller units. Up to a 9x
390 * speed up when alpha-blending and depth-test are enabled
394 fpfn = vram_pages / 2;
398 lpfn = vram_pages / 2;
400 for (i = 0; i < nvbo->placement.num_placement; ++i) {
401 nvbo->placements[i].fpfn = fpfn;
402 nvbo->placements[i].lpfn = lpfn;
404 for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
405 nvbo->busy_placements[i].fpfn = fpfn;
406 nvbo->busy_placements[i].lpfn = lpfn;
412 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t domain,
415 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
416 struct ttm_placement *pl = &nvbo->placement;
417 uint32_t flags = (nvbo->force_coherent ? TTM_PL_FLAG_UNCACHED :
418 TTM_PL_MASK_CACHING) |
419 (nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
421 pl->placement = nvbo->placements;
422 set_placement_list(drm, nvbo->placements, &pl->num_placement,
425 pl->busy_placement = nvbo->busy_placements;
426 set_placement_list(drm, nvbo->busy_placements, &pl->num_busy_placement,
427 domain | busy, flags);
429 set_placement_range(nvbo, domain);
433 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t domain, bool contig)
435 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
436 struct ttm_buffer_object *bo = &nvbo->bo;
437 bool force = false, evict = false;
440 ret = ttm_bo_reserve(bo, false, false, NULL);
444 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
445 domain == NOUVEAU_GEM_DOMAIN_VRAM && contig) {
453 if (nvbo->pin_refcnt) {
456 switch (bo->mem.mem_type) {
458 error |= !(domain & NOUVEAU_GEM_DOMAIN_VRAM);
461 error |= !(domain & NOUVEAU_GEM_DOMAIN_GART);
467 NV_ERROR(drm, "bo %p pinned elsewhere: "
468 "0x%08x vs 0x%08x\n", bo,
469 bo->mem.mem_type, domain);
477 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART, 0);
478 ret = nouveau_bo_validate(nvbo, false, false);
484 nouveau_bo_placement_set(nvbo, domain, 0);
486 /* drop pin_refcnt temporarily, so we don't trip the assertion
487 * in nouveau_bo_move() that makes sure we're not trying to
488 * move a pinned buffer
491 ret = nouveau_bo_validate(nvbo, false, false);
496 switch (bo->mem.mem_type) {
498 drm->gem.vram_available -= bo->mem.size;
501 drm->gem.gart_available -= bo->mem.size;
509 nvbo->contig = false;
510 ttm_bo_unreserve(bo);
515 nouveau_bo_unpin(struct nouveau_bo *nvbo)
517 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
518 struct ttm_buffer_object *bo = &nvbo->bo;
521 ret = ttm_bo_reserve(bo, false, false, NULL);
525 ref = --nvbo->pin_refcnt;
526 WARN_ON_ONCE(ref < 0);
530 switch (bo->mem.mem_type) {
532 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_VRAM, 0);
535 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART, 0);
541 ret = nouveau_bo_validate(nvbo, false, false);
543 switch (bo->mem.mem_type) {
545 drm->gem.vram_available += bo->mem.size;
548 drm->gem.gart_available += bo->mem.size;
556 ttm_bo_unreserve(bo);
561 nouveau_bo_map(struct nouveau_bo *nvbo)
565 ret = ttm_bo_reserve(&nvbo->bo, false, false, NULL);
569 ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
571 ttm_bo_unreserve(&nvbo->bo);
576 nouveau_bo_unmap(struct nouveau_bo *nvbo)
581 ttm_bo_kunmap(&nvbo->kmap);
585 nouveau_bo_sync_for_device(struct nouveau_bo *nvbo)
587 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
588 struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
594 /* Don't waste time looping if the object is coherent */
595 if (nvbo->force_coherent)
598 for (i = 0; i < ttm_dma->ttm.num_pages; i++)
599 dma_sync_single_for_device(drm->dev->dev,
600 ttm_dma->dma_address[i],
601 PAGE_SIZE, DMA_TO_DEVICE);
605 nouveau_bo_sync_for_cpu(struct nouveau_bo *nvbo)
607 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
608 struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
614 /* Don't waste time looping if the object is coherent */
615 if (nvbo->force_coherent)
618 for (i = 0; i < ttm_dma->ttm.num_pages; i++)
619 dma_sync_single_for_cpu(drm->dev->dev, ttm_dma->dma_address[i],
620 PAGE_SIZE, DMA_FROM_DEVICE);
623 void nouveau_bo_add_io_reserve_lru(struct ttm_buffer_object *bo)
625 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
626 struct nouveau_bo *nvbo = nouveau_bo(bo);
628 mutex_lock(&drm->ttm.io_reserve_mutex);
629 list_move_tail(&nvbo->io_reserve_lru, &drm->ttm.io_reserve_lru);
630 mutex_unlock(&drm->ttm.io_reserve_mutex);
633 void nouveau_bo_del_io_reserve_lru(struct ttm_buffer_object *bo)
635 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
636 struct nouveau_bo *nvbo = nouveau_bo(bo);
638 mutex_lock(&drm->ttm.io_reserve_mutex);
639 list_del_init(&nvbo->io_reserve_lru);
640 mutex_unlock(&drm->ttm.io_reserve_mutex);
644 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
647 struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
650 ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement, &ctx);
654 nouveau_bo_sync_for_device(nvbo);
660 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
663 u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
668 iowrite16_native(val, (void __force __iomem *)mem);
674 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
677 u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
682 return ioread32_native((void __force __iomem *)mem);
688 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
691 u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
696 iowrite32_native(val, (void __force __iomem *)mem);
701 static struct ttm_tt *
702 nouveau_ttm_tt_create(struct ttm_buffer_object *bo, uint32_t page_flags)
704 #if IS_ENABLED(CONFIG_AGP)
705 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
707 if (drm->agp.bridge) {
708 return ttm_agp_tt_create(bo, drm->agp.bridge, page_flags);
712 return nouveau_sgdma_create_ttm(bo, page_flags);
716 nouveau_ttm_tt_bind(struct ttm_bo_device *bdev, struct ttm_tt *ttm,
717 struct ttm_resource *reg)
719 #if IS_ENABLED(CONFIG_AGP)
720 struct nouveau_drm *drm = nouveau_bdev(bdev);
724 #if IS_ENABLED(CONFIG_AGP)
726 return ttm_agp_bind(ttm, reg);
728 return nouveau_sgdma_bind(bdev, ttm, reg);
732 nouveau_ttm_tt_unbind(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
734 #if IS_ENABLED(CONFIG_AGP)
735 struct nouveau_drm *drm = nouveau_bdev(bdev);
737 if (drm->agp.bridge) {
742 nouveau_sgdma_unbind(bdev, ttm);
746 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
748 struct nouveau_bo *nvbo = nouveau_bo(bo);
750 switch (bo->mem.mem_type) {
752 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART,
753 NOUVEAU_GEM_DOMAIN_CPU);
756 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_CPU, 0);
760 *pl = nvbo->placement;
764 nouveau_bo_move_prep(struct nouveau_drm *drm, struct ttm_buffer_object *bo,
765 struct ttm_resource *reg)
767 struct nouveau_mem *old_mem = nouveau_mem(&bo->mem);
768 struct nouveau_mem *new_mem = nouveau_mem(reg);
769 struct nvif_vmm *vmm = &drm->client.vmm.vmm;
772 ret = nvif_vmm_get(vmm, LAZY, false, old_mem->mem.page, 0,
773 old_mem->mem.size, &old_mem->vma[0]);
777 ret = nvif_vmm_get(vmm, LAZY, false, new_mem->mem.page, 0,
778 new_mem->mem.size, &old_mem->vma[1]);
782 ret = nouveau_mem_map(old_mem, vmm, &old_mem->vma[0]);
786 ret = nouveau_mem_map(new_mem, vmm, &old_mem->vma[1]);
789 nvif_vmm_put(vmm, &old_mem->vma[1]);
790 nvif_vmm_put(vmm, &old_mem->vma[0]);
796 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
797 bool no_wait_gpu, 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 mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
816 ret = nouveau_fence_sync(nouveau_bo(bo), chan, true, intr);
818 ret = drm->ttm.move(chan, bo, &bo->mem, new_reg);
820 ret = nouveau_fence_new(chan, false, &fence);
822 ret = ttm_bo_move_accel_cleanup(bo,
826 nouveau_fence_unref(&fence);
830 mutex_unlock(&cli->mutex);
835 nouveau_bo_move_init(struct nouveau_drm *drm)
837 static const struct _method_table {
841 int (*exec)(struct nouveau_channel *,
842 struct ttm_buffer_object *,
843 struct ttm_resource *, struct ttm_resource *);
844 int (*init)(struct nouveau_channel *, u32 handle);
846 { "COPY", 4, 0xc5b5, nve0_bo_move_copy, nve0_bo_move_init },
847 { "GRCE", 0, 0xc5b5, nve0_bo_move_copy, nvc0_bo_move_init },
848 { "COPY", 4, 0xc3b5, nve0_bo_move_copy, nve0_bo_move_init },
849 { "GRCE", 0, 0xc3b5, nve0_bo_move_copy, nvc0_bo_move_init },
850 { "COPY", 4, 0xc1b5, nve0_bo_move_copy, nve0_bo_move_init },
851 { "GRCE", 0, 0xc1b5, nve0_bo_move_copy, nvc0_bo_move_init },
852 { "COPY", 4, 0xc0b5, nve0_bo_move_copy, nve0_bo_move_init },
853 { "GRCE", 0, 0xc0b5, nve0_bo_move_copy, nvc0_bo_move_init },
854 { "COPY", 4, 0xb0b5, nve0_bo_move_copy, nve0_bo_move_init },
855 { "GRCE", 0, 0xb0b5, nve0_bo_move_copy, nvc0_bo_move_init },
856 { "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
857 { "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
858 { "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
859 { "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
860 { "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
861 { "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
862 { "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
863 { "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
864 { "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
867 const struct _method_table *mthd = _methods;
868 const char *name = "CPU";
872 struct nouveau_channel *chan;
881 ret = nvif_object_ctor(&chan->user, "ttmBoMove",
882 mthd->oclass | (mthd->engine << 16),
883 mthd->oclass, NULL, 0,
886 ret = mthd->init(chan, drm->ttm.copy.handle);
888 nvif_object_dtor(&drm->ttm.copy);
892 drm->ttm.move = mthd->exec;
893 drm->ttm.chan = chan;
897 } while ((++mthd)->exec);
899 NV_INFO(drm, "MM: using %s for buffer copies\n", name);
903 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
904 bool no_wait_gpu, struct ttm_resource *new_reg)
906 struct ttm_operation_ctx ctx = { intr, no_wait_gpu };
907 struct ttm_place placement_memtype = {
910 .mem_type = TTM_PL_TT,
911 .flags = TTM_PL_MASK_CACHING
913 struct ttm_placement placement;
914 struct ttm_resource tmp_reg;
917 placement.num_placement = placement.num_busy_placement = 1;
918 placement.placement = placement.busy_placement = &placement_memtype;
921 tmp_reg.mm_node = NULL;
922 ret = ttm_bo_mem_space(bo, &placement, &tmp_reg, &ctx);
926 ret = ttm_tt_populate(bo->bdev, bo->ttm, &ctx);
930 ret = ttm_bo_tt_bind(bo, &tmp_reg);
934 ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, &tmp_reg);
938 ret = ttm_bo_move_ttm(bo, &ctx, new_reg);
940 ttm_resource_free(bo, &tmp_reg);
945 nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
946 bool no_wait_gpu, struct ttm_resource *new_reg)
948 struct ttm_operation_ctx ctx = { intr, no_wait_gpu };
949 struct ttm_place placement_memtype = {
952 .mem_type = TTM_PL_TT,
953 .flags = TTM_PL_MASK_CACHING
955 struct ttm_placement placement;
956 struct ttm_resource tmp_reg;
959 placement.num_placement = placement.num_busy_placement = 1;
960 placement.placement = placement.busy_placement = &placement_memtype;
963 tmp_reg.mm_node = NULL;
964 ret = ttm_bo_mem_space(bo, &placement, &tmp_reg, &ctx);
968 ret = ttm_bo_move_ttm(bo, &ctx, &tmp_reg);
972 ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, new_reg);
977 ttm_resource_free(bo, &tmp_reg);
982 nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, bool evict,
983 struct ttm_resource *new_reg)
985 struct nouveau_mem *mem = new_reg ? nouveau_mem(new_reg) : NULL;
986 struct nouveau_bo *nvbo = nouveau_bo(bo);
987 struct nouveau_vma *vma;
989 /* ttm can now (stupidly) pass the driver bos it didn't create... */
990 if (bo->destroy != nouveau_bo_del_ttm)
993 nouveau_bo_del_io_reserve_lru(bo);
995 if (mem && new_reg->mem_type != TTM_PL_SYSTEM &&
996 mem->mem.page == nvbo->page) {
997 list_for_each_entry(vma, &nvbo->vma_list, head) {
998 nouveau_vma_map(vma, mem);
1001 list_for_each_entry(vma, &nvbo->vma_list, head) {
1002 WARN_ON(ttm_bo_wait(bo, false, false));
1003 nouveau_vma_unmap(vma);
1008 if (new_reg->mm_node)
1009 nvbo->offset = (new_reg->start << PAGE_SHIFT);
1017 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_resource *new_reg,
1018 struct nouveau_drm_tile **new_tile)
1020 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1021 struct drm_device *dev = drm->dev;
1022 struct nouveau_bo *nvbo = nouveau_bo(bo);
1023 u64 offset = new_reg->start << PAGE_SHIFT;
1026 if (new_reg->mem_type != TTM_PL_VRAM)
1029 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_CELSIUS) {
1030 *new_tile = nv10_bo_set_tiling(dev, offset, new_reg->size,
1031 nvbo->mode, nvbo->zeta);
1038 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
1039 struct nouveau_drm_tile *new_tile,
1040 struct nouveau_drm_tile **old_tile)
1042 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1043 struct drm_device *dev = drm->dev;
1044 struct dma_fence *fence = dma_resv_get_excl(bo->base.resv);
1046 nv10_bo_put_tile_region(dev, *old_tile, fence);
1047 *old_tile = new_tile;
1051 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict,
1052 struct ttm_operation_ctx *ctx,
1053 struct ttm_resource *new_reg)
1055 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1056 struct nouveau_bo *nvbo = nouveau_bo(bo);
1057 struct ttm_resource *old_reg = &bo->mem;
1058 struct nouveau_drm_tile *new_tile = NULL;
1061 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
1065 if (nvbo->pin_refcnt)
1066 NV_WARN(drm, "Moving pinned object %p!\n", nvbo);
1068 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1069 ret = nouveau_bo_vm_bind(bo, new_reg, &new_tile);
1075 if (old_reg->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
1076 ttm_bo_move_null(bo, new_reg);
1080 /* Hardware assisted copy. */
1081 if (drm->ttm.move) {
1082 if (new_reg->mem_type == TTM_PL_SYSTEM)
1083 ret = nouveau_bo_move_flipd(bo, evict,
1085 ctx->no_wait_gpu, new_reg);
1086 else if (old_reg->mem_type == TTM_PL_SYSTEM)
1087 ret = nouveau_bo_move_flips(bo, evict,
1089 ctx->no_wait_gpu, new_reg);
1091 ret = nouveau_bo_move_m2mf(bo, evict,
1093 ctx->no_wait_gpu, new_reg);
1098 /* Fallback to software copy. */
1099 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
1101 ret = ttm_bo_move_memcpy(bo, ctx, new_reg);
1104 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1106 nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1108 nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1115 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
1117 struct nouveau_bo *nvbo = nouveau_bo(bo);
1119 return drm_vma_node_verify_access(&nvbo->bo.base.vma_node,
1120 filp->private_data);
1124 nouveau_ttm_io_mem_free_locked(struct nouveau_drm *drm,
1125 struct ttm_resource *reg)
1127 struct nouveau_mem *mem = nouveau_mem(reg);
1129 if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1130 switch (reg->mem_type) {
1133 nvif_object_unmap_handle(&mem->mem.object);
1136 nvif_object_unmap_handle(&mem->mem.object);
1145 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_resource *reg)
1147 struct nouveau_drm *drm = nouveau_bdev(bdev);
1148 struct nvkm_device *device = nvxx_device(&drm->client.device);
1149 struct nouveau_mem *mem = nouveau_mem(reg);
1152 mutex_lock(&drm->ttm.io_reserve_mutex);
1154 switch (reg->mem_type) {
1160 #if IS_ENABLED(CONFIG_AGP)
1161 if (drm->agp.bridge) {
1162 reg->bus.offset = (reg->start << PAGE_SHIFT) +
1164 reg->bus.is_iomem = !drm->agp.cma;
1167 if (drm->client.mem->oclass < NVIF_CLASS_MEM_NV50 ||
1173 fallthrough; /* tiled memory */
1175 reg->bus.offset = (reg->start << PAGE_SHIFT) +
1176 device->func->resource_addr(device, 1);
1177 reg->bus.is_iomem = true;
1178 if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1180 struct nv50_mem_map_v0 nv50;
1181 struct gf100_mem_map_v0 gf100;
1186 switch (mem->mem.object.oclass) {
1187 case NVIF_CLASS_MEM_NV50:
1188 args.nv50.version = 0;
1190 args.nv50.kind = mem->kind;
1191 args.nv50.comp = mem->comp;
1192 argc = sizeof(args.nv50);
1194 case NVIF_CLASS_MEM_GF100:
1195 args.gf100.version = 0;
1197 args.gf100.kind = mem->kind;
1198 argc = sizeof(args.gf100);
1205 ret = nvif_object_map_handle(&mem->mem.object,
1209 if (WARN_ON(ret == 0))
1214 reg->bus.offset = handle;
1223 if (ret == -ENOSPC) {
1224 struct nouveau_bo *nvbo;
1226 nvbo = list_first_entry_or_null(&drm->ttm.io_reserve_lru,
1230 list_del_init(&nvbo->io_reserve_lru);
1231 drm_vma_node_unmap(&nvbo->bo.base.vma_node,
1233 nouveau_ttm_io_mem_free_locked(drm, &nvbo->bo.mem);
1238 mutex_unlock(&drm->ttm.io_reserve_mutex);
1243 nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_resource *reg)
1245 struct nouveau_drm *drm = nouveau_bdev(bdev);
1247 mutex_lock(&drm->ttm.io_reserve_mutex);
1248 nouveau_ttm_io_mem_free_locked(drm, reg);
1249 mutex_unlock(&drm->ttm.io_reserve_mutex);
1253 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1255 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1256 struct nouveau_bo *nvbo = nouveau_bo(bo);
1257 struct nvkm_device *device = nvxx_device(&drm->client.device);
1258 u32 mappable = device->func->resource_size(device, 1) >> PAGE_SHIFT;
1261 /* as long as the bo isn't in vram, and isn't tiled, we've got
1262 * nothing to do here.
1264 if (bo->mem.mem_type != TTM_PL_VRAM) {
1265 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA ||
1269 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
1270 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART,
1273 ret = nouveau_bo_validate(nvbo, false, false);
1280 /* make sure bo is in mappable vram */
1281 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA ||
1282 bo->mem.start + bo->mem.num_pages < mappable)
1285 for (i = 0; i < nvbo->placement.num_placement; ++i) {
1286 nvbo->placements[i].fpfn = 0;
1287 nvbo->placements[i].lpfn = mappable;
1290 for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
1291 nvbo->busy_placements[i].fpfn = 0;
1292 nvbo->busy_placements[i].lpfn = mappable;
1295 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_VRAM, 0);
1296 return nouveau_bo_validate(nvbo, false, false);
1300 nouveau_ttm_tt_populate(struct ttm_bo_device *bdev,
1301 struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
1303 struct ttm_dma_tt *ttm_dma = (void *)ttm;
1304 struct nouveau_drm *drm;
1306 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1308 if (ttm_tt_is_populated(ttm))
1311 if (slave && ttm->sg) {
1312 /* make userspace faulting work */
1313 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1314 ttm_dma->dma_address, ttm->num_pages);
1315 ttm_tt_set_populated(ttm);
1319 drm = nouveau_bdev(bdev);
1320 dev = drm->dev->dev;
1322 #if IS_ENABLED(CONFIG_AGP)
1323 if (drm->agp.bridge) {
1324 return ttm_pool_populate(ttm, ctx);
1328 #if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1329 if (swiotlb_nr_tbl()) {
1330 return ttm_dma_populate((void *)ttm, dev, ctx);
1333 return ttm_populate_and_map_pages(dev, ttm_dma, ctx);
1337 nouveau_ttm_tt_unpopulate(struct ttm_bo_device *bdev,
1340 struct ttm_dma_tt *ttm_dma = (void *)ttm;
1341 struct nouveau_drm *drm;
1343 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1348 drm = nouveau_bdev(bdev);
1349 dev = drm->dev->dev;
1351 #if IS_ENABLED(CONFIG_AGP)
1352 if (drm->agp.bridge) {
1353 ttm_pool_unpopulate(ttm);
1358 #if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1359 if (swiotlb_nr_tbl()) {
1360 ttm_dma_unpopulate((void *)ttm, dev);
1365 ttm_unmap_and_unpopulate_pages(dev, ttm_dma);
1369 nouveau_ttm_tt_destroy(struct ttm_bo_device *bdev,
1372 #if IS_ENABLED(CONFIG_AGP)
1373 struct nouveau_drm *drm = nouveau_bdev(bdev);
1374 if (drm->agp.bridge) {
1375 ttm_tt_destroy_common(bdev, ttm);
1376 ttm_agp_destroy(ttm);
1380 nouveau_sgdma_destroy(bdev, ttm);
1384 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence, bool exclusive)
1386 struct dma_resv *resv = nvbo->bo.base.resv;
1389 dma_resv_add_excl_fence(resv, &fence->base);
1391 dma_resv_add_shared_fence(resv, &fence->base);
1394 struct ttm_bo_driver nouveau_bo_driver = {
1395 .ttm_tt_create = &nouveau_ttm_tt_create,
1396 .ttm_tt_populate = &nouveau_ttm_tt_populate,
1397 .ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1398 .ttm_tt_bind = &nouveau_ttm_tt_bind,
1399 .ttm_tt_unbind = &nouveau_ttm_tt_unbind,
1400 .ttm_tt_destroy = &nouveau_ttm_tt_destroy,
1401 .eviction_valuable = ttm_bo_eviction_valuable,
1402 .evict_flags = nouveau_bo_evict_flags,
1403 .move_notify = nouveau_bo_move_ntfy,
1404 .move = nouveau_bo_move,
1405 .verify_access = nouveau_bo_verify_access,
1406 .fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1407 .io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1408 .io_mem_free = &nouveau_ttm_io_mem_free,