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, u32 flags,
163 int *align, u64 *size)
165 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
166 struct nvif_device *device = &drm->client.device;
168 if (device->info.family < NV_DEVICE_INFO_V0_TESLA) {
170 if (device->info.chipset >= 0x40) {
172 *size = roundup_64(*size, 64 * nvbo->mode);
174 } else if (device->info.chipset >= 0x30) {
176 *size = roundup_64(*size, 64 * nvbo->mode);
178 } else if (device->info.chipset >= 0x20) {
180 *size = roundup_64(*size, 64 * nvbo->mode);
182 } else if (device->info.chipset >= 0x10) {
184 *size = roundup_64(*size, 32 * nvbo->mode);
188 *size = roundup_64(*size, (1 << nvbo->page));
189 *align = max((1 << nvbo->page), *align);
192 *size = roundup_64(*size, PAGE_SIZE);
196 nouveau_bo_alloc(struct nouveau_cli *cli, u64 *size, int *align, u32 flags,
197 u32 tile_mode, u32 tile_flags)
199 struct nouveau_drm *drm = cli->drm;
200 struct nouveau_bo *nvbo;
201 struct nvif_mmu *mmu = &cli->mmu;
202 struct nvif_vmm *vmm = cli->svm.cli ? &cli->svm.vmm : &cli->vmm.vmm;
206 NV_WARN(drm, "skipped size %016llx\n", *size);
207 return ERR_PTR(-EINVAL);
210 nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
212 return ERR_PTR(-ENOMEM);
213 INIT_LIST_HEAD(&nvbo->head);
214 INIT_LIST_HEAD(&nvbo->entry);
215 INIT_LIST_HEAD(&nvbo->vma_list);
216 nvbo->bo.bdev = &drm->ttm.bdev;
218 /* This is confusing, and doesn't actually mean we want an uncached
219 * mapping, but is what NOUVEAU_GEM_DOMAIN_COHERENT gets translated
220 * into in nouveau_gem_new().
222 if (flags & TTM_PL_FLAG_UNCACHED) {
223 /* Determine if we can get a cache-coherent map, forcing
224 * uncached mapping if we can't.
226 if (!nouveau_drm_use_coherent_gpu_mapping(drm))
227 nvbo->force_coherent = true;
230 if (cli->device.info.family >= NV_DEVICE_INFO_V0_FERMI) {
231 nvbo->kind = (tile_flags & 0x0000ff00) >> 8;
232 if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
234 return ERR_PTR(-EINVAL);
237 nvbo->comp = mmu->kind[nvbo->kind] != nvbo->kind;
239 if (cli->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
240 nvbo->kind = (tile_flags & 0x00007f00) >> 8;
241 nvbo->comp = (tile_flags & 0x00030000) >> 16;
242 if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
244 return ERR_PTR(-EINVAL);
247 nvbo->zeta = (tile_flags & 0x00000007);
249 nvbo->mode = tile_mode;
250 nvbo->contig = !(tile_flags & NOUVEAU_GEM_TILE_NONCONTIG);
252 /* Determine the desirable target GPU page size for the buffer. */
253 for (i = 0; i < vmm->page_nr; i++) {
254 /* Because we cannot currently allow VMM maps to fail
255 * during buffer migration, we need to determine page
256 * size for the buffer up-front, and pre-allocate its
259 * Skip page sizes that can't support needed domains.
261 if (cli->device.info.family > NV_DEVICE_INFO_V0_CURIE &&
262 (flags & TTM_PL_FLAG_VRAM) && !vmm->page[i].vram)
264 if ((flags & TTM_PL_FLAG_TT) &&
265 (!vmm->page[i].host || vmm->page[i].shift > PAGE_SHIFT))
268 /* Select this page size if it's the first that supports
269 * the potential memory domains, or when it's compatible
270 * with the requested compression settings.
272 if (pi < 0 || !nvbo->comp || vmm->page[i].comp)
275 /* Stop once the buffer is larger than the current page size. */
276 if (*size >= 1ULL << vmm->page[i].shift)
281 return ERR_PTR(-EINVAL);
283 /* Disable compression if suitable settings couldn't be found. */
284 if (nvbo->comp && !vmm->page[pi].comp) {
285 if (mmu->object.oclass >= NVIF_CLASS_MMU_GF100)
286 nvbo->kind = mmu->kind[nvbo->kind];
289 nvbo->page = vmm->page[pi].shift;
291 nouveau_bo_fixup_align(nvbo, flags, align, size);
297 nouveau_bo_init(struct nouveau_bo *nvbo, u64 size, int align, u32 flags,
298 struct sg_table *sg, struct dma_resv *robj)
300 int type = sg ? ttm_bo_type_sg : ttm_bo_type_device;
304 acc_size = ttm_bo_dma_acc_size(nvbo->bo.bdev, size, sizeof(*nvbo));
306 nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
307 nouveau_bo_placement_set(nvbo, flags, 0);
308 INIT_LIST_HEAD(&nvbo->io_reserve_lru);
310 ret = ttm_bo_init(nvbo->bo.bdev, &nvbo->bo, size, type,
311 &nvbo->placement, align >> PAGE_SHIFT, false,
312 acc_size, sg, robj, nouveau_bo_del_ttm);
314 /* ttm will call nouveau_bo_del_ttm if it fails.. */
322 nouveau_bo_new(struct nouveau_cli *cli, u64 size, int align,
323 uint32_t flags, uint32_t tile_mode, uint32_t tile_flags,
324 struct sg_table *sg, struct dma_resv *robj,
325 struct nouveau_bo **pnvbo)
327 struct nouveau_bo *nvbo;
330 nvbo = nouveau_bo_alloc(cli, &size, &align, flags, tile_mode,
333 return PTR_ERR(nvbo);
335 ret = nouveau_bo_init(nvbo, size, align, flags, sg, robj);
344 set_placement_list(struct ttm_place *pl, unsigned *n, uint32_t type, uint32_t flags)
348 if (type & TTM_PL_FLAG_VRAM)
349 pl[(*n)++].flags = TTM_PL_FLAG_VRAM | flags;
350 if (type & TTM_PL_FLAG_TT)
351 pl[(*n)++].flags = TTM_PL_FLAG_TT | flags;
352 if (type & TTM_PL_FLAG_SYSTEM)
353 pl[(*n)++].flags = TTM_PL_FLAG_SYSTEM | flags;
357 set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
359 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
360 u32 vram_pages = drm->client.device.info.ram_size >> PAGE_SHIFT;
361 unsigned i, fpfn, lpfn;
363 if (drm->client.device.info.family == NV_DEVICE_INFO_V0_CELSIUS &&
364 nvbo->mode && (type & TTM_PL_FLAG_VRAM) &&
365 nvbo->bo.mem.num_pages < vram_pages / 4) {
367 * Make sure that the color and depth buffers are handled
368 * by independent memory controller units. Up to a 9x
369 * speed up when alpha-blending and depth-test are enabled
373 fpfn = vram_pages / 2;
377 lpfn = vram_pages / 2;
379 for (i = 0; i < nvbo->placement.num_placement; ++i) {
380 nvbo->placements[i].fpfn = fpfn;
381 nvbo->placements[i].lpfn = lpfn;
383 for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
384 nvbo->busy_placements[i].fpfn = fpfn;
385 nvbo->busy_placements[i].lpfn = lpfn;
391 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
393 struct ttm_placement *pl = &nvbo->placement;
394 uint32_t flags = (nvbo->force_coherent ? TTM_PL_FLAG_UNCACHED :
395 TTM_PL_MASK_CACHING) |
396 (nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
398 pl->placement = nvbo->placements;
399 set_placement_list(nvbo->placements, &pl->num_placement,
402 pl->busy_placement = nvbo->busy_placements;
403 set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
406 set_placement_range(nvbo, type);
410 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype, bool contig)
412 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
413 struct ttm_buffer_object *bo = &nvbo->bo;
414 bool force = false, evict = false;
417 ret = ttm_bo_reserve(bo, false, false, NULL);
421 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
422 memtype == TTM_PL_FLAG_VRAM && contig) {
430 if (nvbo->pin_refcnt) {
431 if (!(memtype & (1 << bo->mem.mem_type)) || evict) {
432 NV_ERROR(drm, "bo %p pinned elsewhere: "
433 "0x%08x vs 0x%08x\n", bo,
434 1 << bo->mem.mem_type, memtype);
442 nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT, 0);
443 ret = nouveau_bo_validate(nvbo, false, false);
449 nouveau_bo_placement_set(nvbo, memtype, 0);
451 /* drop pin_refcnt temporarily, so we don't trip the assertion
452 * in nouveau_bo_move() that makes sure we're not trying to
453 * move a pinned buffer
456 ret = nouveau_bo_validate(nvbo, false, false);
461 switch (bo->mem.mem_type) {
463 drm->gem.vram_available -= bo->mem.size;
466 drm->gem.gart_available -= bo->mem.size;
474 nvbo->contig = false;
475 ttm_bo_unreserve(bo);
480 nouveau_bo_unpin(struct nouveau_bo *nvbo)
482 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
483 struct ttm_buffer_object *bo = &nvbo->bo;
486 ret = ttm_bo_reserve(bo, false, false, NULL);
490 ref = --nvbo->pin_refcnt;
491 WARN_ON_ONCE(ref < 0);
495 nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
497 ret = nouveau_bo_validate(nvbo, false, false);
499 switch (bo->mem.mem_type) {
501 drm->gem.vram_available += bo->mem.size;
504 drm->gem.gart_available += bo->mem.size;
512 ttm_bo_unreserve(bo);
517 nouveau_bo_map(struct nouveau_bo *nvbo)
521 ret = ttm_bo_reserve(&nvbo->bo, false, false, NULL);
525 ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
527 ttm_bo_unreserve(&nvbo->bo);
532 nouveau_bo_unmap(struct nouveau_bo *nvbo)
537 ttm_bo_kunmap(&nvbo->kmap);
541 nouveau_bo_sync_for_device(struct nouveau_bo *nvbo)
543 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
544 struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
550 /* Don't waste time looping if the object is coherent */
551 if (nvbo->force_coherent)
554 for (i = 0; i < ttm_dma->ttm.num_pages; i++)
555 dma_sync_single_for_device(drm->dev->dev,
556 ttm_dma->dma_address[i],
557 PAGE_SIZE, DMA_TO_DEVICE);
561 nouveau_bo_sync_for_cpu(struct nouveau_bo *nvbo)
563 struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
564 struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
570 /* Don't waste time looping if the object is coherent */
571 if (nvbo->force_coherent)
574 for (i = 0; i < ttm_dma->ttm.num_pages; i++)
575 dma_sync_single_for_cpu(drm->dev->dev, ttm_dma->dma_address[i],
576 PAGE_SIZE, DMA_FROM_DEVICE);
579 void nouveau_bo_add_io_reserve_lru(struct ttm_buffer_object *bo)
581 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
582 struct nouveau_bo *nvbo = nouveau_bo(bo);
584 mutex_lock(&drm->ttm.io_reserve_mutex);
585 list_move_tail(&nvbo->io_reserve_lru, &drm->ttm.io_reserve_lru);
586 mutex_unlock(&drm->ttm.io_reserve_mutex);
589 void nouveau_bo_del_io_reserve_lru(struct ttm_buffer_object *bo)
591 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
592 struct nouveau_bo *nvbo = nouveau_bo(bo);
594 mutex_lock(&drm->ttm.io_reserve_mutex);
595 list_del_init(&nvbo->io_reserve_lru);
596 mutex_unlock(&drm->ttm.io_reserve_mutex);
600 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
603 struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
606 ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement, &ctx);
610 nouveau_bo_sync_for_device(nvbo);
616 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
619 u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
624 iowrite16_native(val, (void __force __iomem *)mem);
630 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
633 u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
638 return ioread32_native((void __force __iomem *)mem);
644 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
647 u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
652 iowrite32_native(val, (void __force __iomem *)mem);
657 static struct ttm_tt *
658 nouveau_ttm_tt_create(struct ttm_buffer_object *bo, uint32_t page_flags)
660 #if IS_ENABLED(CONFIG_AGP)
661 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
663 if (drm->agp.bridge) {
664 return ttm_agp_tt_create(bo, drm->agp.bridge, page_flags);
668 return nouveau_sgdma_create_ttm(bo, page_flags);
672 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
674 struct nouveau_bo *nvbo = nouveau_bo(bo);
676 switch (bo->mem.mem_type) {
678 nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
682 nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
686 *pl = nvbo->placement;
690 nouveau_bo_move_prep(struct nouveau_drm *drm, struct ttm_buffer_object *bo,
691 struct ttm_resource *reg)
693 struct nouveau_mem *old_mem = nouveau_mem(&bo->mem);
694 struct nouveau_mem *new_mem = nouveau_mem(reg);
695 struct nvif_vmm *vmm = &drm->client.vmm.vmm;
698 ret = nvif_vmm_get(vmm, LAZY, false, old_mem->mem.page, 0,
699 old_mem->mem.size, &old_mem->vma[0]);
703 ret = nvif_vmm_get(vmm, LAZY, false, new_mem->mem.page, 0,
704 new_mem->mem.size, &old_mem->vma[1]);
708 ret = nouveau_mem_map(old_mem, vmm, &old_mem->vma[0]);
712 ret = nouveau_mem_map(new_mem, vmm, &old_mem->vma[1]);
715 nvif_vmm_put(vmm, &old_mem->vma[1]);
716 nvif_vmm_put(vmm, &old_mem->vma[0]);
722 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
723 bool no_wait_gpu, struct ttm_resource *new_reg)
725 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
726 struct nouveau_channel *chan = drm->ttm.chan;
727 struct nouveau_cli *cli = (void *)chan->user.client;
728 struct nouveau_fence *fence;
731 /* create temporary vmas for the transfer and attach them to the
732 * old nvkm_mem node, these will get cleaned up after ttm has
733 * destroyed the ttm_resource
735 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
736 ret = nouveau_bo_move_prep(drm, bo, new_reg);
741 mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
742 ret = nouveau_fence_sync(nouveau_bo(bo), chan, true, intr);
744 ret = drm->ttm.move(chan, bo, &bo->mem, new_reg);
746 ret = nouveau_fence_new(chan, false, &fence);
748 ret = ttm_bo_move_accel_cleanup(bo,
752 nouveau_fence_unref(&fence);
756 mutex_unlock(&cli->mutex);
761 nouveau_bo_move_init(struct nouveau_drm *drm)
763 static const struct _method_table {
767 int (*exec)(struct nouveau_channel *,
768 struct ttm_buffer_object *,
769 struct ttm_resource *, struct ttm_resource *);
770 int (*init)(struct nouveau_channel *, u32 handle);
772 { "COPY", 4, 0xc5b5, nve0_bo_move_copy, nve0_bo_move_init },
773 { "GRCE", 0, 0xc5b5, nve0_bo_move_copy, nvc0_bo_move_init },
774 { "COPY", 4, 0xc3b5, nve0_bo_move_copy, nve0_bo_move_init },
775 { "GRCE", 0, 0xc3b5, nve0_bo_move_copy, nvc0_bo_move_init },
776 { "COPY", 4, 0xc1b5, nve0_bo_move_copy, nve0_bo_move_init },
777 { "GRCE", 0, 0xc1b5, nve0_bo_move_copy, nvc0_bo_move_init },
778 { "COPY", 4, 0xc0b5, nve0_bo_move_copy, nve0_bo_move_init },
779 { "GRCE", 0, 0xc0b5, nve0_bo_move_copy, nvc0_bo_move_init },
780 { "COPY", 4, 0xb0b5, nve0_bo_move_copy, nve0_bo_move_init },
781 { "GRCE", 0, 0xb0b5, nve0_bo_move_copy, nvc0_bo_move_init },
782 { "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
783 { "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
784 { "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
785 { "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
786 { "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
787 { "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
788 { "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
789 { "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
790 { "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
793 const struct _method_table *mthd = _methods;
794 const char *name = "CPU";
798 struct nouveau_channel *chan;
807 ret = nvif_object_ctor(&chan->user, "ttmBoMove",
808 mthd->oclass | (mthd->engine << 16),
809 mthd->oclass, NULL, 0,
812 ret = mthd->init(chan, drm->ttm.copy.handle);
814 nvif_object_dtor(&drm->ttm.copy);
818 drm->ttm.move = mthd->exec;
819 drm->ttm.chan = chan;
823 } while ((++mthd)->exec);
825 NV_INFO(drm, "MM: using %s for buffer copies\n", name);
829 nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
830 bool no_wait_gpu, struct ttm_resource *new_reg)
832 struct ttm_operation_ctx ctx = { intr, no_wait_gpu };
833 struct ttm_place placement_memtype = {
836 .flags = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING
838 struct ttm_placement placement;
839 struct ttm_resource tmp_reg;
842 placement.num_placement = placement.num_busy_placement = 1;
843 placement.placement = placement.busy_placement = &placement_memtype;
846 tmp_reg.mm_node = NULL;
847 ret = ttm_bo_mem_space(bo, &placement, &tmp_reg, &ctx);
851 ret = ttm_tt_bind(bo->bdev, bo->ttm, &tmp_reg, &ctx);
855 ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, &tmp_reg);
859 ret = ttm_bo_move_ttm(bo, &ctx, new_reg);
861 ttm_resource_free(bo, &tmp_reg);
866 nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
867 bool no_wait_gpu, struct ttm_resource *new_reg)
869 struct ttm_operation_ctx ctx = { intr, no_wait_gpu };
870 struct ttm_place placement_memtype = {
873 .flags = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING
875 struct ttm_placement placement;
876 struct ttm_resource tmp_reg;
879 placement.num_placement = placement.num_busy_placement = 1;
880 placement.placement = placement.busy_placement = &placement_memtype;
883 tmp_reg.mm_node = NULL;
884 ret = ttm_bo_mem_space(bo, &placement, &tmp_reg, &ctx);
888 ret = ttm_bo_move_ttm(bo, &ctx, &tmp_reg);
892 ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, new_reg);
897 ttm_resource_free(bo, &tmp_reg);
902 nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, bool evict,
903 struct ttm_resource *new_reg)
905 struct nouveau_mem *mem = new_reg ? nouveau_mem(new_reg) : NULL;
906 struct nouveau_bo *nvbo = nouveau_bo(bo);
907 struct nouveau_vma *vma;
909 /* ttm can now (stupidly) pass the driver bos it didn't create... */
910 if (bo->destroy != nouveau_bo_del_ttm)
913 nouveau_bo_del_io_reserve_lru(bo);
915 if (mem && new_reg->mem_type != TTM_PL_SYSTEM &&
916 mem->mem.page == nvbo->page) {
917 list_for_each_entry(vma, &nvbo->vma_list, head) {
918 nouveau_vma_map(vma, mem);
921 list_for_each_entry(vma, &nvbo->vma_list, head) {
922 WARN_ON(ttm_bo_wait(bo, false, false));
923 nouveau_vma_unmap(vma);
928 if (new_reg->mm_node)
929 nvbo->offset = (new_reg->start << PAGE_SHIFT);
937 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_resource *new_reg,
938 struct nouveau_drm_tile **new_tile)
940 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
941 struct drm_device *dev = drm->dev;
942 struct nouveau_bo *nvbo = nouveau_bo(bo);
943 u64 offset = new_reg->start << PAGE_SHIFT;
946 if (new_reg->mem_type != TTM_PL_VRAM)
949 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_CELSIUS) {
950 *new_tile = nv10_bo_set_tiling(dev, offset, new_reg->size,
951 nvbo->mode, nvbo->zeta);
958 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
959 struct nouveau_drm_tile *new_tile,
960 struct nouveau_drm_tile **old_tile)
962 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
963 struct drm_device *dev = drm->dev;
964 struct dma_fence *fence = dma_resv_get_excl(bo->base.resv);
966 nv10_bo_put_tile_region(dev, *old_tile, fence);
967 *old_tile = new_tile;
971 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict,
972 struct ttm_operation_ctx *ctx,
973 struct ttm_resource *new_reg)
975 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
976 struct nouveau_bo *nvbo = nouveau_bo(bo);
977 struct ttm_resource *old_reg = &bo->mem;
978 struct nouveau_drm_tile *new_tile = NULL;
981 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
985 if (nvbo->pin_refcnt)
986 NV_WARN(drm, "Moving pinned object %p!\n", nvbo);
988 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
989 ret = nouveau_bo_vm_bind(bo, new_reg, &new_tile);
995 if (old_reg->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
996 ttm_bo_move_null(bo, new_reg);
1000 /* Hardware assisted copy. */
1001 if (drm->ttm.move) {
1002 if (new_reg->mem_type == TTM_PL_SYSTEM)
1003 ret = nouveau_bo_move_flipd(bo, evict,
1005 ctx->no_wait_gpu, new_reg);
1006 else if (old_reg->mem_type == TTM_PL_SYSTEM)
1007 ret = nouveau_bo_move_flips(bo, evict,
1009 ctx->no_wait_gpu, new_reg);
1011 ret = nouveau_bo_move_m2mf(bo, evict,
1013 ctx->no_wait_gpu, new_reg);
1018 /* Fallback to software copy. */
1019 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
1021 ret = ttm_bo_move_memcpy(bo, ctx, new_reg);
1024 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1026 nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1028 nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1035 nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
1037 struct nouveau_bo *nvbo = nouveau_bo(bo);
1039 return drm_vma_node_verify_access(&nvbo->bo.base.vma_node,
1040 filp->private_data);
1044 nouveau_ttm_io_mem_free_locked(struct nouveau_drm *drm,
1045 struct ttm_resource *reg)
1047 struct nouveau_mem *mem = nouveau_mem(reg);
1049 if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1050 switch (reg->mem_type) {
1053 nvif_object_unmap_handle(&mem->mem.object);
1056 nvif_object_unmap_handle(&mem->mem.object);
1065 nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_resource *reg)
1067 struct nouveau_drm *drm = nouveau_bdev(bdev);
1068 struct nvkm_device *device = nvxx_device(&drm->client.device);
1069 struct nouveau_mem *mem = nouveau_mem(reg);
1072 mutex_lock(&drm->ttm.io_reserve_mutex);
1074 switch (reg->mem_type) {
1080 #if IS_ENABLED(CONFIG_AGP)
1081 if (drm->agp.bridge) {
1082 reg->bus.offset = (reg->start << PAGE_SHIFT) +
1084 reg->bus.is_iomem = !drm->agp.cma;
1087 if (drm->client.mem->oclass < NVIF_CLASS_MEM_NV50 ||
1093 fallthrough; /* tiled memory */
1095 reg->bus.offset = (reg->start << PAGE_SHIFT) +
1096 device->func->resource_addr(device, 1);
1097 reg->bus.is_iomem = true;
1098 if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1100 struct nv50_mem_map_v0 nv50;
1101 struct gf100_mem_map_v0 gf100;
1106 switch (mem->mem.object.oclass) {
1107 case NVIF_CLASS_MEM_NV50:
1108 args.nv50.version = 0;
1110 args.nv50.kind = mem->kind;
1111 args.nv50.comp = mem->comp;
1112 argc = sizeof(args.nv50);
1114 case NVIF_CLASS_MEM_GF100:
1115 args.gf100.version = 0;
1117 args.gf100.kind = mem->kind;
1118 argc = sizeof(args.gf100);
1125 ret = nvif_object_map_handle(&mem->mem.object,
1129 if (WARN_ON(ret == 0))
1134 reg->bus.offset = handle;
1143 if (ret == -ENOSPC) {
1144 struct nouveau_bo *nvbo;
1146 nvbo = list_first_entry_or_null(&drm->ttm.io_reserve_lru,
1150 list_del_init(&nvbo->io_reserve_lru);
1151 drm_vma_node_unmap(&nvbo->bo.base.vma_node,
1153 nouveau_ttm_io_mem_free_locked(drm, &nvbo->bo.mem);
1158 mutex_unlock(&drm->ttm.io_reserve_mutex);
1163 nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_resource *reg)
1165 struct nouveau_drm *drm = nouveau_bdev(bdev);
1167 mutex_lock(&drm->ttm.io_reserve_mutex);
1168 nouveau_ttm_io_mem_free_locked(drm, reg);
1169 mutex_unlock(&drm->ttm.io_reserve_mutex);
1173 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1175 struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1176 struct nouveau_bo *nvbo = nouveau_bo(bo);
1177 struct nvkm_device *device = nvxx_device(&drm->client.device);
1178 u32 mappable = device->func->resource_size(device, 1) >> PAGE_SHIFT;
1181 /* as long as the bo isn't in vram, and isn't tiled, we've got
1182 * nothing to do here.
1184 if (bo->mem.mem_type != TTM_PL_VRAM) {
1185 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA ||
1189 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
1190 nouveau_bo_placement_set(nvbo, TTM_PL_TT, 0);
1192 ret = nouveau_bo_validate(nvbo, false, false);
1199 /* make sure bo is in mappable vram */
1200 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA ||
1201 bo->mem.start + bo->mem.num_pages < mappable)
1204 for (i = 0; i < nvbo->placement.num_placement; ++i) {
1205 nvbo->placements[i].fpfn = 0;
1206 nvbo->placements[i].lpfn = mappable;
1209 for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
1210 nvbo->busy_placements[i].fpfn = 0;
1211 nvbo->busy_placements[i].lpfn = mappable;
1214 nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_VRAM, 0);
1215 return nouveau_bo_validate(nvbo, false, false);
1219 nouveau_ttm_tt_populate(struct ttm_bo_device *bdev,
1220 struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
1222 struct ttm_dma_tt *ttm_dma = (void *)ttm;
1223 struct nouveau_drm *drm;
1225 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1227 if (ttm->state != tt_unpopulated)
1230 if (slave && ttm->sg) {
1231 /* make userspace faulting work */
1232 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1233 ttm_dma->dma_address, ttm->num_pages);
1234 ttm->state = tt_unbound;
1238 drm = nouveau_bdev(bdev);
1239 dev = drm->dev->dev;
1241 #if IS_ENABLED(CONFIG_AGP)
1242 if (drm->agp.bridge) {
1243 return ttm_agp_tt_populate(bdev, ttm, ctx);
1247 #if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1248 if (swiotlb_nr_tbl()) {
1249 return ttm_dma_populate((void *)ttm, dev, ctx);
1252 return ttm_populate_and_map_pages(dev, ttm_dma, ctx);
1256 nouveau_ttm_tt_unpopulate(struct ttm_bo_device *bdev,
1259 struct ttm_dma_tt *ttm_dma = (void *)ttm;
1260 struct nouveau_drm *drm;
1262 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1267 drm = nouveau_bdev(bdev);
1268 dev = drm->dev->dev;
1270 #if IS_ENABLED(CONFIG_AGP)
1271 if (drm->agp.bridge) {
1272 ttm_agp_tt_unpopulate(bdev, ttm);
1277 #if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1278 if (swiotlb_nr_tbl()) {
1279 ttm_dma_unpopulate((void *)ttm, dev);
1284 ttm_unmap_and_unpopulate_pages(dev, ttm_dma);
1288 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence, bool exclusive)
1290 struct dma_resv *resv = nvbo->bo.base.resv;
1293 dma_resv_add_excl_fence(resv, &fence->base);
1295 dma_resv_add_shared_fence(resv, &fence->base);
1298 struct ttm_bo_driver nouveau_bo_driver = {
1299 .ttm_tt_create = &nouveau_ttm_tt_create,
1300 .ttm_tt_populate = &nouveau_ttm_tt_populate,
1301 .ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1302 .eviction_valuable = ttm_bo_eviction_valuable,
1303 .evict_flags = nouveau_bo_evict_flags,
1304 .move_notify = nouveau_bo_move_ntfy,
1305 .move = nouveau_bo_move,
1306 .verify_access = nouveau_bo_verify_access,
1307 .fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1308 .io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1309 .io_mem_free = &nouveau_ttm_io_mem_free,