diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c | 193 |
1 files changed, 150 insertions, 43 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c index 0c001bb8fc2b..e22cfa7c6d32 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c @@ -36,12 +36,35 @@ #include "amdgpu_gem.h" #include "amdgpu_dma_buf.h" #include "amdgpu_xgmi.h" +#include "amdgpu_vm.h" #include <drm/amdgpu_drm.h> #include <drm/ttm/ttm_tt.h> #include <linux/dma-buf.h> #include <linux/dma-fence-array.h> #include <linux/pci-p2pdma.h> -#include <linux/pm_runtime.h> + +static const struct dma_buf_attach_ops amdgpu_dma_buf_attach_ops; + +/** + * dma_buf_attach_adev - Helper to get adev of an attachment + * + * @attach: attachment + * + * Returns: + * A struct amdgpu_device * if the attaching device is an amdgpu device or + * partition, NULL otherwise. + */ +static struct amdgpu_device *dma_buf_attach_adev(struct dma_buf_attachment *attach) +{ + if (attach->importer_ops == &amdgpu_dma_buf_attach_ops) { + struct drm_gem_object *obj = attach->importer_priv; + struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj); + + return amdgpu_ttm_adev(bo->tbo.bdev); + } + + return NULL; +} /** * amdgpu_dma_buf_attach - &dma_buf_ops.attach implementation @@ -54,42 +77,49 @@ static int amdgpu_dma_buf_attach(struct dma_buf *dmabuf, struct dma_buf_attachment *attach) { + struct amdgpu_device *attach_adev = dma_buf_attach_adev(attach); struct drm_gem_object *obj = dmabuf->priv; struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj); struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); int r; - if (pci_p2pdma_distance(adev->pdev, attach->dev, false) < 0) + /* + * Disable peer-to-peer access for DCC-enabled VRAM surfaces on GFX12+. + * Such buffers cannot be safely accessed over P2P due to device-local + * compression metadata. Fallback to system-memory path instead. + * Device supports GFX12 (GC 12.x or newer) + * BO was created with the AMDGPU_GEM_CREATE_GFX12_DCC flag + * + */ + if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(12, 0, 0) && + bo->flags & AMDGPU_GEM_CREATE_GFX12_DCC) attach->peer2peer = false; - r = pm_runtime_get_sync(adev_to_drm(adev)->dev); - if (r < 0) - goto out; + /* + * Disable peer-to-peer access for DCC-enabled VRAM surfaces on GFX12+. + * Such buffers cannot be safely accessed over P2P due to device-local + * compression metadata. Fallback to system-memory path instead. + * Device supports GFX12 (GC 12.x or newer) + * BO was created with the AMDGPU_GEM_CREATE_GFX12_DCC flag + * + */ + if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(12, 0, 0) && + bo->flags & AMDGPU_GEM_CREATE_GFX12_DCC) + attach->peer2peer = false; - return 0; + if (!amdgpu_dmabuf_is_xgmi_accessible(attach_adev, bo) && + pci_p2pdma_distance(adev->pdev, attach->dev, false) < 0) + attach->peer2peer = false; -out: - pm_runtime_put_autosuspend(adev_to_drm(adev)->dev); - return r; -} + r = dma_resv_lock(bo->tbo.base.resv, NULL); + if (r) + return r; -/** - * amdgpu_dma_buf_detach - &dma_buf_ops.detach implementation - * - * @dmabuf: DMA-buf where we remove the attachment from - * @attach: the attachment to remove - * - * Called when an attachment is removed from the DMA-buf. - */ -static void amdgpu_dma_buf_detach(struct dma_buf *dmabuf, - struct dma_buf_attachment *attach) -{ - struct drm_gem_object *obj = dmabuf->priv; - struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj); - struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); + amdgpu_vm_bo_update_shared(bo); - pm_runtime_mark_last_busy(adev_to_drm(adev)->dev); - pm_runtime_put_autosuspend(adev_to_drm(adev)->dev); + dma_resv_unlock(bo->tbo.base.resv); + + return 0; } /** @@ -101,11 +131,35 @@ static void amdgpu_dma_buf_detach(struct dma_buf *dmabuf, */ static int amdgpu_dma_buf_pin(struct dma_buf_attachment *attach) { - struct drm_gem_object *obj = attach->dmabuf->priv; - struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj); + struct dma_buf *dmabuf = attach->dmabuf; + struct amdgpu_bo *bo = gem_to_amdgpu_bo(dmabuf->priv); + u32 domains = bo->allowed_domains; + + dma_resv_assert_held(dmabuf->resv); + + /* Try pinning into VRAM to allow P2P with RDMA NICs without ODP + * support if all attachments can do P2P. If any attachment can't do + * P2P just pin into GTT instead. + * + * To avoid with conflicting pinnings between GPUs and RDMA when move + * notifiers are disabled, only allow pinning in VRAM when move + * notiers are enabled. + */ + if (!IS_ENABLED(CONFIG_DMABUF_MOVE_NOTIFY)) { + domains &= ~AMDGPU_GEM_DOMAIN_VRAM; + } else { + list_for_each_entry(attach, &dmabuf->attachments, node) + if (!attach->peer2peer) + domains &= ~AMDGPU_GEM_DOMAIN_VRAM; + } + + if (domains & AMDGPU_GEM_DOMAIN_VRAM) + bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED; - /* pin buffer into GTT */ - return amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT); + if (WARN_ON(!domains)) + return -EINVAL; + + return amdgpu_bo_pin(bo, domains); } /** @@ -149,7 +203,7 @@ static struct sg_table *amdgpu_dma_buf_map(struct dma_buf_attachment *attach, if (!bo->tbo.pin_count) { /* move buffer into GTT or VRAM */ struct ttm_operation_ctx ctx = { false, false }; - unsigned domains = AMDGPU_GEM_DOMAIN_GTT; + unsigned int domains = AMDGPU_GEM_DOMAIN_GTT; if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM && attach->peer2peer) { @@ -160,10 +214,6 @@ static struct sg_table *amdgpu_dma_buf_map(struct dma_buf_attachment *attach, r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); if (r) return ERR_PTR(r); - - } else if (!(amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type) & - AMDGPU_GEM_DOMAIN_GTT)) { - return ERR_PTR(-EBUSY); } switch (bo->tbo.resource->mem_type) { @@ -180,6 +230,11 @@ static struct sg_table *amdgpu_dma_buf_map(struct dma_buf_attachment *attach, break; case TTM_PL_VRAM: + /* XGMI-accessible memory should never be DMA-mapped */ + if (WARN_ON(amdgpu_dmabuf_is_xgmi_accessible( + dma_buf_attach_adev(attach), bo))) + return ERR_PTR(-EINVAL); + r = amdgpu_vram_mgr_alloc_sgt(adev, bo->tbo.resource, 0, bo->tbo.base.size, attach->dev, dir, &sgt); @@ -211,7 +266,7 @@ static void amdgpu_dma_buf_unmap(struct dma_buf_attachment *attach, struct sg_table *sgt, enum dma_data_direction dir) { - if (sgt->sgl->page_link) { + if (sg_page(sgt->sgl)) { dma_unmap_sgtable(attach->dev, sgt, dir, 0); sg_free_table(sgt); kfree(sgt); @@ -261,9 +316,38 @@ static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf, return ret; } +static int amdgpu_dma_buf_vmap(struct dma_buf *dma_buf, struct iosys_map *map) +{ + struct drm_gem_object *obj = dma_buf->priv; + struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj); + int ret; + + /* + * Pin to keep buffer in place while it's vmap'ed. The actual + * domain is not that important as long as it's mapable. Using + * GTT and VRAM should be compatible with most use cases. + */ + ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT | AMDGPU_GEM_DOMAIN_VRAM); + if (ret) + return ret; + ret = drm_gem_dmabuf_vmap(dma_buf, map); + if (ret) + amdgpu_bo_unpin(bo); + + return ret; +} + +static void amdgpu_dma_buf_vunmap(struct dma_buf *dma_buf, struct iosys_map *map) +{ + struct drm_gem_object *obj = dma_buf->priv; + struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj); + + drm_gem_dmabuf_vunmap(dma_buf, map); + amdgpu_bo_unpin(bo); +} + const struct dma_buf_ops amdgpu_dmabuf_ops = { .attach = amdgpu_dma_buf_attach, - .detach = amdgpu_dma_buf_detach, .pin = amdgpu_dma_buf_pin, .unpin = amdgpu_dma_buf_unpin, .map_dma_buf = amdgpu_dma_buf_map, @@ -271,8 +355,8 @@ const struct dma_buf_ops amdgpu_dmabuf_ops = { .release = drm_gem_dmabuf_release, .begin_cpu_access = amdgpu_dma_buf_begin_cpu_access, .mmap = drm_gem_dmabuf_mmap, - .vmap = drm_gem_dmabuf_vmap, - .vunmap = drm_gem_dmabuf_vunmap, + .vmap = amdgpu_dma_buf_vmap, + .vunmap = amdgpu_dma_buf_vunmap, }; /** @@ -290,11 +374,23 @@ struct dma_buf *amdgpu_gem_prime_export(struct drm_gem_object *gobj, { struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj); struct dma_buf *buf; + struct ttm_operation_ctx ctx = { + .interruptible = true, + .no_wait_gpu = true, + /* We opt to avoid OOM on system pages allocations */ + .gfp_retry_mayfail = true, + .allow_res_evict = false, + }; + int ret; if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) || bo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID) return ERR_PTR(-EPERM); + ret = ttm_bo_setup_export(&bo->tbo, &ctx); + if (ret) + return ERR_PTR(ret); + buf = drm_gem_prime_export(gobj, flags); if (!IS_ERR(buf)) buf->ops = &amdgpu_dmabuf_ops; @@ -331,12 +427,13 @@ amdgpu_dma_buf_create_obj(struct drm_device *dev, struct dma_buf *dma_buf) flags |= other->flags & (AMDGPU_GEM_CREATE_CPU_GTT_USWC | AMDGPU_GEM_CREATE_COHERENT | + AMDGPU_GEM_CREATE_EXT_COHERENT | AMDGPU_GEM_CREATE_UNCACHED); } ret = amdgpu_gem_object_create(adev, dma_buf->size, PAGE_SIZE, AMDGPU_GEM_DOMAIN_CPU, flags, - ttm_bo_type_sg, resv, &gobj); + ttm_bo_type_sg, resv, &gobj, 0); if (ret) goto error; @@ -372,6 +469,10 @@ amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach) struct amdgpu_vm_bo_base *bo_base; int r; + /* FIXME: This should be after the "if", but needs a fix to make sure + * DMABuf imports are initialized in the right VM list. + */ + amdgpu_vm_bo_invalidate(bo, false); if (!bo->tbo.resource || bo->tbo.resource->mem_type == TTM_PL_SYSTEM) return; @@ -403,9 +504,12 @@ amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach) continue; } - r = amdgpu_vm_clear_freed(adev, vm, NULL); + /* Reserve fences for two SDMA page table updates */ + r = dma_resv_reserve_fences(resv, 2); + if (!r) + r = amdgpu_vm_clear_freed(adev, vm, NULL); if (!r) - r = amdgpu_vm_handle_moved(adev, vm); + r = amdgpu_vm_handle_moved(adev, vm, ticket); if (r && r != -EBUSY) DRM_ERROR("Failed to invalidate VM page tables (%d))\n", @@ -479,7 +583,10 @@ bool amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device *adev, struct drm_gem_object *obj = &bo->tbo.base; struct drm_gem_object *gobj; - if (obj->import_attach) { + if (!adev) + return false; + + if (drm_gem_is_imported(obj)) { struct dma_buf *dma_buf = obj->import_attach->dmabuf; if (dma_buf->ops != &amdgpu_dmabuf_ops) |