From 70773bef4e091ff6d2a91e3dfb4f29013eb81f1f Mon Sep 17 00:00:00 2001 From: Arvind Yadav Date: Wed, 25 Sep 2024 18:09:49 +0200 Subject: [PATCH 3071/5403] drm/amdgpu: update userqueue BOs and PDs MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit This patch updates the VM_IOCTL to allow userspace to synchronize the mapping/unmapping of a BO in the page table. The major changes are: - it adds a drm_timeline object as an input parameter to the VM IOCTL. - this object is used by the kernel to sync the update of the BO in the page table during the mapping of the object. - the kernel also synchronizes the tlb flush of the page table entry of this object during the unmapping (Added in this series: https://patchwork.freedesktop.org/series/131276/ and https://patchwork.freedesktop.org/patch/584182/) - the userspace can wait on this timeline, and then the BO is ready to be consumed by the GPU. The UAPI for the same has been approved here: https://gitlab.freedesktop.org/mesa/drm/-/merge_requests/392 V2: - remove the eviction fence coupling V3: - added the drm timeline support instead of input/output fence (Christian) V4: - made timeline 64-bit (Christian) - bug fix (Arvind) V5: GLCTS bug fix (Arvind) V6: Rename syncobj_handle -> timeline_syncobj_out Rename point -> timeline_point_in (Marek) V7: Addressed review comments from Christian: - do not send last_update fence in case of vm_clear_freed, instead return the fence from gen_va_update_vm - move the functions to update bo_mapping to amdgpu_gem.c - do not use amdgpu_userq_update_vm anymore in userq_create() V8: Addressed review comments from Christian: - Split amdgpu_gem_update_bo_mapping function. - amdgpu_gem_va_update_vm should return stub for error. V9: Addressed review comments from Christian: - Rename the function amdgpu_gem_update_timeline_node. - amdgpu_gem_update_timeline_node should be void function. - when timeline_point is zero don't allocate a chain and call drm_syncobj_replace_fence() instead of drm_syncobj_add_point(). V11: rebase V12: Fix 32-bit holes issue in sturct drm_amdgpu_gem_va. V13: Fix the review comment by renaming timeline syncobj (Marek) Cc: Alex Deucher Cc: Felix Kuehling Cc: Christian König Reviewed-by: Alex Deucher Reviewed-by: Christian König Signed-off-by: Arvind Yadav Signed-off-by: Shashank Sharma Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 113 ++++++++++++++++++++++-- include/uapi/drm/amdgpu_drm.h | 9 ++ 2 files changed, 113 insertions(+), 9 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c index 69429df09477..542a1b70f2ee 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c @@ -36,6 +36,7 @@ #include #include #include +#include #include "amdgpu.h" #include "amdgpu_display.h" @@ -44,6 +45,75 @@ #include "amdgpu_xgmi.h" #include "amdgpu_vm.h" +static int +amdgpu_gem_update_timeline_node(struct drm_file *filp, + uint32_t syncobj_handle, + uint64_t point, + struct drm_syncobj **syncobj, + struct dma_fence_chain **chain) +{ + if (!syncobj_handle) + return 0; + + /* Find the sync object */ + *syncobj = drm_syncobj_find(filp, syncobj_handle); + if (!*syncobj) + return -ENOENT; + + if (!point) + return 0; + + /* Allocate the chain node */ + *chain = dma_fence_chain_alloc(); + if (!*chain) { + drm_syncobj_put(*syncobj); + return -ENOMEM; + } + + return 0; +} + +static void +amdgpu_gem_update_bo_mapping(struct drm_file *filp, + struct amdgpu_bo_va *bo_va, + uint32_t operation, + uint64_t point, + struct dma_fence *fence, + struct drm_syncobj *syncobj, + struct dma_fence_chain *chain) +{ + struct amdgpu_bo *bo = bo_va ? bo_va->base.bo : NULL; + struct amdgpu_fpriv *fpriv = filp->driver_priv; + struct amdgpu_vm *vm = &fpriv->vm; + struct dma_fence *last_update; + + if (!syncobj) + return; + + /* Find the last update fence */ + switch (operation) { + case AMDGPU_VA_OP_MAP: + case AMDGPU_VA_OP_REPLACE: + if (bo && (bo->tbo.base.resv == vm->root.bo->tbo.base.resv)) + last_update = vm->last_update; + else + last_update = bo_va->last_pt_update; + break; + case AMDGPU_VA_OP_UNMAP: + case AMDGPU_VA_OP_CLEAR: + last_update = fence; + break; + default: + return; + } + + /* Add fence to timeline */ + if (!point) + drm_syncobj_replace_fence(syncobj, last_update); + else + drm_syncobj_add_point(syncobj, chain, last_update, point); +} + static vm_fault_t amdgpu_gem_fault(struct vm_fault *vmf) { struct ttm_buffer_object *bo = vmf->vma->vm_private_data; @@ -638,18 +708,23 @@ int amdgpu_gem_metadata_ioctl(struct drm_device *dev, void *data, * * Update the bo_va directly after setting its address. Errors are not * vital here, so they are not reported back to userspace. + * + * Returns resulting fence if freed BO(s) got cleared from the PT. + * otherwise stub fence in case of error. */ -static void amdgpu_gem_va_update_vm(struct amdgpu_device *adev, - struct amdgpu_vm *vm, - struct amdgpu_bo_va *bo_va, - uint32_t operation) +static struct dma_fence * +amdgpu_gem_va_update_vm(struct amdgpu_device *adev, + struct amdgpu_vm *vm, + struct amdgpu_bo_va *bo_va, + uint32_t operation) { + struct dma_fence *fence = dma_fence_get_stub(); int r; if (!amdgpu_vm_ready(vm)) - return; + return fence; - r = amdgpu_vm_clear_freed(adev, vm, NULL); + r = amdgpu_vm_clear_freed(adev, vm, &fence); if (r) goto error; @@ -665,6 +740,8 @@ static void amdgpu_gem_va_update_vm(struct amdgpu_device *adev, error: if (r && r != -ERESTARTSYS) DRM_ERROR("Couldn't update BO_VA (%d)\n", r); + + return fence; } /** @@ -713,6 +790,9 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data, struct amdgpu_fpriv *fpriv = filp->driver_priv; struct amdgpu_bo *abo; struct amdgpu_bo_va *bo_va; + struct drm_syncobj *timeline_syncobj = NULL; + struct dma_fence_chain *timeline_chain = NULL; + struct dma_fence *fence; struct drm_exec exec; uint64_t va_flags; uint64_t vm_size; @@ -827,9 +907,24 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data, default: break; } - if (!r && !(args->flags & AMDGPU_VM_DELAY_UPDATE) && !adev->debug_vm) - amdgpu_gem_va_update_vm(adev, &fpriv->vm, bo_va, - args->operation); + if (!r && !(args->flags & AMDGPU_VM_DELAY_UPDATE) && !adev->debug_vm) { + + r = amdgpu_gem_update_timeline_node(filp, + args->vm_timeline_syncobj_out, + args->vm_timeline_point, + &timeline_syncobj, + &timeline_chain); + + fence = amdgpu_gem_va_update_vm(adev, &fpriv->vm, bo_va, + args->operation); + + if (!r) + amdgpu_gem_update_bo_mapping(filp, bo_va, + args->operation, + args->vm_timeline_point, + fence, timeline_syncobj, + timeline_chain); + } error: drm_exec_fini(&exec); diff --git a/include/uapi/drm/amdgpu_drm.h b/include/uapi/drm/amdgpu_drm.h index 1a21259cb8c4..ca82935ff93a 100644 --- a/include/uapi/drm/amdgpu_drm.h +++ b/include/uapi/drm/amdgpu_drm.h @@ -857,6 +857,15 @@ struct drm_amdgpu_gem_va { __u64 offset_in_bo; /** Specify mapping size. Must be correctly aligned. */ __u64 map_size; + /** + * vm_timeline_point is a sequence number used to add new timeline point. + */ + __u64 vm_timeline_point; + /** + * The vm page table update fence is installed in given vm_timeline_syncobj_out + * at vm_timeline_point. + */ + __u32 vm_timeline_syncobj_out; }; #define AMDGPU_HW_IP_GFX 0 -- 2.54.0