<html><head> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> </head> <body> <div class="moz-cite-prefix">On 2023-09-11 21:11, Felix Kuehling wrote: </div> <blockquote type="cite" cite="mid:b8b95ee3-e608-a996-db65-cb02c418620a@amd.com">On 2023-09-11 15:55, Philip Yang wrote: <blockquote type="cite">Otherwise GPU may access the stale mapping and generate IOMMU IO_PAGE_FAULT. Move this to inside p->mutex to prevent multiple threads mapping and unmapping concurrently race condition. After kfd_mem_dmaunmap_attachment is removed from unmap_bo_from_gpuvm, kfd_mem_dmaunmap_attachment is called if failed to map to GPUs, and before free the mem attachment in case failed to unmap from GPUs. Signed-off-by: Philip Yang <a class="moz-txt-link-rfc2396E" href="mailto:Philip.Yang@amd.com"><Philip.Yang@amd.com></a> --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 1 + .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 22 +++++++++++++++--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 23 ++++++++++++------- 3 files changed, 35 insertions(+), 11 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index 559f14cc0a99..dff79a623f4a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -304,6 +304,7 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv); int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv); +void amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct amdgpu_vm *vm, struct kgd_mem *mem); </blockquote> For consistency with the other amdgpu_amdkfd APIs, please replace the vm parameter with a drm_priv parameter and do the drm_priv_to_vm translation inside amdgpu_amdkfd_gpuvm_dmaunmap_mem. </blockquote> ok. <blockquote type="cite" cite="mid:b8b95ee3-e608-a996-db65-cb02c418620a@amd.com"> <blockquote type="cite"> int amdgpu_amdkfd_gpuvm_sync_memory( struct amdgpu_device *adev, struct kgd_mem *mem, bool intr); int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index b5b940485059..ae767ad7afa2 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c @@ -1249,8 +1249,6 @@ static void unmap_bo_from_gpuvm(struct kgd_mem *mem, amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update); amdgpu_sync_fence(sync, bo_va->last_pt_update); - - kfd_mem_dmaunmap_attachment(mem, entry); } static int update_gpuvm_pte(struct kgd_mem *mem, @@ -1305,6 +1303,7 @@ static int map_bo_to_gpuvm(struct kgd_mem *mem, update_gpuvm_pte_failed: unmap_bo_from_gpuvm(mem, entry, sync); + kfd_mem_dmaunmap_attachment(mem, entry); return ret; } @@ -1910,8 +1909,10 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( mem->va + bo_size * (1 + mem->aql_queue)); /* Remove from VM internal data structures */ - list_for_each_entry_safe(entry, tmp, &mem->attachments, list) + list_for_each_entry_safe(entry, tmp, &mem->attachments, list) { + kfd_mem_dmaunmap_attachment(mem, entry); </blockquote> In theory this should not be needed, because user mode is required to unmap all mappings before freeing the BO. If </blockquote> amdgpu_amdkfd_gpuvm_sync_memory may return -ERESTARTSYS, or if vm_validate_pt_pd_bos failed, then amdgpu_amdkfd_gpuvm_dmaunmap_mem is not called, this is needed to avoid leaking dma mapping. <blockquote type="cite" cite="mid:b8b95ee3-e608-a996-db65-cb02c418620a@amd.com">(mapped_to_gpu_memory > 0), this function returns -EBUSY. For SG BOs, you may run into this error message: pr_err("SG Table of BO is UNEXPECTEDLY NULL"); If you decide to leave this dmaunmap_attachment here, you may want to turn that error into a pr_debug and remove "UNEXPECTEDLY". </blockquote> Thanks, I missed this, will change it in next version. Regards, Philip <blockquote type="cite" cite="mid:b8b95ee3-e608-a996-db65-cb02c418620a@amd.com"> Regards, Felix <blockquote type="cite"> kfd_mem_detach(entry); + } ret = unreserve_bo_and_vms(&ctx, false, false); @@ -2085,6 +2086,21 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( return ret; } +void amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct amdgpu_vm *vm, struct kgd_mem *mem) +{ + struct kfd_mem_attachment *entry; + + mutex_lock(&mem->lock); + + list_for_each_entry(entry, &mem->attachments, list) { + if (entry->bo_va->base.vm != vm) + continue; + kfd_mem_dmaunmap_attachment(mem, entry); + } + + mutex_unlock(&mem->lock); +} + int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 08687ce0aa8b..645628ff1faf 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -1432,17 +1432,24 @@ static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep, goto sync_memory_failed; } } - mutex_unlock(&p->mutex); - if (flush_tlb) { - /* Flush TLBs after waiting for the page table updates to complete */ - for (i = 0; i < args->n_devices; i++) { - peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]); - if (WARN_ON_ONCE(!peer_pdd)) - continue; + /* Flush TLBs after waiting for the page table updates to complete */ + for (i = 0; i < args->n_devices; i++) { + struct amdgpu_vm *vm; + + peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]); + if (WARN_ON_ONCE(!peer_pdd)) + continue; + if (flush_tlb) kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT); - } + + /* Remove dma mapping after tlb flush to avoid IO_PAGE_FAULT */ + vm = drm_priv_to_vm(peer_pdd->drm_priv); + amdgpu_amdkfd_gpuvm_dmaunmap_mem(vm, mem); } + + mutex_unlock(&p->mutex); + kfree(devices_arr); return 0; </blockquote> </blockquote> </body> </html>