On 2023-09-11 21:11, Felix Kuehling
wrote:
On 2023-09-11 15:55, Philip Yang wrote:ok.
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 <philip.y...@amd.com>
---
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);
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.
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.
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);
In theory this should not be needed, because user mode is required to unmap all mappings before freeing the BO. If
(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".
Thanks, I missed this, will change it in next version.
Regards,
Philip
Regards,
Felix
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;