Am 2023-03-08 um 16:37 schrieb Chia-I Wu:
kgd_mem should be accessed with p->mutex locked, or it could have been freed by kfd_ioctl_free_memory_of_gpu.
Thank you for the patch. It's not just about accessing kgd_mem with p->mutex held. It's also about holding the mutex continuously. I'd update the description to be more explicit about the invariant being broken here:
kgd_mem pointers returned by kfd_process_device_translate_handle are only guaranteed to be valid while p->mutex is held. As soon as the mutex is unlocked, another thread can free the BO.
I can update the description and submit the patch. Reviewed-by: Felix Kuehling <felix.kuehl...@amd.com> Regards, Felix
Signed-off-by: Chia-I Wu <olva...@gmail.com> --- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 6d291aa6386bd..3c630114210d6 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -1293,14 +1293,14 @@ static int kfd_ioctl_map_memory_to_gpu(struct file *filep, args->n_success = i+1; }- mutex_unlock(&p->mutex);- err = amdgpu_amdkfd_gpuvm_sync_memory(dev->adev, (struct kgd_mem *) mem, true); if (err) { pr_debug("Sync memory failed, wait interrupted by user signal\n"); goto sync_memory_failed; }+ mutex_unlock(&p->mutex);+ /* 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]); @@ -1316,9 +1316,9 @@ static int kfd_ioctl_map_memory_to_gpu(struct file *filep, bind_process_to_device_failed: get_mem_obj_from_handle_failed: map_memory_to_gpu_failed: +sync_memory_failed: mutex_unlock(&p->mutex); copy_from_user_failed: -sync_memory_failed: kfree(devices_arr);return err;@@ -1332,6 +1332,7 @@ static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep, void *mem; long err = 0; uint32_t *devices_arr = NULL, i; + bool flush_tlb;if (!args->n_devices) {pr_debug("Device IDs array empty\n"); @@ -1384,16 +1385,19 @@ static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep, } args->n_success = i+1; } - mutex_unlock(&p->mutex);- if (kfd_flush_tlb_after_unmap(pdd->dev)) {+ flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev); + if (flush_tlb) { err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev, (struct kgd_mem *) mem, true); if (err) { pr_debug("Sync memory failed, wait interrupted by user signal\n"); 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]); @@ -1409,9 +1413,9 @@ static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep, bind_process_to_device_failed: get_mem_obj_from_handle_failed: unmap_memory_from_gpu_failed: +sync_memory_failed: mutex_unlock(&p->mutex); copy_from_user_failed: -sync_memory_failed: kfree(devices_arr); return err; }
