Re: [PATCH 2/2] drm/amdkfd: Downgrade mmap lock in SVM restore work

[Felix Kuehling <felix.kuehling@amd.com>]

On 2026-10-20 12:31, Philip Yang wrote:
> svm_range_restore_work holds the mmap write lock across the entire loop
> validating and mapping every invalid range, which blocks all mmap,
> munmap and page fault activity of the process, not only the KFD paths
> waiting for the same locks.
>
> Only the deferred work flush handshake in
> svm_range_list_lock_and_flush_work needs the write lock.
> svm_range_validate_and_map runs under the read lock, as
> svm_range_set_attr already does after its mmap_write_downgrade. The
> flush stays valid because svm_range_deferred_list_work takes the write
> lock, so it cannot run while the read lock is held.
>
> Downgrade the mmap lock before walking the range list.
>
> Signed-off-by: Philip Yang <Philip.Yang@amd.com>

Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>


> ---
>   drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 6 +++++-
>   1 file changed, 5 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> index 28b687133cf8..f63bd8cf4174 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> @@ -1987,6 +1987,10 @@ static void svm_range_restore_work(struct work_struct *work)
>   
>   	mutex_lock(&process_info->lock);
>   	svm_range_list_lock_and_flush_work(svms, mm);
> +	/* Validate and map needs the read lock only, and deferred list work
> +	 * takes the write lock, so the flush above stays valid.
> +	 */
> +	mmap_write_downgrade(mm);
>   	mutex_lock(&svms->lock);
>   
>   	evicted_ranges = atomic_read(&svms->evicted_ranges);
> @@ -2045,7 +2049,7 @@ static void svm_range_restore_work(struct work_struct *work)
>   					       AMDGPU_SVM_RANGE_RESTORE_DELAY_MAX_MS);
>   
>   	mutex_unlock(&svms->lock);
> -	mmap_write_unlock(mm);
> +	mmap_read_unlock(mm);
>   	mutex_unlock(&process_info->lock);
>   
>   	/* If validation failed, reschedule another attempt */