[Chuck Lever 2022-09-12 17:40:52 UTC]
I've noted that the VFS is serializing NFSD write calls (via a mutex). Small writes thus incur the overhead of a synchronization primitive. Large writes, even without a flush/sync, can take longer. Serialization of these requests
means that handling a burst of writes to the same file but different offsets can take much longer than needed.
I'd like to find a way of enabling NFSD's write requests to not be serialized through the VFS.
This was bugzilla.linux-nfs.org 397
[Chuck Lever 2022-09-12 17:40:52 UTC] I've noted that the VFS is serializing NFSD write calls (via a mutex). Small writes thus incur the overhead of a synchronization primitive. Large writes, even without a flush/sync, can take longer. Serialization of these requests means that handling a burst of writes to the same file but different offsets can take much longer than needed.
I'd like to find a way of enabling NFSD's write requests to not be serialized through the VFS.