3.1.2 MxFS-Linux Release Notes

Matrix File Serving Solution Pack for Linux 7

10577 Clients may occasionally experience errors during failover

Due to the inherent unreliable nature of the NFS protocol, NFS

RPC requests or replies may occasionally be lost. The NFS client

will then retry the request. This situation occasionally results in

duplicate NFS RPC requests sent to the NFS server.

Normally, a duplicate NFS RPC is detected in the NFS replay

cache and a response sent immediately without attempting to

execute the command a second time. This prevents unlinks, for

example, from generating spurious file-not-found or stale-file-

handle errors.

However, if an NFS RPC arrives on a node that subsequently

fails before the reply can be sent, the client will retry the NFS

RPC. This retransmission will arrive at the backup (gaining)

node, will not be detected as a replay in the replay cache, and

will be processed a second time. RPCʹs that have non-

idempotent results (such as unlink, rename, and create) will

then generate an error back to the clientʹs application.

Read and write operations will not show this problem.

10646 Invalid link count can be returned to NFS clients

Linux NFS clients running with Linux 2.4.x kernels will return

an invalid link count via the stat() system call interface if a file

with multiple hard links is unlinked by one coincident with a

failover of a Virtual NFS Service. The unlink succeeds correctly

in the matrix server and the correct link count is returned to the

client, but the stat() system call on the NFS client will continue to

return the old link count. This situation may cause problems

with programs that unlink files with multiple hard-links and

then use the stat() system call interface to verify that the link

count is what is expected.

This problem has not been observed on Linux NFS clients with

2.6.x kernels.

Defect Description