Re: NFS client errors with 2.5.74?

[Jamie Lokier]

Trond Myklebust wrote:
> The first one should fix the problem of the kernel missing replies
> while we are busy trying to resend a request.

This by itself doesn't fix the problem of too-fast timeout errors on
soft mounts (e.g. returning EIO within <0.1s).

I am still seeing the fs get into a state where each time a large file
is written, it reports EIO (but writes successfully anyway).  And "ls
-R" still shows EIO errors also.

> The second, solves a problem of resource starvation. The fact that we
> can currently just submit arbitrary numbers of asynchronous requests
> means that we can exhaust resources to the point where the socket
> starts dropping replies.
> This patch limits the number of outstanding asynchronous requests to
> 16 per socket (the maximum number of xprt/transport slots).

I haven't tried this yet.  It doesn't apply to 2.5.74 due to the calls
to io_schedule().

- Jamie