[openib-general] [PATCH] IB/SRP identify QP in error state

[Michael S. Tsirkin]

Quoting r. Ishai Rabinovitz <ishai at mellanox.co.il>:
> Subject: [PATCH] IB/SRP identify QP in error state
> 
> There is a bug in mthca low level driver. 
> A call to ib_post_send that tries to post to a QP that is in error state does
> not return immediately with error. It terminates with errors after a timeout.

Let me rephrase: after post send/receive to QP in error state in mthca,
a completion with error might never get generated.
SRP will then timeout.

To fix mthca, we'd need to change QP state on completion with error
and on modify to error, and add actual code where it now says

/* XXX check that state is OK to post send */
/* XXX check that state is OK to post receive */

I guess the reason we never fixed this was because it did not
seem to actually hurt any real ULPs, and testing QP state will
affect fast path performance.

However, IB spec is quite explicit on this point, and fixing a low
level drivers seems a better approach than adding work-arounds in ULPs.

Roland, what do you think?

-- 
MST