[openib-general] Re: [PATCHv2] fix sdp aio race

[Michael S. Tsirkin]

Quoting r. Libor Michalek <limichal at cisco.com>:
>   I don't think a bit is sufficient, it would need to be a reference
> count, since you'll potentially have multiple outstanding completions.

Please note I stop polling cq once the bit is set.
Therefore there's always exactly 1 outstanding deferred completion.
The rest is polled later in thread context.

Doesnt it make sense?

> Also, the lock.event is modified, deferred flag cleared, inside the
> conn_lock, but it needs to be protected by the conn_lock's irq.
> 
>   I'll look at if it's possible to never get into this ordering issue
> by always forcing iocb defered completions...
> 
> -Libor

But, what about merging get_user_pages first? I'm kind of tired of
maintaining it.

-- 
MST