On 1/26/2012 1:03 PM, Andrew Deason wrote:
> I saw some emails to you today that sounded like someone wanted it :)
> 
> But sure, some people may want something like that. Some people want
> hardmount. (Not saying we need an option for that at this time or
> anything, though.)

That is certainly one interpretation.  The other is that they want the
root cause of the data corruption problems to be fixed.

>>> I believe/assume what is being considered "right" is the latter
>>> option.
>>
>> ?
> 
> I definitely heard some advocacy for this the last time we were arguing
> about this (how the fileserver-side should make this determination,
> which is true, but...). But yeah, this isn't what went in.

The file server now disable keep alives when performing disk I/O.  If
the I/O doesn't return, the clients will timeout.  This is in tree.

> Well, I was talking about idledead stuff here, specifically client
> behavior, which isn't going to screw up another client. I think that
> matters much more when you're talking about server behavior, in which
> case, yes, certainly.

Clients that use idle-dead timeouts to cancel and retry RPCs do have
negative impacts on other clients.  Repeated retries of RPCs that the
file server cannot process because a vnode is already locked result in
the thread pool becoming depleted.  The side effect is that a file
server will enter a state where it can process no requests much sooner
than it otherwise would have.


Attachment: signature.asc
Description: OpenPGP digital signature

Reply via email to