[sheepdog] [PATCH] sheep: fix handling of too old epoch in check_request

[levin li]

On 05/31/2012 04:03 PM, Christoph Hellwig wrote:
> On Thu, May 31, 2012 at 09:59:55AM +0800, levin li wrote:
>> my original plan is that when a request with old epoch comes, we directly
>> send back a response with SD_RES_OLD_NODE_VER in check_request() to make the
>> sender to retry, argee this?
> 
> That's what my patch does if we're on a node that doesn't also act as
> gateway.  If we're doing local I/O on a gateway life obviously isn't
> that simple.
> 
> Talking about that I really hate all the code required to support
> gateways doing local I/O.
> 

Gateway doing local IO indeed confused me a lot before, if separated, the logic
would be much clearer, but doing this may need to modify too much existing code
and it seems not so easy to do that.

thanks,

levin

> Does anyone actually use sheepdog clusters small enough that optimizing
> this case matters?  I'd love to basically queue up a new SD_FLAG_CMD_IO_LOCAL
> I/O in the gateway if if finds a local node and kill all these special
> cases that make life hard.
> 
> In fact I wonder if we should make the gateway and I/O nodes entirely
> separate process, although that would cause quite a bit of trouble for
> users initially.
> 
>> Note that in check_request(), for a gateway request without SD_FLAG_CMD_IO_LOCAL,
>> the epoch of request is always equal to system epoch, for we have set it to
>> system epoch in queue_request() before entering check_request().
>> 	/*
>> 	 * we set epoch for non direct requests here.  Note that we need to
>> 	 * sample sys->epoch before passing requests to worker threads as
>> 	 * it can change anytime we return to processing membership change
>> 	 * events.
>> 	 */
>> 	if (!(hdr->flags & SD_FLAG_CMD_IO_LOCAL))
>> 		hdr->epoch = sys->epoch;
> 
> Indeed, I forgot that we also always reset it when requeing.  That'll
> make life a lot simpler.
>