Hello,

On 08/19/2010 12:32 PM, Kiyoshi Ueda wrote:

quoted

I see but single pending flush and steady write streams w/o saturating
the mempool would be able to stall dm_wait_for_completeion(), no?  Eh
well, it's a separate issue, I guess.

Your understanding is correct, dm_wait_for_completion() for flush
will stall in such cases for request-based dm.
That's why I mentioned below in
https://www.redhat.com/archives/dm-devel/2010-August/msg00026.html.

    In other words, current request-based device-mapper can't handle
    other requests while a flush request is in progress.

In flush request handling, request-based dm uses dm_wait_for_completion()
to wait for the completion of cloned flush requests, depending on
the fact that there should be only flush requests in flight owning
to the block layer sequencing.

I see.  bio based implementation also uses dm_wait_for_completion()
but it also has DMF_QUEUE_IO_TO_THREAD to plug all the follow up bio's
while flush is in progress, which sucks for throughput but
successfully avoids starvation.

It's not a separate issue and we need to resolve it at least.
I'm still considering how I can fix the request-based dm.

Right, I thought you were talking about REQ_FLUSHes not sycnhronized
against barrier write.  Anyways, yeah, it's a problem.  I don't think
not being able to handle multiple flushes concurrently would be a
major issue.  The problem is not being able to process other
bios/requests while a flush is in progress.  All that's necessary is
making the completion detection a bit more fine grained so that it
counts the number of in flight flush bios/requests and completes when
it reaches zero instead of waiting for all outstanding commands.
Shouldn't be too hard.

Thanks.

-- 
tejun