On Mon, 2025-11-24 at 15:11 -0500, Benjamin Marzinski wrote:
> On Fri, Nov 21, 2025 at 02:35:23PM -0800, Brian Bunker wrote:
> 
> > 
> > I added the purge thread because I didn't want to starve the
> > checker thread at a large disconnect volume scale. I noticed
> > that the number of devices if I purged inline with the check that
> > it didn't scale linearly after a point and seemed to be
> > significantly
> > starving the checker thread. Doing the purge in another thread
> > seemed to relieve that pressure.
> 
> That might be because the check thread has safeguards to try to avoid
> starving the other threads. Since a lot of multipathd's work is gated
> by
> the vecs lock, there's only so much parallelism that can happen with
> multiple threads. If deleting the devices is taking a long time, it
> might
> be better for this to get interrupted, so that other threads can run.
> 
> Since purging the paths is fairly low priority, we could continue to
> run
> it in its own thread, but instead of running through all the devices
> at
> once, purgeloop could lock the vecs->lock, handle one device, and
> then
> unlock and loop. This means it would need to search through the list
> from the start each time it regrabbed the lock, since the list could
> have changed while it wasn't holding it. When purgeloop makes it all
> the
> way through the list without finding any paths that are due for a
> delete
> attempt, it can sleep or block waiting for more paths.
> 
> This would mean that paths would need to remember if they had already
> been handled in a cycle. That could be done by purgeloop keeping
> track
> of which cycle it was on, and the paths storing the cycle number
> whenever they were checked.

As I wrote in my other post, wish that this thread wouldn't hold the
vecs lock at all. multipathd could simply use a pipe to write dev_t's
of to-be-purged paths to the purge thread (or process :-) ).

Martin

Reply via email to