Re: [PATCH 6/6] xfs: flush multiple device caches in parallel in xlog_write_iclog

[Christoph Hellwig <hch@lst.de>]

On Wed, Sep 02, 2026 at 09:18:58AM -0700, Darrick J. Wong wrote:
> On Wed, Sep 02, 2026 at 08:49:20AM +0300, Christoph Hellwig wrote:
> > When xlog_write_iclog needs to flush the cache for more than one devices,
> > the current implementations does this sequentially, which adds up the
> > flush latency for all devices.  Switch to kicking off all cache flushes
> > in parallel so that only the longest latency bounds the time of the log
> > I/O.  This removes the REQ_PREFLUSH optimization for the log device,
> > but as that flag is never passed on to the device and just very slightly
> > reduce the latency by queueing the following write from a lower-level
> > context it is trivially shadowed by the latency improvements of the
> > parallel flush commands.
> > 
> > Signed-off-by: Christoph Hellwig <hch@lst.de>
> 
> That looks like a neat trick.  Do you see any performance speedups from
> flushing in parallel?

I haven't really done a benchmark, as all the devices I test multi-device
setups don't expose volatile write caches, so I actually had to create
a synthetic setup to fully test this in qemu.  But I can kick off a run
using say consumer SSDs.  I expect it to significantly speed up fsync
bound workloads.