On Fri, Feb 23, 2024 at 06:18:55PM +0100, Christoph Hellwig wrote:
> On Fri, Feb 23, 2024 at 09:09:53AM -0800, Darrick J. Wong wrote:
> > > - lcounter = (long long)mp->m_resblks_avail + delta;
> > > - if (lcounter >= 0) {
> > > - mp->m_resblks_avail = lcounter;
> > > + xfs_warn_once(mp,
> > > +"Reserve blocks depleted! Consider increasing reserve pool size.");
> >
> > Hmm. This message gets logged if the __percpu_counter_compare above
> > returns a negative value and we don't have a reserve pool. I think
> > that's a change from the current code, where running out of m_frextents
> > returns ENOSPC without generating (misleading) log messages about the
> > reserve pool.
>
> Yes, I'll need to fix that up.
>
> > without all the reserve pool machinery that fdblocks has. The original
> > code to implement m_frextents did exactly that, but Dave really wanted
> > me to optimize this for icache footprint[1] so we ended up with (IMO)
> > harder to understand xfs_mod_freecounter.
>
> FYI, I have uses for reserved frextents a bit down the road, so
> I'll make use of that. Without that splitting them would seem nicer
> to me.
Oh, ok. I agree that keeping the combined function makes more sense if
you'll be adding a reserve pool for rtextents.
--D
