Re: [PATCH 36/42] VFS: export drop_pagecache_sb

[Nick Piggin]

On Friday 14 December 2007 02:24, Erez Zadok wrote:
> In message <200712121638.35167.nickpiggin@xxxxxxxxxxxx>, Nick Piggin writes:
> > On Monday 10 December 2007 13:42, Erez Zadok wrote:
> > > Needed to maintain cache coherency after branch management.
> >
> > Hmm, I'd much prefer to be able to sleep in invalidate_mapping_pages
> > before this function gets exported.
> >
> > As it is, it can cause massive latencies on preemption and the inode_lock
> > so it is pretty much debug-only IMO. I'd rather it didn't escape into the
> > wild as is.
> >
> > Either that or rework your cache coherency somehow.
>
> Nick, thanks for the advice.
>
> We use a generation number after each successful branch configuration
> command, so that ->d_revalidate later on can discover that change, and
> rebuild the union of objects.  At ->remount time, I figured it'd be nice to
> "encourage" that revalidation to happen sooner, by invalidating as many
> upper pages as possible, thus causing ->d_revalidate/->readpage to take
> place sooner.  So we used to call drop_pagecache_sb from our remount code:
> it was the only caller of drop_pagecache_sb.  It wasn't too much of an
> latency issue to call drop_pagecache_sb there: the VFS remount code path is
> already pretty slow (dropping temporarily to readonly mode, and dropping
> other caches), and remount isn't an operation used often, so a little bit
> more latency would probably not have been noticed by users.

Well a large, infrequent spike is the most damaging to latency sensitive
users. And anyway, I guess the infrequency of remount means it doesn't
have to be really efficient with invalidating pagecache either.


> Nevertheless, it was not strictly necessary to call drop_pagecache_sb in
> unionfs_remount, because the objects in question will have gotten
> revalidated sooner or later anyway; the call to drop_pagecache_sb was just
> an optimization (one which I wasn't 100% sure about anyway, as per my long
> "XXX" comment above that call in unionfs_remount).
>
> So I agree with you: if this symbol can be abused by modules and cause
> problems, then exporting it to modules is too risky.  I've reworked my code
> to avoid calling drop_pagecache_sb and I'll [sic] drop that patch.

Thanks, I'd be much happier with that.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/