Re: [PATCH 1/2] SHM_UNLOCK: fix long unpreemptible section

[Hugh Dickins]

On Wed, 18 Jan 2012, Andrew Morton wrote:
> On Sat, 14 Jan 2012 16:18:43 -0800 (PST)
> Hugh Dickins <hughd@xxxxxxxxxx> wrote:
> 
> > scan_mapping_unevictable_pages() is used to make SysV SHM_LOCKed pages
> > evictable again once the shared memory is unlocked.  It does this with
> > pagevec_lookup()s across the whole object (which might occupy most of
> > memory), and takes 300ms to unlock 7GB here.  A cond_resched() every
> > PAGEVEC_SIZE pages would be good.
>... 
> Is -stable backporting really warranted?  AFAICT the only thing we're
> fixing here is a long latency glitch during a rare operation on large
> machines.  Usually it will be on only one CPU, too.

True: I'm not sure if it amounts to -stable material or not.
I see you've taken out its Cc: stable line: that's fine by me, but...

> "[PATCH 2/2] SHM_UNLOCK: fix Unevictable pages stranded after swap"
> does loko like -stable material, so omitting 1/1 will probably screw
> things up :(

Sort of, but they both(?) needed respinning for -stable anyway.
Even against 3.2, there's some little change in vmscan.c that generates
a reject.  Greg has now closed down 3.1.N (which would have been tiresome
to port to, because it was still supporting a second caller of check_move),
and by your argument above it's not worth porting 1/2 back to 2.6.32.  So
I think 2/2 can just go into 3.2.N, dragging 1/2 along in its slipstream
(if you can have a slipstream in front of you).

I ordered them that way because 1/2 fixes an old, and 2/2 a recent, bug.

> > Resend in the hope that it can get into 3.3.
> 
> That we can do ;)

Thank you!

> > +#else
> > +void scan_mapping_unevictable_pages(struct address_space *mapping)
> > +{
> > +}
> > +#endif /* CONFIG_SHMEM */
> 
> Inlining the CONFIG_SHMEM=n stub would have been mroe efficient.

True, though in 2/2 it morphs into shmem_unlock_mapping() over
in shmem.c, and we seem to have the convention that TINY's !SHMEM
stubs live as non-inline functions there - probably no good reason
for that, just reflects their historical origins in tiny-shmem.c.
A grand saving to make some other time ;)

Hugh
--
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/