Re: [PATCH 2/2] mm: compaction: Minimise the time IRQs are disabled while isolating pages for migration

[Andrew Morton <akpm@linux-foundation.org>]

On Wed, 2 Mar 2011 07:22:33 +0900
Minchan Kim <minchan.kim@gmail.com> wrote:

> On Wed, Mar 2, 2011 at 1:19 AM, Andrea Arcangeli <aarcange@redhat.com> wrote:
> > On Wed, Mar 02, 2011 at 12:35:58AM +0900, Minchan Kim wrote:
> >> On Tue, Mar 01, 2011 at 01:49:25PM +0900, KAMEZAWA Hiroyuki wrote:
> >> > On Tue, 1 Mar 2011 13:11:46 +0900
> >> > Minchan Kim <minchan.kim@gmail.com> wrote:
> >> >
>
> ...
>
> > pages freed from irq shouldn't be PageLRU.
> 
> Hmm..
> As looking code, it seems to be no problem and I didn't see the any
> comment about such rule. It should have been written down in
> __page_cache_release.
> Just out of curiosity.
> What kinds of problem happen if we release lru page in irq context?

put_page() from irq context has been permissible for ten years.  I
expect there are a number of sites which do this (via subtle code
paths, often).  It might get messy.

> >
> > deferring freeing to workqueue doesn't look ok. firewall loads runs
> > only from irq and this will cause some more work and a delay in the
> > freeing. I doubt it's worhwhile especially for the lru_lock.
> >
> 
> As you said, if it is for decreasing lock contention in SMP to deliver
> overall better performance, maybe we need to check again how much it
> helps.
> If it doesn't help much, could we remove irq_save/restore of lru_lock?
> Do you know any benchmark to prove it had a benefit at that time or
> any thread discussing about that in lkml?


: commit b10a82b195d63575958872de5721008b0e9bef2d
: Author: akpm <akpm>
: Date:   Thu Aug 15 18:21:05 2002 +0000
: 
:     [PATCH] make pagemap_lru_lock irq-safe
:     
:     It is expensive for a CPU to take an interrupt while holding the page
:     LRU lock, because other CPUs will pile up on the lock while the
:     interrupt runs.
:     
:     Disabling interrupts while holding the lock reduces contention by an
:     additional 30% on 4-way.  This is when the only source of interrupts is
:     disk completion.  The improvement will be higher with more CPUs and it
:     will be higher if there is networking happening.
:     
:     The maximum hold time of this lock is 17 microseconds on 500 MHx PIII,
:     which is well inside the kernel's maximum interrupt latency (which was
:     100 usecs when I last looked, a year ago).
:     
:     This optimisation is not needed on uniprocessor, but the patch disables
:     IRQs while holding pagemap_lru_lock anyway, so it becomes an irq-safe
:     spinlock, and pages can be moved from the LRU in interrupt context.
:     
:     pagemap_lru_lock has been renamed to _pagemap_lru_lock to pick up any
:     missed uses, and to reliably break any out-of-tree patches which may be
:     using the old semantics.
:     
:     BKrev: 3d5bf1110yfdAAur4xqJfiLBDJ2Cqw


Ancient stuff, and not a lot of detail.  But I did measure it.  I
measured everything ;) And, as mentioned, I'd expect that the
contention problems would worsen on higher CPU machines and higher
interrupt frequencies.

I expect we could eliminate the irqsave requirement from
rotate_reclaimable_page() simply by switching to a trylock.  Some pages
will end up at the wrong end of the LRU but the effects may be
negligible.  Or perhaps they may not - disk seeks are costly.