lkml.org 
[lkml]   [2008]   [Dec]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH] cpuset,mm: fix allocating page cache/slab object on the unallowed node when memory spread is set
Date
On Wednesday 31 December 2008 09:28:05 Andrew Morton wrote:
> On Fri, 26 Dec 2008 14:37:07 +0800
>
> Miao Xie <miaox@cn.fujitsu.com> wrote:
> > The task still allocated the page caches on old node after modifying its
> > cpuset's mems when 'memory_spread_page' was set, it is caused by the old
> > mem_allowed_list of the task. Slab has the same problem.
>
> ok...
>
> > diff --git a/mm/filemap.c b/mm/filemap.c
> > index f3e5f89..d978983 100644
> > --- a/mm/filemap.c
> > +++ b/mm/filemap.c
> > @@ -517,6 +517,9 @@ int add_to_page_cache_lru(struct page *page, struct
> > address_space *mapping, #ifdef CONFIG_NUMA
> > struct page *__page_cache_alloc(gfp_t gfp)
> > {
> > + if ((gfp & __GFP_WAIT) && !in_interrupt())
> > + cpuset_update_task_memory_state();
> > +
> > if (cpuset_do_page_mem_spread()) {
> > int n = cpuset_mem_spread_node();
> > return alloc_pages_node(n, gfp, 0);
> > diff --git a/mm/slab.c b/mm/slab.c
> > index 0918751..3b6e3d7 100644
> > --- a/mm/slab.c
> > +++ b/mm/slab.c
> > @@ -3460,6 +3460,9 @@ __cache_alloc(struct kmem_cache *cachep, gfp_t
> > flags, void *caller) if (should_failslab(cachep, flags))
> > return NULL;
> >
> > + if ((flags & __GFP_WAIT) && !in_interrupt())
> > + cpuset_update_task_memory_state();
> > +

These paths are pretty performance critical. Why don't cpusets code do this
work in the slowpath where the cpuset's mems_allowed gets changed rather
than putting these calls all over the place with apparently no real rhyme or
reason :( (this is not against your patch, but just this part of the cpusets
design)

> > cache_alloc_debugcheck_before(cachep, flags);
> > local_irq_save(save_flags);
> > objp = __do_cache_alloc(cachep, flags);
>
> Problems.
>
> a) There's no need to test in_interrupt(). Any caller who passed us
> __GFP_WAIT from interrupt context is horridly buggy and needs to be
> fixed.

Right. There are existing sites that do the same check, which is probably
where it is copied from.


> b) Even if the caller _did_ set __GFP_WAIT, there's no guarantee
> that we're deadlock safe here. Does anyone ever do a __GFP_WAIT
> allocation while holding callback_mutex? If so, it'll deadlock.

It's static to cpuset.c, so I'd hope not.


> c) These are two of the kernel's hottest code paths. We really
> really really really don't want to be polling for some dopey
> userspace admin change on each call to __cache_alloc()!

Yeah, right. Let's try to fix cpuset.c instead...


> d) How does slub handle this problem?

SLUB seems to do a "sloppy" kind of memory policy allocation, where it just
relies on the page allocator to hand us the correct page and AFAIKS does not
exactly obey this stuff all the time.



\
 
 \ /
  Last update: 2008-12-31 04:17    [W:0.058 / U:0.116 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site