Messages in this thread | | | Date | Mon, 26 Mar 2012 16:50:27 +0100 | From | Mel Gorman <> | Subject | Re: [PATCH] cpuset: mm: Reduce large amounts of memory barrier related damage v3 |
| |
On Mon, Mar 26, 2012 at 12:56:24PM +0200, Peter Zijlstra wrote: > On Wed, 2012-03-07 at 18:08 +0000, Mel Gorman wrote: > > + } while (!put_mems_allowed(cpuset_mems_cookie) && !page); > > Sorry for only noticing this now, but wouldn't it be better to first > check page and only then bother with the put_mems_allowed() thing? That > avoids the smp_rmb() and seqcount conditional all together in the likely > case the allocation actually succeeded. > > <SNIP> > > diff --git a/mm/filemap.c b/mm/filemap.c > index c3811bc..3b41553 100644 > --- a/mm/filemap.c > +++ b/mm/filemap.c > @@ -504,7 +504,7 @@ struct page *__page_cache_alloc(gfp_t gfp) > cpuset_mems_cookie = get_mems_allowed(); > n = cpuset_mem_spread_node(); > page = alloc_pages_exact_node(n, gfp, 0); > - } while (!put_mems_allowed(cpuset_mems_cookie) && !page); > + } while (!page && !put_mems_allowed(cpuset_mems_cookie)); > > return page; > }
I think such a change would be better but should also rename the API. If developers see a get_foo type call, they will expect to see a put_foo call or assume it's a bug even though the implementation happens to be ok with that. Any suggestion on what a good new name would be?
How about read_mems_allowed_begin() and read_mems_allowed_retry()?
read_mems_allowed_begin would be a rename of get_mems_allowed(). In an error path, read_mems_allowed_retry() would documented to be *optionally* called when deciding whether to retry the operation or not. In this scheme, !put_mems_allowed would become read_mems_allowed_retry() which might be a bit easier to read overall.
-- Mel Gorman SUSE Labs
| |