cpuset: mm: reduce large amounts of memory barrier related damage v3

Commit c0ff7453bb5c ("cpuset,mm: fix no node to alloc memory when
changing cpuset's mems") wins a super prize for the largest number of
memory barriers entered into fast paths for one commit.

[get|put]_mems_allowed is incredibly heavy with pairs of full memory
barriers inserted into a number of hot paths.  This was detected while
investigating at large page allocator slowdown introduced some time
after 2.6.32.  The largest portion of this overhead was shown by
oprofile to be at an mfence introduced by this commit into the page
allocator hot path.

For extra style points, the commit introduced the use of yield() in an
implementation of what looks like a spinning mutex.

This patch replaces the full memory barriers on both read and write
sides with a sequence counter with just read barriers on the fast path
side.  This is much cheaper on some architectures, including x86.  The
main bulk of the patch is the retry logic if the nodemask changes in a
manner that can cause a false failure.

While updating the nodemask, a check is made to see if a false failure
is a risk.  If it is, the sequence number gets bumped and parallel
allocators will briefly stall while the nodemask update takes place.

In a page fault test microbenchmark, oprofile samples from
__alloc_pages_nodemask went from 4.53% of all samples to 1.15%.  The
actual results were

                             3.3.0-rc3          3.3.0-rc3
                             rc3-vanilla        nobarrier-v2r1
    Clients   1 UserTime       0.07 (  0.00%)   0.08 (-14.19%)
    Clients   2 UserTime       0.07 (  0.00%)   0.07 (  2.72%)
    Clients   4 UserTime       0.08 (  0.00%)   0.07 (  3.29%)
    Clients   1 SysTime        0.70 (  0.00%)   0.65 (  6.65%)
    Clients   2 SysTime        0.85 (  0.00%)   0.82 (  3.65%)
    Clients   4 SysTime        1.41 (  0.00%)   1.41 (  0.32%)
    Clients   1 WallTime       0.77 (  0.00%)   0.74 (  4.19%)
    Clients   2 WallTime       0.47 (  0.00%)   0.45 (  3.73%)
    Clients   4 WallTime       0.38 (  0.00%)   0.37 (  1.58%)
    Clients   1 Flt/sec/cpu  497620.28 (  0.00%) 520294.53 (  4.56%)
    Clients   2 Flt/sec/cpu  414639.05 (  0.00%) 429882.01 (  3.68%)
    Clients   4 Flt/sec/cpu  257959.16 (  0.00%) 258761.48 (  0.31%)
    Clients   1 Flt/sec      495161.39 (  0.00%) 517292.87 (  4.47%)
    Clients   2 Flt/sec      820325.95 (  0.00%) 850289.77 (  3.65%)
    Clients   4 Flt/sec      1020068.93 (  0.00%) 1022674.06 (  0.26%)
    MMTests Statistics: duration
    Sys Time Running Test (seconds)             135.68    132.17
    User+Sys Time Running Test (seconds)         164.2    160.13
    Total Elapsed Time (seconds)                123.46    120.87

The overall improvement is small but the System CPU time is much
improved and roughly in correlation to what oprofile reported (these
performance figures are without profiling so skew is expected).  The
actual number of page faults is noticeably improved.

For benchmarks like kernel builds, the overall benefit is marginal but
the system CPU time is slightly reduced.

To test the actual bug the commit fixed I opened two terminals.  The
first ran within a cpuset and continually ran a small program that
faulted 100M of anonymous data.  In a second window, the nodemask of the
cpuset was continually randomised in a loop.

Without the commit, the program would fail every so often (usually
within 10 seconds) and obviously with the commit everything worked fine.
With this patch applied, it also worked fine so the fix should be
functionally equivalent.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <cl@linux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 673596a..40de685 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -2380,8 +2380,9 @@
 {
 	enum zone_type high_zoneidx = gfp_zone(gfp_mask);
 	struct zone *preferred_zone;
-	struct page *page;
+	struct page *page = NULL;
 	int migratetype = allocflags_to_migratetype(gfp_mask);
+	unsigned int cpuset_mems_cookie;
 
 	gfp_mask &= gfp_allowed_mask;
 
@@ -2400,15 +2401,15 @@
 	if (unlikely(!zonelist->_zonerefs->zone))
 		return NULL;
 
-	get_mems_allowed();
+retry_cpuset:
+	cpuset_mems_cookie = get_mems_allowed();
+
 	/* The preferred zone is used for statistics later */
 	first_zones_zonelist(zonelist, high_zoneidx,
 				nodemask ? : &cpuset_current_mems_allowed,
 				&preferred_zone);
-	if (!preferred_zone) {
-		put_mems_allowed();
-		return NULL;
-	}
+	if (!preferred_zone)
+		goto out;
 
 	/* First allocation attempt */
 	page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
@@ -2418,9 +2419,19 @@
 		page = __alloc_pages_slowpath(gfp_mask, order,
 				zonelist, high_zoneidx, nodemask,
 				preferred_zone, migratetype);
-	put_mems_allowed();
 
 	trace_mm_page_alloc(page, order, gfp_mask, migratetype);
+
+out:
+	/*
+	 * When updating a task's mems_allowed, it is possible to race with
+	 * parallel threads in such a way that an allocation can fail while
+	 * the mask is being updated. If a page allocation is about to fail,
+	 * check if the cpuset changed during allocation and if so, retry.
+	 */
+	if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
+		goto retry_cpuset;
+
 	return page;
 }
 EXPORT_SYMBOL(__alloc_pages_nodemask);
@@ -2634,13 +2645,15 @@
 bool skip_free_areas_node(unsigned int flags, int nid)
 {
 	bool ret = false;
+	unsigned int cpuset_mems_cookie;
 
 	if (!(flags & SHOW_MEM_FILTER_NODES))
 		goto out;
 
-	get_mems_allowed();
-	ret = !node_isset(nid, cpuset_current_mems_allowed);
-	put_mems_allowed();
+	do {
+		cpuset_mems_cookie = get_mems_allowed();
+		ret = !node_isset(nid, cpuset_current_mems_allowed);
+	} while (!put_mems_allowed(cpuset_mems_cookie));
 out:
 	return ret;
 }