Thread (3 messages) 3 messages, 3 authors, 2012-07-27

Re: [PATCH] list corruption by gather_surplus

From: Aneesh Kumar K.V <hidden>
Date: 2012-07-27 17:52:13

Cliff Wickman [off-list ref] writes:
From: Cliff Wickman <redacted>

Gentlemen,
I see that you all have done maintenance on mm/hugetlb.c, so I'm hoping one
or two of you could comment on a problem and proposed fix.


I am seeing list corruption occurring from within gather_surplus_pages()
(mm/hugetlb.c).  The problem occurs under a heavy load, and seems to be
because this function drops the hugetlb_lock.

I have CONFIG_DEBUG_LIST=y, and am running an MPI application with 64 threads
and a library that creates a large heap of hugetlbfs pages for it.

The below patch fixes the problem.
The gist of this patch is that gather_surplus_pages() does not have to drop
the lock if alloc_buddy_huge_page() is told whether the lock is
already held.

But you didn't explain the corruption details right ? What cause the
corruption ? It would be nice to document that in the commit.
quoted hunk ↗ jump to hunk
But I may be missing some reason why gather_surplus_pages() is unlocking and
locking the hugetlb_lock several times (besides around the allocator).

Could you take a look and advise?

Signed-off-by: Cliff Wickman <redacted>
---
 mm/hugetlb.c |   28 +++++++++++++++++-----------
 1 file changed, 17 insertions(+), 11 deletions(-)

Index: linux/mm/hugetlb.c
===================================================================
--- linux.orig/mm/hugetlb.c
+++ linux/mm/hugetlb.c
@@ -747,7 +747,9 @@ static int free_pool_huge_page(struct hs
 	return ret;
 }

-static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
+/* already_locked means the caller has already locked hugetlb_lock */
+static struct page *alloc_buddy_huge_page(struct hstate *h, int nid,
+						int already_locked)
 {
Why ? Why can't we always call this with lock held ?
quoted hunk ↗ jump to hunk
 	struct page *page;
 	unsigned int r_nid;
@@ -778,7 +780,8 @@ static struct page *alloc_buddy_huge_pag
 	 * the node values until we've gotten the hugepage and only the
-aneesh

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help