On Jun 6, 2011, at 10:46 PM, Miklos Szeredi wrote: On Mon, 2011-06-06 at 15:16 -0700, Andrew Morton wrote: On Mon, 30 May 2011 11:37:38 +0200
Jan Kara <jack@xxxxxxx> wrote:
Under heavy memory and filesystem load, users observe the assertion
mapping->nrpages == 0 in end_writeback() trigger. This can be caused
by page reclaim reclaiming the last page from a mapping in the following
race:
CPU0 CPU1
...
shrink_page_list()
__remove_mapping()
__delete_from_page_cache()
radix_tree_delete()
evict_inode()
truncate_inode_pages()
truncate_inode_pages_range()
pagevec_lookup() - finds nothing
end_writeback()
mapping->nrpages != 0 -> BUG
page->mapping = NULL
mapping->nrpages--
Fix the problem by cycling the mapping->tree_lock at the end of
truncate_inode_pages_range() to synchronize with page reclaim.
Analyzed by Jay <jinshan.xiong@xxxxxxxxxxxxx>, lost in LKML, and dug
out by Miklos Szeredi <mszeredi@xxxxxxx>.
CC: Jay <jinshan.xiong@xxxxxxxxxxxxx>
CC: stable@xxxxxxxxxx
Acked-by: Miklos Szeredi <mszeredi@xxxxxxx>
Signed-off-by: Jan Kara <jack@xxxxxxx>
---
mm/truncate.c | 7 +++++++
1 files changed, 7 insertions(+), 0 deletions(-)
Andrew, would you merge this patch please? Thanks.
diff --git a/mm/truncate.c b/mm/truncate.c
index a956675..ec3d292 100644
--- a/mm/truncate.c
+++ b/mm/truncate.c
@@ -291,6 +291,13 @@ void truncate_inode_pages_range(struct address_space *mapping,
pagevec_release(&pvec);
mem_cgroup_uncharge_end();
}
+ /*
+ * Cycle the tree_lock to make sure all __delete_from_page_cache()
+ * calls run from page reclaim have finished as well (this handles the
+ * case when page reclaim took the last page from our range).
+ */
+ spin_lock_irq(&mapping->tree_lock);
+ spin_unlock_irq(&mapping->tree_lock);
}
EXPORT_SYMBOL(truncate_inode_pages_range);
That's one ugly patch.
Perhaps this regression was added by Nick's RCUification of pagecache.
Before that patch, mapping->nrpages and the radix-tree state were
coherent for holders of tree_lock. So pagevec_lookup() would never
return "no pages" while ->nrpages is non-zero.
After that patch, find_get_pages() uses RCU to protect the radix-tree
but I don't think it correctly protects the aggregate (radix-tree +
nrpages).
Yes, that's the case.
If it's not that then I see another possibility.
truncate_inode_pages_range() does
if (mapping->nrpages == 0)
return;
Is there anything to prevent a page getting added to the inode _after_
this test? i_mutex? If not, that would trigger the BUG.
That BUG is in the inode eviction phase, so there's nothing that could be adding a page. And the only thing that could be removing one is page reclaim. Either way, I don't think that the uglypatch expresses a full
understanding of te bug ;)
I don't see a better way, how would we make nrpages update atomically wrt the radix-tree while using only RCU? The question is, does it matter that those two can get temporarily out of sync? In case of inode eviction it does, not only because of that BUG_ON, but because page reclaim must be somehow synchronised with eviction. Otherwise it may access tree_lock on the mapping of an already freed inode.
I tend to think your patch is absolutely ok to fix this problem. However, I think it would be better to move:
spin_lock(&mapping->tree_lock); spin_unlock(&mapping->tree_lock);
into end_writeback(). This is because truncate_inode_pages_range() is a generic function and it will be called somewhere else, maybe unnecessarily to do this extra thing.
Actually, I'd like to hold an inode refcount in page stealing process. The reason is obvious: it makes no sense to steal pages from a to-be-freed inode. However, the problem is the overhead to grab an inode is damned heavy.
Thanks, Jinshan
In other cases? AFAICS it doesn't matter. Most ->nrpages accesses
weren't under tree_lock before Nick's RCUification, so their use were
just optimization.
Thanks,
Miklos
|