tmpfs: fix shmem_swaplist races
authorHugh Dickins <hugh@veritas.com>
Tue, 5 Feb 2008 06:28:55 +0000 (22:28 -0800)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Tue, 5 Feb 2008 17:44:16 +0000 (09:44 -0800)
Intensive swapoff testing shows shmem_unuse spinning on an entry in
shmem_swaplist pointing to itself: how does that come about?  Days pass...

First guess is this: shmem_delete_inode tests list_empty without taking the
global mutex (so the swapping case doesn't slow down the common case); but
there's an instant in shmem_unuse_inode's list_move_tail when the list entry
may appear empty (a rare case, because it's actually moving the head not the
the list member).  So there's a danger of leaving the inode on the swaplist
when it's freed, then reinitialized to point to itself when reused.  Fix that
by skipping the list_move_tail when it's a no-op, which happens to plug this.

But this same spinning then surfaces on another machine.  Ah, I'd never
suspected it, but shmem_writepage's swaplist manipulation is unsafe: though we
still hold page lock, which would hold off inode deletion if the page were in
pagecache, it doesn't hold off once it's in swapcache (free_swap_and_cache
doesn't wait on locked pages).  Hmm: we could put the the inode on swaplist
earlier, but then shmem_unuse_inode could never prune unswapped inodes.

Fix this with an igrab before dropping info->lock, as in shmem_unuse_inode;
though I am a little uneasy about the iput which has to follow - it works, and
I see nothing wrong with it, but it is surprising that shmem inode deletion
may now occur below shmem_writepage.  Revisit this fix later?

And while we're looking at these races: the way shmem_unuse tests swapped
without holding info->lock looks unsafe, if we've more than one swap area: a
racing shmem_writepage on another page of the same inode could be putting it
in swapcache, just as we're deciding to remove the inode from swaplist -
there's a danger of going on swap without being listed, so a later swapoff
would hang, being unable to locate the entry.  Move that test and removal down
into shmem_unuse_inode, once info->lock is held.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/shmem.c

index 530c5033d028f0ba302704083574a94d03f01b0a..ee9024483f6094bac4d6bdac8b798d11e8846fe7 100644 (file)
@@ -833,6 +833,10 @@ static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, s
        idx = 0;
        ptr = info->i_direct;
        spin_lock(&info->lock);
+       if (!info->swapped) {
+               list_del_init(&info->swaplist);
+               goto lost2;
+       }
        limit = info->next_index;
        size = limit;
        if (size > SHMEM_NR_DIRECT)
@@ -894,8 +898,15 @@ found:
        inode = igrab(&info->vfs_inode);
        spin_unlock(&info->lock);
 
-       /* move head to start search for next from here */
-       list_move_tail(&shmem_swaplist, &info->swaplist);
+       /*
+        * Move _head_ to start search for next from here.
+        * But be careful: shmem_delete_inode checks list_empty without taking
+        * mutex, and there's an instant in list_move_tail when info->swaplist
+        * would appear empty, if it were the only one on shmem_swaplist.  We
+        * could avoid doing it if inode NULL; or use this minor optimization.
+        */
+       if (shmem_swaplist.next != &info->swaplist)
+               list_move_tail(&shmem_swaplist, &info->swaplist);
        mutex_unlock(&shmem_swaplist_mutex);
 
        error = 1;
@@ -955,10 +966,7 @@ int shmem_unuse(swp_entry_t entry, struct page *page)
        mutex_lock(&shmem_swaplist_mutex);
        list_for_each_safe(p, next, &shmem_swaplist) {
                info = list_entry(p, struct shmem_inode_info, swaplist);
-               if (info->swapped)
-                       found = shmem_unuse_inode(info, entry, page);
-               else
-                       list_del_init(&info->swaplist);
+               found = shmem_unuse_inode(info, entry, page);
                cond_resched();
                if (found)
                        goto out;
@@ -1021,18 +1029,23 @@ static int shmem_writepage(struct page *page, struct writeback_control *wbc)
                remove_from_page_cache(page);
                shmem_swp_set(info, entry, swap.val);
                shmem_swp_unmap(entry);
+               if (list_empty(&info->swaplist))
+                       inode = igrab(inode);
+               else
+                       inode = NULL;
                spin_unlock(&info->lock);
-               if (list_empty(&info->swaplist)) {
-                       mutex_lock(&shmem_swaplist_mutex);
-                       /* move instead of add in case we're racing */
-                       list_move_tail(&info->swaplist, &shmem_swaplist);
-                       mutex_unlock(&shmem_swaplist_mutex);
-               }
                swap_duplicate(swap);
                BUG_ON(page_mapped(page));
                page_cache_release(page);       /* pagecache ref */
                set_page_dirty(page);
                unlock_page(page);
+               if (inode) {
+                       mutex_lock(&shmem_swaplist_mutex);
+                       /* move instead of add in case we're racing */
+                       list_move_tail(&info->swaplist, &shmem_swaplist);
+                       mutex_unlock(&shmem_swaplist_mutex);
+                       iput(inode);
+               }
                return 0;
        }