[PATCH] Remove semi-softlockup from invalidate_mapping_pages
authorNeilBrown <neilb@suse.de>
Fri, 23 Jun 2006 09:05:48 +0000 (02:05 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Fri, 23 Jun 2006 14:43:07 +0000 (07:43 -0700)
If invalidate_mapping_pages is called to invalidate a very large mapping
(e.g.  a very large block device) and if the only active page in that
device is near the end (or at least, at a very large index), such as, say,
the superblock of an md array, and if that page happens to be locked when
invalidate_mapping_pages is called, then

  pagevec_lookup will return this page and
  as it is locked, 'next' will be incremented and pagevec_lookup
  will be called again. and again. and again.
  while we count from 0 upto a very large number.

We should really always set 'next' to 'page->index+1' before going around
the loop again, not just if the page isn't locked.

Cc: "Steinar H. Gunderson" <sgunderson@bigfoot.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
mm/truncate.c

index 6cb3fff25f670948d1ef519d7609793b54bfdca3..cf1b015df4a7d770d3bccdecb9b8cef62de529d0 100644 (file)
@@ -230,14 +230,24 @@ unsigned long invalidate_mapping_pages(struct address_space *mapping,
                        pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
                for (i = 0; i < pagevec_count(&pvec); i++) {
                        struct page *page = pvec.pages[i];
+                       pgoff_t index;
+                       int lock_failed;
 
-                       if (TestSetPageLocked(page)) {
-                               next++;
-                               continue;
-                       }
-                       if (page->index > next)
-                               next = page->index;
+                       lock_failed = TestSetPageLocked(page);
+
+                       /*
+                        * We really shouldn't be looking at the ->index of an
+                        * unlocked page.  But we're not allowed to lock these
+                        * pages.  So we rely upon nobody altering the ->index
+                        * of this (pinned-by-us) page.
+                        */
+                       index = page->index;
+                       if (index > next)
+                               next = index;
                        next++;
+                       if (lock_failed)
+                               continue;
+
                        if (PageDirty(page) || PageWriteback(page))
                                goto unlock;
                        if (page_mapped(page))