Revert "[PATCH] generic_file_buffered_write(): deadlock on vectored write"
Andrew Morton [Tue, 16 Oct 2007 08:24:54 +0000 (01:24 -0700)]
This reverts commit 6527c2bdf1f833cc18e8f42bd97973d583e4aa83, which
fixed the following bug:

  When prefaulting in the pages in generic_file_buffered_write(), we only
  faulted in the pages for the firts segment of the iovec.  If the second of
  successive segment described a mmapping of the page into which we're
  write()ing, and that page is not up-to-date, the fault handler tries to lock
  the already-locked page (to bring it up to date) and deadlocks.

  An exploit for this bug is in writev-deadlock-demo.c, in
  http://www.zip.com.au/~akpm/linux/patches/stuff/ext3-tools.tar.gz.

  (These demos assume blocksize < PAGE_CACHE_SIZE).

The problem with this fix is that it takes the kernel back to doing a single
prepare_write()/commit_write() per iovec segment.  So in the worst case we'll
run prepare_write+commit_write 1024 times where we previously would have run
it once. The other problem with the fix is that it fix all the locking problems.

<insert numbers obtained via ext3-tools's writev-speed.c here>

And apparently this change killed NFS overwrite performance, because, I
suppose, it talks to the server for each prepare_write+commit_write.

So just back that patch out - we'll be fixing the deadlock by other means.

Nick says: also it only ever actually papered over the bug, because after
faulting in the pages, they might be unmapped or reclaimed.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/filemap.c

index caaaa7a..4bf7d1a 100644 (file)
@@ -1865,21 +1865,14 @@ generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
        do {
                unsigned long index;
                unsigned long offset;
+               unsigned long maxlen;
                size_t copied;
 
                offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
                index = pos >> PAGE_CACHE_SHIFT;
                bytes = PAGE_CACHE_SIZE - offset;
-
-               /* Limit the size of the copy to the caller's write size */
-               bytes = min(bytes, count);
-
-               /*
-                * Limit the size of the copy to that of the current segment,
-                * because fault_in_pages_readable() doesn't know how to walk
-                * segments.
-                */
-               bytes = min(bytes, cur_iov->iov_len - iov_base);
+               if (bytes > count)
+                       bytes = count;
 
                /*
                 * Bring in the user page that we will copy from _first_.
@@ -1887,7 +1880,10 @@ generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
                 * same page as we're writing to, without it being marked
                 * up-to-date.
                 */
-               fault_in_pages_readable(buf, bytes);
+               maxlen = cur_iov->iov_len - iov_base;
+               if (maxlen > bytes)
+                       maxlen = bytes;
+               fault_in_pages_readable(buf, maxlen);
 
                page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
                if (!page) {