xfs: truncate delalloc extents when IO fails in writeback
authorDave Chinner <dchinner@redhat.com>
Fri, 5 Mar 2010 02:00:42 +0000 (02:00 +0000)
committerAlex Elder <aelder@sgi.com>
Fri, 5 Mar 2010 17:01:53 +0000 (11:01 -0600)
We currently use block_invalidatepage() to clean up pages where I/O
fails in ->writepage(). Unfortunately, if the page has delalloc
regions on it, we fail to remove the delalloc regions when we
invalidate the page.  This can result in tripping a BUG() in
xfs_get_blocks() later on if a direct IO read is done on that same
region - the delalloc extent is returned when none is supposed to be
there.

Fix this by truncating away the delalloc regions on the page before
invalidating it. Because they are delalloc, we can do this without
needing a transaction. Indeed - if we get ENOSPC errors, we have to
be able to do this truncation without a transaction as there is
no space left for block reservation (typically why we see a ENOSPC
in writeback).

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
fs/xfs/linux-2.6/xfs_aops.c

index 793908a8a09b11e9297b369638b1f2cf3a344131..9083357f9e444170b5b95124e936bb0e0d430c7f 100644 (file)
@@ -39,6 +39,7 @@
 #include "xfs_iomap.h"
 #include "xfs_vnodeops.h"
 #include "xfs_trace.h"
+#include "xfs_bmap.h"
 #include <linux/mpage.h>
 #include <linux/pagevec.h>
 #include <linux/writeback.h>
@@ -893,6 +894,118 @@ xfs_cluster_write(
        }
 }
 
+STATIC void
+xfs_vm_invalidatepage(
+       struct page             *page,
+       unsigned long           offset)
+{
+       trace_xfs_invalidatepage(page->mapping->host, page, offset);
+       block_invalidatepage(page, offset);
+}
+
+/*
+ * If the page has delalloc buffers on it, we need to punch them out before we
+ * invalidate the page. If we don't, we leave a stale delalloc mapping on the
+ * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
+ * is done on that same region - the delalloc extent is returned when none is
+ * supposed to be there.
+ *
+ * We prevent this by truncating away the delalloc regions on the page before
+ * invalidating it. Because they are delalloc, we can do this without needing a
+ * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
+ * truncation without a transaction as there is no space left for block
+ * reservation (typically why we see a ENOSPC in writeback).
+ *
+ * This is not a performance critical path, so for now just do the punching a
+ * buffer head at a time.
+ */
+STATIC void
+xfs_aops_discard_page(
+       struct page             *page)
+{
+       struct inode            *inode = page->mapping->host;
+       struct xfs_inode        *ip = XFS_I(inode);
+       struct buffer_head      *bh, *head;
+       loff_t                  offset = page_offset(page);
+       ssize_t                 len = 1 << inode->i_blkbits;
+
+       if (!xfs_is_delayed_page(page, IOMAP_DELAY))
+               goto out_invalidate;
+
+       xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
+               "page discard on page %p, inode 0x%llx, offset %llu.",
+                       page, ip->i_ino, offset);
+
+       xfs_ilock(ip, XFS_ILOCK_EXCL);
+       bh = head = page_buffers(page);
+       do {
+               int             done;
+               xfs_fileoff_t   offset_fsb;
+               xfs_bmbt_irec_t imap;
+               int             nimaps = 1;
+               int             error;
+               xfs_fsblock_t   firstblock;
+               xfs_bmap_free_t flist;
+
+               if (!buffer_delay(bh))
+                       goto next_buffer;
+
+               offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
+
+               /*
+                * Map the range first and check that it is a delalloc extent
+                * before trying to unmap the range. Otherwise we will be
+                * trying to remove a real extent (which requires a
+                * transaction) or a hole, which is probably a bad idea...
+                */
+               error = xfs_bmapi(NULL, ip, offset_fsb, 1,
+                               XFS_BMAPI_ENTIRE,  NULL, 0, &imap,
+                               &nimaps, NULL, NULL);
+
+               if (error) {
+                       /* something screwed, just bail */
+                       xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
+                       "page discard failed delalloc mapping lookup.");
+                       break;
+               }
+               if (!nimaps) {
+                       /* nothing there */
+                       goto next_buffer;
+               }
+               if (imap.br_startblock != DELAYSTARTBLOCK) {
+                       /* been converted, ignore */
+                       goto next_buffer;
+               }
+               WARN_ON(imap.br_blockcount == 0);
+
+               /*
+                * Note: while we initialise the firstblock/flist pair, they
+                * should never be used because blocks should never be
+                * allocated or freed for a delalloc extent and hence we need
+                * don't cancel or finish them after the xfs_bunmapi() call.
+                */
+               xfs_bmap_init(&flist, &firstblock);
+               error = xfs_bunmapi(NULL, ip, offset_fsb, 1, 0, 1, &firstblock,
+                                       &flist, NULL, &done);
+
+               ASSERT(!flist.xbf_count && !flist.xbf_first);
+               if (error) {
+                       /* something screwed, just bail */
+                       xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
+                       "page discard unable to remove delalloc mapping.");
+                       break;
+               }
+next_buffer:
+               offset += len;
+
+       } while ((bh = bh->b_this_page) != head);
+
+       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+out_invalidate:
+       xfs_vm_invalidatepage(page, 0);
+       return;
+}
+
 /*
  * Calling this without startio set means we are being asked to make a dirty
  * page ready for freeing it's buffers.  When called with startio set then
@@ -1144,7 +1257,7 @@ error:
         */
        if (err != -EAGAIN) {
                if (!unmapped)
-                       block_invalidatepage(page, 0);
+                       xfs_aops_discard_page(page);
                ClearPageUptodate(page);
        }
        return err;
@@ -1554,15 +1667,6 @@ xfs_vm_readpages(
        return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks);
 }
 
-STATIC void
-xfs_vm_invalidatepage(
-       struct page             *page,
-       unsigned long           offset)
-{
-       trace_xfs_invalidatepage(page->mapping->host, page, offset);
-       block_invalidatepage(page, offset);
-}
-
 const struct address_space_operations xfs_address_space_operations = {
        .readpage               = xfs_vm_readpage,
        .readpages              = xfs_vm_readpages,