@@ -6,6 +6,7 @@
*/
#include "xfs.h"
#include "xfs_shared.h"
+#include "xfs_bit.h"
#include "xfs_format.h"
#include "xfs_log_format.h"
#include "xfs_trans_resv.h"
@@ -301,6 +302,46 @@ xfs_end_bio(
xfs_destroy_ioend(ioend, blk_status_to_errno(bio->bi_status));
}
+#if defined(DEBUG) || defined(XFS_WARN)
+/*
+ * DEBUG mode checks to ensure we never perform writeback to shared blocks. This
+ * is useful because writeback uses the presence of blocks in the COW fork to
+ * determine when COW is necessary. We always expect to have COW blocks covering
+ * shared blocks at writeback time, but if we ever get this wrong we're left
+ * with data corruptions that are difficult to detect. Check the extent state
+ * explicitly in DEBUG mode.
+ */
+static int
+xfs_assert_nonshared(
+ struct xfs_inode *ip,
+ uint64_t offset,
+ struct xfs_bmbt_irec *imap)
+{
+ struct xfs_mount *mp = ip->i_mount;
+ struct xfs_bmbt_irec tmap;
+ xfs_fileoff_t offset_fsb = XFS_B_TO_FSBT(mp, offset);
+ xfs_agblock_t fbno;
+ xfs_extlen_t flen;
+ int error;
+
+ if (!xfs_is_reflink_inode(ip))
+ return 0;
+
+ tmap = *imap; /* struct copy */
+ xfs_trim_extent(&tmap, offset_fsb, 1);
+ error = xfs_reflink_find_shared(mp, NULL,
+ XFS_FSB_TO_AGNO(mp, tmap.br_startblock),
+ XFS_FSB_TO_AGBNO(mp, tmap.br_startblock),
+ 1, &fbno, &flen, false);
+ if (error)
+ return error;
+ ASSERT(fbno == NULLAGBLOCK && flen == 0);
+ return 0;
+}
+#else
+#define xfs_assert_nonshared(ip, offset, imap) do { } while (0);
+#endif
+
STATIC int
xfs_map_blocks(
struct xfs_writepage_ctx *wpc,
@@ -723,6 +764,7 @@ xfs_writepage_map(
break;
if (wpc->io_type == XFS_IO_HOLE)
continue;
+ xfs_assert_nonshared(XFS_I(inode), file_offset, &wpc->imap);
xfs_add_to_ioend(inode, file_offset, page, iop, wpc, wbc,
&submit_list);
count++;
Signed-off-by: Brian Foster <bfoster@redhat.com> --- Here's a quick patch to try and do some explicit shared extent writeback checking. What's interesting so far is that this seems to allow reproducing the original shared/010 problem reliably with generic/127. That goes away with the previously posted patch to call xfs_trim_extent(), but if I reset my hardcoded shared/010 seed and start running that test in a loop I can eventually reproduce another instance of this assert failure that isn't detected by the checksum checks (i.e., shared/010 would pass if not for this particular assert firing). It's not immediately clear to me if this is another problem or a flaw in this patch. I'll have to dig into it further, so I'm just throwing this over the wall for the time being in case anybody has thoughts or finds it useful.. Brian fs/xfs/xfs_aops.c | 42 ++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 42 insertions(+)