libxfs: don't hardcode cycle 1 into unmount op header

The libxfs helper to write a log record after zeroing the log fills much
of the record header and unmount record with dummy data. It also
hardcodes the cycle number into the transaction oh_tid field as the
kernel expects to find the cycle stamped at the top of each block and
the original oh_tid value packed into h_cycle_data of the record header.

The log clearing code requires the ability to format the log to an
arbitrary cycle number to fix v5 superblock log recovery ordering
problems. As a result, the unmount record helper must not hardcode a
cycle of 1.

Fix up libxfs_log_header() to pack the unmount record appropriately, as
is already done for extra blocks that might exist beyond the record. Use
h_cycle_data for the original 32-bit word of the log record data block
and stamp the cycle number in its place. This allows unmount_record() to
work for arbitrary cycle numbers and libxfs_log_header() to pack a cycle
value that matches the lsn used in the record header. Note that this
patch does not change behavior as the lsn is still hardcoded to (1:0).

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>

diff --git a/libxfs/rdwr.c b/libxfs/rdwr.c
index bc776992278e23048c30689e40cb82d4eda95762..282f22d7990357d394013519b0ef89b46972d084 100644 (file)
--- a/libxfs/rdwr.c
+++ b/libxfs/rdwr.c
@@ -122,7 +122,8 @@ static void unmount_record(void *p)
        } magic = { XLOG_UNMOUNT_TYPE, 0, 0 };
 
        memset(p, 0, BBSIZE);
-       op->oh_tid = cpu_to_be32(1);
+       /* dummy tid to mark this as written from userspace */
+       op->oh_tid = cpu_to_be32(0xb0c0d0d0);
        op->oh_len = cpu_to_be32(sizeof(magic));
        op->oh_clientid = XFS_LOG;
        op->oh_flags = XLOG_UNMOUNT_TRANS;
@@ -188,10 +189,6 @@ libxfs_log_header(
 
        len = ((version == 2) && sunit) ? BTOBB(sunit) : 1;
 
-       /* note that oh_tid actually contains the cycle number
-        * and the tid is stored in h_cycle_data[0] - that's the
-        * way things end up on disk.
-        */
        memset(p, 0, BBSIZE);
        head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
        head->h_cycle = cpu_to_be32(1);
@@ -203,7 +200,6 @@ libxfs_log_header(
        head->h_crc = cpu_to_le32(0);
        head->h_prev_block = cpu_to_be32(-1);
        head->h_num_logops = cpu_to_be32(1);
-       head->h_cycle_data[0] = cpu_to_be32(0xb0c0d0d0);
        head->h_fmt = cpu_to_be32(fmt);
        head->h_size = cpu_to_be32(XLOG_HEADER_CYCLE_SIZE);
 
@@ -212,11 +208,25 @@ libxfs_log_header(
 
        memcpy(&head->h_fs_uuid, fs_uuid, sizeof(uuid_t));
 
-       len = MAX(len, 2);
        p = nextfunc(p, BBSIZE, private);
        unmount_record(p);
 
+       /*
+        * The kernel expects to see either a log record header magic or the LSN
+        * cycle at the top of every log block (for example, see
+        * xlog_[un]pack_data() and xlog_get_cycle()). Pack the unmount record
+        * block appropriately here.
+        */
        cycle_lsn = CYCLE_LSN_DISK(head->h_lsn);
+       head->h_cycle_data[0] = *(__be32 *)p;
+       *(__be32 *)p = cycle_lsn;
+
+       /*
+        * Now zero any remaining blocks in the record and stamp with the cycle.
+        * Note that we don't need to swap into h_cycle_data because it has
+        * already been initialized to zero.
+        */
+       len = MAX(len, 2);
        for (i = 2; i < len; i++) {
                p = nextfunc(p, BBSIZE, private);
                memset(p, 0, BBSIZE);