From: Jan Kara <[email protected]>
Update information about locking in JBD revoke code. Inconsistency in comments
found by Lin Tan <[email protected]>.
CC: Lin Tan <[email protected]>
Signed-off-by: Jan Kara <[email protected]>
fs/jbd/revoke.c | 24 +++++++++++++++++++-----
1 files changed, 19 insertions(+), 5 deletions(-)
Andrew, could you please merge the patch? Ted has already merged JBD2
diff --git a/fs/jbd/revoke.c b/fs/jbd/revoke.c
index c7bd649..3e9afc2 100644
@@ -55,6 +55,25 @@
* need do nothing.
* RevokeValid set, Revoked set:
* buffer has been revoked.
+ * Locking rules:
+ * We keep two hash tables of revoke records. One hashtable belongs to the
+ * running transaction (is pointed to by journal->j_revoke), the other one
+ * belongs to the committing transaction. Accesses to the second hash table
+ * happen only from the kjournald and no other thread touches this table. Also
+ * journal_switch_revoke_table() which switches which hashtable belongs to the
+ * running and which to the committing transaction is called only from
+ * kjournald. Therefore we need no locks when accessing the hashtable belonging
+ * to the committing transaction.
+ * All users operating on the hash table belonging to the running transaction
+ * have a handle to the transaction. Therefore they are safe from kjournald
+ * switching hash tables under them. For operations on the lists of entries in
+ * the hash table j_revoke_lock is used.
+ * Finally, also replay code uses the hash tables but at this moment noone else
+ * can touch them (filesystem isn't mounted yet) and hence no locking is
+ * needed.
@@ -402,8 +421,6 @@ int journal_revoke(handle_t *handle, unsigned long blocknr,
* the second time we would still have a pending revoke to cancel. So,
* do not trust the Revoked bit on buffers unless RevokeValid is also
- * The caller must have the journal locked.
int journal_cancel_revoke(handle_t *handle, struct journal_head *jh)
@@ -481,10 +498,7 @@ void journal_switch_revoke_table(journal_t *journal)
* Write revoke records to the journal for all entries in the current
* revoke hash, deleting the entries as we go.
- * Called with the journal lock held.