ext4: fix NULL pointer dereference when journal restart fails
authorLukas Czerner <lczerner@redhat.com>
Thu, 14 May 2015 22:55:18 +0000 (18:55 -0400)
committerTheodore Ts'o <tytso@mit.edu>
Thu, 14 May 2015 22:55:18 +0000 (18:55 -0400)
Currently when journal restart fails, we'll have the h_transaction of
the handle set to NULL to indicate that the handle has been effectively
aborted. We handle this situation quietly in the jbd2_journal_stop() and just
free the handle and exit because everything else has been done before we
attempted (and failed) to restart the journal.

Unfortunately there are a number of problems with that approach
introduced with commit

41a5b913197c "jbd2: invalidate handle if jbd2_journal_restart()
fails"

First of all in ext4 jbd2_journal_stop() will be called through
__ext4_journal_stop() where we would try to get a hold of the superblock
by dereferencing h_transaction which in this case would lead to NULL
pointer dereference and crash.

In addition we're going to free the handle regardless of the refcount
which is bad as well, because others up the call chain will still
reference the handle so we might potentially reference already freed
memory.

Moreover it's expected that we'll get aborted handle as well as detached
handle in some of the journalling function as the error propagates up
the stack, so it's unnecessary to call WARN_ON every time we get
detached handle.

And finally we might leak some memory by forgetting to free reserved
handle in jbd2_journal_stop() in the case where handle was detached from
the transaction (h_transaction is NULL).

Fix the NULL pointer dereference in __ext4_journal_stop() by just
calling jbd2_journal_stop() quietly as suggested by Jan Kara. Also fix
the potential memory leak in jbd2_journal_stop() and use proper
handle refcounting before we attempt to free it to avoid use-after-free
issues.

And finally remove all WARN_ON(!transaction) from the code so that we do
not get random traces when something goes wrong because when journal
restart fails we will get to some of those functions.

Cc: stable@vger.kernel.org
Signed-off-by: Lukas Czerner <lczerner@redhat.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Jan Kara <jack@suse.cz>
fs/ext4/ext4_jbd2.c
fs/jbd2/transaction.c

index 3445035c7e015e9460f2cd3f74b698bf823cabb6..d4184318181878b023931078377ae17c9568a208 100644 (file)
@@ -87,6 +87,12 @@ int __ext4_journal_stop(const char *where, unsigned int line, handle_t *handle)
                ext4_put_nojournal(handle);
                return 0;
        }
+
+       if (!handle->h_transaction) {
+               err = jbd2_journal_stop(handle);
+               return handle->h_err ? handle->h_err : err;
+       }
+
        sb = handle->h_transaction->t_journal->j_private;
        err = handle->h_err;
        rc = jbd2_journal_stop(handle);
index 5f09370c90a8199647a87ef7cb6a5cfdba113d2b..ff2f2e6ad3114664cbcd80c3812b6b875f64807b 100644 (file)
@@ -551,7 +551,6 @@ int jbd2_journal_extend(handle_t *handle, int nblocks)
        int result;
        int wanted;
 
-       WARN_ON(!transaction);
        if (is_handle_aborted(handle))
                return -EROFS;
        journal = transaction->t_journal;
@@ -627,7 +626,6 @@ int jbd2__journal_restart(handle_t *handle, int nblocks, gfp_t gfp_mask)
        tid_t           tid;
        int             need_to_start, ret;
 
-       WARN_ON(!transaction);
        /* If we've had an abort of any type, don't even think about
         * actually doing the restart! */
        if (is_handle_aborted(handle))
@@ -785,7 +783,6 @@ do_get_write_access(handle_t *handle, struct journal_head *jh,
        int need_copy = 0;
        unsigned long start_lock, time_lock;
 
-       WARN_ON(!transaction);
        if (is_handle_aborted(handle))
                return -EROFS;
        journal = transaction->t_journal;
@@ -1051,7 +1048,6 @@ int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
        int err;
 
        jbd_debug(5, "journal_head %p\n", jh);
-       WARN_ON(!transaction);
        err = -EROFS;
        if (is_handle_aborted(handle))
                goto out;
@@ -1266,7 +1262,6 @@ int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
        struct journal_head *jh;
        int ret = 0;
 
-       WARN_ON(!transaction);
        if (is_handle_aborted(handle))
                return -EROFS;
        journal = transaction->t_journal;
@@ -1397,7 +1392,6 @@ int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
        int err = 0;
        int was_modified = 0;
 
-       WARN_ON(!transaction);
        if (is_handle_aborted(handle))
                return -EROFS;
        journal = transaction->t_journal;
@@ -1530,8 +1524,22 @@ int jbd2_journal_stop(handle_t *handle)
        tid_t tid;
        pid_t pid;
 
-       if (!transaction)
-               goto free_and_exit;
+       if (!transaction) {
+               /*
+                * Handle is already detached from the transaction so
+                * there is nothing to do other than decrease a refcount,
+                * or free the handle if refcount drops to zero
+                */
+               if (--handle->h_ref > 0) {
+                       jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
+                                                        handle->h_ref);
+                       return err;
+               } else {
+                       if (handle->h_rsv_handle)
+                               jbd2_free_handle(handle->h_rsv_handle);
+                       goto free_and_exit;
+               }
+       }
        journal = transaction->t_journal;
 
        J_ASSERT(journal_current_handle() == handle);
@@ -2373,7 +2381,6 @@ int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode)
        transaction_t *transaction = handle->h_transaction;
        journal_t *journal;
 
-       WARN_ON(!transaction);
        if (is_handle_aborted(handle))
                return -EROFS;
        journal = transaction->t_journal;