]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
jbd2: drop checkpoint mutex when waiting in __jbd2_log_wait_for_space()
authorPaul Gortmaker <paul.gortmaker@windriver.com>
Thu, 13 Jun 2013 02:47:35 +0000 (22:47 -0400)
committerTheodore Ts'o <tytso@mit.edu>
Thu, 13 Jun 2013 02:47:35 +0000 (22:47 -0400)
While trying to debug an an issue under extreme I/O loading
on preempt-rt kernels, the following backtrace was observed
via SysRQ output:

rm              D ffff8802203afbc0  4600  4878   4748 0x00000000
 ffff8802217bfb78 0000000000000082 ffff88021fc2bb80 ffff88021fc2bb80
 ffff88021fc2bb80 ffff8802217bffd8 ffff8802217bffd8 ffff8802217bffd8
 ffff88021f1d4c80 ffff88021fc2bb80 ffff8802217bfb88 ffff88022437b000
Call Trace:
 [<ffffffff8172dc34>] schedule+0x24/0x70
 [<ffffffff81225b5d>] jbd2_log_wait_commit+0xbd/0x140
 [<ffffffff81060390>] ? __init_waitqueue_head+0x50/0x50
 [<ffffffff81223635>] jbd2_log_do_checkpoint+0xf5/0x520
 [<ffffffff81223b09>] __jbd2_log_wait_for_space+0xa9/0x1f0
 [<ffffffff8121dc40>] start_this_handle.isra.10+0x2e0/0x530
 [<ffffffff81060390>] ? __init_waitqueue_head+0x50/0x50
 [<ffffffff8121e0a3>] jbd2__journal_start+0xc3/0x110
 [<ffffffff811de7ce>] ? ext4_rmdir+0x6e/0x230
 [<ffffffff8121e0fe>] jbd2_journal_start+0xe/0x10
 [<ffffffff811f308b>] ext4_journal_start_sb+0x5b/0x160
 [<ffffffff811de7ce>] ext4_rmdir+0x6e/0x230
 [<ffffffff811435c5>] vfs_rmdir+0xd5/0x140
 [<ffffffff8114370f>] do_rmdir+0xdf/0x120
 [<ffffffff8105c6b4>] ? task_work_run+0x44/0x80
 [<ffffffff81002889>] ? do_notify_resume+0x89/0x100
 [<ffffffff817361ae>] ? int_signal+0x12/0x17
 [<ffffffff81145d85>] sys_unlinkat+0x25/0x40
 [<ffffffff81735f22>] system_call_fastpath+0x16/0x1b

What is interesting here, is that we call log_wait_commit, from
within wait_for_space, but we are still holding the checkpoint_mutex
as it surrounds mostly the whole of wait_for_space.  And then, as we
are waiting, journal_commit_transaction can run, and if the JBD2_FLUSHED
bit is set, then we will also try to take the same checkpoint_mutex.

It seems that we need to drop the checkpoint_mutex while sitting in
jbd2_log_wait_commit, if we want to guarantee that progress can be made
by jbd2_journal_commit_transaction().  There does not seem to be
anything preempt-rt specific about this, other then perhaps increasing
the odds of it happening.

Signed-off-by: Paul Gortmaker <paul.gortmaker@windriver.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
fs/jbd2/checkpoint.c

index 75a15f371b00c6e19a5e953e25b68be0ea391f03..7f34f4716165311f66b62b81dd25a32e899dcf80 100644 (file)
@@ -156,7 +156,15 @@ void __jbd2_log_wait_for_space(journal_t *journal)
                                /* We were able to recover space; yay! */
                                ;
                        } else if (tid) {
+                               /*
+                                * jbd2_journal_commit_transaction() may want
+                                * to take the checkpoint_mutex if JBD2_FLUSHED
+                                * is set.  So we need to temporarily drop it.
+                                */
+                               mutex_unlock(&journal->j_checkpoint_mutex);
                                jbd2_log_wait_commit(journal, tid);
+                               write_lock(&journal->j_state_lock);
+                               continue;
                        } else {
                                printk(KERN_ERR "%s: needed %d blocks and "
                                       "only had %d space available\n",