• Paul Gortmaker's avatar
    jbd2: drop checkpoint mutex when waiting in __jbd2_log_wait_for_space() · 0ef54180
    Paul Gortmaker authored
    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: 's avatarPaul Gortmaker <paul.gortmaker@windriver.com>
    Signed-off-by: 's avatar"Theodore Ts'o" <tytso@mit.edu>
    0ef54180
checkpoint.c 20 KB