• Frederic Weisbecker's avatar
    reiserfs: kill-the-BKL · 8ebc4232
    Frederic Weisbecker authored
    This patch is an attempt to remove the Bkl based locking scheme from
    reiserfs and is intended.
    
    It is a bit inspired from an old attempt by Peter Zijlstra:
    
       http://lkml.indiana.edu/hypermail/linux/kernel/0704.2/2174.html
    
    The bkl is heavily used in this filesystem to prevent from
    concurrent write accesses on the filesystem.
    
    Reiserfs makes a deep use of the specific properties of the Bkl:
    
    - It can be acqquired recursively by a same task
    - It is released on the schedule() calls and reacquired when schedule() returns
    
    The two properties above are a roadmap for the reiserfs write locking so it's
    very hard to simply replace it with a common mutex.
    
    - We need a recursive-able locking unless we want to restructure several blocks
      of the code.
    - We need to identify the sites where the bkl was implictly relaxed
      (schedule, wait, sync, etc...) so that we can in turn release and
      reacquire our new lock explicitly.
      Such implicit releases of the lock are often required to let other
      resources producer/consumer do their job or we can suffer unexpected
      starvations or deadlocks.
    
    So the new lock that replaces the bkl here is a per superblock mutex with a
    specific property: it can be acquired recursively by a same task, like the
    bkl.
    
    For such purpose, we integrate a lock owner and a lock depth field on the
    superblock information structure.
    
    The first axis on this patch is to turn reiserfs_write_(un)lock() function
    into a wrapper to manage this mutex. Also some explicit calls to
    lock_kernel() have been converted to reiserfs_write_lock() helpers.
    
    The second axis is to find the important blocking sites (schedule...(),
    wait_on_buffer(), sync_dirty_buffer(), etc...) and then apply an explicit
    release of the write lock on these locations before blocking. Then we can
    safely wait for those who can give us resources or those who need some.
    Typically this is a fight between the current writer, the reiserfs workqueue
    (aka the async commiter) and the pdflush threads.
    
    The third axis is a consequence of the second. The write lock is usually
    on top of a lock dependency chain which can include the journal lock, the
    flush lock or the commit lock. So it's dangerous to release and trying to
    reacquire the write lock while we still hold other locks.
    
    This is fine with the bkl:
    
          T1                       T2
    
    lock_kernel()
        mutex_lock(A)
        unlock_kernel()
        // do something
                                lock_kernel()
                                    mutex_lock(A) -> already locked by T1
                                    schedule() (and then unlock_kernel())
        lock_kernel()
        mutex_unlock(A)
        ....
    
    This is not fine with a mutex:
    
          T1                       T2
    
    mutex_lock(write)
        mutex_lock(A)
        mutex_unlock(write)
        // do something
                               mutex_lock(write)
                                  mutex_lock(A) -> already locked by T1
                                  schedule()
    
        mutex_lock(write) -> already locked by T2
        deadlock
    
    The solution in this patch is to provide a helper which releases the write
    lock and sleep a bit if we can't lock a mutex that depend on it. It's another
    simulation of the bkl behaviour.
    
    The last axis is to locate the fs callbacks that are called with the bkl held,
    according to Documentation/filesystem/Locking.
    
    Those are:
    
    - reiserfs_remount
    - reiserfs_fill_super
    - reiserfs_put_super
    
    Reiserfs didn't need to explicitly lock because of the context of these callbacks.
    But now we must take care of that with the new locking.
    
    After this patch, reiserfs suffers from a slight performance regression (for now).
    On UP, a high volume write with dd reports an average of 27 MB/s instead
    of 30 MB/s without the patch applied.
    Signed-off-by: default avatarFrederic Weisbecker <fweisbec@gmail.com>
    Reviewed-by: default avatarIngo Molnar <mingo@elte.hu>
    Cc: Jeff Mahoney <jeffm@suse.com>
    Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
    Cc: Bron Gondwana <brong@fastmail.fm>
    Cc: Andrew Morton <akpm@linux-foundation.org>
    Cc: Linus Torvalds <torvalds@linux-foundation.org>
    Cc: Alexander Viro <viro@zeniv.linux.org.uk>
    LKML-Reference: <1239070789-13354-1-git-send-email-fweisbec@gmail.com>
    Signed-off-by: default avatarIngo Molnar <mingo@elte.hu>
    8ebc4232
ioctl.c 5.48 KB