• Takashi Iwai's avatar
    ALSA: pcm: More fine-grained PCM link locking · f57f3df0
    Takashi Iwai authored
    We have currently two global locks, a rwlock and a rwsem, that are
    used for managing linking the PCM streams.  Due to these global locks,
    once when a linked stream is used, the lock granularity suffers a
    lot.
    
    This patch attempts to eliminate the former global lock for atomic
    ops.  The latter rwsem needs remaining because of the loosy way of the
    loop calls in snd_pcm_action_nonatomic(), as well as for avoiding the
    deadlock at linking.  However, these are used far rarely, actually
    only by two actions (prepare and  reset), where both are no timing
    critical ones.  So this can be still seen as a good improvement.
    
    The basic strategy to eliminate the rwlock is to assure group->lock at
    adding or removing a stream to / from the group.  Since we already
    takes the group lock whenever taking the all substream locks under the
    group, this shouldn't be a big problem.  The reference to group
    pointer in snd_pcm_substream object is protected by the stream lock
    itself.
    
    However, there are still pitfalls: a race window at re-locking and the
    lifecycle of group object.  The former is a small race window for
    dereferencing the substream group object opened while snd_pcm_action()
    performs re-locking to avoid ABBA deadlocks.  This includes the unlink
    of group during that window, too.  And the latter is the kfree
    performed after all streams are removed from the group while it's
    still dereferenced.
    
    For addressing these corner cases, two new tricks are introduced:
    - After re-locking, the group assigned to the stream is checked again;
      if the group is changed, we retry the whole procedure.
    - Introduce a refcount to snd_pcm_group object, so that it's freed
      only when it's empty and really no one refers to it.
    
    (Some readers might wonder why not RCU for the latter.  RCU in this
    case would cost more than refcounting, unfortunately.  We take the
    group lock sooner or later, hence the performance improvement by RCU
    would be negligible.  Meanwhile, because we need to deal with
    schedulable context depending on the pcm->nonatomic flag, it'll become
    dynamic RCU/SRCU switch, and the grace period may become too long.)
    
    Along with these changes, there are a significant amount of code
    refactoring.  The complex group re-lock & ref code is factored out to
    snd_pcm_stream_group_ref() function, for example.
    Signed-off-by: default avatarTakashi Iwai <tiwai@suse.de>
    f57f3df0
pcm.h 47.9 KB