The current code deals with the stream start / stop solely via
line6_pcm_acquire() and line6_pcm_release(). This was (supposedly)
intended to avoid the races, but it doesn't work as expected. The
concurrent acquire and release calls can be performed without proper
protections, thus this might result in memory corruption.
Furthermore, we can't take a mutex to protect the whole function
because it can be called from the PCM trigger callback that is an
atomic context. Also spinlock isn't appropriate because the function
allocates with kmalloc with GFP_KERNEL. That is, these function just
lead to singular problems.
This is an attempt to reduce the existing races. First off, separate
both the stream buffer management and the stream URB management. The
former is protected via a newly introduced state_mutex while the
latter is protected via each line6_pcm_stream lock.
Secondly, the stream state are now managed in opened and running bit
flags of each line6_pcm_stream. Not only this a bit clearer than
previous combined bit flags, this also gives a better abstraction.
These rewrites allows us to make common hw_params and hw_free
callbacks for both playback and capture directions.
For the monitor and impulse operations, still line6_pcm_acquire() and
line6_pcm_release() are used. They call internally the corresponding
functions for both playback and capture streams with proper lock or
mutex. Unlike the previous versions, these function don't take the
bit masks but the only single type value. Also they are supposed to
be applied only as duplex operations.
Tested-by: Chris Rorvick <chris@rorvick.com>
Signed-off-by: Takashi Iwai <tiwai@suse.de>
63e20df1e5