| Message ID | 20170820044908.17972-4-o-takashi@sakamocchi.jp (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
diff --git a/sound/core/control.c b/sound/core/control.c index 1c1fc0898afb..249140c15d64 100644 --- a/sound/core/control.c +++ b/sound/core/control.c @@ -967,9 +967,9 @@ static int snd_ctl_elem_write_user(struct snd_ctl_file *file, snd_power_lock(card); result = snd_power_wait(card, SNDRV_CTL_POWER_D0); if (result >= 0) { - down_read(&card->controls_rwsem); + down_write(&card->controls_rwsem); result = snd_ctl_elem_write(card, file, control); - up_read(&card->controls_rwsem); + up_write(&card->controls_rwsem); } snd_power_unlock(card); if (result >= 0)
In ALSA control interface, applications can execute two types of request for value of members on each element; ELEM_READ and ELEM_WRITE. In ALSA control core, these two requests are handled within read lock of a counting semaphore, therefore several processes can run to execute these two requests at the same time. This has an issue because ELEM_WRITE requests have an effect to change state of the target element. Concurrent access should be controlled for each of ELEM_READ/ELEM_WRITE case. This commit uses the counting semaphore as write lock for ELEM_WRITE requests, while use it as read lock for ELEM_READ requests. The state of a target element is maintained exclusively between ELEM_WRITE/ELEM_READ operations. There's a concern. If the counting semaphore is acquired for read lock in implementations of 'struct snd_kcontrol.put()' in each driver, this commit shall cause dead lock. As of v4.13-rc5, 'snd-mixer-oss.ko', 'snd-emu10k1.ko' and 'snd-soc-sst-atom-hifi2-platform.ko' includes codes for read locks, but these are not in a call graph from 'struct snd_kcontrol.put(). Therefore, this commit is safe. In current implementation, the same solution is applied for the other operations to element; e.g. ELEM_LOCK and ELEM_UNLOCK. There's another discussion about an overhead to maintain concurrent access to an element during operating the other elements on the same card instance, because the lock primitive is originally implemented to maintain a list of elements on the card instance. There's a substantial difference between per-element-list lock and per-element lock. Here, let me investigate another idea to add per-element lock to maintain the concurrent accesses with inquiry/change requests to an element. It's not so frequent for applications to operate members on elements, while adding a new lock primitive to structure increases memory footprint for all of element sets somehow. Experimentally, inquiry operation is more frequent than change operation and usage of counting semaphore for the inquiry operation brings no blocking to the other inquiry operations. Thus the overhead is not so critical for usual applications. For the above reasons, in this commit, the per-element lock is not introduced. Signed-off-by: Takashi Sakamoto <o-takashi@sakamocchi.jp> --- sound/core/control.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)