Commit b888a5f7 authored Nov 26, 2018 by Chanho Min Committed by Takashi Iwai Nov 29, 2018

ALSA: pcm: Fix starvation on down_write_nonblock()



Commit 67ec1072 ("ALSA: pcm: Fix rwsem deadlock for non-atomic PCM
stream") fixes deadlock for non-atomic PCM stream. But, This patch
causes antother stuck.
If writer is RT thread and reader is a normal thread, the reader
thread will be difficult to get scheduled. It may not give chance to
release readlocks and writer gets stuck for a long time if they are
pinned to single cpu.

The deadlock described in the previous commit is because the linux
rwsem queues like a FIFO. So, we might need non-FIFO writelock, not
non-block one.

My suggestion is that the writer gives reader a chance to be scheduled
by using the minimum msleep() instaed of spinning without blocking by
writer. Also, The *_nonblock may be changed to *_nonfifo appropriately
to this concept.
In terms of performance, when trylock is failed, this minimum periodic
msleep will have the same performance as the tick-based
schedule()/wake_up_q().

[ Although this has a fairly high performance penalty, the relevant
  code path became already rare due to the previous commit ("ALSA:
  pcm: Call snd_pcm_unlink() conditionally at closing").  That is, now
  this unconditional msleep appears only when using linked streams,
  and this must be a rare case.  So we accept this as a quick
  workaround until finding a more suitable one -- tiwai ]

Fixes: 67ec1072 ("ALSA: pcm: Fix rwsem deadlock for non-atomic PCM stream")
Suggested-by: Wonmin Jung <wonmin.jung@lge.com>
Signed-off-by: Chanho Min <chanho.min@lge.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Takashi Iwai <tiwai@suse.de>

parent b51abed8

sound/core/pcm_native.c

+6 −5

Original line number	Diff line number	Diff line
		@@ -36,6 +36,7 @@
		#include <sound/timer.h>
		#include <sound/minors.h>
		#include <linux/uio.h>
		#include <linux/delay.h>

		#include "pcm_local.h"

		@@ -91,12 +92,12 @@ static DECLARE_RWSEM(snd_pcm_link_rwsem);
		* and this may lead to a deadlock when the code path takes read sem
		* twice (e.g. one in snd_pcm_action_nonatomic() and another in
		* snd_pcm_stream_lock()). As a (suboptimal) workaround, let writer to
		* spin until it gets the lock.
		* sleep until all the readers are completed without blocking by writer.
		*/
		static inline void down_write_nonblock(struct rw_semaphore *lock)
		static inline void down_write_nonfifo(struct rw_semaphore *lock)
		{
		while (!down_write_trylock(lock))
		cond_resched();
		msleep(1);
		}

		#define PCM_LOCK_DEFAULT 0
		@@ -1967,7 +1968,7 @@ static int snd_pcm_link(struct snd_pcm_substream *substream, int fd)
		res = -ENOMEM;
		goto _nolock;
		}
		down_write_nonblock(&snd_pcm_link_rwsem);
		down_write_nonfifo(&snd_pcm_link_rwsem);
		write_lock_irq(&snd_pcm_link_rwlock);
		if (substream->runtime->status->state == SNDRV_PCM_STATE_OPEN \|\|
		substream->runtime->status->state != substream1->runtime->status->state \|\|
		@@ -2014,7 +2015,7 @@ static int snd_pcm_unlink(struct snd_pcm_substream *substream)
		struct snd_pcm_substream *s;
		int res = 0;

		down_write_nonblock(&snd_pcm_link_rwsem);
		down_write_nonfifo(&snd_pcm_link_rwsem);
		write_lock_irq(&snd_pcm_link_rwlock);
		if (!snd_pcm_stream_linked(substream)) {
		res = -EALREADY;

Admin message