sound/pcm/* bugs (was: Re: page fault panic tracked down
(selwakeuppri()) - really sound/pcm/*)
truckman at FreeBSD.org
Wed Jan 7 00:27:45 PST 2004
On 7 Jan, Bruce Evans wrote:
> On Tue, 6 Jan 2004, Don Lewis wrote:
>> [... lots of good stuff]
>> I found another bug, though. Holding a mutex across a malloc() call is
>> not allowed because malloc() can sleep, and sleeping while holding a
>> mutex is not allowed. sndbuf_resize() sndbuf_remalloc() both call
>> malloc() to allocate buffers, and sndbuf_alloc() and sndbuf_setup() call
>> sndbuf_resize(). The problem is that chn_setblocksize() calls
>> sndbuf_remalloc() while the channel mutex is held. We can't just unlock
>> the mutex around the sndbuf_remalloc() call because the channel
>> interrupt service routine could run while the buffer is being
>> re-allocated. It would be best if the channel were to be shut down when
>> re-allocating the buffer. This is likely to be messy since
>> chn_setblocksize() is called from lots of different places.
> Er, its an M_NOWAIT malloc(), so it can't sleep. M_WAITOK malloc()s
> in interrupt handlers would also be bugs, but that is not a problem
> here for the same reason.
Hmn, I totally missed that. That eliminates the problem of holding the
mutex, but probably violates POLA if the sound stops playing because of
a transient memory shortage. I also wonder what happens if the buffer
goes away while the channel is running, but I'm not curious enough to
dive back into the code.
> There should be more checks for potential sleeps while holding a mutex.
More information about the freebsd-current