Re: [PATCH] MPU-401 fixes

[Takashi Iwai <tiwai@suse.de>]

At Mon, 05 May 2003 17:51:37 +0200 (METDST),
Clemens Ladisch wrote:
> 
> Takashi Iwai wrote:
> > Clemens Ladisch wrote:
> > > Jaroslav Kysela wrote:
> > > > On Mon, 5 May 2003, Clemens Ladisch wrote:
> > > >
> > > > > - fix deadlock: when snd_mpu401_uart_input_trigger() is interrupted
> > > > >   immediately after the call to spin_unlock_irqrestore(), the interrupt
> > > > >   handler won't clear the interrupt condition because the RX_LOOP bit is
> > > > >   still set
> > > >
> > > > Thanks. The RX_LOOP bit is cleared before unlock now.
> > >
> > > The same thing will happen when the interrupt occurs immediately before
> > > the call to spin_lock_irqsave().
> > >
> > > The real problem is that the interrupt handler _must_ read the data from
> > > the device to clear the MPU401's interrupt condition. Therefore, there
> > > must not be any condition which can prevent the handler from reading data.
> >
> > hmm, this problem is not easy to solve.
> >
> > originally, the bit check was introduced to prevent double spinlocks
> > in the path:
> >
> > 	irq
> > 	-> snd_mpu401_uart_interrupt()
> > 	-> lock
> > 	-> snd_mpu401_uart_input_read()
> > 		-> snd_rawmidi_receive()
> > 			-> rawmidi->event()
> > 				-> rawmidi->trigger() (in seq_midi.c)
> > 					-> snd_mpu401_uart_input_trigger()
> 
> AFAICS the only protection needed is against a recursive call of
> _input_trigger() but not of _interrupt() because the interrupt routine
> cannot be called inside the spin_lock_irqsave(), i.e., the code being
> interrupted will not be the _input_read() function.
> 
> So a solution would be that _interrupt() sets the RX_LOOP bit, but does
> _not_ check it. But then RX_LOOP can be set twice, so we'd need an
> atomic_t counter instead.

yes, we need to avoid the double entrance of input_read() from
trigger().  but it's not necessary to be a atomic counter if the bit
flag is set/reset inside the spinlock but is only checked in
trigger() function.

the code is on cvs now.  please check it out.


ciao,

Takashi


-------------------------------------------------------
This sf.net email is sponsored by:ThinkGeek
Welcome to geek heaven.
http://thinkgeek.com/sf