Re: SMP status ?

[Michael <macallan%netbsd.org@localhost>]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Hello,

On Jan 10, 2011, at 11:59 AM, matthew green wrote:

> > > i'm curious if anyone else has success with the following change.   
> > > it
> > > has survived at least 3x longer than normal under load for me.
>
> my ss20 crashed after about 5 hours with the same problem, but
> then the next boot crashed almost instantly.. in about 2 mins.

Ah, so it's not just me. Mine booted fine but crashed about 1 minute  
into build.sh.

> > > Index: cpu.c
> > > ===================================================================
> > > RCS file: /cvsroot/src/sys/arch/sparc/sparc/cpu.c,v
> > > retrieving revision 1.223
> > > diff -p -r1.223 cpu.c
> > > *** cpu.c       22 Jun 2010 18:29:02 -0000      1.223
> > > --- cpu.c       10 Jan 2011 10:46:41 -0000
> > > *************** void
> > > *** 500,506 ****
> > > cpu_init_system(void)
> > > {
> > >
> > > !       mutex_init(&xpmsg_mutex, MUTEX_SPIN, IPL_VM);
> > > }
> > >
> > > /*
> > > --- 500,506 ----
> > > cpu_init_system(void)
> > > {
> > >
> > > !       mutex_init(&xpmsg_mutex, MUTEX_SPIN, IPL_SCHED);
> > > }
> > >
> > > /*
> >
> > Doesn't seem to make a difference here ( 2x 125MHz HyperSPARC in an
> > SS20 )
> > Something completely different though - now the dbri driver causes a
> > locking-against-myself error when calling bus_dmamem_map() from
> > dbri_malloc() when attaching audio - that used to work, even on MP
> > kernels. THe panic doesn't happen on UP kernels and the panic is
> > always on cpu1.
>
> what are the addresses with the bug?  ie last locker, what lock, etc.

Let me check...
it's rw_vector_enter: locking against myself
lock address 0xf02b644c
owner and current_lwp point at config_interrupts_thread()
rw_enter was called from vm_map_lock
The only difference to other audio drivers is that dbri defers  
attaching audio using config_interrupts() - the whole thing is very,  
very DMA-centric ( only 8 registers in the whole chip, half of them  
for testing ) and even making it talk to the codec without interrupts  
enabled would be quite a pain. An earlier call to bus_dmamem_map()  
succeeds without trouble, I'll check if pushing it further back makes  
any difference.
And let's see if LOCKDEBUG helps.

have fun
Michael

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.7 (Darwin)

iQEVAwUBTSuWmcpnzkX8Yg2nAQJ5YQf/dbDFlvZ0mWb8Z5aaiXVKEs3Z0cJFv+K8
Tg3kXTaOp7aNysvde1AoxLnTXcUHexFxpA7Ma4wXScVNEkFdhvwF6kWs0VkpAPnA
OCC/0ifDpjg0ko5ETEPZHcexriPGBo2Qc8f4SXX55kcfn0Zv+EaUc/c6x6qXuQXX
sGirCy/M2ZRmXpxE3E5rB/bEz7Pw4nbyXiOX6wzQZq7ILfUXplmWgWDk50O0ae2w
MPSivTnmX57+KYqe/T2IhKA4kyRDpkK2CTBsNd5k8qjVSIy+latePaq9Lu9grMH5
TBy+ovmQc3RjTdgwnCSE4RodICV9vUCLVRfmBYxACOZLsVTbYeKDAA==
=8MZ8
-----END PGP SIGNATURE-----