Re: panic: lock "aiomtx" already initialized

On Wed, Aug 08, 2007 at 03:59:55PM -0400, Kris Kennaway wrote:
> From stress2's random syscall test:
> 
> db> wh
> Tracing pid 45777 tid 100465 td 0xc61b7000
> kdb_enter(c077f50d,2,c0782352,ed0fab84,2,...) at kdb_enter+0x33
> panic(c0782352,c0788b88,c5e1107c,20000,0,...) at panic+0xed
> lock_init(c5e1107c,c07c67c4,c0788b88,0,20000) at lock_init+0x8c
> mtx_init(c5e1107c,c0788b88,0,0,79400d31,...) at mtx_init+0x9f
> aio_init_aioinfo(c61c4ab0,c057d463,c07e6520,9579c960,31a964,...) at aio_init_aioinfo+0x4b
> aio_aqueue(c61b7000,79400d31,0,2,1,...) at aio_aqueue+0x8d
> oaio_read(c61b7000,ed0facf8,4,c078503a,c07c2250,...) at oaio_read+0x32
> syscall(ed0fad38) at syscall+0x14f
> Xint0x80_syscall() at Xint0x80_syscall+0x20
> --- syscall (318, FreeBSD ELF32, oaio_read), eip = 0x280c0969, esp = 0xbfbfe5f0, ebp = 0xbfbfe638 ---
> db> x/s 0xc0782352
> 0xc0782352:     lock "%s" %p already initialized
> db> x/s 0xc0788b88
> 0xc0788b88:     aiomtx
> db> show lock 0xc5e1107c
>  class: sleep mutex
>  name: aiomtx
>  flags: {DEF}
>  state: {UNOWNED}
> 

This patch should fix the problem:

diff --git a/sys/kern/vfs_aio.c b/sys/kern/vfs_aio.c
index 7610da8..47580b6 100644
--- a/sys/kern/vfs_aio.c
+++ b/sys/kern/vfs_aio.c
_at__at_ -719,6 +719,7 _at__at_ restart:
 	}
 	AIO_UNLOCK(ki);
 	taskqueue_drain(taskqueue_aiod_bio, &ki->kaio_task);
+	mtx_destroy(&ki->kaio_mtx);
 	uma_zfree(kaio_zone, ki);
 	p->p_aioinfo = NULL;
 }

It seems that you shall use a lot of quickly exit()ing processes all of
them using aio to reliable reproduce the problem.