vfs-related deadlock(?)

[Havard Eidnes <he%NetBSD.org@localhost>]

Hi,

while trying to report & get fixes for what causes my dual-cpu G4
system to wedge with updated 10.0_BETA code, I got a fix from
rin@ for the first issue i stumbled upon (thanks!).

Next I've stumbled upon something which looks related to file
systems / vfs.  I have two processes stuck in 'tstile', one is
'ld' the other is the 'vdrain' kernel thread.  Backtrace for
both:

crash> bt/a 7fb26680
trace: pid 0 lid 9 at 0x1001fc60
0x1001fcc0: at cpu_switchto+0x28
0x1001fcd0: at mi_switch+0x140
0x1001fd10: at sleepq_block+0xe0
0x1001fd30: at turnstile_block+0x284
0x1001fd90: at rw_enter+0x124
0x1001fdd0: at cache_purge1+0x198
0x1001fe00: at vcache_reclaim+0xc8
0x1001fe80: at vrecycle.part.0+0xc0
0x1001feb0: at vdrain_thread+0x384
0x1001ff20: at cpu_lwp_bootstrap+0xc
saved LR(0x55aa55a6) is invalid.
crash>

for the vdrain thread, while

crash> bt/a 102ee400
trace: pid 4483 lid 4483 at 0x1b623930
0x1b623990: at cpu_switchto+0x28
0x1b6239a0: at mi_switch+0x140
0x1b6239e0: at sleepq_block+0xe0
0x1b623a00: at turnstile_block+0x284
0x1b623a60: at rw_enter+0x124
0x1b623aa0: at cache_lookup+0xe8
0x1b623ad0: at ufs_lookup+0xd4
0x1b623b80: at VOP_LOOKUP+0x4c
0x1b623ba0: at lookup_once+0x1fc
0x1b623bf0: at namei_tryemulroot.constprop.0+0x4a8
0x1b623cc0: at namei+0x58
0x1b623cf0: at vn_open+0xfc
0x1b623e10: at do_open+0xf0
0x1b623e60: at do_sys_openat+0x9c
0x1b623ea0: at sys_open+0x2c
0x1b623ec0: at syscall+0x294
0x1b623f20: user SC trap #5 by 0xfdc70e18: srr1=0xd032
            r1=0xffffda20 cr=0x84000444 xer=0x20000000 ctr=0xfdc70e10
crash>

for the 'ld' process, which is (supposed to be) busy linking an
.so object.

This was triggered by a "build.sh ... release".

More details can be found in

  https://gnats.netbsd.org/57621

Is anyone here able to figure out why these two appear to be
mutually deadlocked?

Regards,

- Håvard