console freeze

[Emmanuel Dreyfus <manu%netbsd.org@localhost>]

Hello

I have one machine that freeze the serial console after a few seconds
of running userland. It happens on 10.0/amd64 but not on 10.0/i386 nor
on 9.3/amd64 (tested with INSTALL kernels).

Once it get frozen, the kernel is still able to output messages, and
the machines quietly carry on booting. It executes /etc/rc to its end, 
which lets me connect through SSH, and then the sh processes running
/etc/rc wait forever:

PID PPID WCHAN  COMMAND
214    1 wait  sh /etc/rc autoboot
223  214 ttyout sh /etc/rc autoboot

It is not possible to reach ddb, but crash can give us the backtrace
of PID 223:
trace: pid 223 lid 223 at 0xffffaa00ae8f2b50
sleepq_block() at sleepq_block+0x13a
cv_timedwait_sig() at cv_timedwait_sig+0xc6
ttysleep() at ttysleep+0x50
ttwrite() at ttwrite+0x96
cdev_write() at cdev_write+0x87
cdev_write() at cdev_write+0x87
spec_write() at spec_write+0x87
VOP_WRITE() at VOP_WRITE+0xa6
vn_write() at vn_write+0x10e
dofilewrite() at dofilewrite+0x80
sys_write() at sys_write+0x49
syscall() at syscall+0x211

fstat confirms this is the console:
USER     CMD          PID   FD  MOUNT         INUM MODE         SZ|DV R/W
root     sh           223   wd  /                2 drwxr-xr-x    1024 r 
root     sh           223    0* pipe 0xfffffe48d89ec038 <- 0x0 r
root     sh           223    1  /         73317127 crw-------  console rw
root     sh           223    2  /         73317127 crw-------  console rw
root     sh           223    7  /         73317127 crw-------  console rw
root     sh           223    8  /         73317127 crw-------  console rw

Disassembling the kernel around ttwrite+0x96 suggests this is this call:
                        /* Sleep awaiting carrier. */
                        error = ttysleep(tp, &tp->t_rawcv, true, 0);

Why is it waiting carrier? This is a local serial line. 

-- 
Emmanuel Dreyfus
manu%netbsd.org@localhost