Re: Any success stories for HAST + ZFS?

On Sat, 26 Mar 2011 10:52:08 -0700 Freddie Cash wrote:

 FC> hastd backtrace is here:
 FC> http://www.sd73.bc.ca/downloads/crash/hast-backtrace.png

It is not a hastd crash, but a kernel crash triggered by hastd process.

I am not sure I got the same crash as you but apparently the race is possible
in g_gate on device creation.

I got the following crash starting many hast providers simultaneously:

fault virtual address   = 0x0

#8  0xc0c11adc in calltrap () at /usr/src/sys/i386/i386/exception.s:168
#9  0xc086ac6b in g_gate_ioctl (dev=0xc6a24300, cmd=3374345472, 
    addr=0xc9fec000 "\002", flags=3, td=0xc7ff0b80)
    at /usr/src/sys/geom/gate/g_gate.c:410
#10 0xc0853c5b in devfs_ioctl_f (fp=0xc9b9e310, com=3374345472, 
    data=0xc9fec000, cred=0xc8c9c200, td=0xc7ff0b80)
    at /usr/src/sys/fs/devfs/devfs_vnops.c:678
#11 0xc09210cd in kern_ioctl (td=0xc7ff0b80, fd=3, com=3374345472, 
    data=0xc9fec000 "\002") at file.h:262
#12 0xc0921254 in ioctl (td=0xc7ff0b80, uap=0xf5edbcec)
    at /usr/src/sys/kern/sys_generic.c:679
#13 0xc0916616 in syscallenter (td=0xc7ff0b80, sa=0xf5edbce4)
    at /usr/src/sys/kern/subr_trap.c:315
#14 0xc0c2b9ff in syscall (frame=0xf5edbd28)
    at /usr/src/sys/i386/i386/trap.c:1086
#15 0xc0c11b71 in Xint0x80_syscall ()
    at /usr/src/sys/i386/i386/exception.s:266

Or just creating many ggate devices simultaneously:

for i in `jot 100`; do
    ./ggiocreate $i&
done

ggiocreate.c is attached.

In my case the kernel crashes in g_gate_create() when checking for name
collisions in strcmp():

	/* Check for name collision. */
	for (unit = 0; unit < g_gate_maxunits; unit++) {
		if (g_gate_units[unit] == NULL)
			continue;
		if (strcmp(name, g_gate_units[unit]->sc_provider->name) != 0)
			continue;
		mtx_unlock(&g_gate_units_lock);
		mtx_destroy(&sc->sc_queue_mtx);
		free(sc, M_GATE);
		return (EEXIST);
	}

I think the issue is the following. When preparing sc we take
g_gate_units_lock, check for name collision, fill sc fields except
sc->sc_provider, and registers sc in g_gate_units[unit]. sc_provider is filled
later, when g_gate_units_lock is released. So the scenario is possible:

1) Thread A registers sc in g_gate_units[unit] with
g_gate_units[unit]->sc_provider still null and releases g_gate_units_lock.

2) Thread B traverses g_gate_units[] when checking for name collision and
craches accessing g_gate_units[unit]->sc_provider->name.

The attached patch fixes the issue in my case.

-- 
Mikolaj Golub