On Tue, Sep 30, 2014 at 10:27:40AM +0800, Fengguang Wu wrote:
> On Fri, Sep 26, 2014 at 12:42:23AM -0700, Paul E. McKenney wrote:
> > On Thu, Sep 18, 2014 at 09:17:51PM +0800, Fengguang Wu wrote:
> > > Hi Paul,
> > > 
> > > > > > > plymouth-upstart-bridge: ply-event-loop.c:497: ply_event_loop_new: Assertion `loop->epoll_fd >= 0' failed.
> > > > > > > /etc/lsb-base-logging.sh: line 5:  2580 Aborted                 plymouth --ping > /dev/null 2>&1
> > > > > > > /etc/lsb-base-logging.sh: line 5:  2585 Aborted                 plymouth --ping > /dev/null 2>&1
> > > > > > > mount: proc has wrong device number or fs type proc not supported
> > > > > > > /etc/lsb-base-logging.sh: line 5:  2601 Aborted                 plymouth --ping > /dev/null 2>&1
> > > > > > > /etc/rc6.d/S40umountfs: line 20: /proc/mounts: No such file or directory
> > > > > > > cat: /proc/1/maps: No such file or directory
> > > > > > > cat: /proc/1/maps: No such file or directory
> > > > > > > cat: /proc/1/maps: No such file or directory
> > > > > > > cat: /proc/1/maps: No such file or directory
> > > > > > > cat: /proc/1/maps: No such file or directory
> > > > > > > cat: /proc/1/maps: No such file or directory
> > > > > > > umount: /var/run: not mounted
> > > > > > > umount: /var/lock: not mounted
> > > > > > > umount: /dev/shm: not mounted
> > > > > > > mount: / is busy
> > > > > > >  * Will now restart
> > > > 
> > > > Are these expected behavior?
> > > 
> > > Yes, because it's randconfig boot tests, the user space may well
> > > complain about random stuff and I'll ignore them all as long as it
> > > will eventually call the shutdown command to finish the test in time.  :)
> > > 
> > > > So again, I can invoke this commit without losing much (sendkey
> > > > alt-sysrq-z is after all my friend), but it is not clear to me that we
> > > > have gotten to the root of this problem.
> > > 
> > > Sorry about that! If you see any debug tricks that I can try, or
> > > information I can collect, please let me know.
> > 
> > Hmmm...
> > 
> > Looks like rcutorture might be starting too soon.  With all the selftests,
> > it is taking 3-4 minutes to boot.
> 
> Sorry my scripts reboot the machine quickly, there is no logic to wait for the
> completion of rcutorture tests.  Looking at the dmesg, the BUG shows up during
> the shutdown stage:
> 
> ==>     Sending all processes the TERM signal...
>         mount: mounting proc on /proc failed: No such device
>         [  121.930088] Dumping ftrace buffer:
>         [  121.930453] ---------------------------------
>         [  121.930865] BUG: unable to handle kernel NULL pointer dereference at           (null)
>         [  121.931644] IP: [<ffffffff8959b074>] print_trace_line+0x2b0/0x38a

Fair point.

> > One approach would be to set
> > rcutorture.stat_interval=200 or whatever the duration of boot is.
> > Another would be to set rcutorture.torture_runnable=0, and to change:
> > 
> > 	int rcutorture_runnable = RCUTORTURE_RUNNABLE_INIT;
> > 	module_param(rcutorture_runnable, int, 0444);
> > 	MODULE_PARM_DESC(rcutorture_runnable, "Start rcutorture at boot");
> > 
> > To:
> > 
> > 	int rcutorture_runnable = RCUTORTURE_RUNNABLE_INIT;
> > 	module_param(rcutorture_runnable, int, 0644);
> > 	MODULE_PARM_DESC(rcutorture_runnable, "Start rcutorture at boot");
> > 
> > In kernel/rcu/rcutorture.c.
> > 
> > Then have your scripts set rcutorture_runnable=1 from sysfs once boot
> > completes.
> 
> That looks suitable to run as a functional test case in LKP.
> 
> I can enable the below options in the LKP test kernel (hope they will not
> bring noticeable runtime overheads when not insmod):
> 
> CONFIG_TORTURE_TEST=m
> CONFIG_RCU_TORTURE_TEST=m
> CONFIG_LOCK_TORTURE_TEST=m
> 
> and write a simple test case for it. To start the torture test,
> it should be as simple as
> 
>          modproble rcutorture

"modprobe", but yes.  ;-)

>          echo 1 > /proc/sys/kernel/rcutorture_runnable
> 
> To determine if the test has finished, will this do the job in the
> normal cases?
> 
>          dmesg | grep "--- End of test:"

Well, rcutorture won't stop unless you tell it to.  The usual way to
tell it to stop is via rmmod.

> Can I reasonably set the max test timeout to 5 or 10 or more minutes?
> When exceeded, LKP will assume the machine is dead and need a force reboot.

One thing would be something like this:

	( sleep 300; rmmod rcutorture ) &

Running both locktorture and rcutorture concurrently is not supported
at the moment because both make the assumption that they have the entire
system at their disposal.

> > Alternatively, if poking sysfs is not reasonable (and it
> > would not be in my test scripts), put a delay just after the
> > rcutorture_record_test_transition() in rcu_torture_init().  For example,
> > schedule_timeout_interruptible(200 * HZ) to delay 200 seconds.
> 
> I'd prefer the boot test to complete sooner than later, which helps
> improve the test efficiency. :)

Good point...  It would not be hard to make this a timeout-based delay,
though.  But your earlier point about this being during shutdown is
a good one here, also.

> The 0day boot tests works by running this in a init.d script
> 
>         run CPU hotplug tests in the background
>         enable tracing events [optional]
>         run trinity for 100 seconds
>         ### if feasible, we can add a wait-for-rcutorture-finish here
>         reboot
> 
> > Another approach would be for me to figure out some way for rcutorture
> > to figure out that boot was not far enough along for it to safely
> > do much, probably enabled by a third value of rcutorture_runnable.
> 
> Will it be *easy* to have a blockable syswrite to "stop rcutorture tests",
> which will return when all tests have been stopped?

I am not eager to handle the races between such a mechanism and rmmod.  ;-)

> Anyway, I'm already happy with the above two options (where the kernel
> printk some test completion message for the user space to grep).

OK, but rmmod already exists and already shuts down rcutorture and waits
for it to complete.

> > One more approach would be to replace DUMP_ALL with DUMP_NONE in
> > kernel/rcu/rcutorture.c's rcutorture_trace_dump() function.  Or
> > to remove the ftrace_dump() statement entirely.  (The question that
> > this might help answer is which part of rcutorture_trace_dump() is
> > causing the problem.)
> 
> It should be unnecessary in testing POV. Unless if it is a bug.

It might be that it is illegal to do ftrace_dump() too far into shutdown.
I should check with Steven Rostedt.

							Thanx, Paul