schedcpu() in /sys/kern/sched_4bsd.c calls thread_lock() on thread with un-initialized td_lock

[Attilio Rao]

2011/3/31 John Baldwin <jhb at freebsd.org>:
> On Thursday, March 31, 2011 12:34:31 pm Attilio Rao wrote:
>> 2011/3/31 John Baldwin <jhb at freebsd.org>:
>> > On Thursday, March 31, 2011 7:32:26 am Svatopluk Kraus wrote:
>> >> Hi,
>> >>
>> >>   I've got a page fault (because of NULL td_lock) in
>> >> thread_lock_flags() called from schedcpu() in /sys/kern/sched_4bsd.c
>> >> file. During process fork, new thread is linked to new process which
>> >> is linked to allproc list and both allproc_lock and new process lock
>> >> are unlocked before sched_fork() is called, where new thread td_lock
>> >> is initialized. Only PRS_NEW process status is on sentry but not
>> >> checked in schedcpu().
>> >
>> > I think this should fix it:
>> >
>> > Index: sched_4bsd.c
>> > ===================================================================
>> > --- sched_4bsd.c        (revision 220190)
>> > +++ sched_4bsd.c        (working copy)
>> > @@ -463,6 +463,10 @@ schedcpu(void)
>> >        sx_slock(&allproc_lock);
>> >        FOREACH_PROC_IN_SYSTEM(p) {
>> >                PROC_LOCK(p);
>> > +               if (p->p_state == PRS_NEW) {
>> > +                       PROC_UNLOCK(p);
>> > +                       continue;
>> > +               }
>> >                FOREACH_THREAD_IN_PROC(p, td) {
>> >                        awake = 0;
>> >                        thread_lock(td);
>> >
>>
>> I don't really think this fix is right because otherwise, when using
>> sched_4bsd anytime we are going to scan the thread list within a proc
>> we need to check for PRS_NEW.
>>
>> We likely need to change the init scheme for the td_lock by having a
>> scheduler primitive setting it and doing that on thread_init() UMA
>> constructor, or similar approach.
>
> But the thread state isn't valid anyway.  4BSD shouldn't be touching the
> thread since it is in an incomplete / undefined state.

Yep, in this case I'd then want to just add the threads to proc once
they are fully initialized.

It is pointless (and dangerous) to replicate this check all over,
besides we want scheduler agnostic code, which means every iterations
of p_threads will need to check for a valid state of threads.

Attilio


-- 
Peace can only be achieved by understanding - A. Einstein