Re: [rfc] "fair" rw spinlocks

[Eric W. Biederman]

"Paul E. McKenney" <paulmck@xxxxxxxxxxxxxxxxxx> writes:

> On Mon, Dec 07, 2009 at 03:19:59PM -0800, Eric W. Biederman wrote:
>> Andi Kleen <andi@xxxxxxxxxxxxxx> writes:
>> 
>> > ebiederm@xxxxxxxxxxxx (Eric W. Biederman) writes:
>> >
>> >> "Paul E. McKenney" <paulmck@xxxxxxxxxxxxxxxxxx> writes:
>> >>>
>> >>> Is it required that all of the processes see the signal before the
>> >>> corresponding interrupt handler returns?  (My guess is "no", which
>> >>> enables a trick or two, but thought I should ask.)
>> >>
>> >> Not that I recall.  I think it is just an I/O completed signal.
>> >
>> > Wasn't there the sysrq SAK too? That one definitely would need
>> > to be careful about synchronicity.
>> 
>> SAK from sysrq is done through schedule work, I seem to recall the
>> locking being impossible otherwise.  There is also send_sig_all and a
>> few others from sysrq.  I expect we could legitimately make them
>> schedule_work as well if needed.
>
> OK, I will chance it...  Here is one possible trick:
>
> o	Maintain a list of ongoing group-signal operations, protected
> 	by some suitable lock.  These could be in a per-chain-locked
> 	hash table, hashed by the signal target (e.g., pgrp).
>
> o	When a task is created, it scans the above list, committing
> 	suicide (or doing whatever the signal requires) if appropriate.
>
> o	When creating a child task, the parent holds an SRCU across
> 	creation.  It acquires SRCU before starting creation, and
> 	releases it when it knows that the child has completed
> 	scanning the above list.
>
> o	The updater does the following:
>
> 	o	Add its request to the above list.
>
> 	o	Wait for an SRCU grace period to elapse.
>
> 	o	Kill off everything currently in the task list,
> 		and then wait for each such task to get to a point
> 		where it can be guaranteed not to spawn additional
> 		tasks.  (This might be mediated via a reference
> 		count in the corresponding list element, or by
> 		rescanning the task list, or any of a number of
> 		similar tricks.)
>
> 		Of course, if the signal is non-fatal, then it is
> 		necessary only to wait until the child has taken
> 		the signal.
>
> 	o	If it is possible for a given task's children to
> 		outlive it, despite the fact that the children must
> 		commit suicide upon finding themselves indicated by the
> 		list, wait for another SRCU grace period to elapse.
> 		(This additional SRCU grace period would be required
> 		for a non-fatal pgrp signal, for example.)
>
> 	o	Remove the element from the list.
>
> Does this approach make sense, or am I misunderstanding the problem?

I think that is about right.  I played with that idea a little bit.
I was thinking of simply having new children return -ERESTARTSYS, and
retry the fork.  I put it down because I decided that seems like a
very twisted implementation of a read/write lock.

If we can scale noticeably better a than tasklist_lock it is
definitely worth doing.  I think it is really easy to tie yourself up
in pretzels thinking about this.

An srcu in the pid structure that we hold while signaling tasks.
Interesting.

> Either way, one additional question...  It seems to me that non-fatal
> signals really don't require the above mechanism, because if a task
> handles the signal, and then spawns a child, one can argue that the
> child came after the signal and should thus be unaffected.  Right?
> Or more confusion on my part?

SIGSTOP also seems pretty important not to escape.  I'm not certain of
the others.  I think I would get a bit upset if job control signals in
the shell stopped working properly.  I think asking the question did
that app do something wrong with SIGTERM or did the kernel drop it
would drive me a bit batty.

It is hard to tell what breaks because most buggy implementations will
work correctly most of the time.

Eric
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/