On 9/2/2016 1:28 PM, Andy Lutomirski wrote:
On Sep 2, 2016 7:04 AM, "Chris Metcalf" <cmetcalf@xxxxxxxxxxxx> wrote:
On 8/30/2016 3:50 PM, Andy Lutomirski wrote:
On Tue, Aug 30, 2016 at 12:37 PM, Chris Metcalf <cmetcalf@xxxxxxxxxxxx> wrote:
On 8/30/2016 2:43 PM, Andy Lutomirski wrote:
What if we did it the other way around: set a percpu flag saying
"going quiescent; disallow new deferred work", then finish all
existing work and return to userspace. Then, on the next entry, clear
that flag. With the flag set, vmstat would just flush anything that
it accumulates immediately, nothing would be added to the LRU list,
etc.
This is an interesting idea!
However, there are a number of implementation ideas that make me
worry that it might be a trickier approach overall.
First, "on the next entry" hides a world of hurt in four simple words.
Some platforms (arm64 and tile, that I'm familiar with) have a common
chunk of code that always runs on every entry to the kernel. It would
not be too hard to poke at the assembly and make those platforms
always run some task-isolation specific code on entry. But x86 scares
me - there seem to be a whole lot of ways to get into the kernel, and
I'm not convinced there is a lot of shared macrology or whatever that
would make it straightforward to intercept all of them.
Just use the context tracking entry hook. It's 100% reliable. The
relevant x86 function is enter_from_user_mode(), but I would just hook
into user_exit() in the common code. (This code is had better be
reliable, because context tracking depends on it, and, if context
tracking doesn't work on a given arch, then isolation isn't going to
work regardless.
This looks a lot cleaner than last time I looked at the x86 code. So yes, I think
we could do an entry-point approach plausibly now.
This is also good for when we want to look at deferring the kernel TLB flush,
since it's the same mechanism that would be required for that.
There's at least one gotcha for the latter: NMIs aren't currently
guaranteed to go through context tracking. Instead they use their own
RCU hooks. Deferred TLB flushes can still be made to work, but a bit
more care will be needed. I would probably approach it with an
additional NMI hook in the same places as rcu_nmi_enter() that does,
more or less:
if (need_tlb_flush) flush();
and then make sure that the normal exit hook looks like:
if (need_tlb_flush) {
flush();
barrier(); /* An NMI must not see !need_tlb_flush if the TLB hasn't
been flushed */
flush the TLB;
}
This is a good point. For now I will continue not trying to include the TLB flush
in the current patch series, so I will sit on this until we're ready to do so.
So to pop up a level, what is your actual concern about the existing
"do it in a loop" model? The macrology currently in use means there
is zero cost if you don't configure TASK_ISOLATION, and the software
maintenance cost seems low since the idioms used for task isolation
in the loop are generally familiar to people reading that code.
My concern is that it's not obvious to readers of the code that the
loop ever terminates. It really ought to, but it's doing something
very odd. Normally we can loop because we get scheduled out, but
actually blocking in the return-to-userspace path, especially blocking
on a condition that doesn't have a wakeup associated with it, is odd.
True, although, comments :-)
Regardless, though, this doesn't seem at all weird to me in the
context of the vmstat and lru stuff, though. It's exactly parallel to
the fact that we loop around on checking need_resched and signal, and
in some cases you could imagine multiple loops around when we schedule
out and get a signal, so loop around again, and then another
reschedule event happens during signal processing so we go around
again, etc. Eventually it settles down. It's the same with the
vmstat/lru stuff.
Only kind of.
When we say, effectively, while (need_resched()) schedule();, we're
not waiting for an event or condition per se. We're runnable (in the
sense that userspace wants to run and we're not blocked on anything)
the entire time -- we're simply yielding to some other thread that is
also runnable. So if that loop runs forever, it either means that
we're at low priority and we genuinely shouldn't be running or that
there's a scheduler bug.
If, on the other hand, we say while (not quiesced) schedule(); (or
equivalent), we're saying that we're *not* really ready to run and
that we're waiting for some condition to change. The condition in
question is fairly complicated and won't wake us when we are ready. I
can also imagine the scheduler getting rather confused, since, as far
as the scheduler knows, we are runnable and we are supposed to be
running.
So, how about a code structure like this?
In the main return-to-userspace loop where we check TIF flags,
we keep the notion of our TIF_TASK_ISOLATION flag that causes
us to invoke a task_isolation_prepare() routine. This routine
does the following things:
1. As you suggested, set a new TIF bit (or equivalent) that says the
system should no longer create deferred work on this core, and then
flush any necessary already-deferred work (currently, the LRU cache
and the vmstat stuff). We never have to go flush the deferred work
again during this task's return to userspace. Note that this bit can
only be set on a core marked for task isolation, so it can't be used
for denial of service type attacks on normal cores that are trying to
multitask normal Linux processes.
2. Check if the dyntick is stopped, and if not, wait on a completion
that will be set when it does stop. This means we may schedule out at
this point, but when we return, the deferred work stuff is still safe
since your bit is still set, and in principle the dyn tick is
stopped.
Then, after we disable interrupts and re-read the thread-info flags,
we check to see if the TIF_TASK_ISOLATION flag is the ONLY flag still
set that would keep us in the loop. This will always end up happening
on each return to userspace, since the only thing that actually clears
the bit is a prctl() call. When that happens we know we are about to
return to userspace, so we call task_isolation_ready(), which now has
two things to do:
1. We check that the dyntick is in fact stopped, since it's possible
that a race condition led to it being somehow restarted by an interrupt.
If it is not stopped, we go around the loop again so we can go back in
to the completion discussed above and wait some more. This may merit
a WARN_ON or other notice since it seems like people aren't convinced
there are things that could restart it, but honestly the dyntick stuff
is complex enough that I think a belt-and-suspenders kind of test here
at the last minute is just defensive programming.
2. Assuming it's stopped, we clear your bit at this point, and
return "true" so the loop code knows to break out of the loop and do
the actual return to userspace. Clearing the bit at this point is
better than waiting until we re-enter the kernel later, since it
avoids having to figure out all the ways we actually can re-enter.
With interrupts disabled, and this late in the return to userspace
process, there's no way additional deferred work can be created.
Also, this cond_resched stuff doesn't worry me too much at a
fundamental level -- if we're really going quiescent, shouldn't we be
able to arrange that there are no other schedulable tasks on the CPU
in question?
We aren't currently planning to enforce things in the scheduler, so if
the application affinitizes another task on top of an existing task
isolation task, by default the task isolation task just dies. (Unless
it's using NOSIG mode, in which case it just ends up stuck in the
kernel trying to wait out the dyntick until you either kill it, or
re-affinitize the offending task.) But I'm reluctant to guarantee
every possible way that you might (perhaps briefly) have some
schedulable task, and the current approach seems pretty robust if that
sort of thing happens.
This kind of waiting out the dyntick scares me. Why is there ever a
dyntick that you're waiting out? If quiescence is to be a supported
mainline feature, shouldn't the scheduler be integrated well enough
with it that you don't need to wait like this?
Well, this is certainly the funkiest piece of the task isolation
stuff. The problem is that the dyntick stuff may, for example, need
one more tick 4us from now (or whatever) just to close out the current
RCU period. We can't return to userspace until that happens. So what
else can we do when the task is ready to return to userspace? We
could punt into the idle task instead of waiting in this task, which
was my earlier schedule_time() suggestion. Do you think that's cleaner?
Unless I'm missing something (which is reasonably likely), couldn't
the isolation code just force or require rcu_nocbs on the isolated
CPUs to avoid this problem entirely.
I admit I still don't understand why the RCU context tracking code
can't just run the callback right away instead of waiting however many
microseconds in general. I feel like paulmck has explained it to me
at least once, but that doesn't mean I remember the answer.
I admit I am not clear on this either. However, since there are a
bunch of reasons why the dyntick might run (not just LRU), I think
fixing LRU may well not be enough to guarantee the dyntick
turns off exactly when we'd like it to.
And, with the structure proposed here, we can always come back
and revisit this by just removing the code that does the completion
waiting and replacing it with a call that just tells the dyntick to
just stop immediately, once we're confident we can make that work.
Then separately, we can also think about removing the code that
re-checks dyntick being stopped as we are about to return to
userspace with interrupts disabled, if we're convinced there's
also no way for the dyntick to get restarted due to an interrupt
being handled after we think the dyntick has been stopped.
I'd argue always leaving a WARN_ON() there would be good, though.
--
Chris Metcalf, Mellanox Technologies
http://www.mellanox.com
--
To unsubscribe from this list: send the line "unsubscribe linux-api" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
