Re: [RFC PATCH 4/7] x86: use exit_lazy_tlb rather than membarrier_mm_sync_core_before_usermode

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

----- On Jul 17, 2020, at 10:51 AM, Alan Stern stern@rowland.harvard.edu wrote:

> On Fri, Jul 17, 2020 at 09:39:25AM -0400, Mathieu Desnoyers wrote:
>> ----- On Jul 16, 2020, at 5:24 PM, Alan Stern stern@rowland.harvard.edu wrote:
>> 
>> > On Thu, Jul 16, 2020 at 02:58:41PM -0400, Mathieu Desnoyers wrote:
>> >> ----- On Jul 16, 2020, at 12:03 PM, Mathieu Desnoyers
>> >> mathieu.desnoyers@efficios.com wrote:
>> >> 
>> >> > ----- On Jul 16, 2020, at 11:46 AM, Mathieu Desnoyers
>> >> > mathieu.desnoyers@efficios.com wrote:
>> >> > 
>> >> >> ----- On Jul 16, 2020, at 12:42 AM, Nicholas Piggin npiggin@gmail.com wrote:
>> >> >>> I should be more complete here, especially since I was complaining
>> >> >>> about unclear barrier comment :)
>> >> >>> 
>> >> >>> 
>> >> >>> CPU0                     CPU1
>> >> >>> a. user stuff            1. user stuff
>> >> >>> b. membarrier()          2. enter kernel
>> >> >>> c. smp_mb()              3. smp_mb__after_spinlock(); // in __schedule
>> >> >>> d. read rq->curr         4. rq->curr switched to kthread
>> >> >>> e. is kthread, skip IPI  5. switch_to kthread
>> >> >>> f. return to user        6. rq->curr switched to user thread
>> >> >>> g. user stuff            7. switch_to user thread
>> >> >>>                         8. exit kernel
>> >> >>>                         9. more user stuff
> 
> ...
> 
>> >> Requiring a memory barrier between update of rq->curr (back to current process's
>> >> thread) and following user-space memory accesses does not seem to guarantee
>> >> anything more than what the initial barrier at the beginning of __schedule
>> >> already
>> >> provides, because the guarantees are only about accesses to user-space memory.
> 
> ...
> 
>> > Is it correct to say that the switch_to operations in 5 and 7 include
>> > memory barriers?  If they do, then skipping the IPI should be okay.
>> > 
>> > The reason is as follows: The guarantee you need to enforce is that
>> > anything written by CPU0 before the membarrier() will be visible to CPU1
>> > after it returns to user mode.  Let's say that a writes to X and 9
>> > reads from X.
>> > 
>> > Then we have an instance of the Store Buffer pattern:
>> > 
>> >	CPU0			CPU1
>> >	a. Write X		6. Write rq->curr for user thread
>> >	c. smp_mb()		7. switch_to memory barrier
>> >	d. Read rq->curr	9. Read X
>> > 
>> > In this pattern, the memory barriers make it impossible for both reads
>> > to miss their corresponding writes.  Since d does fail to read 6 (it
>> > sees the earlier value stored by 4), 9 must read a.
>> > 
>> > The other guarantee you need is that g on CPU0 will observe anything
>> > written by CPU1 in 1.  This is easier to see, using the fact that 3 is a
>> > memory barrier and d reads from 4.
>> 
>> Right, and Nick's reply involving pairs of loads/stores on each side
>> clarifies the situation even further.
> 
> The key part of my reply was the question: "Is it correct to say that
> the switch_to operations in 5 and 7 include memory barriers?"  From the
> text quoted above and from Nick's reply, it seems clear that they do
> not.

I remember that switch_mm implies it, but not switch_to.

The scenario that triggered this discussion is when the scheduler does a
lazy tlb entry/exit, which is basically switch from a user task to
a kernel thread without changing the mm, and usually switching back afterwards.
This optimization means the rq->curr mm temporarily differs, which prevent
IPIs from being sent by membarrier, but without involving a switch_mm.
This requires explicit memory barriers either on entry/exit of lazy tlb
mode, or explicit barriers in the scheduler for those special-cases.

> I agree with Nick: A memory barrier is needed somewhere between the
> assignment at 6 and the return to user mode at 8.  Otherwise you end up
> with the Store Buffer pattern having a memory barrier on only one side,
> and it is well known that this arrangement does not guarantee any
> ordering.

Yes, I see this now. I'm still trying to wrap my head around why the memory
barrier at the end of membarrier() needs to be paired with a scheduler
barrier though.

> One thing I don't understand about all this: Any context switch has to
> include a memory barrier somewhere, but both you and Nick seem to be
> saying that steps 6 and 7 don't include (or don't need) any memory
> barriers.  What am I missing?

All context switch have the smp_mb__before_spinlock at the beginning of
__schedule(), which I suspect is what you refer to. However this barrier
is before the store to rq->curr, not after.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com