Re: [PATCH 2/2] s390/mm,tlb: race of lazy TLB flush vs. recreationof TLB entries

[Martin Schwidefsky]

On Fri, 15 Nov 2013 13:46:07 +0000
Catalin Marinas <catalin.marinas@xxxxxxx> wrote:

> On 15 November 2013 13:29, Martin Schwidefsky <schwidefsky@xxxxxxxxxx> wrote:
> > On Fri, 15 Nov 2013 11:57:01 +0000
> > Catalin Marinas <catalin.marinas@xxxxxxx> wrote:
> >
> >> On Fri, Nov 15, 2013 at 11:17:36AM +0000, Martin Schwidefsky wrote:
> >> > On Fri, 15 Nov 2013 12:10:00 +0100
> >> > Martin Schwidefsky <schwidefsky@xxxxxxxxxx> wrote:
> >> >
> >> > > On Fri, 15 Nov 2013 10:44:37 +0000
> >> > > Catalin Marinas <catalin.marinas@xxxxxxx> wrote:
> >> > > > 1. thread-A running with mm-A
> >> > > > 2. context_switch() to thread-B1 causing a switch_mm(mm-B)
> >> > > > 3. switch_mm(mm-B) sets thread-B1's TIF_TLB_WAIT but does _not_ call
> >> > > >    update_mm(mm-B). Hardware still using mm-A
> >> > > > 4. scheduler unlocks and is about to call finish_mm_switch(mm-B)
> >> > > > 5. interrupt and preemption before finish_mm_switch(mm-B)
> >> > > > 6. context_switch() to thread-B2 causing a switch_mm(mm-B) (note here
> >> > > >    that thread-B1 and thread-B2 have the same mm-B)
> >> > > > 7. switch_mm() as in this patch exits early because prev == next
> >> > > > 8. finish_mm_switch(mm-B) is indeed called but TIF_TLB_WAIT is not set
> >> > > >    for thread-B2, therefore no call to update_mm(mm-B)
> >> > > >
> >> > > > So after point 8, you get thread-B2 running (and possibly returning to
> >> > > > user space) with mm-A. Do you see a problem here?
> >> > >
> >> > > Oh, now I get it. Thanks for the patience, this is indeed a problem.
> >> > > And I concur, a per-mm flag is the 'obvious' solution.
> >> >
> >> > Having said that and looking at the code I find this to be not as obvious
> >> > any more. If you have multiple cpus using a per-mm flag can get you into
> >> > trouble:
> >> >
> >> > 1. cpu #1 calls switch_mm and finds that irqs are disabled.
> >> >    mm->context.switch_pending is set
> >> > 2. cpu #2 calls switch_mm for the same mm and finds that irqs are disabled.
> >> >    mm->context.switch_pending is set again
> >> > 3. cpu #1 reaches finish_arch_post_lock_switch and finds switch_pending == 1
> >> > 4. cpu #1 zeroes mm->switch_pending and calls cpu_switch_mm
> >> > 5. cpu #2 reaches finish_arch_post_lock_switch and finds switch_pending == 0
> >> > 6. cpu #2 continues with the old mm
> >> >
> >> > This is a race, no?
> >>
> >> Yes, but we only use this on ARMv5 and earlier and there is no SMP
> >> support.
> >>
> >> On arm64 however, I need to fix that and you made a good point. In my
> >> (not yet public) patch, the switch_pending is cleared after all the
> >> IPIs have been acknowledged but it needs some more thinking. A solution
> >> could be to always do the cpu_switch_mm() in finish_mm_switch() without
> >> any checks but this requires that any switch_mm() call from the kernel
> >> needs to be paired with finish_mm_switch(). So your first patch comes in
> >> handy (but I still need to figure out a quick arm64 fix for cc stable).
> >
> > I am currently thinking about the following solution for s390: keep the
> > TIF_TLB_FLUSH bit per task but do a preempt_disable() in switch_mm()
> > if the switch is incomplete. This pairs with a preempt_enable() in
> > finish_switch_mm() after the update_mm has been done.
> 
> That's the first thing I tried when I noticed the problem but I got
> weird kernel warnings with preempt_enable/disabling spanning across
> the scheduler unlocking. So doesn't seem safe.
> 
> It may work if instead a simple flag you use atomic_inc/dec for the mm flag.

I have not seen the kernel warnings because the detour over finish_switch_mm
is used only rarely. After forcing the detour I got the fallout, doing the
preempt_disable in switch_mm and preempt_enable in finish_switch_mm does not
work. But what does work is to copy the TIF_TLB_FLUSH bit in the __switch_to
function just like the TIF_MCCK_PENDING. That way the TIF_TLB_FLUSH can not
get "hidden" by a preemptive schedule.

The patch to use finish_arch_post_lock_switch instead of finish_switch_mm
would look like this:
--