lkml.org 
[lkml]   [2015]   [Jul]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC PATCH v2] memory-barriers: remove smp_mb__after_unlock_lock()
On Mon, Jul 13, 2015 at 03:41:53PM -0400, Peter Hurley wrote:
> On 07/13/2015 02:23 PM, Paul E. McKenney wrote:
> > On Mon, Jul 13, 2015 at 05:54:47PM +0200, Peter Zijlstra wrote:
> >> On Mon, Jul 13, 2015 at 03:21:10PM +0100, Will Deacon wrote:
> >>> On Mon, Jul 13, 2015 at 03:09:15PM +0100, Will Deacon wrote:
> >>>> On Mon, Jul 13, 2015 at 02:11:43PM +0100, Peter Zijlstra wrote:
> >>>>> On Mon, Jul 13, 2015 at 01:15:04PM +0100, Will Deacon wrote:
> >>>>>> smp_mb__after_unlock_lock is used to promote an UNLOCK + LOCK sequence
> >>>>>> into a full memory barrier.
> >>>>>>
> >>>>>> However:
> >>>>>
> >>>>>> - The barrier only applies to UNLOCK + LOCK, not general
> >>>>>> RELEASE + ACQUIRE operations
> >>>>>
> >>>>> No it does too; note that on ppc both acquire and release use lwsync and
> >>>>> two lwsyncs do not make a sync.
> >>>>
> >>>> Really? IIUC, that means smp_mb__after_unlock_lock needs to be a full
> >>>> barrier on all architectures implementing smp_store_release as smp_mb() +
> >>>> STORE, otherwise the following isn't ordered:
> >>>>
> >>>> RELEASE X
> >>>> smp_mb__after_unlock_lock()
> >>>> ACQUIRE Y
> >>>>
> >>>> On 32-bit ARM (at least), the ACQUIRE can be observed before the RELEASE.
> >>>
> >>> I knew we'd had this conversation before ;)
> >>>
> >>> http://lkml.kernel.org/r/20150120093443.GA11596@twins.programming.kicks-ass.net
> >>
> >> Ha! yes. And I had indeed forgotten about this argument.
> >>
> >> However I think we should look at the insides of the critical sections;
> >> for example (from Documentation/memory-barriers.txt):
> >>
> >> " *A = a;
> >> RELEASE M
> >> ACQUIRE N
> >> *B = b;
> >>
> >> could occur as:
> >>
> >> ACQUIRE N, STORE *B, STORE *A, RELEASE M"
> >>
> >> This could not in fact happen, even though we could flip M and N, A and
> >> B will remain strongly ordered.
> >>
> >> That said, I don't think this could even happen on PPC because we have
> >> load_acquire and store_release, this means that:
> >>
> >> *A = a
> >> lwsync
> >> store_release M
> >> load_acquire N
> >> lwsync
> >
> > Presumably the lwsync instructions are part of the store_release and
> > load_acquire?
> >
> >> *B = b
> >>
> >> And since the store to M is wrapped inside two lwsync there must be
> >> strong store order, and because the load from N is equally wrapped in
> >> two lwsyncs there must also be strong load order.
> >>
> >> In fact, no store/load can cross from before the first lwsync to after
> >> the latter and the other way around.
> >>
> >> So in that respect it does provide full load-store ordering. What it
> >> does not provide is order for M and N, nor does it provide transitivity,
> >> but looking at our documentation I'm not at all sure we guarantee that
> >> in any case.
> >
> > I have no idea what the other thread is doing, so I put together the
> > following litmus test, guessing reverse order, inverse operations,
> > and full ordering:
> >
> > PPC peterz.2015.07.13a
> > ""
> > {
> > 0:r1=1; 0:r2=a; 0:r3=b; 0:r4=m; 0:r5=n;
> > 1:r1=1; 1:r2=a; 1:r3=b; 1:r4=m; 1:r5=n;
> > }
> > P0 | P1 ;
> > stw r1,0(r2) | lwz r10,0(r3) ;
> > lwsync | sync ;
> > stw r1,0(r4) | stw r1,0(r5) ;
> > lwz r10,0(r5) | sync ;
> > lwsync | lwz r11,0(r4) ;
> > stw r1,0(r3) | sync ;
> > | lwz r12,0(r2) ;
> > exists
> > (0:r10=0 /\ 1:r10=1 /\ 1:r11=1 /\ 1:r12=1)
> >
> > See http://lwn.net/Articles/608550/ and http://lwn.net/Articles/470681/
> > for information on tools that operate on these litmus tests. (Both
> > the herd and ppcmem tools agree, as is usually the case.)
> >
> > Of the 16 possible combinations of values loaded, the following seven
> > can happen:
> >
> > 0:r10=0; 1:r10=0; 1:r11=0; 1:r12=0;
> > 0:r10=0; 1:r10=0; 1:r11=0; 1:r12=1;
> > 0:r10=0; 1:r10=0; 1:r11=1; 1:r12=1;
> > 0:r10=0; 1:r10=1; 1:r11=1; 1:r12=1;
> > 0:r10=1; 1:r10=0; 1:r11=0; 1:r12=0;
> > 0:r10=1; 1:r10=0; 1:r11=0; 1:r12=1;
> > 0:r10=1; 1:r10=0; 1:r11=1; 1:r12=1;
> >
> > P0's store to "m" and load from "n" can clearly be misordered, as there
> > is nothing to order them. And all four possible outcomes for 0:r10 and
> > 1:r11 are seen, as expected.
> >
> > Given that smp_store_release() is only guaranteed to order against prior
> > operations and smp_load_acquire() is only guaranteed to order against
> > subsequent operations, P0's load from "n" can be misordered with its
> > store to "a", and as expected, all four possible outcomes for 0:r10 and
> > 1:r12 are observed.
> >
> > P0's pairs of stores should all be ordered:
> >
> > o "a" and "m" -> 1:r11=1 and 1:r12=0 cannot happen, as expected.
> >
> > o "a" and "b" -> 1:r10=1 and 1:r12=0 cannot happen, as expected.
> >
> > o "m" and "b" -> 1:r10=1 and 1:r11=0 cannot happen, as expected.
> >
> > So smp_load_acquire() orders against all subsequent operations, but not
> > necessarily against any prior ones, and smp_store_release() orders against
> > all prior operations but not necessarily against any subsequent onse.
> > But additional stray orderings are permitted, as is the case here.
> > Which is in fact what these operations are defined to do.
> >
> > Does that answer the question, or am I missing the point?
>
> Yes, it shows that smp_mb__after_unlock_lock() has no purpose, since it
> is defined only for PowerPC and your test above just showed that for
> the sequence
>
> store a
> UNLOCK M
> LOCK N
> store b
>
> a and b is always observed as an ordered pair {a,b}.

Not quite.

This is instead the sequence that is of concern:

store a
unlock M
lock N
load b

> Additionally, the assertion in Documentation/memory_barriers.txt that
> the sequence above can be reordered as
>
> LOCK N
> store b
> store a
> UNLOCK M
>
> is not true on any existing arch in Linux.

It was at one time and might be again.

Thanx, Paul



\
 
 \ /
  Last update: 2015-07-13 22:41    [W:0.183 / U:1.740 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site