* Re: [External] Re: [PATCH] seqlock: Use while instead of if+goto in __read_seqcount_begin
2020-04-14 11:05 ` Peter Zijlstra
@ 2020-04-14 12:01 ` Muchun Song
2020-04-15 11:41 ` Peter Zijlstra
2020-04-14 13:48 ` Will Deacon
2020-04-15 9:37 ` David Laight
2 siblings, 1 reply; 9+ messages in thread
From: Muchun Song @ 2020-04-14 12:01 UTC (permalink / raw)
To: Peter Zijlstra; +Cc: Will Deacon, mingo, mingo, linux-kernel
On Tue, Apr 14, 2020 at 7:05 PM Peter Zijlstra <peterz@infradead.org> wrote:
>
> On Fri, Apr 10, 2020 at 12:56:58PM +0100, Will Deacon wrote:
> > On Thu, Apr 09, 2020 at 09:45:58PM +0800, Muchun Song wrote:
> > > The creators of the C language gave us the while keyword. Let's use
> > > that instead of synthesizing it from if+goto.
> > >
> > > Signed-off-by: Muchun Song <songmuchun@bytedance.com>
> > > ---
> > > include/linux/seqlock.h | 6 +-----
> > > 1 file changed, 1 insertion(+), 5 deletions(-)
> > >
> > > diff --git a/include/linux/seqlock.h b/include/linux/seqlock.h
> > > index 8b97204f35a77..7bdea019814ce 100644
> > > --- a/include/linux/seqlock.h
> > > +++ b/include/linux/seqlock.h
> > > @@ -125,12 +125,8 @@ static inline unsigned __read_seqcount_begin(const seqcount_t *s)
> > > {
> > > unsigned ret;
> > >
> > > -repeat:
> > > - ret = READ_ONCE(s->sequence);
> > > - if (unlikely(ret & 1)) {
> > > + while (unlikely((ret = READ_ONCE(s->sequence)) & 1))
> > > cpu_relax();
> > > - goto repeat;
> > > - }
> > > kcsan_atomic_next(KCSAN_SEQLOCK_REGION_MAX);
> > > return ret;
> >
> > Patch looks fine to me, but I'll leave it to Peter as I don't have a
> > preference either way.
>
> Linus sometimes prefers the goto variant as that better expresses the
> exception model. But like Will, I don't particularly care. That said,
> Will, would it make sense to use smp_cond_load_relaxed() here ?
I have a similar idea. Would it make sense to use smp_cond_load_acquire()
in raw_read_seqcount_begin()?
--
Yours,
Muchun
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [External] Re: [PATCH] seqlock: Use while instead of if+goto in __read_seqcount_begin
2020-04-14 12:01 ` [External] " Muchun Song
@ 2020-04-15 11:41 ` Peter Zijlstra
0 siblings, 0 replies; 9+ messages in thread
From: Peter Zijlstra @ 2020-04-15 11:41 UTC (permalink / raw)
To: Muchun Song; +Cc: Will Deacon, mingo, mingo, linux-kernel
On Tue, Apr 14, 2020 at 08:01:06PM +0800, Muchun Song wrote:
> On Tue, Apr 14, 2020 at 7:05 PM Peter Zijlstra <peterz@infradead.org> wrote:
> > .... That said,
> > Will, would it make sense to use smp_cond_load_relaxed() here ?
>
> I have a similar idea. Would it make sense to use smp_cond_load_acquire()
> in raw_read_seqcount_begin()?
Not sure; I did consider it, but that rmb it has seems more natural in
the over-all ordering scheme here. I mean:
load seqcount inc seqcount
rmb wmb
// load stuff // modify stuff
rmb wmb
compare seqcount inc seqcount
is nice and symmetric, making that upper left rmb an acquire 'works' but
is just weird IMO. And I suppose you can make the lower right wmb a
store-release, which is somewhat better, but then it gets all weird when
you consider things like barrier and latch.
So best to just leave it as is I think.
Those incs do seem to be really wanting a WRITE_ONCE() though.
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH] seqlock: Use while instead of if+goto in __read_seqcount_begin
2020-04-14 11:05 ` Peter Zijlstra
2020-04-14 12:01 ` [External] " Muchun Song
@ 2020-04-14 13:48 ` Will Deacon
2020-04-15 11:44 ` Peter Zijlstra
2020-04-15 9:37 ` David Laight
2 siblings, 1 reply; 9+ messages in thread
From: Will Deacon @ 2020-04-14 13:48 UTC (permalink / raw)
To: Peter Zijlstra; +Cc: Muchun Song, mingo, mingo, linux-kernel
On Tue, Apr 14, 2020 at 01:05:16PM +0200, Peter Zijlstra wrote:
> On Fri, Apr 10, 2020 at 12:56:58PM +0100, Will Deacon wrote:
> > On Thu, Apr 09, 2020 at 09:45:58PM +0800, Muchun Song wrote:
> > > The creators of the C language gave us the while keyword. Let's use
> > > that instead of synthesizing it from if+goto.
> > >
> > > Signed-off-by: Muchun Song <songmuchun@bytedance.com>
> > > ---
> > > include/linux/seqlock.h | 6 +-----
> > > 1 file changed, 1 insertion(+), 5 deletions(-)
> > >
> > > diff --git a/include/linux/seqlock.h b/include/linux/seqlock.h
> > > index 8b97204f35a77..7bdea019814ce 100644
> > > --- a/include/linux/seqlock.h
> > > +++ b/include/linux/seqlock.h
> > > @@ -125,12 +125,8 @@ static inline unsigned __read_seqcount_begin(const seqcount_t *s)
> > > {
> > > unsigned ret;
> > >
> > > -repeat:
> > > - ret = READ_ONCE(s->sequence);
> > > - if (unlikely(ret & 1)) {
> > > + while (unlikely((ret = READ_ONCE(s->sequence)) & 1))
> > > cpu_relax();
> > > - goto repeat;
> > > - }
> > > kcsan_atomic_next(KCSAN_SEQLOCK_REGION_MAX);
> > > return ret;
> >
> > Patch looks fine to me, but I'll leave it to Peter as I don't have a
> > preference either way.
>
> Linus sometimes prefers the goto variant as that better expresses the
> exception model. But like Will, I don't particularly care. That said,
> Will, would it make sense to use smp_cond_load_relaxed() here ?
Oh yeah, good thinking. Didn't spot that one, but it should work well as
long as smp_cond_load_relaxed() always implies a control dependency (surely
it has to?)
Will
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH] seqlock: Use while instead of if+goto in __read_seqcount_begin
2020-04-14 13:48 ` Will Deacon
@ 2020-04-15 11:44 ` Peter Zijlstra
0 siblings, 0 replies; 9+ messages in thread
From: Peter Zijlstra @ 2020-04-15 11:44 UTC (permalink / raw)
To: Will Deacon; +Cc: Muchun Song, mingo, mingo, linux-kernel
On Tue, Apr 14, 2020 at 02:48:31PM +0100, Will Deacon wrote:
> Oh yeah, good thinking. Didn't spot that one, but it should work well as
> long as smp_cond_load_relaxed() always implies a control dependency (surely
> it has to?)
Well, yeah, but then you know as well as I do that compilers are dodgy
pieces of crap which are out to get you. Still, I too can't see how it
could mess this up.
^ permalink raw reply [flat|nested] 9+ messages in thread
* RE: [PATCH] seqlock: Use while instead of if+goto in __read_seqcount_begin
2020-04-14 11:05 ` Peter Zijlstra
2020-04-14 12:01 ` [External] " Muchun Song
2020-04-14 13:48 ` Will Deacon
@ 2020-04-15 9:37 ` David Laight
2 siblings, 0 replies; 9+ messages in thread
From: David Laight @ 2020-04-15 9:37 UTC (permalink / raw)
To: 'Peter Zijlstra', Will Deacon
Cc: Muchun Song, mingo, mingo, linux-kernel
From: Peter Zijlstra
> Sent: 14 April 2020 12:05
> On Fri, Apr 10, 2020 at 12:56:58PM +0100, Will Deacon wrote:
> > On Thu, Apr 09, 2020 at 09:45:58PM +0800, Muchun Song wrote:
> > > The creators of the C language gave us the while keyword. Let's use
> > > that instead of synthesizing it from if+goto.
> > >
> > > Signed-off-by: Muchun Song <songmuchun@bytedance.com>
> > > ---
> > > include/linux/seqlock.h | 6 +-----
> > > 1 file changed, 1 insertion(+), 5 deletions(-)
> > >
> > > diff --git a/include/linux/seqlock.h b/include/linux/seqlock.h
> > > index 8b97204f35a77..7bdea019814ce 100644
> > > --- a/include/linux/seqlock.h
> > > +++ b/include/linux/seqlock.h
> > > @@ -125,12 +125,8 @@ static inline unsigned __read_seqcount_begin(const seqcount_t *s)
> > > {
> > > unsigned ret;
> > >
> > > -repeat:
> > > - ret = READ_ONCE(s->sequence);
> > > - if (unlikely(ret & 1)) {
> > > + while (unlikely((ret = READ_ONCE(s->sequence)) & 1))
> > > cpu_relax();
> > > - goto repeat;
> > > - }
> > > kcsan_atomic_next(KCSAN_SEQLOCK_REGION_MAX);
> > > return ret;
> >
> > Patch looks fine to me, but I'll leave it to Peter as I don't have a
> > preference either way.
>
> Linus sometimes prefers the goto variant as that better expresses the
> exception model. But like Will, I don't particularly care. That said,
> Will, would it make sense to use smp_cond_load_relaxed() here ?
gcc also has a nasty habit of converting:
while (foo)
bar;
into:
if (foo) {
do
bar;
while (foo);
}
with all the code bloat that entails - especially when 'foo'
is non-trivial.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)
^ permalink raw reply [flat|nested] 9+ messages in thread