Message-ID: <20160812053008.GA18611@tardis.cn.ibm.com>
Date:	Fri, 12 Aug 2016 13:30:15 +0800
From:	Boqun Feng <boqun.feng@...il.com>
To:	Mathieu Desnoyers <mathieu.desnoyers@...icios.com>
Cc:	Dave Watson <davejwatson@...com>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Russell King <linux@....linux.org.uk>,
	Thomas Gleixner <tglx@...utronix.de>,
	Ingo Molnar <mingo@...hat.com>,
	"H. Peter Anvin" <hpa@...or.com>,
	linux-kernel <linux-kernel@...r.kernel.org>,
	linux-api <linux-api@...r.kernel.org>,
	Paul Turner <pjt@...gle.com>, Andrew Hunter <ahh@...gle.com>,
	Peter Zijlstra <peterz@...radead.org>,
	Andy Lutomirski <luto@...capital.net>,
	Andi Kleen <andi@...stfloor.org>, Chris Lameter <cl@...ux.com>,
	Ben Maurer <bmaurer@...com>, rostedt <rostedt@...dmis.org>,
	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>,
	Josh Triplett <josh@...htriplett.org>,
	Linus Torvalds <torvalds@...ux-foundation.org>,
	Catalin Marinas <catalin.marinas@....com>,
	Will Deacon <will.deacon@....com>,
	Michael Kerrisk <mtk.manpages@...il.com>
Subject: Re: [RFC PATCH v7 7/7] Restartable sequences: self-tests

On Fri, Aug 12, 2016 at 03:10:38AM +0000, Mathieu Desnoyers wrote:
> ----- On Aug 11, 2016, at 9:28 PM, Boqun Feng boqun.feng@...il.com wrote:
> 
> > On Thu, Aug 11, 2016 at 11:26:30PM +0000, Mathieu Desnoyers wrote:
> >> ----- On Jul 24, 2016, at 2:01 PM, Dave Watson davejwatson@...com wrote:
> >> 
> >> >>> +static inline __attribute__((always_inline))
> >> >>> +bool rseq_finish(struct rseq_lock *rlock,
> >> >>> + intptr_t *p, intptr_t to_write,
> >> >>> + struct rseq_state start_value)
> >> > 
> >> >>> This ABI looks like it will work fine for our use case. I don't think it
> >> >>> has been mentioned yet, but we may still need multiple asm blocks
> >> >>> for differing numbers of writes. For example, an array-based freelist push:
> >> > 
> >> >>> void push(void *obj) {
> >> >>> if (index < maxlen) {
> >> >>> freelist[index++] = obj;
> >> >>> }
> >> >>> }
> >> > 
> >> >>> would be more efficiently implemented with a two-write rseq_finish:
> >> > 
> >> >>> rseq_finish2(&freelist[index], obj, // first write
> >> >>> &index, index + 1, // second write
> >> >>> ...);
> >> > 
> >> >> Would pairing one rseq_start with two rseq_finish do the trick
> >> >> there ?
> >> > 
> >> > Yes, two rseq_finish works, as long as the extra rseq management overhead
> >> > is not substantial.
> >> 
> >> I've added a commit implementing rseq_finish2() in my rseq volatile
> >> dev branch. You can fetch it at:
> >> 
> >> https://github.com/compudj/linux-percpu-dev/tree/rseq-fallback
> >> 
> >> I also have a separate test and benchmark tree in addition to the
> >> kernel selftests here:
> >> 
> >> https://github.com/compudj/rseq-test
> >> 
> >> I named the first write a "speculative" write, and the second write
> >> the "final" write.
> >> 
> > 
> > Maybe I miss something subtle, but if the first write is only a
> > "speculative" write, why can't we put it in the rseq critical section
> > rather than asm block? Like this:
> > 
> >	do_rseq(..., result, targetptr, newval
> >		{
> >			newval = index;
> >			targetptr = &index;
> >			if (newval < maxlen)
> >				freelist[newval++] = obj;
> >			else
> >				result = false;
> >		}
> > 
> > No extra rseq_finish() is needed here, but maybe a little more
> > "speculative" writes?
> 
> This won't work unfortunately. The speculative stores need to be
> between the rseq_event_counter comparison instruction in the rseq_finish
> asm sequence and the final store. The ip fixup is really needed for
> correctness of speculative stores. The sequence number scheme only works
> for loads.
> 
> Putting it in the C code between rseq_start and rseq_finish would lead
> to races such as:
> 
> thread A                                thread B
> rseq_start
> <preempted>
>                                         <sched in>
>                                         rseq_start
>                                         freelist[offset + 1] = obj
>                                         rseq_finish
>                                            offset++
>                                         <preempted>
> <sched in>
> freelist[newval + 1] = obj  <--- corrupts the list content.
> 

Ah, right!

We couldn't do any "global"(real global or percpu) update in the rseq
critical section(code between rseq_start and rseq_finish), because
without an ip fixup, we cannot abort the critical section immediately,
we have to compare the event_counter in rseq_finish, but that's too late
for speculates stores.

> <snip>
> 
> > Besides, do we allow userspace programs do read-only access to the
> > memory objects modified by do_rseq(). If so, we have a problem when
> > there are two writes in a do_rseq()(either in the rseq critical section
> > or in the asm block), because in current implemetation, these two writes
> > are unordered, which makes the readers outside a do_rseq() could observe
> > the ordering of writes differently.
> > 
> > For rseq_finish2(), a simple solution would be making the "final" write
> > a RELEASE.
> 
> Indeed, we would need a release semantic for the final store here if this
> is the common use. Or we could duplicate the "flavors" of rseq_finish2 and
> add a rseq_finish2_release. We should find a way to eliminate code duplication

I'm in favor of a separate rseq_finish2_release().

> there. I suspect we'll end up doing macros.
> 

Me too. Lemme have a try ;-)

Regards,
Boqun

> Thanks,
> 
> Mathieu
> 
> > 
> > Regards,
> > Boqun
> > 
> >> Thanks,
> >> 
> >> Mathieu
> >> 
> >> --
> >> Mathieu Desnoyers
> >> EfficiOS Inc.
> > > http://www.efficios.com
> 
> -- 
> Mathieu Desnoyers
> EfficiOS Inc.
> http://www.efficios.com

Download attachment "signature.asc" of type "application/pgp-signature" (474 bytes)

Powered by blists - more mailing lists