Re: [PATCH 8/8] blk-mq: drain I/O when all CPUs in a hctx are offline

[Paul E. McKenney]

On Thu, May 28, 2020 at 06:37:47AM -0700, Bart Van Assche wrote:
> On 2020-05-27 22:19, Ming Lei wrote:
> > On Wed, May 27, 2020 at 08:33:48PM -0700, Bart Van Assche wrote:
> >> My understanding is that operations that have acquire semantics pair
> >> with operations that have release semantics. I haven't been able to find
> >> any documentation that shows that smp_mb__after_atomic() has release
> >> semantics. So I looked up its definition. This is what I found:
> >>
> >> $ git grep -nH 'define __smp_mb__after_atomic'
> >> arch/ia64/include/asm/barrier.h:49:#define __smp_mb__after_atomic()
> >> barrier()
> >> arch/mips/include/asm/barrier.h:133:#define __smp_mb__after_atomic()
> >> smp_llsc_mb()
> >> arch/s390/include/asm/barrier.h:50:#define __smp_mb__after_atomic()
> >> barrier()
> >> arch/sparc/include/asm/barrier_64.h:57:#define __smp_mb__after_atomic()
> >> barrier()
> >> arch/x86/include/asm/barrier.h:83:#define __smp_mb__after_atomic()	do {
> >> } while (0)
> >> arch/xtensa/include/asm/barrier.h:20:#define __smp_mb__after_atomic()	
> >> barrier()
> >> include/asm-generic/barrier.h:116:#define __smp_mb__after_atomic()
> >> __smp_mb()
> >>
> >> My interpretation of the above is that not all smp_mb__after_atomic()
> >> implementations have release semantics. Do you agree with this conclusion?
> > 
> > I understand smp_mb__after_atomic() orders set_bit(BLK_MQ_S_INACTIVE)
> > and reading the tag bit which is done in blk_mq_all_tag_iter().
> > 
> > So the two pair of OPs are ordered:
> > 
> > 1) if one request(tag bit) is allocated before setting BLK_MQ_S_INACTIVE,
> > the tag bit will be observed in blk_mq_all_tag_iter() from blk_mq_hctx_has_requests(),
> > so the request will be drained.
> > 
> > OR
> > 
> > 2) if one request(tag bit) is allocated after setting BLK_MQ_S_INACTIVE,
> > the request(tag bit) will be released and retried on another CPU
> > finally, see __blk_mq_alloc_request().
> > 
> > Cc Paul and linux-kernel list.
> 
> I do not agree with the above conclusion. My understanding of
> acquire/release labels is that if the following holds:
> (1) A store operation that stores the value V into memory location M has
> a release label.
> (2) A load operation that reads memory location M has an acquire label.
> (3) The load operation (2) retrieves the value V that was stored by (1).
> 
> that the following ordering property holds: all load and store
> instructions that happened before the store instruction (1) in program
> order are guaranteed to happen before the load and store instructions
> that follow (2) in program order.
> 
> In the ARM manual these semantics have been described as follows: "A
> Store-Release instruction is multicopy atomic when observed with a
> Load-Acquire instruction".
> 
> In this case the load-acquire operation is the
> "test_and_set_bit_lock(nr, word)" statement from the sbitmap code. That
> code is executed indirectly by blk_mq_get_tag(). Since there is no
> matching store-release instruction in __blk_mq_alloc_request() for
> 'word', ordering of the &data->hctx->state and 'tag' memory locations is
> not guaranteed by the acquire property of the "test_and_set_bit_lock(nr,
> word)" statement from the sbitmap code.

I feel like I just parachuted into the middle of the conversation,
so let me start by giving a (silly) example illustrating the limits of
smp_mb__{before,after}_atomic() that might be tangling things up.

But please please please avoid doing this in real code unless you have
an extremely good reason included in a comment.

void t1(void)
{
	WRITE_ONCE(a, 1);
	smp_mb__before_atomic();
	WRITE_ONCE(b, 1);  // Just Say No to code here!!!
	atomic_inc(&c);
	WRITE_ONCE(d, 1);  // Just Say No to code here!!!
	smp_mb__after_atomic();
	WRITE_ONCE(e, 1);
}

void t2(void)
{
	r1 = READ_ONCE(e);
	smp_mb();
	r2 = READ_ONCE(d);
	smp_mb();
	r3 = READ_ONCE(c);
	smp_mb();
	r4 = READ_ONCE(b);
	smp_mb();
	r5 = READ_ONCE(a);
}

Each platform must provide strong ordering for either atomic_inc()
on the one hand (as ia64 does) or for smp_mb__{before,after}_atomic()
on the other (as powerpc does).  Note that both ia64 and powerpc are
weakly ordered.

So ia64 could see (r1 == 1 && r2 == 0) on the one hand as well as (r4 ==
1 && r5 == 0).  So clearly smp_mb_{before,after}_atomic() need not have
any ordering properties whatsoever.

Similarly, powerpc could see (r3 == 1 && r4 == 0) on the one hand as well
as (r2 == 1 && r3 == 0) on the other.  Or even both at the same time.
So clearly atomic_inc() need not have any ordering properties whatsoever.

But the combination of smp_mb__before_atomic() and the later atomic_inc()
does provide full ordering, so that no architecture can see (r3 == 1 &&
r5 == 0), and either of r1 or r2 can be substituted for r3.

Similarly, atomic_inc() and the late4r smp_mb__after_atomic() also
provide full ordering, so that no architecture can see (r1 == 1 && r3 ==
0), and either r4 or r5 can be substituted for r3.


So a call to set_bit() followed by a call to smp_mb__after_atomic() will
provide a full memory barrier (implying release semantics) for any write
access after the smp_mb__after_atomic() with respect to the set_bit() or
any access preceding it.  But the set_bit() by itself won't have release
semantics, nor will the smp_mb__after_atomic(), only their combination
further combined with some write following the smp_mb__after_atomic().

More generally, there will be the equivalent of smp_mb() somewhere between
the set_bit() and every access following the smp_mb__after_atomic().

Does that help, or am I missing the point?

							Thanx, Paul