Re: [PATCH 05/10] staging: erofs: add a full barrier in erofs_workgroup_unfreeze

From: Andrea Parri
Date: Fri Nov 23 2018 - 04:51:40 EST


On Fri, Nov 23, 2018 at 10:51:33AM +0800, Gao Xiang wrote:
> Hi Andrea,
>
> On 2018/11/23 2:50, Andrea Parri wrote:
> > On Thu, Nov 22, 2018 at 06:56:32PM +0800, Gao Xiang wrote:
> >> Hi Greg,
> >>
> >> On 2018/11/22 18:22, Greg Kroah-Hartman wrote:
> >>> Please document this memory barrier. It does not make much sense to
> >>> me...
> >>
> >> Because we need to make the other observers noticing the latest values modified
> >> in this locking period before unfreezing the whole workgroup, one way is to use
> >> a memory barrier and the other way is to use ACQUIRE and RELEASE. we selected
> >> the first one.
> >>
> >> Hmmm...ok, I will add a simple message to explain this, but I think that is
> >> plain enough for a lock...
> >
> > Sympathizing with Greg's request, let me add some specific suggestions:
> >
> > 1. It wouldn't hurt to indicate a pair of memory accesses which are
> > intended to be "ordered" by the memory barrier in question (yes,
> > this pair might not be unique, but you should be able to provide
> > an example).
> >
> > 2. Memory barriers always come matched by other memory barriers, or
> > dependencies (it really does not make sense to talk about a full
> > barrier "in isolation"): please also indicate (an instance of) a
> > matching barrier or the matching barriers.
> >
> > 3. How do the hardware threads communicate? In the acquire/release
> > pattern you mentioned above, the load-acquire *reads from* a/the
> > previous store-release, a memory access that follows the acquire
> > somehow communicate with a memory access preceding the release...
> >
> > 4. It is a good practice to include the above information within an
> > (inline) comment accompanying the added memory barrier (in fact,
> > IIRC, checkpatch.pl gives you a "memory barrier without comment"
> > warning when you omit to do so); not just in the commit message.
> >
> > Hope this helps. Please let me know if something I wrote is unclear,
>
> Thanks for taking time on the detailed explanation. I think it is helpful for me. :)
> And you are right, barriers should be in pairs, and I think I need to explain more:
>
> 255 static inline bool erofs_workgroup_get(struct erofs_workgroup *grp, int *ocnt)
> 256 {
> 257 int o;
> 258
> 259 repeat:
> 260 o = erofs_wait_on_workgroup_freezed(grp);
> 261
> 262 if (unlikely(o <= 0))
> 263 return -1;
> 264
> 265 if (unlikely(atomic_cmpxchg(&grp->refcount, o, o + 1) != o)) <- *
> 266 goto repeat;
> imply a memory barrier here
> 267
> 268 *ocnt = o;
> 269 return 0;
> 270 }
>
> I think atomic_cmpxchg implies a memory barrier semantics when the value comparison (*) succeeds...

Correct. This is informally documented in Documentation/atomic_t.txt
and formalized within tools/memory-model/.


>
> I don't know whether my understanding is correct, If I am wrong..please correct me, or
> I need to add more detailed code comments to explain in the code?

Yes, please; please review the above points (including 1. and 3.) and
try to address them with inline comments. Maybe (if that matches the
*behavior*/guarantee you have in mind...) something like:

[in erofs_workgroup_unfreeze()]

/*
* Orders the store/load to/from [???] and the store to
* ->refcount performed by the atomic_set() below.
*
* Matches the atomic_cmpxchg() in erofs_workgroup_get().
*
* Guarantees that if a successful atomic_cmpxchg() reads
* the value stored by the atomic_set() then [???].
*/
smp_mb();
atomic_set(&grp->refcount, v);


[in erofs_workgroup_get()]

/*
* Orders the load from ->refcount performed by the
* atomic_cmpxchg() below and the store/load [???].
*
* See the comment for the smp_mb() in
* erofs_workgroup_unfreeze().
*/
if (unlikely(atomic_cmpxchg(&grp->refcount, o, o + 1) != o))
goto repeat;

Thanks,
Andrea


>
> Thanks,
> Gao Xiang
>
>
> >
> > Andrea
> >
> >
> >>
> >> Thanks,
> >> Gao Xiang
> >>
> >>>
> >>> thanks,
> >>>
> >>> greg k-h