Re: [PATCH 3/5] userns: Don't read extents twice in m_start

[Christian Brauner <christian.brauner@canonical.com>]

On Wed, Nov 01, 2017 at 03:16:54PM +0100, Peter Zijlstra wrote:
> On Wed, Nov 01, 2017 at 03:01:45PM +0100, Christian Brauner wrote:
> > Tbf, this isn't solely Eric's fault. I'm to blame here too since I didn't
> > document the already existing smb_rmb()s and the new one I introduced when
> > writing the patches. I didn't know that there was a hard-set requirement to
> > document those. I also didn't see anything in the kernel coding style or the
> > memory barriers documentation (But it has been some time since I read those.).
> 
> There's too many documents to read.. I'm not sure we changed
> coding-style, and I suspect that'll just end up being another bike-shed
> in any case.
> 
> We did get checkpatch changed though, which is a strong enough clue that
> something needs to happen.
> 
> But What Nikolay said; memory ordering is hard enough if you're clear on
> what exactly you intend to do. But if you later try and reconstruct
> without comments, its nearly impossible.

Yeah, agreed. I was happy to see that Eric explained his smp_wmb() in detail.
That was quite helpful in figuring this out!

> 
> It gets even better if someone changes the ordering requirements over
> time and you grow hidden and non-obvious dependencies :/
> 
> > > Also, you probably want READ_ONCE() here and WRITE_ONCE() in
> > > map_write(), the compiler is free to do unordered byte loads/stores
> > > without it.
> > > 
> > > And finally, did you want to use smp_store_release() and
> > > smp_load_acquire() instead?
> > 
> > Maybe a stupid question but do you suspect this is a real problem in
> > this case since you're phrasing it as a question?
> 
> Rhetorical question mostly, I suspect its just what you meant to do, as
> per the proposed patch.
> 
> > Iirc, *_acquire() operations include
> > locking operations and might come with a greater performance impact then
> > smb_{rmb,wmb}(). Given that this is a very performance critical path we should
> > be sure.
> 
> No locking what so ever. LOAD-ACQUIRE and STORE-RELEASE are memory ordering
> flavours that are paired with the respective memory operation.

Ah right, now I remember I was confused by a part of the memory barriers
documentation that referenced locks. Acquire operations include locks and
smp_load_acquire().  Right, should've remembered that. Thanks!

> 
> It is true that locking ops provide these exact orderings, but that
> doesn't imply the reverse.
> 
> In short, store-release is a store that ensures all prior load _and_
> stores happen-before this store. A load-acquire is a load which
> happens-before any subsequent load or stores.
> 
> But a release does not constrain later loads or stores, and an acquire
> does not constrain prior load or stores.
> 
>