Re: [PATCH 7/7] microblaze: Do atomic operations by using exclusive ops

[Peter Zijlstra]

On Thu, Feb 13, 2020 at 10:04:03AM +0000, Will Deacon wrote:
> On Thu, Feb 13, 2020 at 10:16:51AM +0100, Peter Zijlstra wrote:
> > On Thu, Feb 13, 2020 at 09:58:49AM +0100, Peter Zijlstra wrote:
> > 
> > > The thing is, your bog standard LL/SC _SHOULD_ fail the SC if someone
> > > else does a regular store to the same variable. See the example in
> > > Documentation/atomic_t.txt.
> > > 
> > > That is, a competing SW/SWI should result in the interconnect responding
> > > with something other than EXOKAY, the SWX should fail and MSR[C] <- 1.
> > 
> > The thing is; we have code that relies on this behaviour. There are a
> > few crusty SMP archs that sorta-kinda limp along (mostly by disabling
> > some of the code and praying the rest doesn't trigger too often), but we
> > really should not allow more broken SMP archs.
> 
> I did find this in the linked pdf:
> 
>   | If the store [swx] is successful, the sequence of instructions from
>   | the semaphore load to the semaphore store appear to be executed
>   | atomically - no other device modified the semaphore location between
>   | the read and the update.
> 
> which sounds like we're ok, although it could be better worded.
> 
> One part I haven't figured out is what happens if you take an interrupt
> between the lwx and the swx and whether you can end up succeeding thanks
> to somebody else's reservation. Also, the manual is silent about the
> interaction with TLB invalidation and just refers to "address" when
> talking about the reservation. What happens if a user thread triggers
> CoW while another is in the middle of a lwx/swx?

Page 79, Table 2-40 has the note:

"All of these events will clear the reservation bit, used together with
the LWX and SWX instructions to implement mutual exclusion,..."