Re: [PATCH v14 for 4.1] sys_membarrier(): system-wide memory barrier (x86)

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

----- Original Message -----
> On Tue, 14 Apr 2015, Mathieu Desnoyers wrote:
> > Thinking about it a bit more, one reason for doing the QUERY along
> > with the exact set of flags queried allow us to do more than just
> > returning which flags are supported: it allows us to tell userspace
> > whether the combination of flags used is valid or not.
> > 
> > For instance, if we add a MEMBARRIER_PRIVATE flag in a future release
> > to issue memory barriers only to other threads from the same process,
> > and we add a MEMBARRIER_EXPEDITED which uses IPIs to issue those
> > barriers, we could very well have a situation where using
> > 
> >   EXPEDITED | PRIVATE  would be valid (only sending IPIs to CPUs
> >   running threads from the same process)
> > 
> > but
> > 
> >   EXPEDITED alone would be invalid (-EINVAL), until we figure out
> >   how to expedite memory barriers to all processors without impacting
> >   other processes, if at all possible.
> > 
> > Using QUERY with an empty set of flags could however return the set of
> > flags supported, which could be a nice feature. Anyway, I think
> > the "0" flag should be the basic always correct configuration that
> > is always supported, otherwise we'd have -ENOSYS. Therefore, querying
> > whether the empty set of flags is supported has little value, other
> > than checking for -ENOSYS.
> > 
> > So considering the above, the typical use of this query method from
> > library initialization would be:
> > 
> > int supported_flags = sys_membarrier(MEMBARRIER_QUERY);
> > 
> > ... check for -ENOSYS ....
> > ... check whether the flags we need are supported ...
> > 
> > if (sys_membarrier(MEMBARRIER_QUERY | flag1 | flag2))
> >   goto error;
> > 
> > then we are guaranteed that using sys_membarrier(flag1 | flag2)
> > will always succeed within the application, without needing to
> > handle errors every time it is used. This property is useful
> > to implement a synchronize_rcu() that returns "void" and simplify
> > error handling within the application.
> 
> So how many of these "flags" are you planning to implement and how
> many valid combinations are going to exist?
> 
> I doubt it's more than a dozen. So I prefer explicit operation modes
> for the valid ones rather than having a random pile of "flags".

I don't expect many, so indeed your approach would allow
listing the valid flags, and using them as "one-hot".

If we go for a single active flag at a time, I would call that
"cmd" rather than "flags". Each command would be a power
of two. Only one cmd could be passed as argument (no "or" mask).
QUERY would return a mask of the supported commands.

Thoughts ?

Thanks,

Mathieu

> 
> Thanks,
> 
> 	tglx
> 
> 
> 

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com