Re: [RFC PATCH 4/4] Fix: sched/membarrier: p->mm->membarrier_state racy load (v2)

[Will Deacon]

Hi Mathieu,

Sorry for the delay in responding.

On Fri, Sep 13, 2019 at 10:22:28AM -0400, Mathieu Desnoyers wrote:
> ----- On Sep 12, 2019, at 11:47 AM, Will Deacon will@xxxxxxxxxx wrote:
> 
> > On Thu, Sep 12, 2019 at 03:24:35PM +0100, Linus Torvalds wrote:
> >> On Thu, Sep 12, 2019 at 2:48 PM Will Deacon <will@xxxxxxxxxx> wrote:
> >> >
> >> > So the man page for sys_membarrier states that the expedited variants "never
> >> > block", which feels pretty strong. Do any other system calls claim to
> >> > provide this guarantee without a failure path if blocking is necessary?
> >> 
> >> The traditional semantics for "we don't block" is that "we block on
> >> memory allocations and locking and user accesses etc, but  we don't
> >> wait for our own IO".
> >> 
> >> So there may be new IO started (and waited on) as part of allocating
> >> new memory etc, or in just paging in user memory, but the IO that the
> >> operation _itself_ explicitly starts is not waited on.
> > 
> > Thanks, that makes sense, and I'd be inclined to suggest an update to the
> > sys_membarrier manpage to make this more clear since the "never blocks"
> > phrasing doesn't seem to be used like this for other system calls.
> 
> The current wording from membarrier(2) is:
> 
>               The  "expedited" commands complete faster than the non-expedited
>               ones; they never block, but have the downside of  causing  extra
>               overhead.
> 
> We could simply remove the "; they never block" part then ?

I think so, yes. That or, "; they do not voluntarily block" or something
like that. Maybe look at other man pages for inspiration ;)

> >> No system call should ever be considered "atomic" in any sense. If
> >> you're doing RT, you should maybe expect "getpid()" to not ever block,
> >> but that's just about the exclusive list of truly nonblocking system
> >> calls, and even that can be preempted.
> > 
> > In which case, why can't we just use GFP_KERNEL for the cpumask allocation
> > instead of GFP_NOWAIT and then remove the failure path altogether? Mathieu?
> 
> Looking at:
> 
> #define GFP_KERNEL      (__GFP_RECLAIM | __GFP_IO | __GFP_FS)
> 
> I notice that it does not include __GFP_NOFAIL. What prevents GFP_KERNEL from
> failing, and where is this guarantee documented ?

There was an lwn article a little while ago about this:

https://lwn.net/Articles/723317/

I'm not sure what (if anything) has changed in this regard since then,
however.

> Regarding __GFP_NOFAIL, its use seems to be discouraged in linux/gfp.h:
> 
>  * %__GFP_NOFAIL: The VM implementation _must_ retry infinitely: the caller
>  * cannot handle allocation failures. The allocation could block
>  * indefinitely but will never return with failure. Testing for
>  * failure is pointless.
>  * New users should be evaluated carefully (and the flag should be
>  * used only when there is no reasonable failure policy) but it is
>  * definitely preferable to use the flag rather than opencode endless
>  * loop around allocator.
>  * Using this flag for costly allocations is _highly_ discouraged.
> 
> So I am reluctant to use it.
> 
> But if we can agree on the right combination of flags that guarantees there
> is no failure, I would be perfectly fine with using them to remove the fallback
> code.

I reckon you'll be fine using GFP_KERNEL and returning -ENOMEM on allocation
failure. This shouldn't happen in practice and it removes the fallback
path.

Will