----- On Jan 29, 2018, at 2:09 PM, Peter Zijlstra peterz@xxxxxxxxxxxxx wrote:
> On Mon, Jan 29, 2018 at 06:36:05PM +0000, Mathieu Desnoyers wrote:
>> ----- On Jan 29, 2018, at 1:15 PM, Peter Zijlstra peterz@xxxxxxxxxxxxx wrote:
>
>> > Aaah, its the case where we do not pass through switch_mm(), the partial
>> > comment got to me. I only realized after reading the next patch.
>>
>> Indeed, if we read the entire comment, it's made clear that this case is for
>> when switch_mm is not invoked, where the current mm is changed without going
>> through switch_mm(), when scheduling between uthread->kthread->uthread for
>> instance.
>>
>> /*
>> * When transitioning from a kernel thread to a userspace
>> * thread, mmdrop()'s implicit full barrier is required by the
>> * membarrier system call, because the current active_mm can
>> * become the current mm without going through switch_mm().
>> * membarrier also requires a core serializing instruction
>> * before going back to user-space after storing to rq->curr.
>> */
>>
>> Is there something I should improve in the wording of this added
>> sentence to make it clearer ?
>
> Can be improved I think, its got two unqualified "membarrier"s in and
> its a bit mixed up. I'm having a major case of the mondays (brain just
> won't start today), but maybe something like:
>
> When we switched through a kernel thread, the loop in
> membarrier_{private,global}_expedited() can have observed that
> kernel thread and not issued an IPI. We will also not pass
> through switch_mm(). Membarrier requires a barrier after writing
> rq->curr and returning to userspace, so provide them here:
>
> - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED
> - a sync_core for SYNC_CORE
Editing to remove use of "we" and clarify, which ends up as:
/*
* When switching through a kernel thread, the loop in
* membarrier_{private,global}_expedited() may have observed that
* kernel thread and not issued an IPI. It is therefore possible to
* schedule between user->kernel->user threads without passing though
* switch_mm(). Membarrier requires a barrier after storing to
* rq->curr, before returning to userspace, so provide them here:
*
* - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
* provided by mmdrop(),
* - a sync_core for SYNC_CORE.
*/
>
> Also I think changing the changlog to state where we need core-sync
> would be good. Currently the x86 patch does that, but not this one,
> while this introduces the feature.
Planning to add this:
Architectures selecting this feature need to either document that
they issue core serializing instructions when returning to user-space,
or implement their architecture-specific sync_core_before_usermode().
Thanks,
Mathieu
--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
