On Wed, Sep 30, 2020 at 08:00:59PM +0200, Thomas Gleixner wrote: > On Wed, Sep 30 2020 at 19:03, Peter Zijlstra wrote: > > On Wed, Sep 30, 2020 at 05:40:08PM +0200, Thomas Gleixner wrote: > > Also, that preempt_disable() in there doesn't actually do anything. > > Worse, preempt_disable(); for_each_cpu(); is an anti-pattern. It mixes > > static_cpu_has() and boot_cpu_has() in the same bloody condition and has > > a pointless ret variable. Also, I forgot to add, it accesses ->cpus_mask without the proper locking, so it could be reading intermediate state from whatever cpumask operation that's in progress. > I absolutely agree and I really missed it when looking at it before > merging. cpus_read_lock()/unlock() is the right thing to do if at all. > > > It's shoddy code, that only works if you align the planets right. We > > really shouldn't provide interfaces that are this bad. > > > > It's correct operation is only by accident. > > True :( > > I understand Balbirs problem and it makes some sense to provide a > solution. We can: > > 1) reject set_affinity() if the task has that flush muck enabled > and user space tries to move it to a SMT enabled core > > 2) disable the muck if if detects that it is runs on a SMT enabled > core suddenly (hotplug says hello) > > This one is nasty because there is no feedback to user space > about the wreckage. That's and, right, not or. because 1) deals with sched_setffinity() and 2) deals wit hotplug. Now 1) requires an arch hook in sched_setaffinity(), something I'm not keen on providing, once we provide it, who knows what strange and wonderful things archs will dream up. And 2) also happens on hot-un-plug, when the task's affinity gets forced because it became empty. No user feedback there either, and information is lost. I suppose we can do 2) but send a signal. That would cover all cases and keep it in arch code. But yes, that's pretty terrible too.
