On Wed, Oct 11, 2023 at 05:05:04AM +0000, Joel Fernandes wrote:
> On Tue, Oct 10, 2023 at 06:34:35PM -0700, Paul E. McKenney wrote:
> [...]
> > > > > > > > It's also worth noting that the bug this fixes wasn't exposed until the
> > > > > > > > maple tree (added in v6.1) was used for the IRQ descriptors (added in
> > > > > > > > v6.5).
> > > > > > >
> > > > > > > Lots of latent bugs, to be sure, even with rcutorture. :-/
> > > > > >
> > > > > > The Right Thing is to fix the bug all the way back to the introduction,
> > > > > > but what fallout makes the backport less desirable than living with the
> > > > > > unexposed bug?
> > > > >
> > > > > You are quite right that it is possible for the risk of a backport to
> > > > > exceed the risk of the original bug.
> > > > >
> > > > > I defer to Joel (CCed) on how best to resolve this in -stable.
> > > >
> > > > Maybe I am missing something but this issue should also be happening
> > > > in mainline right?
> > > >
> > > > Even though mainline has 897ba84dc5aa ("rcu-tasks: Handle idle tasks
> > > > for recently offlined CPUs") , the warning should still be happening
> > > > due to Liam's "kernel/sched: Modify initial boot task idle setup"
> > > > because the warning is just rearranged a bit but essentially the same.
> > > >
> > > > IMHO, the right thing to do then is to drop Liam's patch from 5.15 and
> > > > fix it in mainline (using the ideas described in this thread), then
> > > > backport both that new fix and Liam's patch to 5.15.
> > > >
> > > > Or is there a reason this warning does not show up on the mainline?
> >
> > There is not a whole lot of commonality between the v5.15.134 version of
> > RCU Tasks Trace and that of mainline. In theory, in mainline, CPU hotplug
> > is supposed to be disabled across all calls to trc_inspect_reader(),
> > which means that there would not be any CPU coming or going.
> >
> > But there could potentially be some time between when a CPU was
> > marked as online and its idle task was marked PF_IDLE. And in
> > fact x86 start_secondary() invokes set_cpu_online() before it calls
> > cpu_startup_entry(), and it is the latter than sets PF_IDLE.
> >
> > The same is true of alpha, arc, arm, arm64, csky, ia64, loongarch, mips,
> > openrisc, parisc, powerpc, riscv, s390, sh, sparc32, sparc64, x86 xen,
> > and xtensa, which is everybody.
> >
> > One reason why my testing did not reproduce this is because I was running
> > against v6.6-rc1, and cff9b2332ab7 ("kernel/sched: Modify initial boot
> > task idle setup") went into v6.6-rc3. An initial run merging in current
> > mainline also failed to reproduce this, but I am running overnight.
> > If that doesn't reproduce, I will try inserting delays between the
> > set_cpu_online() and the cpu_startup_entry().
>
> I thought the warning happens before set_cpu_online() is even called, because
> under such situation, ofl == true and the task is not set to PF_IDLE yet:
>
> WARN_ON_ONCE(ofl && task_curr(t) && !is_idle_task(t));
That case is supposed to be excluded by the cpus_read_lock() calls.
Yes, key phrase "supposed to be". ;-)
> > If this problem is real, fixes include:
> >
> > o Revert Liam's patch and make Tiny RCU's call_rcu() deal with
> > the problem. This is overhead and non-tinyness, but to Joel's
> > point, it might be best.
> >
> > o Go back to something more like Liam's original patch, which
> > cleared PF_IDLE only for the boot CPU.
> >
> > o Set PF_IDLE before calling set_cpu_online(). This would work,
> > but it would also be rather ugly, reaching into each and every
> > architecture.
> >
> > o Move the call to set_cpu_online() into cpu_startup_entry().
> > This would require some serious inspection to prove that it is
> > safe, assuming that it is in fact safe.
> >
> > o Drop the WARN_ON_ONCE() from trc_inspect_reader(). Not all
> > that excited by losing this diagnostic, but then again it
> > has been awhile since it has caught anything.
> >
> > o Make the WARN_ON_ONCE() condition in trc_inspect_reader() instead
> > to a "return false" to retry later. Ditto, also not liking the
> > possibility of indefinite deferral with no warning.
>
> Just for completeness,
>
> o Since it just a warning, checking for task_struct::pid == 0 instead of is_idle_task()?
> Though PF_IDLE is also set in play_idle_precise().
>
> o Change warning to:
> WARN_ON_ONCE(ofl && task_curr(t) && (!is_idle_task(t) && t->pid != 0));
This change does look promising, thank you!
Thanx, Paul
