Josh Poimboeuf <jpoimboe@xxxxxxxxxx> writes:
> On Thu, Nov 09, 2023 at 12:31:47PM -0500, Steven Rostedt wrote:
>> On Thu, 9 Nov 2023 09:26:37 -0800
>> Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
>>
>> > On Tue, Nov 07, 2023 at 06:16:09PM -0500, Steven Rostedt wrote:
>> > > On Tue, 7 Nov 2023 13:56:53 -0800
>> > > Ankur Arora <ankur.a.arora@xxxxxxxxxx> wrote:
>> > >
>> > > > This reverts commit e3ff7c609f39671d1aaff4fb4a8594e14f3e03f8.
>> > > >
>> > > > Note that removing this commit reintroduces "live patches failing to
>> > > > complete within a reasonable amount of time due to CPU-bound kthreads."
>> > > >
>> > > > Unfortunately this fix depends quite critically on PREEMPT_DYNAMIC and
>> > > > existence of cond_resched() so this will need an alternate fix.
>> >
>> > We definitely don't want to introduce a regression, something will need
>> > to be figured out before removing cond_resched().
>> >
>> > We could hook into preempt_schedule_irq(), but that wouldn't work for
>> > non-ORC.
>> >
>> > Another option would be to hook into schedule(). Then livepatch could
>> > set TIF_NEED_RESCHED on remaining unpatched tasks. But again if they go
>> > through the preemption path then we have the same problem for non-ORC.
>> >
>> > Worst case we'll need to sprinkle cond_livepatch() everywhere :-/
>> >
>>
>> I guess I'm not fully understanding what the cond rescheds are for. But
>> would an IPI to all CPUs setting NEED_RESCHED, fix it?
Yeah. We could just temporarily toggle to full preemption, when
NEED_RESCHED_LAZY is always upgraded to NEED_RESCHED which will
then send IPIs.
> If all livepatch arches had the ORC unwinder, yes.
>
> The problem is that frame pointer (and similar) unwinders can't reliably
> unwind past an interrupt frame.
Ah, I wonder if we could just disable the preempt_schedule_irq() path
temporarily? Hooking into schedule() alongside something like this:
@@ -379,7 +379,7 @@ noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
void irqentry_exit_cond_resched(void)
{
- if (!preempt_count()) {
+ if (klp_cond_resched_disable() && !preempt_count()) {
The problem would be tasks that don't go through any preemptible
sections.
--
ankur
