Re: [PATCH v5 00/25] context_tracking,x86: Defer some IPIs until a user->kernel transition

[Valentin Schneider]

On 02/05/25 06:53, Dave Hansen wrote:
> On 5/2/25 02:55, Valentin Schneider wrote:
>> My gripe with that was having two separate mechanisms
>> - super early entry around SWITCH_TO_KERNEL_CR3)
>> - later entry at context tracking
>
> What do you mean by "later entry"?
>

I meant the point at which the deferred operation is run in the current
patches, i.e. ct_kernel_enter() - kernel entry from the PoV of context
tracking.

> All of the paths to enter the kernel from userspace have some
> SWITCH_TO_KERNEL_CR3 variant. If they didn't, the userspace that they
> entered from could have attacked the kernel with Meltdown.
>
> I'm theorizing that if this is _just_ about avoiding TLB flush IPIs that
> you can get away with a single mechanism.

So right now there would indeed be the TLB flush IPIs, but also the
text_poke() ones (sync_core() after patching text).

These are the two NOHZ-breaking IPIs that show up on my HP box, and that I
also got reports for from folks using NOHZ_FULL + CPU isolation in
production, mostly on SPR "edge enhanced" type of systems.

There's been some other sources of IPIs that have been fixed with an ad-hoc
solution - disable the mechanism for NOHZ_FULL CPUs or do it differently
such that an IPI isn't required, e.g.

  https://lore.kernel.org/lkml/ZJtBrybavtb1x45V@tpad/

While I don't expect the list to grow much, it's unfortunately not just the
TLB flush IPIs.