Re: [PATCH] x86/split_lock: Restore warn mode (and add a new one) to avoid userspace regression

[Lukas Straub <lukasstraub2@web.de>]

[-- Attachment #1: Type: text/plain, Size: 1935 bytes --]

On Thu, 29 Sep 2022 15:37:55 +0000
"Luck, Tony" <tony.luck@intel.com> wrote:

> >> I have a revert removing the misery ready and tested, let me know if I
> >> should submit it.  
> >
> > I'm a bit of a late arrival to the split lock party, so I'm a bit
> > hesitant to merge any changes immediately.
> >
> > How about we give it a few weeks and see if the current behavior impacts
> > anyone else?  Maybe the best route will be more clear then.  
> 
> Applying "misery" to the processes that are executing split-lock flows saves
> the rest of the system from a different level of misery (for the duration of the
> split lock other logical CPUs and I/O devices have access to memory blocked).
> 
> So the "misery" serves a very useful purpose on multi-user systems.

Hello Everyone,
How about the following: The kernel traps the split-lock, but instead
of punishing the process artificially it emulates it in a different way
that won't harm the system as a whole. Of course this still will be
slower than a non-split-lock but surely won't take 10ms.

For example, you could emulate the instruction without atomic semantics,
but protected under a single global mutex for all split-lock operations
instead. This is how atomics are done on on alpha AFAIK, which doesn't
have atomic instructions.

This is not as simple as the current solution. But I see the current
solution more like a quick and dirty workaround for this security/DoS
issue, until a proper solution (like my proposal) is implemented.

Regards,
Lukas Straub

> Maybe the decision of which mode to use could be dynamic based on
> number of online CPUs? Laptops/desktops with low counts (<50???)
> could just "warn", while servers could default to the "seq" mode.
> 
> Or perhaps there is some other heuristic to distinguish single-user
> systems where the split-locks are not causing pain to other users?
> 
> -Tony



-- 


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 833 bytes --]