Re: [PATCH v2 1/3] x86/mce/inject: Remove call to mce_notify_irq()

[Nikolay Borisov]

On 16.02.25 г. 18:10 ч., Borislav Petkov wrote:
> On Mon, Feb 10, 2025 at 05:47:04PM +0200, Nikolay Borisov wrote:
> > The call is actually a noop because when the MCE is raised the early
> > notifier is the only call site that correctly calls mce_notify_irq()
> > because it also sets mce_need_notify. Remove this call and as a result
> > make mce_notify_irq() static
> >
> > Signed-off-by: Nikolay Borisov <nik.borisov@xxxxxxxx>
> > ---
> >   arch/x86/include/asm/mce.h       |  2 --
> >   arch/x86/kernel/cpu/mce/core.c   | 44 ++++++++++++++++----------------
> >   arch/x86/kernel/cpu/mce/inject.c |  1 -
> >   3 files changed, 22 insertions(+), 25 deletions(-)
>
> So what you're looking at are the remnants of the old soft-inject of MCE
> errors. And it seems that we lost some of that functionality along the way and
> no one has noticed because, well, it seems no one uses it anymore.
>
> In order to understand how this thing was supposed to work, checkout
>
> ea149b36c7f5 ("x86, mce: add basic error injection infrastructure")


The original code in  ea149b36c7f5 was setting the notify_user bit via

raise_mce()->machine_check_poll()->mce_log(),
 mce_notify_user() - consumes notify_user set in mce_log above.


 subsequently in 011d82611172 ("RAS: Add a Corrected Errors Collector") 
you factored out the code from mce_log() mce_first_notifier, where the 
bit setting remains to this day, but since it's been removed from 
mce_log() it made the call in raise_local() defunct.


Considering that no one complained about this after all these years and 
that the dev-mcelog is deprecated I think it still makes sense to make 
mce_notify_irq() private

> and follow what raise_mce() does and pay attention to notify_user which is
> what mce_need_notify was called back then.
>
> Remember to have fun :-P