Re: [PATCH 2/2] x86/MCE: Add command line option to extend MCE Records pool

[Paul E. McKenney]

On Mon, Feb 12, 2024 at 09:10:38PM +0100, Borislav Petkov wrote:
> On Mon, Feb 12, 2024 at 07:49:43PM +0000, Luck, Tony wrote:
> > Yes. The question is whether a #MC that come in the middle of list_rcu_add()
> > can safely do list_for_each_entry_rcu() on the same list.
> > 
> > RCU is black magic ... maybe it can do this? Adding Paul.
> 
> Yeah, the list traversal might be ok as this is what that list_add does
> - you can't encounter an inconsistent list - but we still take
>  a spinlock on addition and the commit which added it:
> 
> 7f184275aa30 ("lib, Make gen_pool memory allocator lockless")
> 
> says
> 
>     The lockless operation only works if there is enough memory available.
>     If new memory is added to the pool a lock has to be still taken.  So
>     any user relying on locklessness has to ensure that sufficient memory
>     is preallocated.
> 
> and this is exactly what we're doing - adding new memory.

Is the #MC adding new memory, or is the interrupted context adding new
memory?

> So, until we're absolutely sure that it is ok to interrupt a context
> holding a spinlock with a #MC which is non-maskable, I don't think we
> wanna do this.

If it is the #MC adding new memory, agreed.

If the #MC is simply traversing the list, and the interrupted context
was in the midst of adding a new element, this should be no worse than
some other CPU traversing the list while this CPU is in the midst of
adding a new element.

Or am I missing a turn in here somewhere?

							Thanx, Paul