On Fri, Oct 07, 2022, Sean Christopherson wrote:
> On Thu, Oct 06, 2022, Mingwei Zhang wrote:
> > I have a limited knowledge on MM, but from my observations, it looks
> > like the property of a page being "PINNED" is very unreliable (or
> > expensive), i.e., anyone can jump in and pin the page. So it is hard
> > to see whether a page is truly "PINNED" or maybe just someone is
> > "working" on it without holding the lock.
>
> mm/ differentiates between various types of pins, e.g. elevated refcount vs. pin
> vs. longterm pin. See the comments in include/linux/mm.h for FOLL_PIN.
Ah, after catching up on the off-list thread, I suspect you're referring to the
the fact that even longterm pins don't prevent zapping[*].
NUMA balancing - already discussed
PMD splitting - If necessary, solvable by introducing MMU_NOTIFY_SPLIT, as a true
split doesn't reclaim memory, i.e. KVM doesn't need to do WBINVD.
Unfortunately, not straightforward as __split_huge_pmd_locked() may
zap instead of split, but it is solvable.
KSM - absolutely should not be turned on for SEV guests
[*] https://lore.kernel.org/linux-arm-kernel/YuEMkKY2RU%2F2KiZW@monolith.localdoman
