On Fri, Jun 07, 2024 at 06:14:28PM +0300, Kirill A. Shutemov wrote:

  I was able to address this issue by switching cpa_lock to a mutex.
  However, this solution will only work if the callers for set_memory
  interfaces are not called from an atomic context. I need to verify if
  this is the case.

Dunno, I'd be nervous about this. Althouth from looking at

   ad5ca55f6bdb ("x86, cpa: srlz cpa(), global flush tlb after splitting big page and before doing cpa")

I don't see how "So that we don't allow any other cpu" can't be done
with a mutex. Perhaps the set_memory* interfaces should be usable in as
many contexts as possible.

Have you run this with lockdep enabled?

- The function __flush_tlb_all() in kernel_(un)map_pages_in_pgd() must be
  called with preemption disabled. Once again, I am unsure why this has
  not caused issues in the EFI case.

It could be because EFI does all that setup on the BSP only before the
others have arrived but I don't remember anymore... It is more than
a decade ago when I did this...

Thx.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette