On 05/09/2025 11:46, Alexander Gordeev wrote:

On Thu, Sep 04, 2025 at 01:57:29PM +0100, Kevin Brodsky wrote:

Hi Kevin,

quoted

When the lazy MMU mode was introduced eons ago, it wasn't made clear
whether such a sequence was legal:

	arch_enter_lazy_mmu_mode()
	...
		arch_enter_lazy_mmu_mode()
		...
		arch_leave_lazy_mmu_mode()
	...
	arch_leave_lazy_mmu_mode()

I did not take too deep - sorry if you already answered this.
Quick question - whether a concern Ryan expressed is addressed
in general case?

The short answer is yes - it's good that you're asking because I failed
to clarify this in the cover letter!

https://lore.kernel.org/all/3cad01ea-b704-4156-807e-7a83643917a8@arm.com/ (local)

	enter_lazy_mmu
		for_each_pte {
			read/modify-write pte

			alloc_page
				enter_lazy_mmu
					make page valid
				exit_lazy_mmu

			write_to_page
		}
	exit_lazy_mmu

<quote>
This example only works because lazy_mmu doesn't support nesting. The "make page
valid" operation is completed by the time of the inner exit_lazy_mmu so that the
page can be accessed in write_to_page. If nesting was supported, the inner
exit_lazy_mmu would become a nop and write_to_page would explode.
</quote>

Further down in the cover letter I refer to the approach Catalin
suggested [4]. This was in fact in response to this concern from Ryan.
The key point is: leave() keeps the lazy MMU mode enabled if it is
nested, but it flushes any batched state *unconditionally*, regardless
of nesting level. See patch 3-6 on the practical implementation of this;
patch 7 also spells it out in the documentation.

Hope that clarifies the situation!

- Kevin

[4] https://lore.kernel.org/all/aEhKSq0zVaUJkomX@arm.com/ (local)