Some comments from my side:

quoted

  static inline void arch_enter_lazy_mmu_mode(void)
  {
-	/*
-	 * lazy_mmu_mode is not supposed to permit nesting. But in practice this
-	 * does happen with CONFIG_DEBUG_PAGEALLOC, where a page allocation
-	 * inside a lazy_mmu_mode section (such as zap_pte_range()) will change
-	 * permissions on the linear map with apply_to_page_range(), which
-	 * re-enters lazy_mmu_mode. So we tolerate nesting in our
-	 * implementation. The first call to arch_leave_lazy_mmu_mode() will
-	 * flush and clear the flag such that the remainder of the work in the
-	 * outer nest behaves as if outside of lazy mmu mode. This is safe and
-	 * keeps tracking simple.
-	 */
-
  	set_thread_flag(TIF_LAZY_MMU);>  }

Should not platform specific changes be deferred to subsequent patches until
nesting is completely enabled in generic first ? Although no problem as such
but would be bit cleaner.

This could indeed be done in a separate patch. But I also don't see a 
problem with updating the doc in this patch.

quoted

diff --git a/include/linux/mm_types_task.h b/include/linux/mm_types_task.h
index a82aa80c0ba4..11bf319d78ec 100644
--- a/include/linux/mm_types_task.h
+++ b/include/linux/mm_types_task.h

@@ -88,4 +88,9 @@ struct tlbflush_unmap_batch {
  #endif
  };
  
+struct lazy_mmu_state {
+	u8 enable_count;
+	u8 pause_count;
+};
+

Should not this be wrapped with CONFIG_ARCH_HAS_LAZY_MMU_MODE as the task_struct
element 'lazy_mmu_state' is only available with the feature.

No strong opinion; the compiler will ignore it either way. And less 
ifdef is good, right? :)

... and there is nothing magical in there that would result in other 
dependencies.

Besides, is a depth
of 256 really expected here ? 4 bits for each element would not be sufficient for
a depth of 16 ?

We could indeed use something like

struct lazy_mmu_state {
	u8 enable_count : 4;
	u8 pause_count : 4;
};

but then, the individual operations on enable_count/pause_count need 
more instructions.

Further, as discussed, this 1 additional byte barely matters given the 
existing size of the task struct.

No strong opinion.

quoted

   */
  #ifdef CONFIG_ARCH_HAS_LAZY_MMU_MODE
+/**
+ * lazy_mmu_mode_enable() - Enable the lazy MMU mode.
+ *
+ * Enters a new lazy MMU mode section; if the mode was not already enabled,
+ * enables it and calls arch_enter_lazy_mmu_mode().
+ *
+ * Must be paired with a call to lazy_mmu_mode_disable().
+ *
+ * Has no effect if called:
+ * - While paused - see lazy_mmu_mode_pause()
+ * - In interrupt context
+ */
  static inline void lazy_mmu_mode_enable(void)
  {
-	if (in_interrupt())
+	struct lazy_mmu_state *state = &current->lazy_mmu_state;
+
+	if (in_interrupt() || state->pause_count > 0)
  		return;
  
-	arch_enter_lazy_mmu_mode();
+	VM_WARN_ON_ONCE(state->enable_count == U8_MAX);
+
+	if (state->enable_count++ == 0)
+		arch_enter_lazy_mmu_mode();

When lazy_mmu_mode_enable() gets called for the first time with state->enable_count as 0,
then arch_enter_lazy_mmu_mode() will not get called ? Bit confused.

state->enable_count++ returns the old value (0). Are you thinking of
++state->enable_count?

But maybe I misudnerstood your concern.

[...]

quoted

+/**
+ * lazy_mmu_mode_pause() - Resume the lazy MMU mode.
+ *
+ * Resumes the lazy MMU mode; if it was active at the point where the matching
+ * call to lazy_mmu_mode_pause() was made, re-enables it and calls
+ * arch_enter_lazy_mmu_mode().
+ *
+ * Must match a call to lazy_mmu_mode_pause().
+ *
+ * Has no effect if called:
+ * - While paused (inside another pause()/resume() pair)
+ * - In interrupt context
+ */
  static inline void lazy_mmu_mode_resume(void)
  {
+	struct lazy_mmu_state *state = &current->lazy_mmu_state;
+
  	if (in_interrupt())
  		return;
  
-	arch_enter_lazy_mmu_mode();
+	VM_WARN_ON_ONCE(state->pause_count == 0);
+
+	if (--state->pause_count == 0 && state->enable_count > 0)
+		arch_enter_lazy_mmu_mode();
  }

Should not state->pause/enable_count tests and increment/decrement be handled
inside include/linux/sched via helpers like in_lazy_mmu_mode() ? This is will
ensure cleaner abstraction with respect to task_struct.

I don't think this is required given that this code here implements
CONFIG_ARCH_HAS_LAZY_MMU_MODE support.

-- 
Cheers

David