Ah, this is a lockdep thing.
The raw_local_irq_save/restore in __native_flush_tlb_global does not update hardirqs_enabled.
When we call into ftrace, when we cross page bounderies, we disable interrupts using the normal local_irq_save/restore calls.
But when we restore, it detects that interrupts are not going to be enabled, and keeps hardirqs_enabled off.
The printk solved the issue because it called local_irq_restore, which set the variable back.
I guess there's two solutions here. One, we can change the raw_local_irq_enable/disable variants in __native_flush_tlb_global to the non-raw type (it should protect against recursion).
or we can try to make the ring buffer use the raw_local_irq variants too. I tried this once before, and it did cause issues.
Note, there's a "check_flags" in lockdep, but it is only called on locking, it is not called when we only disable/enable interrupts.