[RFC] mm/page_owner: Use per-CPU temporal storage for saving stack trace

[Kanishka Solanki]

Hi,

I am looking into resolving the following TODO comment in mm/page_owner.c:

/*
 * TODO: teach PAGE_OWNER_STACK_DEPTH (__dump_page_owner and save_stack)
 * to use off stack temporal storage
 */

Currently, save_stack() allocates a local array on the thread stack to temporarily hold instruction pointers before passing them to Stack Depot:

unsigned long entries[PAGE_OWNER_STACK_DEPTH];

Keeping this array on the kernel thread stack risks stack exhaustion.

To move this scratchpad off the thread stack, I am considering using a per-CPU buffer, similar to how kernel/trace/trace.c manages ftrace_stacks:

struct page_owner_stack {
	unsigned long entries[PAGE_OWNER_STACK_DEPTH];
};
static DEFINE_PER_CPU(struct page_owner_stack, page_owner_stacks);

Is a per-CPU scratchpad the preferred design for this off stack storage in page_owner?

Thanks,
Kanishka