Re: [PATCH 1/3] lib/stackdepot: Add a refcount field in stack_record

[Oscar Salvador]

On Thu, Sep 01, 2022 at 10:24:58AM +0200, Marco Elver wrote:
> On Thu, Sep 01, 2022 at 06:42AM +0200, Oscar Salvador wrote:
> >  include/linux/stackdepot.h | 13 ++++++-
> >  lib/stackdepot.c           | 79 +++++++++++++++++++++++++++++++-------
> >  mm/kasan/common.c          |  3 +-
> 
> +Cc other kasan maintainers

Yeah, sorry about that, I should have CCed you guys.

> > +typedef enum stack_action {
> > +	STACK_ACTION_NONE,
> > +	STACK_ACTION_INC,
> > +}stack_action_t;
> > +
> 
> missing space after '}'. But please no unnecessary typedef, just 'enum
> stack_action' (and spelling out 'enum stack_action' elsewhere) is just
> fine.

Sure, will re-name it.

> 
> This is in the global namespace, so I'd call this
> stack_depot_action+STACK_DEPOT_ACTION_*.
> 
> However, .._ACTION_INC doesn't really say what's incremented. As an
> analog to stack_depot_dec_count(), perhaps .._ACTION_COUNT?

I guess we can go "STACK_DEPOT_ACTION_COUNT", or "STACK_DEPOT_ACTION_REF_INC",
but the latter seems rather baroque for my taste.

> In general it'd be nicer if there was stack_depot_inc_count() instead of
> this additional argument, but I see that for performance reasons you
> might not like that?

Yes, the first prototypes didn't have this stack_action_t thing,
but that implied that we had to look for the stack twice
in the __set_page_owner() case.

This way we only do that in the __reset_page_owner() case.

So yes, it's a trade-off performance vs LOC.

> > --- a/lib/stackdepot.c
> > +++ b/lib/stackdepot.c
> > @@ -63,6 +63,7 @@ struct stack_record {
> >  	u32 hash;			/* Hash in the hastable */
> >  	u32 size;			/* Number of frames in the stack */
> >  	union handle_parts handle;
> > +	refcount_t count;		/* Number of the same repeated stacks */
> 
> This will increase stack_record size for every user, even if they don't
> care about the count.
> 
> Is there a way to store this out-of-line somewhere?

That would require having some kind of e.g: dynamic struct and allocating
new links to stacks as they were created and increase the refcount there.

But that would be too much of complexity, I think.

As I read in your other thread, we can probably live with that, but
it is worth spelling out in the changelog.

> > +void stack_depot_dec_count(depot_stack_handle_t handle)
> > +{
> > +	struct stack_record *stack = NULL;
> > +
> > +	stack = stack_depot_getstack(handle);
> > +	if (stack) {
> > +	/*
> > +	 * page_owner creates some stacks via create_dummy_stack().
> > +	 * We are not interested in those, so make sure we only decrement
> > +	 * "valid" stacks.
> > +	 */
> 
> Comment indent is wrong.

Will fix it.

Thanks for taking the time to review the code Marco!


-- 
Oscar Salvador
SUSE Labs