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

[Marco Elver <elver@google.com>]

On Thu, 1 Sept 2022 at 12:01, Michal Hocko <mhocko@suse.com> wrote:
>
> On Thu 01-09-22 11:18:19, Marco Elver wrote:
> > On Thu, 1 Sept 2022 at 10:38, Michal Hocko <mhocko@suse.com> wrote:
> > >
> > > On Thu 01-09-22 10:24:58, Marco Elver wrote:
> > > > On Thu, Sep 01, 2022 at 06:42AM +0200, Oscar Salvador wrote:
> > > [...]
> > > > > diff --git a/lib/stackdepot.c b/lib/stackdepot.c
> > > > > index 5ca0d086ef4a..aeb59d3557e2 100644
> > > > > --- 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.
> > >
> > > Couldn't this be used for garbage collection?
> >
> > Only if we can precisely figure out at which point a stack is no
> > longer going to be needed.
> >
> > But more realistically, stack depot was designed to be simple. Right
> > now it can allocate new stacks (from an internal pool), but giving the
> > memory back to that pool isn't supported. Doing garbage collection
> > would effectively be a redesign of stack depot.
>
> Fair argument.
>
> > And for the purpose
> > for which stack depot was designed (debugging tools), memory has never
> > been an issue (note that stack depot also has a fixed upper bound on
> > memory usage).
>
> Is the increased size really a blocker then? I see how it sucks to
> maintain a counter when it is not used by anything but page_owner but
> storing that counte externally would just add more complexity AFAICS
> (more allocations, more tracking etc.).

Right, I think keeping it simple is better.

> Maybe the counter can be conditional on the page_owner which would add
> some complexity as well (variable size structure) but at least the
> external allocation stuff could be avoided.

Not sure it's needed - I just checked the size of stack_record on a
x86-64 build, and it's 24 bytes. Because 'handle_parts' is 4 bytes,
and refcount_t is 4 bytes, and the alignment of 'entries' being 8
bytes, even with the refcount_t, stack_record is still 24 bytes. :-)

And for me that's good enough. Maybe mentioning this in the commit
message is worthwhile. Of course 32-bit builds still suffer a little,
but I think we can live with that.