Re: [PATCH v2 00/11] unwind, perf: sframe user space unwinding, deferred perf callchains

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Sat, Sep 14, 2024 at 08:12:46AM -0400, Steven Rostedt wrote:
> I think the unwinder should have an interface itself that provides the
> deferred unwinding, instead of having all the tracers to implement
> their own.
> 
> The user space unwinder (regardless of sframe or not) should provide a
> way to say "I want a user space stack trace for here, but don't do it
> yet. Just give me a cookie and then call my callback function with the
> stack and cookie before going back to user space".

We (Steven, Mathieu and I) have been discussing this at GNU Cauldron and
I think we're in basic agreement on this.

I think the biggest tweak we decided on is that the context id (aka
"cookie") would be percpu.  Its initial value is (cpuid << 48).  It gets
incremented for every entry from user space.

> That is, we should have an interface like:
> 
> typedef void (unwinder_callback_t)(struct user_space_stack *, u64 cookie);

> struct unwinder_client {
> 	struct list_head	list;
> 	unwinder_callback_t	callback;

I assume we want to allow tracers to pick sframes or FP (or auto).  If
so we would need to add a user_unwind_type enum to this struct.

Then the question is, what to do if tracer A wants sframe and tracer B
wants FP?  I'm thinking it's fine to allow that.  I assume the "multiple
tracers unwinding user space" case isn't realistic so any extra overhead
from these cases is the user's fault?

The unwinder would need two sets of callbacks and task work functions:
one for sframe, one for FP.

The tracer would need to pass its &my_unwinder struct to
unwinder_trigger() so the unwinder knows which task work function to
activate.


I thinking we also need a 'max_entries' field.  No need to keep
unwinding if we've already reached the tracer's max.  If there are
multiple callbacks then we'd have to get the max of the maxes, but
that's easy enough.


Mathieu had requested passing an opaque void * from the NMI to the
callback, but I don't think that's possible with this scheme since the
unwinder wouldn't know which callback to give it to.  Instead the tracer
would have to keep track of its own data associated with the given
cookie.

-- 
Josh