Re: [RFC PATCH v2 1/3] kunit: Add kunit_add_action() to defer a call until test exit

[David Gow]

On Fri, 14 Apr 2023 at 19:00, Benjamin Berg <benjamin@xxxxxxxxxxxxxxxx> wrote:
>
> Hi,
>
> On Fri, 2023-04-14 at 12:01 +0200, maxime@xxxxxxxxxx wrote:
> > Hi David,
> >
> > On Fri, Mar 31, 2023 at 04:04:09PM +0800, David Gow wrote:
> > > Many uses of the KUnit resource system are intended to simply defer
> > > calling a function until the test exits (be it due to success or
> > > failure). The existing kunit_alloc_resource() function is often used for
> > > this, but was awkward to use (requiring passing NULL init functions, etc),
> > > and returned a resource without incrementing its reference count, which
> > > -- while okay for this use-case -- could cause problems in others.
> > >
> > > Instead, introduce a simple kunit_add_action() API: a simple function
> > > (returning nothing, accepting a single void* argument) can be scheduled
> > > to be called when the test exits. Deferred actions are called in the
> > > opposite order to that which they were registered.
> > >
> > > This mimics the devres API, devm_add_action(), and also provides
> > > kunit_remove_action(), to cancel a deferred action, and
> > > kunit_release_action() to trigger one early.
> > >
> > > This is implemented as a resource under the hood, so the ordering
> > > between resource cleanup and deferred functions is maintained.
> > >
> > > Signed-off-by: David Gow <davidgow@xxxxxxxxxx>
> > > ---
> > >
> > > Changes since RFC v1:
> > > https://lore.kernel.org/linux-kselftest/20230325043104.3761770-2-davidgow@xxxxxxxxxx/
> > > - Rename functions to better match the devm_* APIs. (Thanks Maxime)
> > > - Embed the kunit_resource in struct kunit_action_ctx to avoid an extra
> > >   allocation (Thanks Benjamin)
> > > - Use 'struct kunit_action_ctx' as the type for cancellation tokens
> > >   (Thanks Benjamin)
> > > - Add tests.
> > >
> > > ---
> > >  include/kunit/resource.h |  89 ++++++++++++++++++++++++++++
> > >  lib/kunit/kunit-test.c   | 123 ++++++++++++++++++++++++++++++++++++++-
> > >  lib/kunit/resource.c     |  99 +++++++++++++++++++++++++++++++
> > >  3 files changed, 310 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/include/kunit/resource.h b/include/kunit/resource.h
> > > index c0d88b318e90..15efd8924666 100644
> > > --- a/include/kunit/resource.h
> > > +++ b/include/kunit/resource.h
> > > @@ -387,4 +387,93 @@ static inline int kunit_destroy_named_resource(struct kunit *test,
> > >   */
> > >  void kunit_remove_resource(struct kunit *test, struct kunit_resource *res);
> > >
> > > +typedef void (*kunit_defer_function_t)(void *ctx);
> > > +
> > > +/* An opaque token to a deferred action. */
> > > +struct kunit_action_ctx;
> > > +
> > > +/**
> > > + * kunit_add_action() - Defer an 'action' (function call) until the test ends.
> > > + * @test: Test case to associate the action with.
> > > + * @func: The function to run on test exit
> > > + * @ctx: Data passed into @func
> > > + * @internal_gfp: gfp to use for internal allocations, if unsure, use GFP_KERNEL
> > > + *
> > > + * Defer the execution of a function until the test exits, either normally or
> > > + * due to a failure.  @ctx is passed as additional context. All functions
> > > + * registered with kunit_add_action() will execute in the opposite order to that
> > > + * they were registered in.
> > > + *
> > > + * This is useful for cleaning up allocated memory and resources.
> > > + *
> > > + * Returns:
> > > + *   An opaque "cancellation token", or NULL on error. Pass this token to
> > > + *   kunit_remove_action_token() in order to cancel the deferred execution of
> > > + *   func().
> > > + */
> > > +struct kunit_action_ctx *kunit_add_action(struct kunit *test, kunit_defer_function_t func,
> > > +                     void *ctx, gfp_t internal_gfp);
> >
> > I've tried to leverage kunit_add_action() today, and I'm wondering if
> > passing the struct kunit pointer to the deferred function would help.
> >
> > The code I'm struggling with is something like:
> >
> > > static int test_init(struct kunit *test)
> > > {
> > >         priv = kunit_kzalloc(sizeof(*priv), GFP_KERNEL);
> > >         KUNIT_ASSERT_NOT_NULL(test, priv);
> > >         test->priv = priv;
> > >
> > >         priv->dev = alloc_device();
> > >
> > >         return 0;
> > > }
> >
> > and then in the test itself:
> >
> > > static void actual_test(struct kunit *test)
> > > {
> > >         struct test_priv *priv = test->priv;
> > >
> > >         id = allocate_buffer(priv->dev);
> > >
> > >         KUNIT_EXPECT_EQ(test, id, 42);
> > >
> > >         free_buffer(priv->dev, id);
> > > }
> >
> > I'd like to turn free_buffer an action registered right after allocate
> > buffer. However, since it takes several arguments and kunit_add_action
> > expects a single pointer, we would need to create a structure for it,
> > allocate it, fill it, and then free it when the action has ran.
> >
> > It creates a lot of boilerplate, while if we were passing the pointer to
> > struct kunit we could access the context of the test as well, and things
> > would be much simpler.
>
> The question seems to be what about the typical use-case. I was always
> imagining calling functions like kfree/kfree_skb which often only
> require a single argument.

Yeah, my thought was that having just the one argument would be
easiest for re-using existing functions. That being said, implementing
a simple wrapper which just discards the 'test' argument is probably
more ergonomic than having to write all the struct manipulation stuff,
so it depends a bit on what proves the more common case.

> For arbitrary arguments, a struct and custom free function will be
> needed. At that point, maybe it is fair to assume that API users will
> use the resource API directly, doing the same trick as kunit_add_action
> and storing the arguments together with struct kunit_resource.
>
At this point, I'd still probably use the kunit_add_action() API, just
because the resource one adds yet more complication with the 'init'
function and the reference counting. I have some vague plans to
simplify it a bit, but still definitely wouldn't rule out using the
action API here, even if it involves managing structs.


> That said, maybe one could add it as a second argument? It is a little
> bit weird API wise, but it would allow simply casting single-argument
> functions in order to ignore "struct kunit *" argument.
>

Ooh... that's evil in a particularly fun way. :-) It'd definitely be
convenient in both cases (we usually need to cast kfree() anyway for
const reasons), so I'm a bit tempted. Do we know that this would work
with the calling convention on all architectures? I'm not aware of the
kernel using anything like stdcall where the callee pops the stack on
return, but there definitely could be some architecture which does...

Cheers,

-- David

Attachment: smime.p7s
Description: S/MIME Cryptographic Signature