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

[David Gow]

On Wed, 5 Apr 2023 at 01:55, Benjamin Berg <benjamin@xxxxxxxxxxxxxxxx> wrote:
>
> Hi,
>
> On Tue, 2023-04-04 at 15:32 +0200, Maxime Ripard wrote:
> > [SNIP]
> > > +/**
> > > + * 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);
> >
> > Do we expect any other context than GFP_KERNEL?
> >
> > If so, then maybe we can have kunit_add_action() assume GFP_KERNEL and
> > add a variant for the odd case where we would actually need a different
> > GFP flag.
>
> Does anything other than GFP_KERNEL make sense? I would assume these
> functions should only ever be called from a kunit context, i.e. the
> passed test is guaranteed to be identical to the value returned by
> kunit_get_current_test().

That's not strictly-speaking guaranteed. (Indeed, we have some, albeit
contrived, counterexamples in the test.)

The theoretical use-case here is if the kunit context pointer is
passed to another thread or workqueue or something.

There aren't any such users, yet (apart from, possibly,
kunit_kmalloc_array()), though. So we could use GFP_KERNEL by default
for now, and add a variant if such a use-case appears.

>
> That said, I am happy with merging this in this form. I feel the right
> thing here is a patch (with corresponding spatch) that changes all of
> the related APIs to remove the gfp argument.
>
> > > +/**
> > > + * kunit_remove_action_token() - Cancel a deferred action.
> > > + * @test: Test case the action is associated with.
> > > + * @cancel_token: The cancellation token returned by kunit_add_action()
> > > + *
> > > + * Prevent an action deferred using kunit_add_action() from executing when the
> > > + * test ends.
> > > + *
> > > + * You can also use the (test, function, context) triplet to remove an action
> > > + * with kunit_remove_action().
> > > + */
> > > +void kunit_remove_action_token(struct kunit *test, struct kunit_action_ctx *cancel_token);
> >
> > It's not clear to me why we still need the token. If
> > kunit_remove_action() works fine, why would we need to store the token?
> >
> > Can't we just make kunit_add_action() return an int to indicate whether
> > it failed or not, and that's it?
> >
> > > [SNIP]
> >
> > One thing worth pointing is that if kunit_add_action() fails, the
> > cleanup function passed as an argument won't run.
> >
> > So, if the kzalloc call ever fails, patch 2 will leak its res->data()
> > resource for example.
> >
> > devm (and drmm) handles this using a variant called
> > devm_add_action_or_reset, we should either provide the same variant or
> > just go for that behavior by default.
>
> Both version of the function would need a return value. An alternative
> might be to make assertions part of the API. But as with dropping the
> gfp argument, that seems like a more intrusive change that needs to
> happen independently.
>
> Anyway, I am fine with action_or_reset as the default and possibly the
> only behaviour. I expect that every API user will want an assertion
> that checks for failure here anyway.
>

I'm tempted to just have both kunit_add_action() and
kunit_add_action_or_reset(), just to keep things matching the devm_
API to minimise any confusion.

And if we're not too worried about proliferating variants of these
(and, personally, I quite like them), we could have a
kunit_add_action_or_asserrt() version as well.

> Benjamin
>
>
> If kunit_* functions can assert in error conditions, then the example
>
> void test_func(struct kunit *test)
> {
>   char u8 *buf = kunit_kzalloc(test, 1024, GFP_KERNEL);
>   struct sk_buff *skb_a;
>   struct sk_buff *skb_b;
>   /* Further variables */
>
>   KUNIT_ASSERT_NOT_NULL(test, buf);
>
>   skb_a = skb_alloc(1024, GFP_KERNEL);
>   KUNIT_ASSERT_NOT_NULL(test, skb_a);
>   if (kunit_add_cleanup(test, (kunit_defer_function_t) kfree_skb, skb_a))
>     KUNIT_ASSERT_FAILURE("Failed to add cleanup");
>
>   /* Or, maybe: */
>   skb_b = skb_alloc(1024, GFP_KERNEL);
>   KUNIT_ASSERT_NOT_NULL(test, skb_b);
>   KUNIT_ASSERT_EQ(test, 0,
>                   kunit_add_cleanup(test,
>                                     (kunit_defer_function_t) kfree_skb,
>                                     skb_b));
>
>   /* run code that may assert */
> }
>
>
> could be shortened to (with a trivial kunit_skb_alloc helper)
>
> void test_func(struct kunit *test)
> {
>   char u8 *buf = kunit_kzalloc(test, 1024, GFP_KERNEL);
>   struct sk_buff *skb_a = kunit_skb_alloc(1024, GFP_KERNEL);
>   struct sk_buff *skb_b = kunit_skb_alloc(1024, GFP_KERNEL);
>   /* Further variables */
>
>   /* run code that may assert */
> }
>
> I should just post a patch for the existing API and see what people say
> then ...

We definitely already have some functions (e.g.
__kunit_activate_static_stub()) which just assert on failure. In
general, we've avoided doing so where we think there might be a good
reason to handle failures separately (or it makes the API diverge a
lot from a function we're imitating), but I'm open to using them more.
Specialised handling of allocation failures in a test is likely to be
rare enough that making those who need it write their own helpers
wouldn't be a disaster. Or we could have an _or_assert() variant.

In any case, I think your example pretty comfortably speaks for itself.

Cheers,
-- David

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