Re: [RFC v1 00/12] kunit: introduce class mocking support.

[Brendan Higgins]

On Tue, Sep 22, 2020 at 5:24 PM Daniel Latypov <dlatypov@xxxxxxxxxx> wrote:
>
> On Fri, Sep 18, 2020 at 11:31 AM Daniel Latypov <dlatypov@xxxxxxxxxx> wrote:
> >
> > # Background
> > KUnit currently lacks any first-class support for mocking.
> > For an overview and discussion on the pros and cons, see
> > https://martinfowler.com/articles/mocksArentStubs.html
> >
> > This patch set introduces the basic machinery needed for mocking:
> > setting and validating expectations, setting default actions, etc.
> >
> > Using that basic infrastructure, we add macros for "class mocking", as
> > it's probably the easiest type of mocking to start with.
> >
> > ## Class mocking
> >
> > By "class mocking", we're referring mocking out function pointers stored
> > in structs like:
> >   struct sender {
> >         int (*send)(struct sender *sender, int data);
> >   };
>
> Discussed this offline a bit with Brendan and David.
> The requirement that the first argument is a `sender *` means this
> doesn't work for a few common cases.
>
> E.g. ops structs don't usually have funcs that take `XXX_ops *`
> `pci_ops` all take a `struct pci_bus *`, which at least contains a
> `struct pci_ops*`.
>
> Why does this matter?
> We need to stash a  `struct mock` somewhere to store expectations.
> For this version of class mocking (in patch 10), we've assumed we could have
>
> struct MOCK(sender) {
>          struct mock ctrl;
>          struct sender trgt; //&trgt assumed to be the first param
> }
>
> Then we can always use `container_of(sender)` to get the MOCK(sender)
> which holds `ctrl`.
> But if the first parameter isn't a `struct sender*`, this trick
> obviously doesn't work.
>
> So to support something like pci_ops, we'd need another approach to
> stash `ctrl`.
>
> After thinking a bit more, we could perhaps decouple the first param
> from the mocked struct.
> Using pci_ops as the example:
>
> struct MOCK(pci_ops) {
>          struct mock ctrl;
>          struct pci_bus *self; // this is the first param. Using
> python terminology here.
>          struct pci_ops trgt; // continue to store this, as this holds
> the function pointers
> }
>
> // Name of this func is currently based on the class we're mocking
> static inline struct mock *from_pci_ops_to_mock(
>   const struct pci_bus *self)
> {
>           return mock_get_ctrl(container_of(self, struct MOCK(pci_ops), self));
> }
>
> // then in tests, we'd need something like
> struct pci_bus bus;
> bus.ops = mock_get_trgt(mock_ops);
> mock_ops.self = &bus;
>
> Do others have thoughts/opinions?

In the above example, wouldn't we actually want to mock struct
pci_bus, not struct pci_ops?

I think pci_bus is what actually gets implemented. Consider the
Rockchip PCIe host controller:

Rockchip provides it's own pci_ops struct:

https://elixir.bootlin.com/linux/latest/source/drivers/pci/controller/pcie-rockchip-host.c#L256

If you look at one of the implemented methods, say
rockchip_pcie_rd_conf(), you will see:

...
struct rockchip_pcie *rockchip = bus->sysdata;
...
(This function signature is:

int (*read)(struct pci_bus *bus, unsigned int devfn, int where, int
size, u32 *val);

).

Thus, conceptually struct pci_bus is what is being manipulated; it
best fits the candidate for the *self pointer.

In fact - if I am not mistaken - I think if you were to mock pci_bus
and just pretend that the methods were on pci_bus, not pci_ops, I
think it might just work.

The bigger problem is that it seems pci_bus does not want the user to
allocate it: in the case of rockchip_pcie, it is allocated by
devm_pci_alloc_host_bridge(). Thus, embedding pci_bus inside of a
MOCK(pci_bus) would probably not work, so you would need to have
different tooling around that and would then need to provide a custom
definition of from_pci_bus_to_mock() (generated by
DECLARE_STRUCT_CLASS_MOCK_CONVERTER()).

> After grepping around for examples, I'm struck by how the number of
> outliers there are.
> E.g. most take a `struct thing *` as the first param, but cfspi_ifc in
> caif_spi.h opts to take that as the last parameter.

That's not a problem, just use the
DECLARE_STRUCT_CLASS_MOCK_HANDLE_INDEX() variant to create your mock
(as opposed to DECLARE_STRUCT_CLASS_MOCK()).

For example let's say you have the following struct that you want to mock:

struct example {
    ...
    int (*foo)(int bar, char baz, struct example *self);
};

You could mock the function with:

DECLARE_STRUCT_CLASS_MOCK_HANDLE_INDEX(
    METHOD(foo), CLASS(example),
    2, /* The "handle" is in position 2. */
    RETURNS(int),
    PARAMS(int, char, struct example *));

> And others take a different mix of structs for each function.
>
> But it feels like a decoupled self / struct-holding-func-pointers
> approach should be generic enough, as far as I can tell.
> The more exotic types will probably have to remain unsupported.

I think the code is pretty much here to handle any situation in which
one of the parameters input to the function can be used to look up a
mock object; we just need to expose the from_<class>_to_mock()
function to the user to override. The problem I see with what we have
now is the assumption that the user gets to create the object that
they want to mock. Your example has illustrated that that is not a
safe assumption.

But yes, sufficiently exoctic cases will have to be unsupported.

BTW, sorry for not sharing the above in our offline discussion; since
we were talking without concrete examples in front of us, the problem
sounded worse than it looks here.

[...]