Re: [PATCH 07/10] test: dm: add SCMI base protocol test

[AKASHI Takahiro <takahiro.akashi@linaro.org>]

Hi Simon,

On Tue, Jul 11, 2023 at 12:41:58PM -0600, Simon Glass wrote:
> Hi Takahiro,
> 
> On Mon, 10 Jul 2023 at 19:02, AKASHI Takahiro
> <takahiro.akashi@linaro.org> wrote:
> >
> > Hi Simon,
> >
> > On Mon, Jul 10, 2023 at 01:45:58PM -0600, Simon Glass wrote:
> > > Hi,
> > >
> > > On Sun, 9 Jul 2023 at 20:04, AKASHI Takahiro <takahiro.akashi@linaro.org> wrote:
> > > >
> > > > On Fri, Jul 07, 2023 at 11:35:49AM -0600, Simon Glass wrote:
> > > > > Hi,
> > > > >
> > > > > On Tue, 4 Jul 2023 at 03:35, AKASHI Takahiro <takahiro.akashi@linaro.org> wrote:
> > > > > >
> > > > > > Hi Simon,
> > > > > >
> > > > > > On Mon, Jul 03, 2023 at 02:30:57PM +0100, Simon Glass wrote:
> > > > > > > Hi,
> > > > > > >
> > > > > > > On Mon, 3 Jul 2023 at 01:57, AKASHI Takahiro <takahiro.akashi@linaro.org> wrote:
> > > > > > > >
> > > > > > > > On Thu, Jun 29, 2023 at 08:09:58PM +0100, Simon Glass wrote:
> > > > > > > > > Hi AKASHI,
> > > > > > > > >
> > > > > > > > > On Wed, 28 Jun 2023 at 01:49, AKASHI Takahiro
> > > > > > > > > <takahiro.akashi@linaro.org> wrote:
> > > > > > > > > >
> > > > > > > > > > Added is a new unit test for SCMI base protocol, which will exercise all
> > > > > > > > > > the commands provided by the protocol, except SCMI_BASE_NOTIFY_ERRORS.
> > > > > > > > > >   $ ut dm scmi_base
> > > > > > > > > > It is assumed that test.dtb is used as sandbox's device tree.
> > > > > > > > > >
> > > > > > > > > > Signed-off-by: AKASHI Takahiro <takahiro.akashi@linaro.org>
> > > > > > > > > > ---
> > > > > > > > > >  test/dm/scmi.c | 112 +++++++++++++++++++++++++++++++++++++++++++++++++
> > > > > > > > > >  1 file changed, 112 insertions(+)
> > > > > > > > > >
> > > > > > > > > > diff --git a/test/dm/scmi.c b/test/dm/scmi.c
> > > > > > > > > > index 881be3171b7c..563017bb63e0 100644
> > > > > > > > > > --- a/test/dm/scmi.c
> > > > > > > > > > +++ b/test/dm/scmi.c
> > > > > > > > > > @@ -16,6 +16,9 @@
> > > > > > > > > >  #include <clk.h>
> > > > > > > > > >  #include <dm.h>
> > > > > > > > > >  #include <reset.h>
> > > > > > > > > > +#include <scmi_agent.h>
> > > > > > > > > > +#include <scmi_agent-uclass.h>
> > > > > > > > > > +#include <scmi_protocols.h>
> > > > > > > > > >  #include <asm/scmi_test.h>
> > > > > > > > > >  #include <dm/device-internal.h>
> > > > > > > > > >  #include <dm/test.h>
> > > > > > > > > > @@ -95,6 +98,115 @@ static int dm_test_scmi_sandbox_agent(struct unit_test_state *uts)
> > > > > > > > > >  }
> > > > > > > > > >  DM_TEST(dm_test_scmi_sandbox_agent, UT_TESTF_SCAN_FDT);
> > > > > > > > > >
> > > > > > > > > > +static int dm_test_scmi_base(struct unit_test_state *uts)
> > > > > > > > > > +{
> > > > > > > > > > +       struct udevice *agent_dev, *base;
> > > > > > > > > > +       struct scmi_agent_priv *priv;
> > > > > > > > > > +       const struct scmi_base_ops *ops;
> > > > > > > > > > +       u32 version, num_agents, num_protocols, impl_version;
> > > > > > > > > > +       u32 attributes, agent_id;
> > > > > > > > > > +       char vendor[SCMI_BASE_NAME_LENGTH_MAX],
> > > > > > > > > > +            agent_name[SCMI_BASE_NAME_LENGTH_MAX];
> > > > > > > > > > +       u8 *protocols;
> > > > > > > > > > +       int ret;
> > > > > > > > > > +
> > > > > > > > > > +       /* preparation */
> > > > > > > > > > +       ut_assertok(uclass_get_device_by_name(UCLASS_SCMI_AGENT, "scmi",
> > > > > > > > > > +                                             &agent_dev));
> > > > > > > > > > +       ut_assertnonnull(agent_dev);
> > > > > > > > > > +       ut_assertnonnull(priv = dev_get_uclass_plat(agent_dev));
> > > > > > > > > > +       ut_assertnonnull(base = scmi_get_protocol(agent_dev,
> > > > > > > > > > +                                                 SCMI_PROTOCOL_ID_BASE));
> > > > > > > > > > +       ut_assertnonnull(ops = dev_get_driver_ops(base));
> > > > > > > > > > +
> > > > > > > > > > +       /* version */
> > > > > > > > > > +       ret = (*ops->protocol_version)(base, &version);
> > > > > > > > >
> > > > > > > > > Can you add uclass helpers to call each of the methods? That is how it
> > > > > > > > > is commonly done. You should not be calling ops->xxx directly here.
> > > > > > > >
> > > > > > > > Yes, I will add inline functions instead.
> > > > > > >
> > > > > > > I don't mean inline...see all the other uclasses which define a
> > > > > >
> > > > > > Okay, I will *real* functions.
> > > > > >
> > > > > > > function which is implemented in the uclass. It is confusing when one
> > > > > > > uclass does something different. People might copy this style and then
> > > > > > > the code base diverges. Did you not notice this when looking around
> > > > > > > the source tree?
> > > > > >
> > > > > > But one concern came up in my mind.
> > > > > > Contrary to ordinary "device controllers", there exists only a single
> > > > > > implementation of driver for each of "udevice"'s associated with SCMI
> > > > > > protocols including the base protocol.
> > > > > >
> > > > > > So if I follow your suggestion, the code (base.c) might look like:
> > > > > > ===
> > > > > > static int __scmi_base_discover_vendor(struct udevice *dev, u8 *vendor)
> > > > > > {
> > > > > >   ...
> > > > > > }
> > > > > >
> > > > > > struct scmi_base_ops scmi_base_ops = {
> > > > > >
> > > > > >   .base_discover_vendor = __scmi_base_discover_vendor,
> > > > > >
> > > > > > }
> > > > > >
> > > > > > int scmi_base_discover_vendor(struct udevice *dev, u8 *vendor)
> > > > > > {
> > > > > >   struct scmi_base_ops *ops;
> > > > > >
> > > > > >   ops = scmi_base_dev_ops(dev);
> > > > > >
> > > > > >   return ops->base_discover_vendor(dev, vendor);
> > > > > > }
> > > > > > ===
> > > > > >
> > > > > > We will have to have similar definitions for every operation in ops.
> > > > > > It looks quite weird to me as there are always pairs of functions,
> > > > > > like __scmi_base_discover_vendor() and scmi_base_discover_vendor().
> > > > >
> > > > > Yes I understand that you only have one driver at present. Is there
> > > > > not a sandbox driver?
> > > >
> > > > No.
> > > > Please remember that SCMI protocol drivers on U-Boot are nothing but
> > > > stubs that makes a call to SCMI servers, supporting common communication
> > > > channel interfaces for different transports (either OP-TEE, SMCCC or mailbox).
> > > >
> > > > Sandbox driver, if is properly named, is also implemented as a sort of
> > > > transport layer, where a invocation is replaced with a function call which
> > > > mimicks one of specific commands in SCMI protocol on behalf of a real SCMI server.
> > > >
> > > > In this sense, there will exist only a single driver under the current
> > > > form of framework forever.
> > >
> > > OK, so driver model is used for the transport but not the top-level
> > > driver? I see.
> >
> > I'm not sure if you fully understand.
> > Yes, transports, or interchangeably named as a channel, are modeled
> > as U-Boot devices as you see in drivers/firmware/scmi/*_agent.c
> > (Their names, *_agent, are misleading in my opinion as their functionality
> > is purely a transport method to SCMI server. An agent means, in SCMI jargon,
> > a user or client of SCMI server.)
> >
> > On top of that, each SCMI protocol, the base protocol in my patch set,
> > is also modeled as a U-Boot device.
> > You can see another example, say, in drivers/firmware/clk/clk_scmi.c.
> >
> > Since there is no corresponding uclass for the base protocol, I create
> > a new one (UCLASS_SCMI_BASE) even though it may not be seen an concrete
> > device object.
> 
> It doesn't need to be 'concrete', whatever that means. A uclass is a
> model of hardware or something else. We have lots of drivers that
> don't deal with a hardware device.

Yes, I know.

> >
> > > >
> > > > >
> > > > > >
> > > > > > We can avoid this redundant code easily by eliminating "ops" abstraction.
> > > > > > But as far as I remember, you insist that every driver that complies
> > > > > > to U-Boot driver model should have a "ops".
> > > > > >
> > > > > > What do you make of this?
> > > > >
> > > > > Well there are some exceptions, but yes that is the idea. Operations
> > > > > should be in a 'ops' struct and documented and implemented in a
> > > > > consistent way.
> > > >
> > > > Is it your choice that I should keep "ops" structure in this specific
> > > > implementation?
> > >
> > > I can't actually find this patch on patchwork.
> >
> > Indeed (why?), but you have already seen it.
> > https://lists.denx.de/pipermail/u-boot/2023-June/521288.html
> 
> I mean patchwork.

Yes, I know.
But even a bunch of patches listed in patchwork are left "new" state forever.

> >
> > > But yes, you do need a function for each ops call. They should be used
> > > in the tests, which should not directly call functions using
> > > ops->xxx()
> >
> > Even though there is no practical benefit to do so. Right?
> 
> I don't have a useful answer to that question. If you want to
> contribute to U-Boot, please follow its conventions. It is just
> frustrating for people doing code review, with limited time.

I simply wanted to double-check the rule since you mention,
in another thread, that there are some exceptions.

> While xxx may be unique in U-Boot, I very much doubt it always is, so
> using 'ops' is not helping people looking through the code, nor is it
> obvious what is happening without looking up 'ops'. The helper
> functions should be used in all cases.
> 
> The thing is, this convention has been in place since the start of
> driver model, so I wonder how you have missed it?

Well, probably I was confused with many uses of function pointers
for boot and runtime services in UEFI code.

That said, after rethinking, I decided to remove udevice-related
code from my patch here, i.e. there will be no U-Boot device
for the base protocol as it is a protocol not a concrete device.
Its sole purpose is to get *meta* information about SCMI server
and the services, then manipulate them.

The analogy is that HTTP or FTP is a protocol over an ethernet
device to access a remote application (server), but is not a
device object in any sense.

> Please also fix the other uclasses, as I see for example that
> devm_scmi_process_msg() calls ops->process_msg() when it should have
> an exported function in that file.

I'm not sure that this is the case. devm_scmi_process_mesg()
is just a wrapper, as the name suggests, to ops->process_msg()
like other uclass driver operations, say blk_read/blk_write().
The only difference, if any, "ops" doesn't derive from the device
itself, but it parent (or grand^* parent).

If you really want, I will fix it (but in a separate patch).

-Takahiro Akashi


> Every uclass should have its
> operations in a struct, clearly documented, as well as helper
> functions to call each operation.
> 
> Regards,
> Simon