[PATCH] firmware: arm_ffa: Handle compatibility with different firmware versions

[Sudeep Holla]

On Fri, Oct 15, 2021 at 02:49:02PM +0200, Jens Wiklander wrote:
> On Fri, Oct 15, 2021 at 1:23 PM Sudeep Holla <sudeep.holla at arm.com> wrote:
> >
> > On Thu, Oct 14, 2021 at 10:55:01AM +0200, Jens Wiklander wrote:
> > > On Wed, Oct 13, 2021 at 6:08 PM Sudeep Holla <sudeep.holla at arm.com> wrote:
> > > >
> > > > On Wed, Oct 13, 2021 at 03:10:50PM +0200, Jens Wiklander wrote:
> > > > > On Wed, Oct 13, 2021 at 11:11 AM Sudeep Holla <sudeep.holla at arm.com> wrote:
> > > > > >
> > > > > > The driver currently just support v1.0 of Arm FFA specification. It also
> > > > > > expects the firmware implementation to match the same and bail out if it
> > > > > > doesn't match. This is causing issue when running with higher version of
> > > > > > firmware implementation(e.g. v1.1 which will released soon).
> > > > > >
> > > > > > In order to support compatibility with different firmware versions, let
> > > > > > us add additional checks and find the compatible version the driver can
> > > > > > work with.
> > > > > >
> > > > > > Signed-off-by: Sudeep Holla <sudeep.holla at arm.com>
> > > > > > ---
> > > > > >  drivers/firmware/arm_ffa/driver.c | 37 ++++++++++++++++++++++++++-----
> > > > > >  1 file changed, 32 insertions(+), 5 deletions(-)
> > > > > >
> >
> > [...]
> >
> > > > > > +static u32 ffa_compatible_version_find(u32 version)
> > > > > > +{
> > > > > > +       u32 compat_version;
> > > > > > +       u16 major = MAJOR_VERSION(version), minor = MINOR_VERSION(version);
> > > > > > +       u16 drv_major = MAJOR_VERSION(FFA_DRIVER_VERSION);
> > > > > > +       u16 drv_minor = MINOR_VERSION(FFA_DRIVER_VERSION);
> > > > > > +
> > > > > > +       if ((major < drv_major) || (major == drv_major && minor <= drv_minor))
> > > > > > +               return version;
> > > > >
> > > > > A mismatch in the major version number makes the versions
> > > > > incompatible. There's no recovery from this, an error code must be
> > > > > returned instead.
> > > > >
> > > >
> > > > The point is if the firmware can operate at reduced functionality, why just
> > > > return error. Most other specs operate with that assumption.
> > > > e.g. if we just add few extra functionality in say v2.0 vs v1.y, would
> > > > you not want the v1.y driver and firmware v2.0 work together ? Especially
> > > > if we give the flexibility for the firmware to decide that and return
> > > > the version as v2.0 in response to v1.y by the caller.
> > > >
> > > > If firmware is completely incompatible, it still has option top return
> > > > NOT_SUPPORTED.
> > >
> > > OK, how do you anticipate that this will work in practice? I mean how
> > > will the driver know which functions are compatible in a v2.0 api when
> > > the driver itself only is at v1.1? Which parts are available at
> > > reduced functionality?
> > >
> >
> > Sorry I wasn't clear earlier. The driver is just aware of v1.1 and nothing
> > more. The firmware implementing FF-A has been upgraded to v2.0 for example
> > and must have knowledge that it implements everything mandatory(not
> > discoverable) functionality as per v1.1. If it does, it can advertise it
> > supports v2.0 in response to the caller specifying v1.1 which for the
> > caller means the firmware can operate well with v1.1 even though from
> > firmware perspective it is reduced functionality.
> 
> In this case it seems that v2.0 is backwards compatible. Why was the
> major version number increased then? It seems more natural to only
> increase the major version number when it's actually needed.
>

Do you mean to say if the caller(driver in this case) passes v1.x and 
even if firmware implements v2.0, it must return what caller asked ?
Sounds good enough but at-least my opinion here is more to let know that
driver can be upgraded if available and required.

E.g. vendor may be stuck with some old version of the kernel but
the firmware is latest. This way we can advertise the version of firmware
so that one can upgrade if the driver changes are available in future
kernel versions for example.

> >
> > On the other hand, if v2.0 spec changes removed some of the mandatory
> > functionality in v1.1 or v1.x, then firmware can return NOT_SUPPORTED
> > to the caller with v1.x as argument or it may implement all the necessary
> > APIs and return its version to indicate that it is fine to interact and
> > provide functionality(though all functionality may not be used by the
> > caller as caller is unware in this case).
> 
> I get the incompatible case, but not the latter when it's fine to
> interact. Is the latter the same case as described further up with a
> backwards compatible change?
>

Not sure if we are aligned with the configurations we are talking here.
I see 3 possibility(not considering major and minor version details for
simplicity)
1. Firmware <= Driver
	Driver must match and operate at firmware version
2. Firmware > Driver and Firmware OK with Driver version
	Returns implemented version but driver uses its higher supported
	version which was used in FFA_VERSION call
3. Firmware > Driver and Firmware not OK(KO) with Driver version
	Returns NOT_SUPPORTED and driver halts there

Let me know if this improves the confusion.

> >
> > Hope I didn't add more confusions and this clarifies what is the intention
> > here.
> 
> Is this about allowing to increase the major number even when it's
> backwards compatible with the previous version?
> 

Yes, to let the caller know the actual firmware version(which may not be
of much use!)

-- 
Regards,
Sudeep