[PATCH v4] tee: optee: rework TA bus scanning code

[Simon Glass]

Hi Etienne,

On Thu, 22 Sept 2022 at 10:52, Etienne Carriere
<etienne.carriere at linaro.org> wrote:
>
> Hello Patrick and all,
>
> On Mon, 19 Sept 2022 at 16:49, Patrick DELAUNAY
> <patrick.delaunay at foss.st.com> wrote:
> >
> >
> > Hi Simon,
> >
> > On 9/12/22 20:31, Simon Glass wrote:
> > > Hi Ilias,
> > >
> > > On Wed, 7 Sept 2022 at 15:32, Ilias Apalodimas
> > > <ilias.apalodimas at linaro.org> wrote:
> > >> Hi Simon,
> > >>
> > >> On Thu, 8 Sept 2022 at 00:11, Simon Glass <sjg at chromium.org> wrote:
> > >>> Hi Ilias,
> > >>>
> > >>> On Tue, 6 Sept 2022 at 15:23, Ilias Apalodimas
> > >>> <ilias.apalodimas at linaro.org> wrote:
> > >>>> Hi Simon,
> > >>>>
> > >>>> On Tue, Sep 06, 2022 at 03:18:28PM -0600, Simon Glass wrote:
> > >>>>> Hi,
> > >>>>>
> > >>>>> On Tue, 6 Sept 2022 at 03:37, Ilias Apalodimas
> > >>>>> <ilias.apalodimas at linaro.org> wrote:
> > >>>>>> Late versions of OP-TEE support a pseudo bus. TAs that behave as
> > >>>>>> hardware blocks (e.g TPM, RNG etc) present themselves on a bus
> > >>>>>> whichwe can
> > >>>>>> scan. Unfortunately U-Boot doesn't support that yet. It's worth
> > >>>>>> noting
> > >>>>>> that we already have a workaround for RNG. The details are in
> > >>>>>> commit 70812bb83da6 ("tee: optee: bind rng optee driver")
> > >>>>>>
> > >>>>>> So let's add a list of devices based on U-Boot Kconfig options
> > >>>>>> that we will
> > >>>>>> scan until we properly implement the tee-bus functionality.
> > >>>>>>
> > >>>>>> While at it change the behaviour of the tee core itself wrt to device
> > >>>>>> binding. If some device binding fails, print a warning instead of
> > >>>>>> disabling OP-TEE.
> > >>>>>>
> > >>>>>> Signed-off-by: Ilias Apalodimas <ilias.apalodimas at linaro.org>
> > >>>>>> Reviewed-by: Jens Wiklander <jens.wiklander at linaro.org>
> > >>>>>> Reviewed-by: Etienne Carriere <etienne.carriere at linaro.org>
> > >>>>>> ---
> > >>>>>> Changes since v3:
> > >>>>>> - Use NULL instead of a child ptr on device_bind_driver(), since
> > >>>>>> it's not
> > >>>>>> really needed
> > >>>>>> - Changed the style of the optee_bus_probe[] definition to
> > >>>>>> {.drv_name = xxx, .dev_name = yyy }
> > >>>>>>
> > >>>>>> Changes since v2:
> > >>>>>> - Fixed typo on driver name ftpm-tee -> ftpm_tee
> > >>>>>>
> > >>>>>> Changes since v1:
> > >>>>>> - remove a macro and use ARRAY_SIZE directly
> > >>>>>> drivers/tee/optee/core.c | 24 +++++++++++++++++++-----
> > >>>>>> 1 file changed, 19 insertions(+), 5 deletions(-)
> > >>>>>>
> > >>>>>> diff --git a/drivers/tee/optee/core.c b/drivers/tee/optee/core.c
> > >>>>>> index a89d62aaf0b3..c201a4635e6b 100644
> > >>>>>> --- a/drivers/tee/optee/core.c
> > >>>>>> +++ b/drivers/tee/optee/core.c
> > >>>>>> @@ -31,6 +31,18 @@ struct optee_pdata {
> > >>>>>> optee_invoke_fn *invoke_fn;
> > >>>>>> };
> > >>>>>>
> > >>>>>> +static const struct {
> > >>>>>> + const char *drv_name;
> > >>>>>> + const char *dev_name;
> > >>>>>> +} optee_bus_probe[] = {
> > >>>>>> +#ifdef CONFIG_RNG_OPTEE
> > >>>>>> + { .drv_name = "optee-rng", .dev_name = "optee-rng" },
> > >>>>>> +#endif
> > >>>>>> +#ifdef CONFIG_TPM2_FTPM_TEE
> > >>>>>> + { .drv_name = "ftpm_tee", .dev_name = "ftpm_tee" },
> > >>>>>> +#endif
> > >>>>>> +};
> > >>>>>> +
> > >>>>>> struct rpc_param {
> > >>>>>> u32 a0;
> > >>>>>> u32 a1;
> > >>>>>> @@ -642,8 +654,7 @@ static int optee_probe(struct udevice *dev)
> > >>>>>> {
> > >>>>>> struct optee_pdata *pdata = dev_get_plat(dev);
> > >>>>>> u32 sec_caps;
> > >>>>>> - struct udevice *child;
> > >>>>>> - int ret;
> > >>>>>> + int ret, i;
> > >>>>>>
> > >>>>>> if (!is_optee_api(pdata->invoke_fn)) {
> > >>>>>> dev_err(dev, "OP-TEE api uid mismatch\n");
> > >>>>>> @@ -672,10 +683,13 @@ static int optee_probe(struct udevice *dev)
> > >>>>>> * in U-Boot, the discovery of TA on the TEE bus is not supported:
> > >>>>>> * only bind the drivers associated to the supported OP-TEETA
> > >>>>>> */
> > >>>>>> - if (IS_ENABLED(CONFIG_RNG_OPTEE)) {
> > >>>>>> - ret = device_bind_driver(dev, "optee-rng", "optee-rng", &child);
> > >>>>>> +
> > >>>>>> + for (i = 0; i < ARRAY_SIZE(optee_bus_probe); i++) {
> > >>>>>> + ret = device_bind_driver(dev, optee_bus_probe[i].drv_name,
> > >>>>>> + optee_bus_probe[i].dev_name, NULL);
> > >>>>>> if (ret)
> > >>>>>> - return ret;
> > >>>>>> + dev_warn(dev, "Failed to bind device %s\n",
> > >>>>>> + optee_bus_probe[i].dev_name);
> > >>>>> Please add device tree nodes for these and all this code can go away.
> > >>>> That's the exact opposite of what the commit message describes. OP-TEE
> > >>>> supports a scannable bus ifor TAs that behave like hardware blocks and
> > >>>> doesn't need a DT entry. Since it's really the TAs compilation decision
> > >>>> to support that or not having them as a DT node is not always the right
> > >>>> choice.
> > >>> This is continuing the perversion of how things are supposed to work
> > >>> in driver model.
> > >> Which is not the only thing we need to keep in mind though.
> > >>
> > >>> We need to talk about this because it is simply the wrong way to be
> > >>> approaching this.
> > >> This is already part of other software components though, e.g it's
> > >> already in the kernel. So I don't think it's the wrong approach.
> > >>
> > >>> There is nothing wrong with putting things in the DT
> > >>> and this is how U-Boot works. For now, please create a binding and get
> > >>> it reviewed. You don't need all the internal objects but you do need
> > >>> an OP-TEE driver and node, as we have with PCI.
> > >> Some things *are* working without a DT entry. You had similar
> > >> concerns on FF-A (where you requested a DT node again) and people gave
> > >> the exact same response. As long as a bus is scanable in any way,
> > >> it's preferable to than adding a DT entry. Moreover this code does
> > >> not prevent anyone from adding a DT entry.
> > >>
> > >> To make things even worse if the TA is compiled as 'scanable' and has
> > >> a DT entry, it might cause issues down the road when being probed by
> > >> the kernel. So really this is just a patch that makes u-boot behave
> > >> and plug in properly to the rest of the ecosystem
> > > Calling device_bind() is supposed to be used in extremis. I don't see
> > > any scanning of an OP-TEE bus here. I just see it binding two child
> > > devices which are hard-coded in U-Boot. What am I missing?
> >
> >
> > The tee bus is supported in Linux kernel (each TA have a UUID and
> > is discoverable by the TEE driver).
> >
> > see drivers/tee/optee/core.c::optee_bus_scan()
> > and "struct tee_client_driver" with TA UUID
> > It wasn't supported in U-Boot is the first TEE/OP-TEE driver implementation
> >
> > => TA support was hardcoded, under the associated CONFIG
> >        and the probe failed when the TA is not present.
> >        for example, I add this binding for TA_RNG in drivers/rng/optee_rng.c
> >
> > The TEE bus feature is added by the Etienne in the serie [1].
> > This bus is more flexible and avoid OP-TEE to dynamically modify the
> > device tree
> > to add/remove the supported SW component (TA support are activated
> > during OP-TEE
> > compilation) as the binding is managed dynamically in OP-TEE as it is
> > done in Linux.
> >
> > For information, on STM32MP15 platform, I have the trace "can't open
> > session:" for
> > RNG TA for each 'rng' command when this TA is not supported in OP-TEE but
> > OP-TEE RNG driver is activated in U-Boot, because the driver is binding.
> >
> > [1] drivers: tee: optee: remove unused probe local variable
> >
> > http://patchwork.ozlabs.org/project/uboot/list/?series=311351&state=*
> >
> > > This appears to be a Linaro binding, so you should be able to update
> > > it easily enough.
> > >
>
> Discussing with Patrick, he made a suggestion and showed me I was
> wrong in OP-TEE tee-supplicant enumeration constraints in U-Boot.
> OP-TEE exposes 2 levels of service discovery, so-called devices
> enumeration and device-with-supplicant enumeration. The later are
> OP-TEE services that depend on RPC service exposed to OP-TEE by in the
> caller OS (U-Boot or Linux kernel). The former are services without
> such dependencies. When i posted OP-TEE services discovery in U-Boot,
> I made U-Boot to enumerate OP-TEE "devices" (without tee-suppl.
> dependencies).
> I made it intentionally as U-Boot tee-supplicant does not implement
> all OP-TEE RPC services as Linux kernel. Since FTPM TA service relies
> on tee-supplicant support, it is not enumerated/discovered.
>
> The point is the U-Boot tee-suppl. does implement the few RPC services
> FTPM TA needs (that are memory allocation and RPMB access).
> So Patrick argued that U-Boot can as well enumerate OP-TEE service
> *with* tee-suppl. devices. The optee ftpm driver can register to this
> service discovery and will operate properly.
> What puzzled me was that discovery of OP-TEE services that require
> tee-suppl. services not available in U-Boot would end in a failure of
> the service, but as Patrick rightly said that it makes no sense for
> one of add implement u-boot a driver for an OP-TEE service if that
> service lacks some U-Boot tee-suppl. supports.
>
> All in one, my apologies Ilias for this mistake. A change in
> tee/optee/core.c to also bind services enumerated by OP-TEE command
> PTA_CMD_GET_DEVICES_SUPPL should enable full dynamic discovery of
> functional FTPM TA service. I'll post a change for that.

Thank you for the update.

I talked with Ilias about this at osfc. The problem I think is that it
is not possible to reference a particular random-number generator
unless it is present in the device tree.

Here are some things that can produce random numbers, i.e. are in the
RNG uclass (which we should rename to RAND or RANDOM to fit with the
other naming there):
- software
- SoC
- TPM
- OP-TEE

For choosing which one to use, we can use aliases. so rand0 is the
first device. That alias can point into OP-TEE if desitred.

If we need a particular device to specify which random-number
generator device it should use, that can be done with a phandle, like:

board-sysinfo {
   compatible = "vendor,sysinfo-board";
   ramdom = <&optee_rng>;
}

So it should work like PCI, where we scan the bus and attach the
correct device tree subnode to the OP-TEE devices. The PCI code for
this is at
pci_bind_bus_devices() where it calls pci_bus_find_devfn().

This means we need to have compatible strings and bindings for any
devices on the bus which can be used as above. If the subnodes are
missing, then scanning will result in binding them in any case.

Regards,
Simon