Re: [PATCH 1/2] tee: system invocation

[Etienne Carriere]

On Thu, 9 Feb 2023 at 10:19, Jens Wiklander <jens.wiklander@xxxxxxxxxx> wrote:
>
> Hi,
>
> On Thu, Feb 09, 2023 at 09:11:53AM +0100, Etienne Carriere wrote:
> > Hi Jens,
> >
> >
> > On Thu, 9 Feb 2023 at 08:14, Jens Wiklander <jens.wiklander@xxxxxxxxxx> wrote:
> > >
> > > Hi Etienne,
> > >
> > > On Wed, Feb 08, 2023 at 06:09:17PM +0100, Etienne Carriere wrote:
> > > > Hello Sumit, Jens,
> > > >
> > > [snip]
> > > > > > > > >
> > > > > > > > >         if  (rpc_arg && tee_shm_is_dynamic(shm)) {
> > > > > > > > > -               param.a0 = OPTEE_SMC_CALL_WITH_REGD_ARG;
> > > > > > > > > +               if (ctx->sys_service &&
> > > > > > > > > +                   (optee->smc.sec_caps & OPTEE_SMC_SEC_CAP_SYSTEM_THREAD))
> > > > > > > > > +                       param.a0 = OPTEE_SMC_CALL_SYSTEM_WITH_REGD_ARG;
> > > > > > > > > +               else
> > > > > > > > > +                       param.a0 = OPTEE_SMC_CALL_WITH_REGD_ARG;
> > > > > > > >
> > > > > > > > This system thread flag should also be applicable to platforms without
> > > > > > > > registered arguments support. IOW, we need similar equivalents for
> > > > > > > > OPTEE_SMC_FUNCID_CALL_WITH_ARG and OPTEE_SMC_FUNCID_CALL_WITH_RPC_ARG
> > > > > > > > too. So I would rather suggest that we add following flag to all 3
> > > > > > > > call types:
> > > > > > > >
> > > > > > > > #define OPTEE_SMC_CALL_SYSTEM_THREAD_FLAG    0x8000
> > > > > > >
> > > > > > > The main reason platforms don't support registered arguments is that
> > > > > > > they haven't been updated since this was introduced. So if a platform
> > > > > > > needs system threads it could update to use registered arguments too.
> > > > > >
> > > > > > Are we hinting at deprecating reserved shared memory support? If yes,
> > > > > > wouldn't it be better to be explicit about it with a boot time warning
> > > > > > message about its deprecation?
> > > > > >
> > > > > > Otherwise it will be difficult to debug for the end user to find out
> > > > > > why system thread support isn't activated.
> > > > > >
> > > > > > > The Linux kernel already supports registered arguments. An advantage
> > > > > > > with the current approach is that the ABI is easier to implement
> > > > > > > since we have distinct SMC IDs for each function.
> > > > > >
> > > > > > I see your point but my initial thought was that we don't end up
> > > > > > making that list too large that it becomes cumbersome to maintain,
> > > > > > involving all the combinatorial.
> > > > >
> > > > > You have a point. Etienne, do you think we could give it a try at
> > > > > https://github.com/OP-TEE/optee_os/pull/5789 to better see how this
> > > > > would play out?
> > > > >
> > > >
> > > > Indeed I miss that...
> > > > With the patch proposed here, indeed if OP-TEE does not support
> > > > dynamic shared memory then Linux will never use the provisioned TEE
> > > > thread. This is weird as in such a case OP-TEE provisions resources
> > > > that will never be used, which is the exact opposite goal of this
> > > > feature. Verified on our qemu-arm setup.
> > > >
> > > > For simplicity, I think this system invocation should require OP-TEE
> > > > supports dyn shm.
> > >
> > > It's not obvious to me that this will easier to implement and maintain.
> > > Looking at the code in optee_os it looks like using a flag bit as
> > > proposed by Sumit would be quite easy to handle.
> >
> > OP-TEE could auto disable thread provis when dyn shm is disabled, right.
> > Will it be sufficient? We will still face cases where an OP-TEE
> > provisions thread but Linux kernel is not aware (older vanilla kernel
> > used with a recent OP-TEE OS). Not much platforms are really affected
> > I guess but those executing with pager in small RAMs where a 4kB
> > thread context costs.
>
> When you add exceptions you make it more complicated. Now we must
> remember to always use dyn shm if we are to succeed in configuring with
> system threads. What if both dyn shm and static shm is configured in
> OP-TEE, but the kernel only uses static shm?
>
> > > > If OP-TEE could know when Linux does not support TEE system
> > > > invocation, then OP-TEE could let any invocation use these provisioned
> > > > resources so that they are not wasted.
> > > > I think a good way would be Linux to expose if it supports this
> > > > capability, during capabilities exchange.
> > > > Would you agree with this approach?
> > >
> > > No, I'm not so keen on adding that side effect to
> > > OPTEE_SMC_EXCHANGE_CAPABILITIES.
> >
> > It is a capability REE would exchanges with TEE.
> > What kind of side effects do you fear?
>
> I was hoping to keep it stateless. One thing less to keep track of when
> handing over from a boot stage to the kernel.

Or from a kernel VM unload/reload.

>
> > > The way you're describing the problem it sounds like it's a normal world
> > > problem to know how many system threads are needed. How about adding a
> > > fast call where normal world can request how many system threads should
> > > be reserved?  If none are requested, none will be reserved.
> >
> > Well, could be. With caps exchange, we have an SMC funcID to REE to
> > say to TEE: "reserved the default configured number of sys thread". I
> > think it is simpler.
>
> Until you realize the that the default number of system threads doesn't
> match what you need.

Ok, I see your point. Indeed, Linux drivers requiring system context
could issue a fastcall SMC to request dynamic provisioning of TEE
context resources, and release their request upon driver unload. I
agree it would better scale in the long term. I'll propose something
in a v2.

>
> >
> > With REE calling TEE to provision thread, we would need another call
> > to release the reservation. Whe caps exchange, we have a single SMC to
> > reconfig the negotiated caps.
>
> A single SMC with growing complexity in its arguments.

:)  fair.

>
> Cheers,
> Jens