Date:   Thu, 9 Feb 2023 08:14:14 +0100
From:   Jens Wiklander <jens.wiklander@linaro.org>
To:     Etienne Carriere <etienne.carriere@linaro.org>
Cc:     Sumit Garg <sumit.garg@linaro.org>, linux-kernel@vger.kernel.org,
        linux-arm-kernel@lists.infradead.org,
        Sudeep Holla <sudeep.holla@arm.com>,
        Cristian Marussi <cristian.marussi@arm.com>
Subject: Re: [PATCH 1/2] tee: system invocation
Message-ID: <Y+SdRrwSq/a9OgGr@jade>
References: <20230130094157.1082712-1-etienne.carriere@linaro.org>
 <CAFA6WYPxzB45pgWkkh++jRaY-d0eHWnrjzu+Z9059PjK5=M-mQ@mail.gmail.com>
 <CAHUa44G3jCqyiXtdZAAEw=8WPC+m5fD8tqsPfGc3MkV3JjRDCA@mail.gmail.com>
 <CAFA6WYN7i+7riJGPH4BEUFK77-kAx4J89Tn3=oX=g6rFUFtDkQ@mail.gmail.com>
 <CAHUa44G1yQtgF1eAUJVA+wtctKHfqYFBhs0PBnpoN-gD8_x8eg@mail.gmail.com>
 <CAN5uoS_M4uMWkf=Q8XFLCrNSvyUdjLgCPixqonKv3mRwRTr-nQ@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
In-Reply-To: <CAN5uoS_M4uMWkf=Q8XFLCrNSvyUdjLgCPixqonKv3mRwRTr-nQ@mail.gmail.com>
Precedence: bulk

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.

> 
> 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.

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.

Cheers,
Jens