Message-ID: <20200717103120.GC24501@bogus>
Date:   Fri, 17 Jul 2020 11:31:20 +0100
From:   Sudeep Holla <sudeep.holla@....com>
To:     Florian Fainelli <f.fainelli@...il.com>
Cc:     Daniele Alessandrelli <daniele.alessandrelli@...ux.intel.com>,
        linux-arm-kernel@...ts.infradead.org, Peng Fan <peng.fan@....com>,
        "Paul J. Murphy" <paul.j.murphy@...ux.intel.com>,
        "Paul J. Murphy" <paul.j.murphy@...el.com>,
        linux-kernel@...r.kernel.org,
        Daniele Alessandrelli <daniele.alessandrelli@...el.com>,
        Sudeep Holla <sudeep.holla@....com>
Subject: Re: [PATCH] firmware: arm_scmi: Pass shmem address to SMCCC call

On Thu, Jul 16, 2020 at 12:57:23PM -0700, Florian Fainelli wrote:
>
>
> On 7/16/2020 7:13 AM, Daniele Alessandrelli wrote:
> > Hi Florian,
> >
> > Thanks for you feedback.
> >
> > On Wed, 2020-07-15 at 15:43 -0700, Florian Fainelli wrote:
> >>
> >> On 7/15/2020 9:55 AM, Daniele Alessandrelli wrote:
> >>> From: Daniele Alessandrelli <daniele.alessandrelli@...el.com>
> >>>
> >>> Currently, when SMC/HVC is used as transport, the base address of
> >>> the
> >>> shared memory used for communication is not passed to the SMCCC
> >>> call.
> >>> This means that such an address must be hard-coded into the
> >>> bootloader.
> >>>
> >>> In order to increase flexibility and allow the memory layout to be
> >>> changed without modifying the bootloader, this patch adds the
> >>> shared
> >>> memory base address to the a1 argument of the SMCCC call.
> >>>
> >>> On the Secure Monitor side, the service call implementation can
> >>> therefore read the a1 argument in order to know the location of the
> >>> shared memory to use. This change is backward compatible to
> >>> existing
> >>> service call implementations as long as they don't check for a1 to
> >>> be
> >>> zero.
> >>
> >> resource_size_t being defined after phys_addr_t, its size is
> >> different
> >> between 32-bit, 32-bit with PAE and 64-bit so it would probably make
> >> more sense to define an physical address alignment, or maybe an
> >> address
> >> that is in multiple of 4KBytes so you can address up to 36-bits of
> >> physical address even on a 32-bit only system?
> >
> > I see your point. After a quick look, I think that, practically, the
> > issue is with ARM32 LPAE addresses, for which phys_addr_t is a u64. So,
> > basically, for AArch32 systems with LPAE the 64-bit shmem_paddr gets
> > truncated to 32-bit when it's passed to the SMC32/HVC32 call.
> >
> > To solve that, I would prefer splitting the address between two SMC
> > parameters (a1 = addr_lo, a2 = addr_hi), instead of imposing an
> > arbitrary alignment. Would that be reasonable?
>
> The low/high part would only be relevant on a 32-bit LPAE platform which
> is probably a corner case, I would just pass the shmem_paddr / 4096
> since that is the smallest granule size and alignment possible and it
> still allows you to map up to 36-bits of physical address, which is the
> maximum that the long descriptor in LPAE can support. For 64-bit we have
> no such problems since we have the full register width.
>

OK, I will check if 32-bit identifier can be relaxed in the spec in which
case we can avoid having DT binding for the identifier. If that is possible
we could use addr/4k page size.

> >>
> >> What discovery mechanism does the OS have that the specified address
> >> within the SMCCC call has been accepted by the firmware given the
> >> return value of that SMCCC call does not appear to be used or checked? Do
> >> we just expect a timeout initializing the SCMI subsystem?
> >
> > The return code is actually checked at the end of the function:
> > https://elixir.bootlin.com/linux/v5.8-rc4/source/drivers/firmware/arm_scmi/smc.c#L118
> >
> > But in the meantime scmi_rx_callback() has already been called. Not
> > sure if that's intentional or a possible bug.
> >
> >>
> >> Given that the kernel must somehow reserve this memory as a shared
> >> memory area for obvious reasons, and the trusted firmware must also
> >> ensure it treats this memory region with specific permissions in its
> >> translation regime, does it really make sense to give that much
> >> flexibility?

I share same opinion as Florian here.

[...]

> >> If your boot loader has FDT patching capability, maybe it can also do
> >> a SMC call to provide the address to your trusted firmware, prior to
> >> loading the Linux kernel, and then they both agree, prior to boot
> >> about the shared memory address?
> >
> > Yes, that's a possible solution, but it looks more complicated to me,
> > since it adds an additional component (the boot loader) to the
> > equation, while the goal of this patch was to reduce the coupling
> > between components (namely the DT/kernel and the trusted firmware).
> >
> > I guess my question is: if we fix the handling of LPAE addresses and
> > the SMC return code, what is the drawback of having the shmem address
> > passed to the SMC?
>
> My only concern is that if somehow Linux gets assigned a shared memory
> range that is completely outside of what the trusted firmware has
> already mapped, or is capable of addressing, or any combination thereof,
> it could be challenging to debug what is going on, especially if INVALID
> PARAMETER must not be returned (assuming this is to avoid Linux
> discovering where other shared memory areas pertaining to the firmware
> reside?).
>

Valid point. Again, I was planning to use this as identifier and didn't
think of the usecase Daniele mentions here.

> The other concern I have is that we are not documenting the various
> SMCCC calling conventions, soon enough it will be come out of control,
> and we are already allowing people to define their own function IDs and
> parameters to call into the trusted firmware. This sounds like something
> that is so basic that it should be standardized from the top, by ARM.
>

Completely agreed. I have failed to achieve that myself. Unless there is
demand from partners/vendors, it will be dismissed by architects. All
I managed to get in was the identifier as the argument for SMC/HVC to
support multiple channels. However 32-bit id will be issue to be used
with 64-bit address. I will see if we can drop that 32-bit requirement
from the spec.

> >
> > Anyway, I should have mentioned this in the commit message (sorry for
> > not doing so), but I submitted this patch because initial feedback from
> > Sudeep was positive [1]; but if there is no consensus around it I'm
> > fine with dropping it.
> >
> > [1] https://lore.kernel.org/lkml/20200710075931.GB1189@bogus/
>
> My review is by no means authoritative however in deploying SCMI on our
> Broadcom STB platforms some experience was gained in the process which
> is how it piqued my interest. Thanks for providing more background to
> this patch, this does help.
>

Your knowledge is important and please always share the same and happy
to take all your review into consideration always. End goal is to help
all or most of the users of the driver if possible.

> We have opted for a solution where the boot loader knows about all
> possible reserved regions prior to booting/loading the trusted firmware
> as well as the kernel, therefore it can pass that information to both
> and we never really had a situation where the two need to evolve in an
> uncoordinated way.

I agree and uncoordinated evolution may not be as simple as it is
presented by Daniele earlier and I am not confident to support that.
The main reason for giving positive response for this initially is
for the addition of identifier which I always thought is required for
supporting multiple channels. E.g. DVFS/Perf has dedicated channel.

--
Regards,
Sudeep

Powered by blists - more mailing lists