[PATCH v1 3/3] arm64: dts: add Hi6220 mailbox node

[Mark Rutland]

On Tue, Aug 25, 2015 at 11:15:10AM +0100, Haojian Zhuang wrote:
> On Tue, 2015-08-25 at 10:46 +0100, Leif Lindholm wrote:
> > On Tue, Aug 25, 2015 at 04:13:47PM +0800, Haojian Zhuang wrote:
> > > On Mon, 2015-08-24 at 12:49 +0100, Leif Lindholm wrote:
> > > > On Mon, Aug 24, 2015 at 06:19:56PM +0800, Haojian Zhuang wrote:
> > > > > > If your EFI memory map describes the memory as mappable, it is wrong.
> > > > > 
> > > > > When kernel is working, kernel will create its own page table based on
> > > > > UEFI memory map. Since it's reserved in DTS file as Leo's patch, it'll
> > > > > be moved to reserved memblock. Why is it wrong?
> > > > > 
> > > > > In the second, UEFI is firmware. When it's stable, nobody should change
> > > > > it without any reason.
> > > > 
> > > > Much like the memory map.
> > > > 
> > > > > These reserved memory are used in mailbox driver.
> > > > > Look. It's driver, so it could be changed at any time.
> > > > 
> > > > No, it is a set of regions of memory set aside for use by a different
> > > > master in the system as well as communications with that master.
> > > > 
> > > > The fact that there is a driver somewhere that is aware of this is
> > > > entirely beside the point. All agents in the system must adher to this
> > > > protocol.
> > > > 
> > > > > Why do you want
> > > > > to UEFI knowing this memory range? Do you hope UEFI to change when
> > > > > mailbox driver is changed?
> > > > 
> > > > Yes.
> > > > 
> > > > UEFI is a runtime environment. Having random magic areas not to be
> > > > touched will cause random pieces of software running under it to break
> > > > horribly or break other things horribly.
> > > > Unless you mark them as reserved in the UEFI memory map.
> > > > At which point the Linux kernel will automatically ignore them, and
> > > > the proposed patch is redundant.
> > > > 
> > > > So, yes, if you want a system that can boot reliably, run testsuites
> > > > (like SCT or FWTS), run applications (like fastboot ... or the EFI
> > > > stub kernel itself), then any memory regions that is reserved for
> > > > mailbox communication (or other masters in the system) _must_ be
> > > > marked in the EFI memory map.
> > > 
> > > 1. We need support both UEFI and uboot. So the reserved buffer have to
> > > be declared in DTB since they are used by kernel driver, not UEFI.
> > 
> > The buffer may need to be declared in DTB also, but it most certanily
> > needs to be declared in UEFI.
> > 
> > And for the U-Boot case, since it is not memory available to Linux, it
> > should not be declared as "memory".
> 
> Something are messed at here. We have these buffer are used in mailbox.
> They should be allocated as non-cacheable.
> 
> If these buffers are contained in memory memblock in kernel, it means
> that they exist in kernel page table with cachable property. When it's
> used in mailbox driver with non-cachable property, it'll only cause
> cache maintenance issue. So Leo declared these buffers as reserved
> in DT with "no-map" property. It's the key. It could avoid the cache
> maintenance issue.

The better solution is to never describe the memory to the kernel as
memory, by never placing it in a memory node, and ensuring that if it is
in the UEFI memory map, its attributes do not allow it to be mapped.

That way a driver can map it as non-cacheable if it wishes, but nothing
else can possibly touch that memory.

That is all you need to do.

> > > 2. UEFI just loads grub. It's no time to run any other custom EFI
> > > application.
> > 
> > Apart from being completely irrelevant, how are you intending to
> > validate that GRUB never touches these memory regions?
> > 
> 
> GRUB is just a part of bootloader. When linux kernel is running,
> who cares GRUB? GRUB's lifetime is already finished.

If GRUB temporarily maps memory as cacheable, or hands it to a device,
then your statements above about cache maintenance are broken.

An EFI application like GRUB might leave something resident in memory
after it's done (consider the UEFI shim), or it could even load the
kernel into the region that you care about having reserved, because as
far as it's concerned it's just memory. That could leave you with a
conflict for that region of memory.

You _must_ care about GRUB (and other EFI applications) doing the right
thing. To get them to avoid a region of memory, it must not be described
as being usable by them in the UEFI memory map.

> By the way, UEFI code region is at [0x3Dxx_xxxx, 0x3DFF_FFFF]. Those
> mailbox buffer is in [0x05e0_xxxx, 0x06f0_xxxx]. Then I can make sure
> UEFI won't touch the reserved buffer. Even if UEFI touched the reserved
> buffer, is it an issue? Definitely it's not.

It definitely is, due to the possibility of stale cache lines being left
in the region from when UEFI may have mapped it with cacheable
attributes.

> > Build a version once, test it, and hope the results remain valid
> > forever? And then when you move the regions and the previously working
> > GRUB now tramples all over them? Or when something changes in upstream
> > GRUB and its memory allocations drifts into the secretly untouchable
> > regions?
> 
> As I said above, UEFI won't touch it. And even UEFI touch it, kernel
> doesn't care since UEFI's lifetime is end.

If EFI touches it there may be stale cache lines left around, which you
don't seem to expect.

> > Are you then going to hack GRUB, release a special HiKey version of
> > GRUB, not support any other versions, and still can your firmware
> > UEFI?
> 
> I don't need to hack GRUB at all.

Then it is working for you by pure chance alone.

Please listen to the advice you are being given here; we're trying to
ensure that your platform functions (and continues to function) as best
it can.

Thanks,
Mark.