Re: [PATCH 2/2] riscv: T-Head: Test availability bit before enabling MAEE errata

[Conor Dooley]

On Fri, Mar 29, 2024 at 12:22:51PM +0100, Christoph Müllner wrote:
> On Thu, Mar 28, 2024 at 4:43 PM Alexandre Ghiti <alex@xxxxxxxx> wrote:
> >
> > Hi Christoph,
> >
> > On 28/03/2024 15:18, Christoph Müllner wrote:
> > > On Wed, Mar 27, 2024 at 1:41 PM Andrew Jones <ajones@xxxxxxxxxxxxxxxx> wrote:
> > >> On Wed, Mar 27, 2024 at 11:03:06AM +0000, Conor Dooley wrote:
> > >>> On Wed, Mar 27, 2024 at 11:31:30AM +0100, Christoph Müllner wrote:
> > >>>> T-Head's MAEE mechanism (non-compatible equivalent of RVI's Svpbmt)
> > >>>> is currently assumed for all T-Head harts. However, QEMU recently
> > >>>> decided to drop acceptance of guests that write reserved bits in PTEs.
> > >>>> As MAEE uses reserved bits in PTEs and Linux applies the MAEE errata
> > >>>> for all T-Head harts, this broke the Linux startup on QEMU emulations
> > >>>> of the C906 emulation.
> > >>>>
> > >>>> This patch attempts to address this issue by testing the MAEE bit
> > >>>> in TH_MXSTATUS CSR. As the TH_MXSTATUS CSR is only accessible in M-mode
> > >>>> this patch depends on M-mode firmware that handles this for us
> > >>>> transparently.
> > >>>>
> > >>>> As this patch breaks Linux bootup on all C9xx machines with MAEE,
> > >>>> which don't have M-mode firmware that handles the access to the
> > >>>> TH_MXSTATUS CSR, this patch is marked as RFC.
> > >> Can we wrap the csr access in a _ASM_EXTABLE()? If firmware handles it,
> > >> then we return true/false based on the value. If firmware doesn't handle
> > >> it, and we get an illegal instruction exception, then we assume the bit
> > >> is set, which is the current behavior.
> > >>
> > >>> I think this is gonna be unacceptable in its current state given that it
> > >>> causes problems for every other version of the firmware. Breaking real
> > >>> systems for the sake of emulation isn't something we can reasonably do.
> > >>>
> > >>> To make this sort of change acceptable, you're gonna have to add some way
> > >>> to differentiate between systems that do and do not support reading this
> > >>> CSR. I think we either a) need to check the version of the SBI
> > >>> implementation to see if it hits the threshold for supporting this
> > >>> feature, or b) add a specific SBI call for this so that we can
> > >>> differentiate between firmware not supporting the function and the
> > >> The FWFT SBI extension is being developed as a mechanism for S-mode to ask
> > >> M-mode things like this, but I think that extension should be used for
> > >> features that have potential to be changed by S-mode (even if not
> > >> everything will be changeable on all platforms), whereas anything that's
> > >> read-only would be better with...
> > >>
> > >>> quote-unquote "hardware" not supporting it. I don't really like option a)
> > >>> as it could grow to several different options (each for a different SBI
> > >>> implementation) and support for reading the CSR would need to be
> > >>> unconditional. I have a feeling that I am missing something though,
> > >>> that'd make it doable without introducing a new call.
> > >>>
> > >>> Thanks,
> > >>> Conor.
> > >>>
> > >>> If only we'd made enabling this be controlled by a specific DT property,
> > >>> then disabling it in QEMU would be as simple as not setting that
> > >>> property :(
> > >> ...this, where "DT property" is "ISA extension name". I wonder if we
> > >> shouldn't start considering the invention of xlinux_vendor_xyz type
> > >> extension names which firmware could add to the ISA string / array,
> > >> in order to communicate read-only information like this?
> > >>
> > >> Thanks,
> > >> drew
> > > Hi Conor and Drew,
> > >
> > > Thank you for your hints.
> > > I fully agree with all your statements and concerns.
> > >
> > > Switching from th.mxstatus to th.sxstatus should address all mentioned concerns:
> > > * no dependency on OpenSBI changes
> > > * no break of functionality
> > > * no need for graceful handling of CSR read failures
> > > * no need to differentiate between HW and emulation (assuming QEMU
> > > accepts the emulation of th.sxstatus)
> > >
> > > Also note that DT handling would be difficult, because we need to probe before
> > > setting up the page table.
> >
> >
> > We already parse the DT before setting the page table to disable KASLR
> > and to parse "no4lvl" or "no5lvl" command line parameters. Take a look
> > at the kernel/pi directory and setup_vm() in mm/init.c.
> 
> Ah, I see. So, this can be done with a function similar to
> get_kaslr_seed() in arch/riscv/kernel/pi/fdt_early.c.
> And the Makefile will apply the necessary steps to get this working.
> The downside is that depending on new information in the DT, it will not be
> backward compatible. So, I don't see a way around probing th.sxstatus.MAEE.

Aye, you're right here. My suggestion about using DT was only for if
there were T-Head CPUs that turned up in the future with non-zero values
for marchid or mimpid. Requiring it for the CPUs we're talking about at
the moment (e.g. c906) would, as you pointed out yourself, cause the same
sort of regression that relying on an updated firmware would.
> Independent of that, there is work to be done for the T-Head extension
> discovery in the Linux kernel:
> * XThead* extensions are not in the DTS
> * XThead* extensions are not parsed during bootup
> * XThead* extensions don't trigger optimizations (string ops) or errata (MAEE)
> * XThead* extensions are not exported via hwprobe

> However, I think this is independent of addressing the MAEE issue.
> So, I will send out a V2 with the th.sxstatus.MAEE probing only.

th.sxstatus seems perfect here since it solves the problem you're
looking to deal with in emulation while being not introducing
regressions for real devices.

Thanks,
Conor.

Attachment: signature.asc
Description: PGP signature