Re: [RFC PATCH-for-8.0 2/3] hw/ppc/spapr: Replace tswap64(HPTE) by cpu_to_be64(HPTE)

[David Gibson <david@gibson.dropbear.id.au>]

[-- Attachment #1: Type: text/plain, Size: 4409 bytes --]

On Fri, Dec 16, 2022 at 09:39:19PM +0000, Peter Maydell wrote:
> On Fri, 16 Dec 2022 at 19:11, Daniel Henrique Barboza
> <danielhb413@gmail.com> wrote:
> >
> >
> >
> > On 12/13/22 10:51, Peter Maydell wrote:
> > > On Tue, 13 Dec 2022 at 12:52, Philippe Mathieu-Daudé <philmd@linaro.org> wrote:
> > >>
> > >> The tswap64() calls introduced in commit 4be21d561d ("pseries:
> > >> savevm support for pseries machine") are used to store the HTAB
> > >> in the migration stream (see savevm_htab_handlers) and are in
> > >> big-endian format.
> > >
> > > I think they're reading the run-time spapr->htab data structure
> > > (some of which is stuck onto the wire as a stream-of-bytes buffer
> > > and some of which is not). But either way, it's a target-endian
> > > data structure, because the code in hw/ppc/spapr_softmmu.c which
> > > reads and writes entries in it is using ldq_p() and stq_p(),
> > > and the current in-tree version of these macros is doing a
> > > "read host 64-bit and convert to/from target endianness wih tswap64".
> > >
> > >>   #define HPTE(_table, _i)   (void *)(((uint64_t *)(_table)) + ((_i) * 2))
> > >> -#define HPTE_VALID(_hpte)  (tswap64(*((uint64_t *)(_hpte))) & HPTE64_V_VALID)
> > >> -#define HPTE_DIRTY(_hpte)  (tswap64(*((uint64_t *)(_hpte))) & HPTE64_V_HPTE_DIRTY)
> > >> -#define CLEAN_HPTE(_hpte)  ((*(uint64_t *)(_hpte)) &= tswap64(~HPTE64_V_HPTE_DIRTY))
> > >> -#define DIRTY_HPTE(_hpte)  ((*(uint64_t *)(_hpte)) |= tswap64(HPTE64_V_HPTE_DIRTY))
> > >> +#define HPTE_VALID(_hpte)  (be64_to_cpu(*((uint64_t *)(_hpte))) & HPTE64_V_VALID)
> > >> +#define HPTE_DIRTY(_hpte)  (be64_to_cpu(*((uint64_t *)(_hpte))) & HPTE64_V_HPTE_DIRTY)
> > >> +#define CLEAN_HPTE(_hpte)  ((*(uint64_t *)(_hpte)) &= cpu_to_be64(~HPTE64_V_HPTE_DIRTY))
> > >> +#define DIRTY_HPTE(_hpte)  ((*(uint64_t *)(_hpte)) |= cpu_to_be64(HPTE64_V_HPTE_DIRTY))
> > >
> > > This means we now have one file that's accessing this data structure
> > > as "this is target-endian", and one file that's accessing it as
> > > "this is big-endian". It happens that that ends up meaning the same
> > > thing because PPC is always TARGET_BIG_ENDIAN, but it seems a bit
> > > inconsistent.
> > >
> > > We should decide whether we're thinking of the data structure
> > > as target-endian or big-endian and change all the accessors
> > > appropriately (or none of them -- currently we're completely
> > > consistent about treating it as "target endian", I think).
> >
> > Yes, most if not all accesses are being handled as "target endian", even
> > though the target is always big endian.

So "target is always big endian" is pretty misleading for POWER.  We
always define "TARGET_BIG_ENDIAN" in qemu, but for at least 10 years
the CPUs have been capable of running in either big endian or little
endian mode (selected at runtime).  Some variants can choose
endianness on a per-page basis.  Since the creation of the ISA it's
had "byte reversed" load and store instructions that let it use little
endian for specific memory accesses.

Really the whole notion of an ISA having an "endianness" doesn't make
a lot of sense - it's an individual load or store to memory that has
an endianness which can depend on a bunch of factors.  When these
macros were created, an ISA nearly always used the same endianness,
but that's not really true any more - not just for POWER, but for a
bunch of targets.  So from that point of view, I think getting rid of
tswap() - particularly one that has compile time semantics, rather
than behaviour which can depend on cpu mode/state is a good idea.

I believe that even when running in little-endian mode, the hash page
tables are encoded in big-endian, so I think the proposed change makes
sense.

> > IIUC the idea behind Phil's cleanups is exactly to replace uses of
> > "target-something" if the endianess of the host is irrelevant, which
> > is the case for ppc64. We would then change the semantics of the code
> > gradually to make it consistent again.
> 
> I would be happier if we just did all the functions that read and
> write this byte array at once -- there are not many of them.
> 
> thanks
> -- PMM
> 

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]