On Thu, 29 Jan 2026, Peter Maydell wrote:
On Thu, 29 Jan 2026 at 17:20, BALATON Zoltan <[email protected]> wrote:
On Thu, 29 Jan 2026, Peter Maydell wrote:
On Thu, 29 Jan 2026 at 16:21, BALATON Zoltan <[email protected]> wrote:
This is a migration compatibility break, because the
memory_region_init_rom() function registers the MR
for migration via vmstate_register_ram(), which picks
an ID string for the memory that includes the path
of the device, whereas vmstate_register_ram_global()
picks an ID string for the memory that does not include
the path of any device. It is this difference that is the
reason why they're still using the _nomigrate functions.
I thought it might be the case but wasn't sure. How was this handled in
all other machines where they were converted to not use _nomigrate? Only
some Sun machines still seem to use this and this may be a good
opportunity to bring them inline with all other machines.
If I remember correctly, we converted all the machines where
we were happy at the time to have a compat break. The remainder
are not only Sun machines -- you can see in your patch 4 that xtensa
and vga are also affected.
Well, almost all of the usages are because of the Sun machines and those
could be easily converted accepting one time migration break. The VGA one
seems to be because of isa cards which may not have been QOM'ified so
could not be converted before but maybe could be now but I haven't checked
and I have no idea about xtensa but if only those few are left that's much
better. There's one more _nomigrate in hw/m68k/next-cube.c which seems to
be some shared memory region (not sure why it can't be migrated) and the
one in backends/hostmem-ram.c which may be the only real place that needs
this. But those could be taken care of later, converting the Sun machines
would only leave 2 or 3 usages.
I think it's reasonable to accept a compat break for the purposes
of cleaning up the codebase, because these are used on sparc machine
types and we don't version those, but the commit message needs to
note that this is a compat break.
If we accept that then the other memory_region_init_ram_nomigrate usages
in Sun machines could also be converted. If Mark is not open to that I'm
inclined to convert these two rom_nomigrate usages to
memory_region_init_ram_flags_nomigrate + memory_region_set_readonly
instead
My preference would be to convert all these Sun devices. I think
we're a bit more clear these days that migration compat matters
for versioned machine types but can be (cleanly) broken for
non-versioned machine types: back then it was a bit less of a
well-defined rule.
I agree but wait for the maintainer's opinion (and maybe some more review
on other patches) before doing a v2. (Mark may have some reason like old
savevm files he may use for testing and does not want to redo them or
similar to want to keep this but if that's the only reason for using this
old stuff it might be time to refresh those testing images.)
Regards,
BALATON Zoltan
