Re: [Qemu-devel] [PATCH] migration: avoid copying ignore-shared ramblock when in incoming migration

[Peter Xu]

On Wed, Mar 20, 2019 at 04:05:58PM +0800, chenxi He wrote:
> On Wed, 20 Mar 2019 at 13:07, Peter Xu <address@hidden> wrote:
> >
> > On Tue, Mar 19, 2019 at 11:49:22AM -0400, Catherine Ho wrote:
> > > Commit 18269069c310 ("migration: Introduce ignore-shared capability")
> > > addes a ignore-shared capability to bypass the shared ramblock (e,g,
> > > membackend + numa node). It does good to live migration.
> > >
> > > This commit expected that QEMU doesn't write to guest RAM until
> > > VM starts, but it does on aarch64 qemu:
> > > Backtrace:
> > > 1  0x000055f4a296dd84 in address_space_write_rom_internal () at 
> > > exec.c:3458
> > > 2  0x000055f4a296de3a in address_space_write_rom () at exec.c:3479
> > > 3  0x000055f4a2d519ff in rom_reset () at hw/core/loader.c:1101
> > > 4  0x000055f4a2d475ec in qemu_devices_reset () at hw/core/reset.c:69
> > > 5  0x000055f4a2c90a28 in qemu_system_reset () at vl.c:1675
> > > 6  0x000055f4a2c9851d in main () at vl.c:4552
> > >
> > > In address_space_write_rom_internal, if the ramblock is RAM of
> > > ramblock_is_ignored(), memcpy this ramblock will cause the memory-backend
> > > file corruption.
> > > But in normal case (!in_incoming_migration), this logic should be 
> > > reserved.
> >
> > (I am sorry if these questions are silly...)
> >
> > Could I ask when a ROM will be changed during execution of the guest,
> > and why?
> Sure :),  as told by Peter
> "as part of reset, we write the contents of ROM blobs, ELF files loaded 
> through
> -kernel, etc, to the guest memory."

I see...

I tried to dig on how x86 implemented the "-kernel" parameter and I
noticed that it's dumping the kernel image into fw_cfg and probably
that's also the reason why this should not be a problem for x86
because rom reset on x86 won't overwrite the image.  Meanwhile, it
seems totally different comparing to what has been done by ARM, which
should probably be using rom_add_elf_program() to load the image if my
reading is correct, so the kernel image is a ROM on ARM even if it can
be written... Is my understanding correct?

With that, I still feel that hacking into the memory write functions
are tricky and I feel like it could bring hard to debug issues.  Would
it be possible that we identify somehow that this ROM is a fake one
(since it can actually be written) and we ignore the reset of it when
proper (e.g., the initial reset on destination)?

Thanks,

-- 
Peter Xu