qemu-stable
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [PATCH] hw/display/ramfb: plug slight guest-triggerable leak on mode


From: Laszlo Ersek
Subject: Re: [PATCH] hw/display/ramfb: plug slight guest-triggerable leak on mode setting
Date: Mon, 2 Oct 2023 04:02:45 +0200

On 10/1/23 18:07, Marc-André Lureau wrote:
> Hi Laszlo
> 
> On Sun, Oct 1, 2023 at 4:20 AM Laszlo Ersek <lersek@redhat.com> wrote:
>>
>> On 10/1/23 00:14, Laszlo Ersek wrote:
>>> On 9/29/23 13:17, Marc-André Lureau wrote:
> [..]
>>>> fwiw, my migration support patch is still unreviewed:
>>>> https://patchew.org/QEMU/20230920082651.3349712-1-marcandre.lureau@redhat.com/
>>>>
>>>
>>> I don't have a copy of that patch (not subscribed, sorry...), but how
>>> simply you did it surprises me. I did expect to simulate an fw_cfg write
>>> in post_load, but I thought we'd approach the device (for the sake of
>>> including it in the migration stream) from the higher level device's
>>> vmstate descriptors (dc->vmsd) that set up / depend on ramfb in the
>>> first place. I didn't know that raw vmstate_register() was still accepted.
>>>
>>> If it is, then, for that patch (with Gerd's comment addressed):
>>>
>>> Acked-by: Laszlo Ersek <lersek@redhat.com>
>>
>> I think I may have found one issue with that patch.
>>
>> The fields that we save into the migration stream are integer members of
>> the RAMFBCfg structure that lives directly in the fw_cfg file. The ramfb
>> device specifies those fields for the guest as big endian. This means
>> that when ramfb_fw_cfg_write() runs, triggered by the guest, then on big
>> endian hosts, be32_to_cpu() and be64_to_cpu() will be no-ops, as the
>> integer representation inside the fw_cfg file will match the host CPU at
>> once. And on little endian hosts, these functions will byte swap. In
>> both cases, ramfb_create_display_surface() will have to be called with
>> identical host-side integers. This means that *before* the be32_to_cpu()
>> and be64_to_cpu() calls, the host side *values* read out from the fw_cfg
>> file members *differ* between big endian and little endian hosts.
>>
>> And the problem is that we write precisely those values into the
>> migration stream, via "vmstate_ramfb_cfg". The migration stream
>> represents integers in big endian regardless of host endianness, but the
>> question is the *values* that we encode in BE for the stream. And the
>> values (from fw_cg) will differ between little endian and big endian hosts.
>>
>> Thus, I think we should just use
>>
>>   VMSTATE_BUFFER(cfg, RAMFBState)
>>
>> in "vmstate_ramfb", and remove "vmstate_ramfb_cfg". For migration
>> purposes, we should treat the fw_cfg file as an opaque blob.
> 
> I think I see your point. Using VMSTATE_BUFFER like that doesn't work
> though.

Why not?

(I mean -- does it compile but misbehave, or it doesn't even compile (an
invalid use of the VMSTATE_BUFFER macro)?)

> It's also more error-prone if fields are added in the struct,
> imho.

The structure is effectively the guest-visible register block for the
device. We probably can't add any fields, and even if we do, the new
fields are going to be part of the fw_cfg blob (writeable file), so they
should be covered by VMSTATE_BUFFER just the same.

> 
> However, we could simply have a post-load to convert the values to BE.

post_load itself is not enough; if we want to go this route, then we
need pre_save too. Without a pre_save, the host endianness influences
the data serialized to the migration stream, and there's no way to know
how to recover (the source host's endianness is unknown at load time).

pre_save could work though, if it performed the same BE to host
conversions (to a separate buffer I guess!) as the fw_cfg write callback
does.

> I wonder if new macros couldn't be introduced too.
> 
>>>
>>> BTW: can you please assign
>>> <https://bugzilla.redhat.com/show_bug.cgi?id=1859424> to yourself and
>>> link your patch into it? The reason we ended up duplicating work here is
>>> that noone took RHBZ 1859424 before.
> 
> I thought I did that.
> 
> https://bugzilla.redhat.com/show_bug.cgi?id=1859424#c17

Ouch, sorry. That must have happened since I last looked, and I was
foolish enough not to CC myself on the BZ early on. My mistake!

> 
>>>
>>> ... Well, the ticket is RHEL-7478 in JIRA now, in fact. :/
> 
> :/
> 
> 

Laszlo




reply via email to

[Prev in Thread] Current Thread [Next in Thread]