Re: [edk2-discuss] Load Option passing. Either bugs or my confusion.

[Hou Qiming]

Very good point, I did neglect ramfb resolution changes... But there is one important thing: it *can* cause a QEMU crash, a potentially exploitable one, not always a guest crash. That's what motivated my heavy-handed approach since allowing resolution change would have necessitated a good deal of security checks. It has crashed my host *kernel* quite a few times.

The point is, while the QemuRamfbDxe driver may behave properly, nothing prevents the guest from writing garbage or *malicious* values to the ramfb config space. Then the values are sent to the display component without any sanity check. For some GUI frontends, this means allocating an OpenGL texture with guest-supplied dimensions and uploading guest memory content to it, which means that guest memory content goes straight into a *kernel driver*, *completely unchecked*. Some integer overflow and a lenient GPU driver later, and the guest escapes straight to kernel.

The proper way to enable ramfb resolution change again is adding sanity checks for ramfb resolution / pointer / etc. on the QEMU side. We have to make sure it doesn't exceed what the host GPU driver supports. Maybe clamp both width and height to between 1 and 2048? We also need to validate that OpenGL texture dimension update succeeds. Note that OpenGL is not obliged to validate anything and everything has to be checked on the QEMU side.

Qiming

On Wed, Apr 15, 2020 at 11:05 PM Laszlo Ersek <address@hidden> wrote:

(CC Gerd, Qiming, Marcel, qemu-devel for ramfb:)

On 04/14/20 23:20, valerij zaporogeci wrote:

[snip]

> There is a Boot Manager UI display problem, I don't know if this is
> qemu problem, but with the ARM (both 32 and 64 bits, the qemu version
> is 4.2.0, the OVMF is fresh), and using "ramfb" device, the Boot
> Manager has troubles with drawing - it's interfase looks entirely
> broken, like this (I'll try to attach the screenshot). UEFI shell
> doesn't have this problem. switching to "serial" (which is -serial vc)
> doesn't produce it too. Only when ramfb is chosen, the Boot Manager UI
> gets smeared. But it takes input and presumable works properly, except
> displaying things. qemu writes these messages in the command prompt:
> ramfb_fw_cfg_write: 640x480 @ 0x4bd00000
> ramfb_fw_cfg_write: resolution locked, change rejected
> ramfb_fw_cfg_write: 800x600 @ 0x4bd00000
> ramfb_fw_cfg_write: resolution locked, change rejected

Gerd contributed the OVMF QemuRamfbDxe driver in edk2 commit
1d25ff51af5c ("OvmfPkg: add QemuRamfbDxe", 2018-06-14). Note the date:
June 2018.

The then-latest (released) QEMU version was v2.12.0, and v2.12.1 /
v3.0.0 were in the making.

At that time, the resolution change definitely worked -- note my
"Tested-by" on the edk2 commit message.


Running "git blame" on the QEMU source, I now find commit a9e0cb67b7f4
("hw/display/ramfb: lock guest resolution after it's set", 2019-05-24).

Again, note the date: May 2019 (and this commit was released with QEMU
v4.1.0).

So I would say that the symptom you see is a QEMU v4.1.0 regression. The
QemuRamfbGraphicsOutputSetMode() function in the OVMF ramfb driver
certainly needs the QemuFwCfgWriteBytes() call to work, for changing the
resolution.


Now, I'm not familiar with the reasons behind QEMU commit a9e0cb67b7f4.
It says it intends to "prevent[] a crash when the guest writes garbage
to the configuration space (e.g. when rebooting)".

But I don't understand why locking the resolution was necessary for
preventing "a crash":

(1) Registering a device reset handler in QEMU seems sufficient, so that
QEMU forget about the currently shared RAMFB area at platform reset.

(2) The crash in question is apparently not a *QEMU* crash -- which
might otherwise justify a heavy-handed approach. Instead, it is a
*guest* crash. See the references below:

(2a) http://mid.mail-archive.com/CABSdmrmU7FK90Bupq_ySowcc9Uk=address@hidden
     https://lists.gnu.org/archive/html/qemu-devel/2019-05/msg02299.html

(2b) https://github.com/intel/gvt-linux/issues/23#issuecomment-483651476

Therefore, I don't think that locking the resolution was justified!

Importantly:

- The QemuRamfbDxe driver allocates the framebuffer in *reserved*
memory, therefore any well-behaving OS will *never* touch the
framebuffer.

- The QemuRamfbDxe driver allocates the framebuffer memory only once,
namely for such a resolution that needs the largest amount of
framebuffer memory. Therefore, framebuffer re-allocations in the guest
driver -- and thereby guest RAM *re-mapping* in QEMU -- are *not*
necessary, upon resolution change.

The ramfb device reset handler in QEMU is justified (for unmapping /
forgetting the previously shared RAMFB area).

The resolution locking is *NOT* justified, and it breaks the OVMF
driver. I suggest backing out the resolution locking from QEMU.

Reference (2a) above indicates 'It could be a misguided attempt to
"resize ramfb" by the guest Intel driver'. If that is the case, then
please fix the Intel guest driver, without regressing the QEMU device
model.

I'm sad that the QEMU device model change was not regression-tested
against the *upstream* OVMF driver (which, by then, had been upstream
for almost a year).

Laszlo