On 14.12.2018 21:45, Kevin O'Connor wrote:
On Fri, Dec 14, 2018 at 01:08:46PM +0300, Евгений Яковлев wrote:
I think we have a problem right now if during emulation first pci config write in __make_bios_writeable_intel we decide to issue qmp system_reset. Based on what i could gather from crashed instances we have something like this:
- seabios issues pci_config_writel to reprogram 0xF0000-0xFFFFF and 0xD0000
- 0xC7FFF. VCPU thread exits into qemu usermode to emulate this.
- At this time we issue system_reset through qmp. When VCPU thread decided
to return to KVM it releases BQL, main event loop sees system_reset, stops cpus and calls reset handlers. So, we're doing a soft reset.
Q35 ICH9 reset emulation does not reset PAMs to default values
Upon re-entering reset vector seabios checks for PAM0 & 0x10 and decides
that we have enabled ram previously and does not jump to __make_bios_writeable_intel in high memory relying on the fact that code is in low memory already. However this assumption only holds true for 0xC8FFFF
- 0xEFFFF because we didn't have a chance to reprogram those PAMs during
previous runs. We also didn't memcpy anything previously, so we're now may execute junk from F-segment.
What do you think about this scenario? I would be happy to fix this but i would be happy to get your advise on how to proceed and if all of my assumptions are correct.
Interesting. So, to summarize, there is a race condition with a reset signal if it arrives between the point SeaBIOS sets the pam registers and completes its memcpy.
This is my current understanding, yes.
Not sure how to best address that. Is it even possible to close that race? (That is, if QEMU doesn't reset the pam registers, could the reboot immediately start execution at junk f-segment code?)
Is this only a problem if the reset signal occurs within the first few milliseconds of the the very first boot?
Yes, exactly between first and second call to pci_config_writel in __make_bios_writable_intel to be specific.
I will be making attempts to fix this because this is a problem for us. I am not that familiar with seabios code though, but i've been working with Tiano a lot and the way they do it now is to unlock PAM segments for writing first, modify ram and then enable read decode from ram. I am not sure (yet) if this works on q35 in qemu, but i will verify.
I was thinking previously to just reset PAMs in qemu in q35 and fx440 mch reset handler, however Intel spec is not specific on whether conforming devices should actually do this on soft reset signal.. I don't actually think they should.
Also, maybe original fix is still useful since we're not relying on all jumps/calls to be relative anymore?
My personal opinion is that the x86 segment/offset stuff is so arcane that it's best to avoid it. I don't think gcc will (in practice) emit an absolute function call.
That's reasonable. I would have still coded __make_bios_writable_intel completely in assembly *juts to make sure* :) But this is your call, Kevin.
-Kevin