On Tue, Dec 08 2020 at 18:25, Maxim Levitsky wrote: > On Tue, 2020-12-08 at 17:02 +0100, Thomas Gleixner wrote: >> For one I have no idea which bug you are talking about and if the bug is >> caused by the VMM then why would you "fix" it in the guest kernel. > > The "bug" is that if VMM moves a hardware time counter (tsc or anything else) > forward by large enough value in one go, > then the guest kernel will supposingly have an overflow in the time code. > I don't consider this to be a buggy VMM behavior, but rather a kernel > bug that should be fixed (if this bug actually exists) Well, that's debatable. The kernel has a safe guard in place for each clocksource which calculates the maximum time before an update needs to take place. That limit comes from: 1) Hardware counter wraparound time 2) Math limitation #1 is a non-issue on TSC, but it is on pm-timer, hpet and lots of other non-x86 devices #2 The overflow surely can happen if you're long enough out. For TSC it's ~ 800 / f [seconds/GHz TSC frequency], i.e. 200 seconds for a 4Ghz TSC. > Purely in theory this can even happen on real hardware if for example > SMM handler blocks a CPU from running for a long duration, or hardware > debugging interface does, or some other hardware transparent sleep > mechanism kicks in and blocks a CPU from running. (We do handle this > gracefully for S3/S4) We had this discussion before. People got upset the stuff didn't work when they resumed debugging after leaving the box in the breakpoint over the weekend. *Shrug* If SMM goes out for lunch for > 200 seconds it's broken. End of story, really. There are bigger problems than timekeeping when that happens. Hardware transparent sleep mechanisms which are doing this behind the kernels back without giving it a mechanism to configure it is pretty much like SMM: It's broken. So now life migration comes a long time after timekeeping had set the limits and just because it's virt it expects that everything works and it just can ignore these limits. TBH. That's not any different than SMM or hard/firmware taking the machine out for lunch. It's exactly the same: It's broken. And of course since that migration muck started _nobody_ bothered until today to talk to me about that. It's not a kernel bug. The kernel works as designed for the purpose and the design clearly had these goals: 1) Correctness 2) Performance 3) Scalability and for that we introduced limitations which were perfectly reasonable at the time because SMM and hardware/firmware wreckage definitely cannot be the limiting factor and for the fast wrapping stuff there is no design at all. These limitations are still reasonable because lifting them hurts performance and depending on the length has effects on correctness as well. Timekeeping is a complex problem. It's a virt bug caused by pure ignorance of the underlying and already existing technology and the unwillingness to talk to people who actually understand it. I don't even want to know what kind of magic workarounds VMMs have dreamed up for that. I'm seriously grumpy that more than 10 years after I reported that time can be observed going backwards this is still not fixed and that has absolutely nothing to do with guest migration. Ignoring the simple and trivial requirement for timekeeping correctness in the first place and then having the chuzpah to claim that the kernel is buggy because virt decided it can do what it wants is beyond my comprehension and yet another proof for the theorem that virt creates more problems than it solves. </rant> The question how it can be made work is a different problem. I carefully said 'made work' because you can't 'fix' it. - It can't be fixed at the VMM side at all - It can't be fixed for fast wrapping clock sources by just fiddling with the timekeeping and time accessor code at all. - Even for TSC it can't be just fixed without imposing overhead on every time read including VDSO. And just fixing it for x86 and TSC does not cut it. There is a whole world outside of x86 and we are not going to impose any x86/TSC specific insanity on everybody else. We are neither going to make generic code have TSC specific hoops and loops just to deal with that. This needs orchestration and collaboration from both the VMM and the guest kernel to make this work proper and reliably. There are two ways to do that: 1) Suspend / resume the guest kernel 2) Have a protocol which is safe under all circumstances. If #2 is not there then #1 is the only correct option unless the VMM can guarantee that the guest is restarted _before_ time goes south. Doing #2 correctly is not rocket science either. The kernel has mechanisms to deal with such problems already. All it requires is to expose and utilize them. The only requirement there is to bring the kernel into a state where no CPU can observe that the time goes backwards. The kernel has two mechanisms for that: 1) Suspend / resume. Trivial because all CPUs except the (usually) boot CPU are unplugged or in a state which does not matter 2) Switching clocksources. A runtime operation which is safe and correct utilizing stop_machine() If you really think about it then this migration problem is nothing else than switching the underlying clocksource. The only difference is that in case of a regular clocksource switch the previous clocksource is still accessible up to the point where the switchover happens which is obviously not the case for VM migration. But instead of switching the clocksource via stop machine we can just use the same mechanism to update the current clocksource so that the time jump does not matter and cannot be observed. That needs a few things to be done: VMM: - Disable NMI delivery to the guest - Inject a magic VMM to guest IPI which starts the operation Guest: - Schedule work from the IPI - work handles nested VMs if necessary - work invokes stop_machine(update_guest_clocksource, NULL, NULL) - When all vCPUs rendevouzed then the one vCPU which actually runs update_guest_clocksource() and reports to the VMM via hypercall or whatever that it reached the state and spin waits on a hyperpage shared between guest and VMM to wait for the VMM to signal to proceed. At that point it's 100% safe to freeze it. Probably not only from a timekeeping POV. VMM: - Freeze VM and send image to destination VMM New VMM: - Install the image - Setup the vCPU TSC muck - Store necessary information in the hyperpage and then flip the bit which makes the guest waiting in update_guest_clocksource() proceed. - Schedule the guest vCPUs Guest: - The one vCPU waiting in update_guest_clocksource() observes the GO bit, updates timekeeping and returns. - All CPUs leave stomp_machine() and everything is fine - work resumes and handles nested VMs - work tells VMM that everything is done All the bits and pieces are there already except for the VMM/guest contract and the extra 20 lines of code in the timekeeping core. There is one caveat vs. the NMI safe time keeper, but we have the mechanism to deal with that in place already for suspend so that's just another 5 lines of code to deal with at the core side. Now you combine this with a proper mechanism to deal with the TSC offset as I outlined before and your problems are pretty much gone in a very clean and understandable way. Thanks, tglx