Hi gengdongjiu, On 16/12/17 04:47, gengdongjiu wrote: > [...] >> >>> + case ESR_ELx_AET_UER: /* The error has not been propagated */ >>> + /* >>> + * Userspace only handle the guest SError Interrupt(SEI) if the >>> + * error has not been propagated >>> + */ >>> + run->exit_reason = KVM_EXIT_EXCEPTION; >>> + run->ex.exception = ESR_ELx_EC_SERROR; >>> + run->ex.error_code = KVM_SEI_SEV_RECOVERABLE; >>> + return 0; >> >> We should not pass RAS notifications to user space. The kernel either handles >> them, or it panics(). User space shouldn't even know if the kernel supports RAS > > For the ESR_ELx_AET_UER(Recoverable error), let us see its definition > below, which get from [0] [..] > so we can see the exception is precise and PE can recover execution > from the preferred return address of the exception, > so let guest handling it is > better, for example, if it is guest application RAS error, we can kill > the guest application instead of panic whole OS; if it is guest kernel > RAS error, guest will panic. If the kernel takes an unhandled RAS error it should panic - we don't know where the error is. I understand you want to kill-off guest tasks as a result of RAS errors, but this needs to go through the whole APEI->memory_failure()->sigbus machinery so that the kernel knows the kernel can keep running. This saves us signalling user-space when we don't need to. An example: code-corruption. Linux can happily re-read affected user-space executables from disk, there is absolutely nothing user-space can do about it. Handling errors first in the kernel allows us to do recovery for all the affected processes, not just the one that happens to be running right now. > Host does not know which application of guest has error, so host can > not handle it, It has to work this out, otherwise the errors we can handle never get a chance. This kernel is expected to look at the error description, (which for some reason we aren't talking about here), e.g. the CPER records, and determine what recovery action is necessary for this error. For memory errors this may be re-reading from disk, or at the worst case, unmapping from all user-space users (including KVM's stage2) and raining signals on all affected processes. For a memory error the important piece of information is the physical address. Only the kernel can do anything with this, it determines who owns the affected memory and what needs doing to recover from the error. If you pass the notification to user-space, all it can do is signal the guest to "stop doing whatever it is you're doing". The guest may have been able to re-read pages from disk, or otherwise handle the error. Has the error been handled? No: The error remains latent in the system. > panic OS is not a good choice for the Recoverable error. If we don't know where the error is, and we can't make progress, its the only sane choice. This code is never expected to run! (why are we arguing about it?) We should get RAS errors as GHES notifications from firmware via some mechanism. If those are NOTIFY_SEI then APEI should claim the notification and kick off the appropriate handling based on the CPER records. If/when we get kernel-first, that can claim the SError. What we're left with is RAS notifications that no-one claimed because there was no error-description found. James
