Hi Christoffer, On 07/06/17 10:41, James Morse wrote: > I evidently stopped before I got to the bottom of this, the commit message is > based on the way I first hit this I've worked out where I went wrong with this. memory_failure()/hwpoison has two 'modes', early and late. My testing was broken for 'early', but caused both to happen at the same time, leading to this confusion. The affected page was mapped and found in the rmap, it then gets unmapped by memory_failure(), which then skipped the early notification because the flags were wrong. Meanwhile the late notification fires at the same time on another CPU. So, from the top: -----%<----- memory_failure() has two modes, early and late. Early is used by machine-managers like Qemu to receive a notification when a memory error is notified to the host. These can then be relayed to the guest before the affected page is accessed. To enable this, the process must set PR_MCE_KILL_EARLY in PR_MCE_KILL_SET using the prctl() syscall. Once the early notification has been handled, nothing stops the machine-manager or guest from accessing the affected page. If the machine-manager does this the page will fail to be mapped and SIGBUS will be sent. This patch adds the equivalent path for when the guest accesses the page, sending SIGBUS to the machine-manager. These two signals can be distinguished by the machine-manager using their si_code: BUS_MCEERR_AO for 'action optional' early notifications, and BUS_MCEERR_AR for 'action required' synchronous/late notifications. -----%<----- If this clears everything up I will post a v3 with the above as the commit message. Thanks! James _______________________________________________ kvmarm mailing list kvmarm@xxxxxxxxxxxxxxxxxxxxx https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
