From: Keith Busch <kbusch@xxxxxxxxxx> The suggestion from Sean appears to be successful, so sending out a new version for consideration. Background: The crosvm VMM might send signals to its threads that have entered KVM_RUN. The signal specifically is SIGRTRMIN from here: https://github.com/google/crosvm/blob/main/src/crosvm/sys/linux/vcpu.rs#L651 If this happens to occur when the huge page recovery is trying to create its vhost task, that will fail with ERESTARTNOINTR. Once this happens, all KVM_RUN calls will fail with ENOMEM despite memory not being the problem. This series propogates the error up so we can distinguish that from the current defaulting to ENOMEM and replaces the call_once since we need to be able to call it repeatedly due to this condition. Changes from the v1 (prefixed as an "RFC", really) patch: Instead of using a VM-wide mutex, update the call_once pattern to complete only if what it calls is successful (from Sean). Keith Busch (1): vhost: return task creation error instead of NULL Sean Christopherson (1): kvm: retry nx_huge_page_recovery_thread creation arch/x86/kvm/mmu/mmu.c | 12 +++++------- drivers/vhost/vhost.c | 2 +- include/linux/call_once.h | 16 +++++++++++----- kernel/vhost_task.c | 4 ++-- 4 files changed, 19 insertions(+), 15 deletions(-) -- 2.43.5
