Re: [libvirt] [PATCH v2 0/5] Workaround mdev uevent race affecting node device driver

v1 here https://www.redhat.com/archives/libvir-list/2017-June/msg00826.html This series addresses [1]. It builds on top of [2], series which introduces a few small fixes and improvements. Even though that one hasn't gotten a review yet, you can fetch and build this from my github branch [3]. Changes since v1: - moved the device handling routine into a detached worker thread, so the event loop wouldn't block - adjusted virFileAccess according to the reviewer's comments Notes: Merging the nodedev obj's refactor storing the list of nodedev objects in a self-lockable object made it possible to drop the driver locks at certain spots either completely or at least rearrange them according to the needs. The worker thread therefore doesn't need to hold the driver lock while waiting on the device's attributes to appear (which would be wrong anyway, since it would still block the event loop) and only acquires the driver lock when doing sanity checks on the udev monitor. [1] https://bugzilla.redhat.com/show_bug.cgi?id=1463285 [2] https://www.redhat.com/archives/libvir-list/2017-July/msg01077.html [3] https://github.com/eskultety/libvirt/commits/mdev-nodedev-next Erik Skultety (5): nodedev: Introduce udevCheckMonitorFD helper function udev: Split udevEventHandleCallback in two functions udev: Convert udevEventHandleThread to an actual thread routine util: Introduce virFileWaitForAccess nodedev: Work around the uevent race by hooking up virFileWaitForAccess src/libvirt_private.syms | 1 + src/node_device/node_device_udev.c | 194 +++++++++++++++++++++++++++++++------ src/util/virfile.c | 29 ++++++ src/util/virfile.h | 2 + 4 files changed, 194 insertions(+), 32 deletions(-)