On 05.11.2011 02:49, Eric Blake wrote:
qemu/THREADS.txt is explicit that the driver lock must not be held for long lengths of time, as it blocks all attempts to manage any other vms on the same connection. We were violating this by sleep()ing while waiting for a qemu child process to finish execution during actions such as destroy.
* src/qemu/qemu_process.h (qemuProcessKill): Alter signature. * src/qemu/qemu_process.c (qemuProcessKill): Add parameter. (qemuProcessFakeReboot, qemuProcessShutdownOrReboot) (qemuProcessStop): Update callers. * src/qemu/qemu_driver.c (qemuDomainDestroyFlags): Likewise. ---
Without this patch, problems in halting one domain could lock out actions on all other domains for more than 3 seconds, which is awfully long.
This doesn't solve all the problems - it is still possible to have a stuck NFS server being the reason for difficulties in stopping a domain, such as an lstat() call while attempting to relabel file systems, and those calls are still done while the driver lock is held; but I'll be submitting further patches as I try and reduce the critical section sizes.
I'm not sure whether this qualifies for 0.9.7 or should wait for post-release.
I'd feel more comfortable to take this in after release. Hopefully, 0.9.7 is going to be stable release (at least by features/bug fixes ratio) even if this is a small and good looking patch.
src/qemu/qemu_driver.c | 2 +- src/qemu/qemu_process.c | 35 ++++++++++++++++++++++++++++------- src/qemu/qemu_process.h | 3 ++- 3 files changed, 31 insertions(+), 9 deletions(-)
ACK Michal