On 2012年11月07日 18:38, Hu Tao wrote:
The reason to disable cgroup cpuset by default is: currently, sub-directories of cgroup cpuset won't be auto-updated if cpu hot-plug happens. For example, if we now have 0-3,5-7 pcpus online, the value of /sys/fs/cgroup/cpuset/libvirt/qemu/mydom/cpuset.cpus is 0-3,5-7, then pcpu 4 is hot-plugged, but the value of cpuset.cpus is still 0-3,5-7. This will cause a problem that we can't pin to a hot-plugged pcpu
IMHO it's a kernel bug and should be fixed in kernel instead.
Users can still enable cpuset by editing qemu.conf.
And per users can enable/disable the cpuset cgroup in qemu.conf, why do we need to disable it by default? On one hand, it has to be enabled again once the kernel bug is fixed; On the other hand, disabling it affects many functions, it's not deserved to disable it with the only problem (actually the libvirt pinning just behaves correctly, as the hotplug pCPU is not visible yet, the root cause is in kernel) on pinning from my p.o.v.
Hu Tao (6): cgroup: Add a flag VIR_CGROUP_DISABLE_CPUSET create cgroup controllers for driver according to config create cgroup controllers for domain according to config create cgroup controllers for vcpu according to config create cgroup controllers for emulator according to config disable cgroup cpuset by default src/lxc/lxc_cgroup.c | 4 +-- src/lxc/lxc_driver.c | 22 ++++++------ src/lxc/lxc_process.c | 4 +-- src/qemu/qemu.conf | 2 +- src/qemu/qemu_cgroup.c | 15 ++++---- src/qemu/qemu_conf.c | 1 - src/qemu/qemu_driver.c | 92 +++++++++++++++++++++++++++++------------------ src/qemu/qemu_hotplug.c | 9 +++-- src/qemu/qemu_migration.c | 2 +- src/util/cgroup.c | 49 +++++++++++++++++++------ src/util/cgroup.h | 12 ++++--- 11 files changed, 136 insertions(+), 76 deletions(-)
-- libvir-list mailing list libvir-list@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/libvir-list