On a Friday in 2020, Martin Kletzander wrote:
On Fri, Sep 04, 2020 at 01:32:24PM +0100, Daniel P. Berrangé wrote:On Fri, Sep 04, 2020 at 02:22:14PM +0200, Martin Kletzander wrote:At least in a particular scenario described in the code. Basically when libvirtd is running without CAP_SYS_NICE (e.g. in a container) and it is trying to set QEMU affinity to all CPUs (because there is no setting requested in the XML) it fails. But if we ignore the failure in this particular case than you can limit the CPUs used by controlling the affinity for libvirtd itself. In any other case (anything requested in the XML, pinning a live domain, etc.) the call is still considered fatal and the action errors out. Resolves: https://bugzilla.redhat.com/1819801I'd prefer if this commit message outlined the reason why this change is ok, instead of just pointing to the BZ eg add the following text: Consider a host with 8 CPUs. There are the following possible scenarios 1. Bare metal; libvirtd has affinity of 8 CPUs; QEMU should get 8 CPUs 2. Bare metal; libvirtd has affinity of 2 CPUs; QEMU should get 8 CPUs 3. Container has affinity of 8 CPUs; libvirtd has affinity of 8 CPus; QEMU should get 8 CPUs 4. Container has affinity of 8 CPUs; libvirtd has affinity of 2 CPus; QEMU should get 8 CPUs 5. Container has affinity of 4 CPUs; libvirtd has affinity of 4 CPus; QEMU should get 4 CPUs 6. Container has affinity of 4 CPUs; libvirtd has affinity of 2 CPus; QEMU should get 4 CPUs Scenarios 1 & 2 always work unless systemd restricted libvirtd privs. Scenario 3 works because libvirt checks current affinity first and skips the sched_setaffinity call, avoiding the SYS_NICE issue Scenario 4 works only if CAP_SYS_NICE is availalbe Scenarios 5 & 6 works only if CAP_SYS_NICE is present *AND* the cgroups cpuset is not set on the container. If libvirt blindly ignores the sched_setaffinity failure, then scenarios 4, 5 and 6 should all work, but with caveat in case 4 and 6, that QEMU will only get 2 CPUs instead of the possible 8 and 4 respectively. This is still better than failing. Therefore libvirt can blindly ignore the setaffinity failure, but *ONLY* ignore it when there was no affinity specified in the XML config. If user specified affinity explicitly, libvirt must report an error if it can't be honoured.I replaced my commit message with yours.Suggested-by: Daniel P. Berrangé <berrange@xxxxxxxxxx> Signed-off-by: Martin Kletzander <mkletzan@xxxxxxxxxx> --- src/qemu/qemu_process.c | 41 ++++++++++++++++++++++++++++++++++++++--- 1 file changed, 38 insertions(+), 3 deletions(-) diff --git a/src/qemu/qemu_process.c b/src/qemu/qemu_process.c index cfe09d632633..270bb37d3682 100644 --- a/src/qemu/qemu_process.c +++ b/src/qemu/qemu_process.c @@ -2528,6 +2528,7 @@ qemuProcessGetAllCpuAffinity(virBitmapPtr *cpumapRet) static int qemuProcessInitCpuAffinity(virDomainObjPtr vm) { + bool settingAll = false; g_autoptr(virBitmap) cpumapToSet = NULL; virDomainNumatuneMemMode mem_mode; qemuDomainObjPrivatePtr priv = vm->privateData; @@ -2566,13 +2567,30 @@ qemuProcessInitCpuAffinity(virDomainObjPtr vm) if (!(cpumapToSet = virBitmapNewCopy(vm->def->cputune.emulatorpin))) return -1; } else { + settingAll = true; if (qemuProcessGetAllCpuAffinity(&cpumapToSet) < 0) return -1; } if (cpumapToSet && virProcessSetAffinity(vm->pid, cpumapToSet) < 0) { - return -1; + /* + * We only want to error out if we failed to set the affinity to + * user-requested mapping. If we are just trying to reset the affinity + * to all CPUs and this fails it can only be an issue if: + * 1) libvirtd does not have CAP_SYS_NICE + * 2) libvirtd does not run on all CPUs + * + * However since this scenario is very improbable, we rather skip + * reporting the error because it helps running libvirtd in a a scenario + * where pinning is handled by someone else.
The patch as-is reports the error in all cases, it is merely ignored in some cases. virResetLastError clears the thread-local error structure, but it has no power to remove the error from the log. It mostly should not be used outside of public APIs where it makes sure we don't leave an error from a previous API invocation there. Reporting harmless errors confuses users and we have made changes in the past specifically to suppress harmless errors when libvirt is run in a container.
I wouldn't call this scenario "improbably" - it is entirely expected by some of our users. Replace these three lines with "This scenario can easily occurr when libvirtd is run
*occur
inside a container with restrictive permissions and CPU pinning"Yeah, I meant "improbable on bare metal". I replaced them with your explanation.
Note that in the version pushed to master, you only replaced the text in one of the two comments. Jano
With the text changes Reviewed-by: Daniel P. Berrangé <berrange@xxxxxxxxxx>Thanks, pushed.Regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
Attachment:
signature.asc
Description: PGP signature
