The unit that cpuset CGroups controller works with is a
thread/process, not individual memory allocations. Therefore,
after we've set cpuset.mems for emulator (after previous commit
it's set to union of all host NUMA nodes allowed for given
domain), and as we try to set up cpuset.mems for vCPUs/IOThreads,
memory is migrated to selected NUMA node(s). We are effectively
saying: "this thread (vCPU thread) can have memory only from
these NUMA node(s)".

That's not really what we want though. The cpuset controller
doesn't differentiate memory "belonging" to the emulator thread
and vCPU thread or IOThread even.

Therefore, set union of all allowed host NUMA nodes, just like
we're doing for the emulator thread.

Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=2138150
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
---
 src/qemu/qemu_process.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/src/qemu/qemu_process.c b/src/qemu/qemu_process.c
index 1d3cdeff9a..d9269e37a1 100644
--- a/src/qemu/qemu_process.c
+++ b/src/qemu/qemu_process.c
@@ -2598,7 +2598,6 @@ qemuProcessSetupPid(virDomainObj *vm,
             /* QEMU allocates its memory from the emulator thread. Thus it
              * needs to access union of all host nodes configured. */
             if (unionMems &&
-                nameval == VIR_CGROUP_THREAD_EMULATOR &&
                 mem_mode != VIR_DOMAIN_NUMATUNE_MEM_RESTRICTIVE) {
                 qemuDomainNumatuneMaybeFormatNodesetUnion(vm, NULL, &mem_mask);
             } else {
@@ -5911,7 +5910,7 @@ qemuProcessSetupVcpu(virDomainObj *vm,
                             vm->def->cputune.period,
                             vm->def->cputune.quota,
                             &vcpu->sched,
-                            false) < 0)
+                            true) < 0)
         return -1;
 
     if (schedCore &&
@@ -6067,7 +6066,7 @@ qemuProcessSetupIOThread(virDomainObj *vm,
                                vm->def->cputune.iothread_period,
                                vm->def->cputune.iothread_quota,
                                &iothread->sched,
-                               false);
+                               true);
 }
 
 
-- 
2.39.3

On Wed, Jun 07, 2023 at 04:41:00PM +0200, Michal Privoznik wrote:
>The unit that cpuset CGroups controller works with is a
>thread/process, not individual memory allocations. Therefore,
>after we've set cpuset.mems for emulator (after previous commit
>it's set to union of all host NUMA nodes allowed for given
>domain), and as we try to set up cpuset.mems for vCPUs/IOThreads,
>memory is migrated to selected NUMA node(s). We are effectively
>saying: "this thread (vCPU thread) can have memory only from
>these NUMA node(s)".
>
>That's not really what we want though. The cpuset controller
>doesn't differentiate memory "belonging" to the emulator thread
>and vCPU thread or IOThread even.
>
>Therefore, set union of all allowed host NUMA nodes, just like
>we're doing for the emulator thread.
>
>Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=2138150
>Signed-off-by: Michal Privoznik <mprivozn@redhat.com>

Reviewed-by: Martin Kletzander <mkletzan@redhat.com>