On 6/10/20 8:35 PM, Daniel Henrique Barboza wrote:
Libvirt allows the user to define an incomplete NUMA topology, where
the sum of all CPUs in each cell is less than the total of VCPUs.
What ends up happening is that QEMU allocates the non-enumerated CPUs
in the first NUMA node. This behavior is being flagged as 'to be
deprecated' at least since QEMU commit ec78f8114bc4 ("numa: use
possible_cpus for not mapped CPUs check").
In [1], Maxiwell suggested that we forbid the user to define such
topologies. In his review [2], Peter Krempa pointed out that we can't
break existing guests, and suggested that Libvirt should emulate the
QEMU behavior of putting the remaining vCPUs in the first NUMA node
in these cases.
This patch implements Peter Krempa's suggestion. Since we're going
to most likely end up with disjointed NUMA configuration in node 0
after the auto-fill, we're making auto-fill dependent on QEMU_CAPS_NUMA.
A following patch will update the documentation not just to inform
about the auto-fill mechanic with incomplete NUMA topologies, but also
to discourage the user to create such topologies in the future. This
approach also makes Libvirt independent of whether QEMU changes
its current behavior since we're either auto-filling the CPUs in
node 0 or the user (hopefully) is aware that incomplete topologies,
although supported in Libvirt, are to be avoided.
[1] https://www.redhat.com/archives/libvir-list/2019-June/msg00224.html
[2] https://www.redhat.com/archives/libvir-list/2019-June/msg00263.html
Signed-off-by: Daniel Henrique Barboza <danielhb413@xxxxxxxxx>
---
src/qemu/qemu_domain.c | 47 ++++++++++++++++++++++++++++++++++++++++++
src/qemu/qemu_domain.h | 4 ++++
src/qemu/qemu_driver.c | 9 ++++++++
3 files changed, 60 insertions(+)
diff --git a/src/qemu/qemu_domain.c b/src/qemu/qemu_domain.c
index 2dad823a86..76191e028b 100644
--- a/src/qemu/qemu_domain.c
+++ b/src/qemu/qemu_domain.c
@@ -4963,6 +4963,50 @@ qemuDomainDefTsegPostParse(virDomainDefPtr def,
}
+/**
+ * qemuDomainDefNumaCPUsRectify:
+ * @numa: pointer to numa definition
+ * @maxCpus: number of CPUs this numa is supposed to have
+ *
+ * This function emulates the (to be deprecated) behavior of filling
+ * up in node0 with the remaining CPUs, in case of an incomplete NUMA
+ * setup, up to getVcpusMax.
+ *
+ * Returns: 0 on success, -1 on error
+ */
+int
+qemuDomainDefNumaCPUsRectify(virDomainDefPtr def, virQEMUCapsPtr qemuCaps)
+{
+ unsigned int vcpusMax, numacpus;
+
+ /* QEMU_CAPS_NUMA tells us if QEMU is able to handle disjointed
+ * NUMA CPU ranges. The filling process will create a disjointed
+ * setup in node0 most of the time. Do not proceed if QEMU
+ * can't handle it.*/
+ if (virDomainNumaGetNodeCount(def->numa) == 0 ||
+ !virQEMUCapsGet(qemuCaps, QEMU_CAPS_NUMA))
+ return 0;
+
+ vcpusMax = virDomainDefGetVcpusMax(def);
+ numacpus = virDomainNumaGetCPUCountTotal(def->numa);
+
+ if (numacpus < vcpusMax) {
+ if (virDomainNumaFillCPUsInNode(def->numa, 0, vcpusMax) < 0)
+ return -1;
+ }
+
+ return 0;
+}
+
+
+static int
+qemuDomainDefNumaCPUsPostParse(virDomainDefPtr def,
+ virQEMUCapsPtr qemuCaps)
+{
+ return qemuDomainDefNumaCPUsRectify(def, qemuCaps);
+}
+
+
static int
qemuDomainDefPostParseBasic(virDomainDefPtr def,
void *opaque G_GNUC_UNUSED)
@@ -5049,6 +5093,9 @@ qemuDomainDefPostParse(virDomainDefPtr def,
if (qemuDomainDefTsegPostParse(def, qemuCaps) < 0)
return -1;
+ if (qemuDomainDefNumaCPUsPostParse(def, qemuCaps) < 0)
+ return -1;
+
return 0;
}
diff --git a/src/qemu/qemu_domain.h b/src/qemu/qemu_domain.h
index 41d3f1561d..e78a2b935d 100644
--- a/src/qemu/qemu_domain.h
+++ b/src/qemu/qemu_domain.h
@@ -1297,3 +1297,7 @@ qemuDomainInitializePflashStorageSource(virDomainObjPtr vm);
bool
qemuDomainDiskBlockJobIsSupported(virDomainObjPtr vm,
virDomainDiskDefPtr disk);
+
+int
+qemuDomainDefNumaCPUsRectify(virDomainDefPtr def,
+ virQEMUCapsPtr qemuCaps);
diff --git a/src/qemu/qemu_driver.c b/src/qemu/qemu_driver.c
index 88517ba6a7..ff9414f3c4 100644
--- a/src/qemu/qemu_driver.c
+++ b/src/qemu/qemu_driver.c
@@ -4999,6 +4999,7 @@ qemuDomainSetVcpusMax(virQEMUDriverPtr driver,
unsigned int nvcpus)
{
g_autoptr(virQEMUDriverConfig) cfg = virQEMUDriverGetConfig(driver);
+ g_autoptr(virQEMUCaps) qemuCaps = NULL;
unsigned int topologycpus;
if (def) {
@@ -5029,6 +5030,14 @@ qemuDomainSetVcpusMax(virQEMUDriverPtr driver,
if (virDomainDefSetVcpusMax(persistentDef, nvcpus, driver->xmlopt) < 0)
return -1;
+ /* re-adjust NUMA nodes if needed */
+ if (!(qemuCaps = virQEMUCapsCacheLookup(driver->qemuCapsCache,
+ persistentDef->emulator)))
This will not do what you think it will do. Although, perhaps for
QEMU_CAPS_NUMA (introduced in what QEMU 2.1.0? point is ancient QEMU) it
doesn't matter because the likely hood is very small. But,
virQEMUCapsCacheLookup() will return caps as detected now, by current
version of libvirt. But what we need is to compare the capability
against the set of caps when the domain was started (vm->priv->qemuCaps).
For instance (and this shows how unlikely this corner case is, but still):
1) a domain is started with qemu-1.5.0 (or what the minimal version is)
2) yum update qemu
3) virQEMUCapsCacheLookup() will now fetch new caps for the updated QEMU.
Michal
