Inside of qemuProcessSetupPid() there's @numatune variable which is set to vm->def->numa, but it lives only in one block. In the rest of places the expanded form (vm->def->numa) is used instead. Move the variable declaration at the beginning of the function and use it instead of the expanded form. Signed-off-by: Michal Privoznik <mprivozn(a)redhat.com&gt; --- src/qemu/qemu_process.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/src/qemu/qemu_process.c b/src/qemu/qemu_process.c index ce5ed6389a..57c3ea2dbf 100644 --- a/src/qemu/qemu_process.c +++ b/src/qemu/qemu_process.c @@ -2553,6 +2553,7 @@ qemuProcessSetupPid(virDomainObj *vm, virDomainThreadSchedParam *sched) { qemuDomainObjPrivate *priv = vm->privateData; + virDomainNuma *numatune = vm->def->numa; virDomainNumatuneMemMode mem_mode; virCgroup *cgroup = NULL; virBitmap *use_cpumask = NULL; @@ -2589,10 +2590,10 @@ qemuProcessSetupPid(virDomainObj *vm, if (virCgroupHasController(priv->cgroup, VIR_CGROUP_CONTROLLER_CPU) || virCgroupHasController(priv->cgroup, VIR_CGROUP_CONTROLLER_CPUSET)) { - if (virDomainNumatuneGetMode(vm->def->numa, -1, &mem_mode) == 0 && + if (virDomainNumatuneGetMode(numatune, -1, &mem_mode) == 0 && (mem_mode == VIR_DOMAIN_NUMATUNE_MEM_STRICT || mem_mode == VIR_DOMAIN_NUMATUNE_MEM_RESTRICTIVE) && - virDomainNumatuneMaybeFormatNodeset(vm->def->numa, + virDomainNumatuneMaybeFormatNodeset(numatune, priv->autoNodeset, &mem_mask, -1) < 0) goto cleanup; @@ -2601,8 +2602,6 @@ qemuProcessSetupPid(virDomainObj *vm, * threads based on the node they are in as there is nothing else uses * for such restriction (e.g. numa_set_membind). */ if (nameval == VIR_CGROUP_THREAD_VCPU) { - virDomainNuma *numatune = vm->def->numa; - /* Look for the guest NUMA node of this vCPU */ for (i = 0; i < virDomainNumaGetNodeCount(numatune); i++) { virBitmap *node_cpus = virDomainNumaGetNodeCpumask(numatune, i); -- 2.39.3

Consider a domain with two guest NUMA nodes and the following <numatune/> setting : <numatune> <memory mode="strict" nodeset="0"/> <memnode cellid="0" mode="strict" nodeset="1"/> </numatune> What this means is the emulator thread is pinned onto host NUMA node #0 (by setting corresponding cpuset.mems to "0"), and two memory-backend-* objects are created: -object '{"qom-type":"memory-backend-ram","id":"ram-node0", .., "host-nodes":[1],"policy":"bind"}' \ -numa node,nodeid=0,cpus=0-1,memdev=ram-node0 \ -object '{"qom-type":"memory-backend-ram","id":"ram-node1", .., "host-nodes":[0],"policy":"bind"}' \ -numa node,nodeid=1,cpus=2-3,memdev=ram-node1 \ Note, the emulator thread is pinned well before QEMU is even exec()-ed. Now, the way memory allocation works in QEMU is: the emulator thread calls mmap() followed by mbind() (which is sane, that's how everybody should do it). BUT, because the thread is already restricted by CGroups to just NUMA node #0, calling: mbind(host-nodes:[1]); /* made up syntax (TM) */ fails. This is expected though. Kernel was instructed to place the memory at NUMA node "0" and yet, process is trying to place it elsewhere. We used to solve this by not restricting emulator thread at all initially, and only after it's done initializing (i.e. we got the QMP greeting) we placed it onto desired nodes. But this had its own problems (e.g. QEMU might have locked pieces of its memory which were then unable to migrate onto different NUMA nodes). Therefore, in v5.1.0-rc1~282 we've changed this and set CGropus upfront (even before exec()-ing QEMU). And this used to work, but something has changed (I can't really put my finger anywhere). Therefore, for the initialization start the thread with union of all configured host NUMA nodes ("0-1" in our example) and fix the placement only after QEMU is started. NB, the memory hotplug suffers the same problem, but that might be fixed later. Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=2138150 Signed-off-by: Michal Privoznik <mprivozn(a)redhat.com&gt; --- src/qemu/qemu_domain.c | 26 +++++++++++++++++++++++++ src/qemu/qemu_domain.h | 5 +++++ src/qemu/qemu_process.c | 42 ++++++++++++++++++++++++++++++----------- 3 files changed, 62 insertions(+), 11 deletions(-) diff --git a/src/qemu/qemu_domain.c b/src/qemu/qemu_domain.c index 35d03515c7..d97a0ce24e 100644 --- a/src/qemu/qemu_domain.c +++ b/src/qemu/qemu_domain.c @@ -12674,3 +12674,29 @@ qemuDomainEvaluateCPUMask(const virDomainDef *def, return NULL; } + + +void +qemuDomainNumatuneMaybeFormatNodesetUnion(virDomainObj *vm, + virBitmap **nodeset, + char **nodesetStr) +{ + virDomainNuma *numatune = vm->def->numa; + qemuDomainObjPrivate *priv = vm->privateData; + g_autoptr(virBitmap) unionMask = virBitmapNew(0); + ssize_t i; + + for (i = -1; i < (ssize_t)virDomainNumaGetNodeCount(numatune); i++) { + virBitmap *tmp; + + tmp = virDomainNumatuneGetNodeset(numatune, priv->autoNodeset, i); + if (tmp) + virBitmapUnion(unionMask, tmp); + } + + if (nodesetStr) + *nodesetStr = virBitmapFormat(unionMask); + + if (nodeset) + *nodeset = g_steal_pointer(&unionMask); +} diff --git a/src/qemu/qemu_domain.h b/src/qemu/qemu_domain.h index ec9ae75bce..999190e381 100644 --- a/src/qemu/qemu_domain.h +++ b/src/qemu/qemu_domain.h @@ -1140,3 +1140,8 @@ virBitmap * qemuDomainEvaluateCPUMask(const virDomainDef *def, virBitmap *cpumask, virBitmap *autoCpuset); + +void +qemuDomainNumatuneMaybeFormatNodesetUnion(virDomainObj *vm, + virBitmap **nodeset, + char **nodesetStr); diff --git a/src/qemu/qemu_process.c b/src/qemu/qemu_process.c index 57c3ea2dbf..6b85b7cee7 100644 --- a/src/qemu/qemu_process.c +++ b/src/qemu/qemu_process.c @@ -2550,7 +2550,8 @@ qemuProcessSetupPid(virDomainObj *vm, virBitmap *cpumask, unsigned long long period, long long quota, - virDomainThreadSchedParam *sched) + virDomainThreadSchedParam *sched, + bool unionMems) { qemuDomainObjPrivate *priv = vm->privateData; virDomainNuma *numatune = vm->def->numa; @@ -2592,11 +2593,22 @@ qemuProcessSetupPid(virDomainObj *vm, if (virDomainNumatuneGetMode(numatune, -1, &mem_mode) == 0 && (mem_mode == VIR_DOMAIN_NUMATUNE_MEM_STRICT || - mem_mode == VIR_DOMAIN_NUMATUNE_MEM_RESTRICTIVE) && - virDomainNumatuneMaybeFormatNodeset(numatune, - priv->autoNodeset, - &mem_mask, -1) < 0) - goto cleanup; + mem_mode == VIR_DOMAIN_NUMATUNE_MEM_RESTRICTIVE)) { + + /* QEMU allocates its memory from the emulator thread. Thus it + * needs to access union of all host nodes configured. This is + * going to be replaced with proper value later in the startup + * process. */ + if (unionMems && + nameval == VIR_CGROUP_THREAD_EMULATOR) { + qemuDomainNumatuneMaybeFormatNodesetUnion(vm, NULL, &mem_mask); + } else { + if (virDomainNumatuneMaybeFormatNodeset(numatune, + priv->autoNodeset, + &mem_mask, -1) < 0) + goto cleanup; + } + } /* For restrictive numatune mode we need to set cpuset.mems for vCPU * threads based on the node they are in as there is nothing else uses @@ -2689,13 +2701,15 @@ qemuProcessSetupPid(virDomainObj *vm, static int -qemuProcessSetupEmulator(virDomainObj *vm) +qemuProcessSetupEmulator(virDomainObj *vm, + bool unionMems) { return qemuProcessSetupPid(vm, vm->pid, VIR_CGROUP_THREAD_EMULATOR, 0, vm->def->cputune.emulatorpin, vm->def->cputune.emulator_period, vm->def->cputune.emulator_quota, - vm->def->cputune.emulatorsched); + vm->def->cputune.emulatorsched, + unionMems); } @@ -5891,7 +5905,8 @@ qemuProcessSetupVcpu(virDomainObj *vm, vcpuid, vcpu->cpumask, vm->def->cputune.period, vm->def->cputune.quota, - &vcpu->sched) < 0) + &vcpu->sched, + false) < 0) return -1; if (schedCore && @@ -6046,7 +6061,8 @@ qemuProcessSetupIOThread(virDomainObj *vm, iothread->cpumask, vm->def->cputune.iothread_period, vm->def->cputune.iothread_quota, - &iothread->sched); + &iothread->sched, + false); } @@ -7746,7 +7762,7 @@ qemuProcessLaunch(virConnectPtr conn, goto cleanup; VIR_DEBUG("Setting emulator tuning/settings"); - if (qemuProcessSetupEmulator(vm) < 0) + if (qemuProcessSetupEmulator(vm, true) < 0) goto cleanup; VIR_DEBUG("Setting cgroup for external devices (if required)"); @@ -7809,6 +7825,10 @@ qemuProcessLaunch(virConnectPtr conn, if (qemuConnectAgent(driver, vm) < 0) goto cleanup; + VIR_DEBUG("Fixing up emulator tuning/settings"); + if (qemuProcessSetupEmulator(vm, false) < 0) + goto cleanup; + VIR_DEBUG("setting up hotpluggable cpus"); if (qemuDomainHasHotpluggableStartupVcpus(vm->def)) { if (qemuDomainRefreshVcpuInfo(vm, asyncJob, false) < 0) -- 2.39.3

This is very similar to the previous commit, but fixes the case where there are no CGroups (or are disabled by admin) and thus libvirt calls mbind() before exec()-ing QEMU. Signed-off-by: Michal Privoznik <mprivozn(a)redhat.com&gt; --- src/qemu/qemu_process.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/src/qemu/qemu_process.c b/src/qemu/qemu_process.c index 6b85b7cee7..0fafda0beb 100644 --- a/src/qemu/qemu_process.c +++ b/src/qemu/qemu_process.c @@ -3106,10 +3106,8 @@ struct qemuProcessHookData { static int qemuProcessHook(void *data) { struct qemuProcessHookData *h = data; - qemuDomainObjPrivate *priv = h->vm->privateData; int ret = -1; int fd; - virBitmap *nodeset = NULL; virDomainNumatuneMemMode mode; /* This method cannot use any mutexes, which are not @@ -3154,8 +3152,9 @@ static int qemuProcessHook(void *data) * and enabled in the config) rather than virNuma*. */ VIR_DEBUG("Relying on CGroups for memory binding"); } else { - nodeset = virDomainNumatuneGetNodeset(h->vm->def->numa, - priv->autoNodeset, -1); + g_autoptr(virBitmap) nodeset = NULL; + + qemuDomainNumatuneMaybeFormatNodesetUnion(h->vm, &nodeset, NULL); if (virNumaSetupMemoryPolicy(mode, nodeset) < 0) goto cleanup; -- 2.39.3

Again, this fixes the same problem as one of previous commits, but this time for memory hotplug. Long story short, if there's a domain running and the emulator thread is restricted to a subset of host NUMA nodes, but the memory that's about to be hotplugged requires memory from a host NUMA node that's not in the set we need to allow emulator thread to access the node, temporarily. Signed-off-by: Michal Privoznik <mprivozn(a)redhat.com&gt; --- src/qemu/qemu_hotplug.c | 8 ++++++++ src/qemu/qemu_process.c | 2 +- src/qemu/qemu_process.h | 3 +++ 3 files changed, 12 insertions(+), 1 deletion(-) diff --git a/src/qemu/qemu_hotplug.c b/src/qemu/qemu_hotplug.c index d5148f5815..972df572a7 100644 --- a/src/qemu/qemu_hotplug.c +++ b/src/qemu/qemu_hotplug.c @@ -2247,6 +2247,7 @@ qemuDomainAttachMemory(virQEMUDriver *driver, bool teardownlabel = false; bool teardowncgroup = false; bool teardowndevice = false; + bool restoreemulatorcgroup = false; g_autoptr(virJSONValue) props = NULL; virObjectEvent *event; int id; @@ -2294,6 +2295,10 @@ qemuDomainAttachMemory(virQEMUDriver *driver, if (qemuDomainAdjustMaxMemLock(vm) < 0) goto removedef; + if (qemuProcessSetupEmulator(vm, true) < 0) + goto removedef; + restoreemulatorcgroup = true; + qemuDomainObjEnterMonitor(vm); if (qemuMonitorAddObject(priv->mon, &props, NULL) < 0) goto exit_monitor; @@ -2333,6 +2338,9 @@ qemuDomainAttachMemory(virQEMUDriver *driver, qemuDomainReleaseMemoryDeviceSlot(vm, mem); } + if (restoreemulatorcgroup) + qemuProcessSetupEmulator(vm, false); + virDomainMemoryDefFree(mem); return ret; diff --git a/src/qemu/qemu_process.c b/src/qemu/qemu_process.c index 0fafda0beb..803bf9b0a3 100644 --- a/src/qemu/qemu_process.c +++ b/src/qemu/qemu_process.c @@ -2700,7 +2700,7 @@ qemuProcessSetupPid(virDomainObj *vm, } -static int +int qemuProcessSetupEmulator(virDomainObj *vm, bool unionMems) { diff --git a/src/qemu/qemu_process.h b/src/qemu/qemu_process.h index e0cbafda1d..1c4c0678ab 100644 --- a/src/qemu/qemu_process.h +++ b/src/qemu/qemu_process.h @@ -235,3 +235,6 @@ void qemuProcessCleanupMigrationJob(virQEMUDriver *driver, void qemuProcessRefreshDiskProps(virDomainDiskDef *disk, struct qemuDomainDiskInfo *info); + +int qemuProcessSetupEmulator(virDomainObj *vm, + bool unionMems); -- 2.39.3