From dc27c829fd5909394e69ed253979f19b47644569 Mon Sep 17 00:00:00 2001

Message-Id: <dc27c829fd5909394e69ed253979f19b47644569@dist-git>

From: Michal Privoznik <mprivozn@redhat.com>

Date: Wed, 5 Jun 2019 11:33:28 +0200

Subject: [PATCH] qemu: Rework setting process affinity

MIME-Version: 1.0

Content-Type: text/plain; charset=UTF-8

Content-Transfer-Encoding: 8bit

https://bugzilla.redhat.com/show_bug.cgi?id=1503284

The way we currently start qemu from CPU affinity POV is as

follows:

  1) the child process is set affinity to all online CPUs (unless

  some vcpu pinning was given in the domain XML)

  2) Once qemu is running, cpuset cgroup is configured taking

  memory pinning into account

Problem is that we let qemu allocate its memory just anywhere in

1) and then rely in 2) to be able to move the memory to

configured NUMA nodes. This might not be always possible (e.g.

qemu might lock some parts of its memory) and is very suboptimal

(copying large memory between NUMA nodes takes significant amount

of time).

The solution is to set affinity to one of (in priority order):

  - The CPUs associated with NUMA memory affinity mask

  - The CPUs associated with emulator pinning

  - All online host CPUs

Later (once QEMU has allocated its memory) we then change this

again to (again in priority order):

  - The CPUs associated with emulator pinning

  - The CPUs returned by numad

  - The CPUs associated with vCPU pinning

  - All online host CPUs

Signed-off-by: Michal Privoznik <mprivozn@redhat.com>

Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>

(cherry picked from commit f136b83139c63f20de0df3285d9e82df2fb97bfc)

RHEL-8.1.0: https://bugzilla.redhat.com/show_bug.cgi?id=1716943

Signed-off-by: Michal Privoznik <mprivozn@redhat.com>

Message-Id: <c5f31a30daef2be65dc404ab0f1fbfb15be0d062.1559727075.git.mprivozn@redhat.com>

Reviewed-by: Andrea Bolognani <abologna@redhat.com>

---

 src/qemu/qemu_process.c | 132 +++++++++++++++++++---------------------

 1 file changed, 63 insertions(+), 69 deletions(-)

diff --git a/src/qemu/qemu_process.c b/src/qemu/qemu_process.c

index 2d2954ba18..6071b3ba3d 100644

--- a/src/qemu/qemu_process.c

+++ b/src/qemu/qemu_process.c

@@ -2335,6 +2335,21 @@ qemuProcessDetectIOThreadPIDs(virQEMUDriverPtr driver,

 }

+static int

+qemuProcessGetAllCpuAffinity(virBitmapPtr *cpumapRet)

+{

+    *cpumapRet = NULL;

+

+    if (!virHostCPUHasBitmap())

+        return 0;

+

+    if (!(*cpumapRet = virHostCPUGetOnlineBitmap()))

+        return -1;

+

+    return 0;

+}

+

+

 /*

  * To be run between fork/exec of QEMU only

  */

@@ -2342,9 +2357,9 @@ static int

 qemuProcessInitCpuAffinity(virDomainObjPtr vm)

 {

     int ret = -1;

-    virBitmapPtr cpumap = NULL;

     virBitmapPtr cpumapToSet = NULL;

-    virBitmapPtr hostcpumap = NULL;

+    VIR_AUTOPTR(virBitmap) hostcpumap = NULL;

+    virDomainNumatuneMemMode mem_mode;

     qemuDomainObjPrivatePtr priv = vm->privateData;

     if (!vm->pid) {

@@ -2353,59 +2368,39 @@ qemuProcessInitCpuAffinity(virDomainObjPtr vm)

         return -1;

     }

-    if (vm->def->placement_mode == VIR_DOMAIN_CPU_PLACEMENT_MODE_AUTO) {

-        VIR_DEBUG("Set CPU affinity with advisory nodeset from numad");

-        cpumapToSet = priv->autoCpuset;

+    /* Here is the deal, we can't set cpuset.mems before qemu is

+     * started as it clashes with KVM allocation. Therefore, we

+     * used to let qemu allocate its memory anywhere as we would

+     * then move the memory to desired NUMA node via CGroups.

+     * However, that might not be always possible because qemu

+     * might lock some parts of its memory (e.g. due to VFIO).

+     * Even if it possible, memory has to be copied between NUMA

+     * nodes which is suboptimal.

+     * Solution is to set affinity that matches the best what we

+     * would have set in CGroups and then fix it later, once qemu

+     * is already running. */

+    if (virDomainNumaGetNodeCount(vm->def->numa) <= 1 &&

+        virDomainNumatuneGetMode(vm->def->numa, -1, &mem_mode) == 0 &&

+        mem_mode == VIR_DOMAIN_NUMATUNE_MEM_STRICT) {

+        if (virDomainNumatuneMaybeGetNodeset(vm->def->numa,

+                                             priv->autoNodeset,

+                                             &cpumapToSet,

+                                             -1) < 0)

+            goto cleanup;

+    } else if (vm->def->cputune.emulatorpin) {

+        cpumapToSet = vm->def->cputune.emulatorpin;

     } else {

-        VIR_DEBUG("Set CPU affinity with specified cpuset");

-        if (vm->def->cpumask) {

-            cpumapToSet = vm->def->cpumask;

-        } else {

-            /* You may think this is redundant, but we can't assume libvirtd

-             * itself is running on all pCPUs, so we need to explicitly set

-             * the spawned QEMU instance to all pCPUs if no map is given in

-             * its config file */

-            int hostcpus;

-

-            if (virHostCPUHasBitmap()) {

-                hostcpumap = virHostCPUGetOnlineBitmap();

-                cpumap = virProcessGetAffinity(vm->pid);

-            }

-

-            if (hostcpumap && cpumap && virBitmapEqual(hostcpumap, cpumap)) {

-                /* we're using all available CPUs, no reason to set

-                 * mask. If libvirtd is running without explicit

-                 * affinity, we can use hotplugged CPUs for this VM */

-                ret = 0;

-                goto cleanup;

-            } else {

-                /* setaffinity fails if you set bits for CPUs which

-                 * aren't present, so we have to limit ourselves */

-                if ((hostcpus = virHostCPUGetCount()) < 0)

-                    goto cleanup;

-

-                if (hostcpus > QEMUD_CPUMASK_LEN)

-                    hostcpus = QEMUD_CPUMASK_LEN;

-

-                virBitmapFree(cpumap);

-                if (!(cpumap = virBitmapNew(hostcpus)))

-                    goto cleanup;

-

-                virBitmapSetAll(cpumap);

-

-                cpumapToSet = cpumap;

-            }

-        }

+        if (qemuProcessGetAllCpuAffinity(&hostcpumap) < 0)

+            goto cleanup;

+        cpumapToSet = hostcpumap;

     }

-    if (virProcessSetAffinity(vm->pid, cpumapToSet) < 0)

+    if (cpumapToSet &&

+        virProcessSetAffinity(vm->pid, cpumapToSet) < 0)

         goto cleanup;

     ret = 0;

-

  cleanup:

-    virBitmapFree(cpumap);

-    virBitmapFree(hostcpumap);

     return ret;

 }

@@ -2478,7 +2473,8 @@ qemuProcessSetupPid(virDomainObjPtr vm,

     qemuDomainObjPrivatePtr priv = vm->privateData;

     virDomainNumatuneMemMode mem_mode;

     virCgroupPtr cgroup = NULL;

-    virBitmapPtr use_cpumask;

+    virBitmapPtr use_cpumask = NULL;

+    VIR_AUTOPTR(virBitmap) hostcpumap = NULL;

     char *mem_mask = NULL;

     int ret = -1;

@@ -2490,12 +2486,21 @@ qemuProcessSetupPid(virDomainObjPtr vm,

     }

     /* Infer which cpumask shall be used. */

-    if (cpumask)

+    if (cpumask) {

         use_cpumask = cpumask;

-    else if (vm->def->placement_mode == VIR_DOMAIN_CPU_PLACEMENT_MODE_AUTO)

+    } else if (vm->def->placement_mode == VIR_DOMAIN_CPU_PLACEMENT_MODE_AUTO) {

         use_cpumask = priv->autoCpuset;

-    else

+    } else if (vm->def->cpumask) {

         use_cpumask = vm->def->cpumask;

+    } else {

+        /* You may think this is redundant, but we can't assume libvirtd

+         * itself is running on all pCPUs, so we need to explicitly set

+         * the spawned QEMU instance to all pCPUs if no map is given in

+         * its config file */

+        if (qemuProcessGetAllCpuAffinity(&hostcpumap) < 0)

+            goto cleanup;

+        use_cpumask = hostcpumap;

+    }

     /*

      * If CPU cgroup controller is not initialized here, then we need

@@ -2520,13 +2525,7 @@ qemuProcessSetupPid(virDomainObjPtr vm,

                 qemuSetupCgroupCpusetCpus(cgroup, use_cpumask) < 0)

                 goto cleanup;

-            /*

-             * Don't setup cpuset.mems for the emulator, they need to

-             * be set up after initialization in order for kvm

-             * allocations to succeed.

-             */

-            if (nameval != VIR_CGROUP_THREAD_EMULATOR &&

-                mem_mask && virCgroupSetCpusetMems(cgroup, mem_mask) < 0)

+            if (mem_mask && virCgroupSetCpusetMems(cgroup, mem_mask) < 0)

                 goto cleanup;

         }

@@ -6440,12 +6439,7 @@ qemuProcessLaunch(virConnectPtr conn,

     /* This must be done after cgroup placement to avoid resetting CPU

      * affinity */

-    if (!vm->def->cputune.emulatorpin &&

-        qemuProcessInitCpuAffinity(vm) < 0)

-        goto cleanup;

-

-    VIR_DEBUG("Setting emulator tuning/settings");

-    if (qemuProcessSetupEmulator(vm) < 0)

+    if (qemuProcessInitCpuAffinity(vm) < 0)

         goto cleanup;

     VIR_DEBUG("Setting cgroup for external devices (if required)");

@@ -6514,10 +6508,6 @@ qemuProcessLaunch(virConnectPtr conn,

     if (qemuProcessUpdateAndVerifyCPU(driver, vm, asyncJob) < 0)

         goto cleanup;

-    VIR_DEBUG("Setting up post-init cgroup restrictions");

-    if (qemuSetupCpusetMems(vm) < 0)

-        goto cleanup;

-

     VIR_DEBUG("setting up hotpluggable cpus");

     if (qemuDomainHasHotpluggableStartupVcpus(vm->def)) {

         if (qemuDomainRefreshVcpuInfo(driver, vm, asyncJob, false) < 0)

@@ -6543,6 +6533,10 @@ qemuProcessLaunch(virConnectPtr conn,

     if (qemuProcessDetectIOThreadPIDs(driver, vm, asyncJob) < 0)

         goto cleanup;

+    VIR_DEBUG("Setting emulator tuning/settings");

+    if (qemuProcessSetupEmulator(vm) < 0)

+        goto cleanup;

+

     VIR_DEBUG("Setting global CPU cgroup (if required)");

     if (qemuSetupGlobalCpuCgroup(vm) < 0)

         goto cleanup;

-- 

2.22.0