threading: support thread autopinning and interface-specific affinity

Using the new configuration format, it is now possible to set CPU affinity
settings per interface.

The threading.autopin option has been added to automatically use CPUs from the
same NUMA node as the interface. The autopin option requires
hwloc-devel / hwloc-dev to be installed and --enable-hwloc flag in configure
script.

Ticket: 7036

configure.ac

@@ -741,6 +741,33 @@
         exit 1
     fi
 
+    LIBHWLOC=""
+    AC_ARG_ENABLE(hwloc,
+            AS_HELP_STRING([--enable-hwloc], [Enable hwloc support [default=no]]),
+                        [enable_hwloc=$enableval],[enable_hwloc=no])
+    AS_IF([test "x$enable_hwloc" = "xyes"], [
+        PKG_CHECK_MODULES([HWLOC], [hwloc >= 2.0.0],
+            [AC_DEFINE([HAVE_HWLOC], [1], [Define if hwloc library is present and meets version requirements])],
+            LIBHWLOC="no")
+
+        if test "$LIBHWLOC" = "no"; then
+            echo
+            echo "   ERROR!  hwloc library version > 2.0.0 not found, go get it"
+            echo "   from https://www.open-mpi.org/projects/hwloc/ "
+            echo "   or your distribution:"
+            echo
+            echo "   Ubuntu: apt-get install hwloc libhwloc-dev"
+            echo "   Fedora: dnf install hwloc hwloc-devel"
+            echo "   CentOS/RHEL: yum install hwloc hwloc-devel"
+            echo
+            exit 1
+        else
+            CFLAGS="${CFLAGS} ${HWLOC_CFLAGS}"
+            LDFLAGS="${LDFLAGS} ${HWLOC_LIBS}"
+            enable_hwloc="yes"
+        fi
+    ])
+
   # libpthread
     AC_ARG_WITH(libpthread_includes,
             [  --with-libpthread-includes=DIR  libpthread include directory],
@@ -2561,6 +2588,7 @@ SURICATA_BUILD_CONF="Suricata Configuration:
   JA4 support:                             ${enable_ja4}
   Non-bundled htp:                         ${enable_non_bundled_htp}
   Hyperscan support:                       ${enable_hyperscan}
+  Hwloc support:                           ${enable_hwloc}
   Libnet support:                          ${enable_libnet}
   liblz4 support:                          ${enable_liblz4}
   Landlock support:                        ${enable_landlock}

doc/userguide/configuration/suricata-yaml.rst

@@ -924,6 +924,7 @@ per available CPU/CPU core.
 
     threading:
       set-cpu-affinity: yes
+      autopin: no
       cpu-affinity:
         management-cpu-set:
           cpu: [ 0 ]  # include only these cpus in affinity settings
@@ -940,6 +941,13 @@ per available CPU/CPU core.
             medium: [ "1-2" ]
             high: [ 3 ]
             default: "medium"
+          interface-specific-cpu-set:
+            - interface: "enp4s0f0" # 0000:3b:00.0 # net_bonding0 # ens1f0
+              cpu: [ 1,3,5,7,9 ]
+              mode: "exclusive"
+              prio:
+                high: [ "all" ]
+                default: "medium"
         verdict-cpu-set:
           cpu: [ 0 ]
           prio:
@@ -976,6 +984,80 @@ Runmode Workers::
 	worker-cpu-set - used for receive,streamtcp,decode,detect,output(logging),respond/reject, verdict
 
 
+Interface-specific CPU affinity settings
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Using the new configuration format introduced in Suricata 8.0 it is possible
+to set CPU affinity settings per interface. This can be useful
+when you have multiple interfaces and you want to dedicate specific CPU cores
+to specific interfaces. This can be useful for example when Suricata runs on
+multiple NUMA nodes and reads from interfaces on each NUMA node.
+
+Interface-specific affinity settings can be configured for the worker-cpu-set
+and the receive-cpu-set (only used in autofp mode).
+This feature is available for capture modes which work with interfaces
+(af-packet, dpdk, etc.). The value of the interface key can be the kernel
+interface name (e.g. eth0 for af-packet), the PCI address of the interface
+(e.g. 0000:3b:00.0 for DPDK capture mode), or the name of the virtual device
+interface (e.g. net_bonding0 for DPDK capture mode).
+The interface names needs to be unique and be located under the capture mode
+configuration.
+
+The interface-specific settings will override the global settings for the
+worker-cpu-set and receive-cpu-set. The CPUs do not need to be contained in
+the parent node settings. If the interface-specific settings are not defined,
+the global settings will be used.
+
+::
+
+  threading:
+    set-cpu-affinity: yes
+    cpu-affinity:
+      worker-cpu-set:
+        interface-specific-cpu-set:
+          - interface: "eth0" # 0000:3b:00.0 # net_bonding0
+            cpu: [ 1,3,5,7,9 ]
+            mode: "exclusive"
+            prio:
+              high: [ "all" ]
+              default: "medium"
+
+Automatic NUMA-aware CPU core pinning
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+When Suricata is running on a system with multiple NUMA nodes, it is possible
+to automatically use CPUs from the same NUMA node as the network capture
+interface.
+CPU cores on the same NUMA nodes as the network capture interface have
+reduced memory access latency and increased the performance of Suricata.
+This is enabled by setting the `autopin` option to `yes` in the threading
+section. This option is available for worker-cpu-set and receive-cpu-set.
+
+::
+
+  threading:
+    set-cpu-affinity: yes
+    autopin: yes
+    cpu-affinity:
+      worker-cpu-set:
+        cpu: [ "all" ]
+        mode: "exclusive"
+        prio:
+          high: [ "all" ]
+
+Consider 2 interfaces defined in the capture mode configuration, one on each
+NUMA node. The `autopin` option is enabled to automatically use CPUs from the
+same NUMA node as the interface. The worker-cpu-set is set to use all CPUs.
+When interface on the first NUMA node is used, the worker threads will be
+pinned to CPUs on the first NUMA node. When interface on the second NUMA node
+is used, the worker threads will be pinned to CPUs on the second NUMA node.
+If the number of CPU cores on a given NUMA node is exhausted then the worker
+threads will be pinned to CPUs on the other NUMA node.
+
+The option `threading.autopin` can be combined with the interface-specific CPU
+affinity settings.
+To use the `autopin` option, the system must have the `hwloc`
+dependency installed and pass `--enable-hwloc` to the configure script.
 
 IP Defrag
 ---------

doc/userguide/upgrade.rst

@@ -99,6 +99,20 @@ Major changes
     +     worker-cpu-set:
     +       cpu: [0, 1]
 
+  - The `threading.cpu-affinity` configuration has been extended to support
+    interface-specific CPU affinity settings. This allows you to specify
+    CPU affinity settings for each interface separately.
+    The new configuration format is described in :ref:`suricata-yaml-threading`.
+    The old configuration format does not support this extension and will be
+    removed in Suricata 9.0.
+  - The `threading.cpu-affinity` configuration now supports autopinning
+    worker or receive threads to the same NUMA node as the network capture
+    interface is located on.
+    This can be enabled by setting `threading.autopin` to `yes`.
+    See :ref:`suricata-yaml-threading` for more information.
+    This requires hwloc dependency to be installed and `--enable-hwloc`
+    to be passed to configure script.
+
 Removals
 ~~~~~~~~
 - The ssh keywords ``ssh.protoversion`` and ``ssh.softwareversion`` have been removed.

src/runmode-dpdk.c

@@ -368,12 +368,17 @@ static int ConfigSetThreads(DPDKIfaceConfig *iconf, const char *entry_str)
         SCReturnInt(-EINVAL);
     }
 
-    ThreadsAffinityType *wtaf = GetAffinityTypeFromName("worker-cpu-set");
+    bool wtaf_periface = true;
+    ThreadsAffinityType *wtaf = GetAffinityTypeForNameAndIface("worker-cpu-set", iconf->iface);
     if (wtaf == NULL) {
-        SCLogError("Specify worker-cpu-set list in the threading section");
-        SCReturnInt(-EINVAL);
+        wtaf_periface = false;
+        wtaf = GetAffinityTypeForNameAndIface("worker-cpu-set", NULL); // mandatory
+        if (wtaf == NULL) {
+            SCLogError("Specify worker-cpu-set list in the threading section");
+            SCReturnInt(-EINVAL);
+        }
     }
-    ThreadsAffinityType *mtaf = GetAffinityTypeFromName("management-cpu-set");
+    ThreadsAffinityType *mtaf = GetAffinityTypeForNameAndIface("management-cpu-set", NULL);
     if (mtaf == NULL) {
         SCLogError("Specify management-cpu-set list in the threading section");
         SCReturnInt(-EINVAL);
@@ -406,7 +411,12 @@ static int ConfigSetThreads(DPDKIfaceConfig *iconf, const char *entry_str)
     }
 
     if (strcmp(entry_str, "auto") == 0) {
-        iconf->threads = (uint16_t)sched_cpus / LiveGetDeviceCount();
+        if (wtaf_periface) {
+            iconf->threads = (uint16_t)sched_cpus;
+            SCLogConfig("%s: auto-assigned %u threads", iconf->iface, iconf->threads);
+            SCReturnInt(0);
+        }
+        iconf->threads = (uint16_t)sched_cpus / LiveGetDeviceCountWithoutAssignedThreading();
         if (iconf->threads == 0) {
             SCLogError("Not enough worker CPU cores with affinity were configured");
             SCReturnInt(-ERANGE);
@@ -416,7 +426,8 @@ static int ConfigSetThreads(DPDKIfaceConfig *iconf, const char *entry_str)
             iconf->threads++;
             remaining_auto_cpus--;
         } else if (remaining_auto_cpus == -1) {
-            remaining_auto_cpus = (int32_t)sched_cpus % LiveGetDeviceCount();
+            remaining_auto_cpus =
+                    (int32_t)sched_cpus % LiveGetDeviceCountWithoutAssignedThreading();
             if (remaining_auto_cpus > 0) {
                 iconf->threads++;
                 remaining_auto_cpus--;
@@ -844,23 +855,46 @@ static int ConfigLoad(DPDKIfaceConfig *iconf, const char *iface)
     SCReturnInt(0);
 }
 
-static int32_t ConfigValidateThreads(uint16_t iface_threads)
+static bool ConfigThreadsGenericIsValid(uint16_t iface_threads, ThreadsAffinityType *wtaf)
 {
     static uint32_t total_cpus = 0;
     total_cpus += iface_threads;
-    ThreadsAffinityType *wtaf = GetAffinityTypeFromName("worker-cpu-set");
     if (wtaf == NULL) {
         SCLogError("Specify worker-cpu-set list in the threading section");
-        return -1;
+        return false;
     }
     if (total_cpus > UtilAffinityGetAffinedCPUNum(wtaf)) {
-        SCLogError("Interfaces requested more cores than configured in the threading section "
-                   "(requested %d configured %d",
+        SCLogError("Interfaces requested more cores than configured in the worker-cpu-set "
+                   "threading section (requested %d configured %d",
                 total_cpus, UtilAffinityGetAffinedCPUNum(wtaf));
-        return -1;
+        return false;
     }
 
-    return 0;
+    return true;
+}
+
+static bool ConfigThreadsInterfaceIsValid(uint16_t iface_threads, ThreadsAffinityType *itaf)
+{
+    if (iface_threads > UtilAffinityGetAffinedCPUNum(itaf)) {
+        SCLogError("Interface requested more cores than configured in the interface-specific "
+                   "threading section (requested %d configured %d",
+                iface_threads, UtilAffinityGetAffinedCPUNum(itaf));
+        return false;
+    }
+
+    return true;
+}
+
+static bool ConfigIsThreadingValid(uint16_t iface_threads, const char *iface)
+{
+    ThreadsAffinityType *itaf = GetAffinityTypeForNameAndIface("worker-cpu-set", iface);
+    ThreadsAffinityType *wtaf = GetAffinityTypeForNameAndIface("worker-cpu-set", NULL);
+    if (itaf && !ConfigThreadsInterfaceIsValid(iface_threads, itaf)) {
+        return false;
+    } else if (itaf == NULL && !ConfigThreadsGenericIsValid(iface_threads, wtaf)) {
+        return false;
+    }
+    return true;
 }
 
 static DPDKIfaceConfig *ConfigParse(const char *iface)
@@ -873,7 +907,7 @@ static DPDKIfaceConfig *ConfigParse(const char *iface)
 
     ConfigInit(&iconf);
     retval = ConfigLoad(iconf, iface);
-    if (retval < 0 || ConfigValidateThreads(iconf->threads) != 0) {
+    if (retval < 0 || !ConfigIsThreadingValid(iconf->threads, iface)) {
         iconf->DerefFunc(iconf);
         SCReturnPtr(NULL, "void *");
     }

src/suricata.c

@@ -111,6 +111,7 @@
 #include "tmqh-packetpool.h"
 #include "tm-queuehandlers.h"
 
+#include "util-affinity.h"
 #include "util-byte.h"
 #include "util-conf.h"
 #include "util-coredump-config.h"
@@ -2297,6 +2298,9 @@ void PostRunDeinit(const int runmode, struct timeval *start_time)
     StreamTcpFreeConfig(STREAM_VERBOSE);
     DefragDestroy();
     HttpRangeContainersDestroy();
+#ifdef HAVE_HWLOC
+    TopologyDestroy();
+#endif /* HAVE_HWLOC */
 
     TmqResetQueues();
 #ifdef PROFILING

src/threadvars.h

@@ -136,6 +136,9 @@ typedef struct ThreadVars_ {
     struct FlowQueue_ *flow_queue;
     bool break_loop;
 
+    /** Interface-specific thread affinity */
+    char *iface_name;
+
     Storage storage[];
 } ThreadVars;
 

src/tm-threads.c

@@ -865,8 +865,24 @@ TmEcode TmThreadSetupOptions(ThreadVars *tv)
         TmThreadSetPrio(tv);
     if (tv->thread_setup_flags & THREAD_SET_AFFTYPE) {
         ThreadsAffinityType *taf = &thread_affinity[tv->cpu_affinity];
+        bool use_iface_affinity = RunmodeIsAutofp() && tv->cpu_affinity == RECEIVE_CPU_SET &&
+                                  FindAffinityByInterface(taf, tv->iface_name) != NULL;
+        use_iface_affinity |= RunmodeIsWorkers() && tv->cpu_affinity == WORKER_CPU_SET &&
+                              FindAffinityByInterface(taf, tv->iface_name) != NULL;
+
+        if (use_iface_affinity) {
+            taf = FindAffinityByInterface(taf, tv->iface_name);
+        }
+
+        if (UtilAffinityGetAffinedCPUNum(taf) == 0) {
+            if (!taf->nocpu_warned) {
+                SCLogWarning("No CPU affinity set for %s", AffinityGetYamlPath(taf));
+                taf->nocpu_warned = true;
+            }
+        }
+
         if (taf->mode_flag == EXCLUSIVE_AFFINITY) {
-            uint16_t cpu = AffinityGetNextCPU(taf);
+            uint16_t cpu = AffinityGetNextCPU(tv, taf);
             SetCPUAffinity(cpu);
             /* If CPU is in a set overwrite the default thread prio */
             if (CPU_ISSET(cpu, &taf->lowprio_cpu)) {
@@ -1600,6 +1616,10 @@ static void TmThreadFree(ThreadVars *tv)
         SCFree(tv->printable_name);
     }
 
+    if (tv->iface_name) {
+        SCFree(tv->iface_name);
+    }
+
     if (tv->stream_pq_local) {
         BUG_ON(tv->stream_pq_local->len);
         SCMutexDestroy(&tv->stream_pq_local->mutex_q);