Merge pull request #20018 from vespa-engine/arnej/add-feature-flag-to-ignore-stack-sizes

ignore requested stack sizes

7 files changed, 28 insertions, 1 deletions

diff --git a/config-model-api/src/main/java/com/yahoo/config/model/api/ModelContext.java b/config-model-api/src/main/java/com/yahoo/config/model/api/ModelContext.java
index c5781c2805d..2232297d296 100644
--- a/config-model-api/src/main/java/com/yahoo/config/model/api/ModelContext.java
+++ b/config-model-api/src/main/java/com/yahoo/config/model/api/ModelContext.java
@@ -105,6 +105,7 @@ public interface ModelContext {
         @ModelFeatureFlag(owners = {"vekterli", "geirst"}) default boolean distributorEnhancedMaintenanceScheduling() { return false; }
         @ModelFeatureFlag(owners = {"arnej"}) default boolean forwardIssuesAsErrors() { return true; }
         @ModelFeatureFlag(owners = {"geirst", "vekterli"}) default boolean asyncApplyBucketDiff() { return false; }
+        @ModelFeatureFlag(owners = {"arnej"}) default boolean ignoreThreadStackSizes() { return false; }
     }
 
     /** Warning: As elsewhere in this package, do not make backwards incompatible changes that will break old config models! */
diff --git a/config-model/src/main/java/com/yahoo/vespa/model/ConfigSentinel.java b/config-model/src/main/java/com/yahoo/vespa/model/ConfigSentinel.java
index c86cda0bbd1..217f6cff778 100644
--- a/config-model/src/main/java/com/yahoo/vespa/model/ConfigSentinel.java
+++ b/config-model/src/main/java/com/yahoo/vespa/model/ConfigSentinel.java
@@ -19,6 +19,7 @@ public class ConfigSentinel extends AbstractService implements SentinelConfig.Pr
 
     private final ApplicationId applicationId;
     private final Zone zone;
+    private final boolean ignoreRequestedStackSizes;
 
     /**
      * Constructs a new ConfigSentinel for the given host.
@@ -31,6 +32,7 @@ public class ConfigSentinel extends AbstractService implements SentinelConfig.Pr
         super(host, "sentinel");
         this.applicationId = applicationId;
         this.zone = zone;
+        this.ignoreRequestedStackSizes = featureFlags.ignoreThreadStackSizes();
         portsMeta.on(0).tag("rpc").tag("admin");
         portsMeta.on(1).tag("telnet").tag("interactive").tag("http").tag("state");
         setProp("clustertype", "hosts");
@@ -73,6 +75,7 @@ public class ConfigSentinel extends AbstractService implements SentinelConfig.Pr
     @Override
     public void getConfig(SentinelConfig.Builder builder) {
         builder.application(getApplicationConfig());
+        builder.ignoreRequestedStackSizes(ignoreRequestedStackSizes);
         for (Service s : getHostResource().getServices()) {
             if (s.getStartupCommand() != null) {
                 builder.service(getServiceConfig(s));
diff --git a/configd/src/apps/sentinel/manager.cpp b/configd/src/apps/sentinel/manager.cpp
index 631f63febd5..ffa9c9281fa 100644
--- a/configd/src/apps/sentinel/manager.cpp
+++ b/configd/src/apps/sentinel/manager.cpp
@@ -79,6 +79,11 @@ Manager::doConfigure()
     LOG_ASSERT(_env.configOwner().hasConfig());
     const SentinelConfig& config(_env.configOwner().getConfig());
 
+    if (config.ignoreRequestedStackSizes) {
+        setenv("VESPA_IGNORE_REQUESTED_STACK_SIZES", "true", 1);
+    } else {
+        unsetenv("VESPA_IGNORE_REQUESTED_STACK_SIZES");
+    }
     _env.rpcPort(config.port.rpc);
     _env.statePort(config.port.telnet);
 
diff --git a/configdefinitions/src/vespa/sentinel.def b/configdefinitions/src/vespa/sentinel.def
index cccf931938a..7bd4a000055 100644
--- a/configdefinitions/src/vespa/sentinel.def
+++ b/configdefinitions/src/vespa/sentinel.def
@@ -28,6 +28,9 @@ connectivity.minOkPercent int default=50
 ## Absolute number of nodes with confirmed network connectivity problems, maximum
 connectivity.maxBadCount int default=1
 
+## Ask low-level thread creation to honor requested stack size
+ignoreRequestedStackSizes bool default=false restart
+
 ## The command to run. This will be run by sh -c, and the following
 ## environment variables are defined: $ROOT, $VESPA_SERVICE_NAME,
 ## $VESPA_CONFIG_ID
diff --git a/configserver/src/main/java/com/yahoo/vespa/config/server/deploy/ModelContextImpl.java b/configserver/src/main/java/com/yahoo/vespa/config/server/deploy/ModelContextImpl.java
index 0b55c4bb53e..90c13364ac0 100644
--- a/configserver/src/main/java/com/yahoo/vespa/config/server/deploy/ModelContextImpl.java
+++ b/configserver/src/main/java/com/yahoo/vespa/config/server/deploy/ModelContextImpl.java
@@ -198,6 +198,7 @@ public class ModelContextImpl implements ModelContext {
         private final int maxUnCommittedMemory;
         private final boolean forwardIssuesAsErrors;
         private final boolean asyncApplyBucketDiff;
+        private final boolean ignoreThreadStackSizes;
 
         public FeatureFlags(FlagSource source, ApplicationId appId) {
             this.defaultTermwiseLimit = flagValue(source, appId, Flags.DEFAULT_TERM_WISE_LIMIT);
@@ -234,6 +235,7 @@ public class ModelContextImpl implements ModelContext {
             this.maxUnCommittedMemory = flagValue(source, appId, Flags.MAX_UNCOMMITTED_MEMORY);;
             this.forwardIssuesAsErrors = flagValue(source, appId, PermanentFlags.FORWARD_ISSUES_AS_ERRORS);
             this.asyncApplyBucketDiff = flagValue(source, appId, Flags.ASYNC_APPLY_BUCKET_DIFF);
+            this.ignoreThreadStackSizes = flagValue(source, appId, Flags.IGNORE_THREAD_STACK_SIZES);
         }
 
         @Override public double defaultTermwiseLimit() { return defaultTermwiseLimit; }
@@ -272,6 +274,7 @@ public class ModelContextImpl implements ModelContext {
         @Override public int maxUnCommittedMemory() { return maxUnCommittedMemory; }
         @Override public boolean forwardIssuesAsErrors() { return forwardIssuesAsErrors; }
         @Override public boolean asyncApplyBucketDiff() { return asyncApplyBucketDiff; }
+        @Override public boolean ignoreThreadStackSizes() { return ignoreThreadStackSizes; }
 
         private static <V> V flagValue(FlagSource source, ApplicationId appId, UnboundFlag<? extends V, ?, ?> flag) {
             return flag.bindTo(source)
diff --git a/fastos/src/vespa/fastos/unix_thread.cpp b/fastos/src/vespa/fastos/unix_thread.cpp
index d7a0ed879c0..a45d90426ef 100644
--- a/fastos/src/vespa/fastos/unix_thread.cpp
+++ b/fastos/src/vespa/fastos/unix_thread.cpp
@@ -71,7 +71,12 @@ bool FastOS_UNIX_Thread::Initialize (int stackSize, int stackGuardSize)
 #else
     adjusted_stack_size += PTHREAD_STACK_MIN;
 #endif
-    pthread_attr_setstacksize(&attr, adjusted_stack_size);
+    if (getenv("VESPA_IGNORE_REQUESTED_STACK_SIZES") == nullptr) {
+        //fprintf(stderr, "pthread_create: using adjusted stack size %zd\n", adjusted_stack_size);
+        pthread_attr_setstacksize(&attr, adjusted_stack_size);
+    } else {
+        //fprintf(stderr, "pthread_create: ignoring requested stack size %d\n", stackSize);
+    }
 
     rc = (0 == pthread_create(&_handle, &attr, FastOS_ThreadHook, this));
     if (rc)
diff --git a/flags/src/main/java/com/yahoo/vespa/flags/Flags.java b/flags/src/main/java/com/yahoo/vespa/flags/Flags.java
index 3cbdb9e1c56..74e0ab2c3d9 100644
--- a/flags/src/main/java/com/yahoo/vespa/flags/Flags.java
+++ b/flags/src/main/java/com/yahoo/vespa/flags/Flags.java
@@ -363,6 +363,13 @@ public class Flags {
             HOSTNAME,
             NODE_TYPE);
 
+    public static final UnboundBooleanFlag IGNORE_THREAD_STACK_SIZES = defineFeatureFlag(
+            "ignore-thread-stack-sizes", false,
+            List.of("arnej"), "2021-11-12", "2022-01-31",
+            "Whether C++ thread creation should ignore any requested stack size",
+            "Triggers restart, takes effect immediately",
+            ZONE_ID, APPLICATION_ID);
+
     public static final UnboundBooleanFlag USE_FILE_DISTRIBUTION_CONNECTION_POOL = defineFeatureFlag(
             "use-file-distribution-connection-pool", false,
             List.of("hmusum"), "2021-11-16", "2021-12-16",