Light, fluffy, and always free - The AWS Local Emulator alternative
fix(codeartifact): close the remaining register-vs-reset gap in PerKeyContainerPool (#4447)
Priamo Germosen3da68ff50506
diff --git a/src/main/java/io/github/hectorvent/floci/core/common/docker/PerKeyContainerPool.java b/src/main/java/io/github/hectorvent/floci/core/common/docker/PerKeyContainerPool.javaindex e17c6edfe7..bed495191b 100644--- a/src/main/java/io/github/hectorvent/floci/core/common/docker/PerKeyContainerPool.java+++ b/src/main/java/io/github/hectorvent/floci/core/common/docker/PerKeyContainerPool.java@@ -76,17 +76,26 @@ public class PerKeyContainerPool {}long startedGeneration = generation.get();StartedContainer started = starter.start();+ // Registered before checking the generation, not after: checking first and only then+ // registering leaves a gap between the two where a concurrent stopAll() can run and+ // see nothing (this container isn't in the map yet) without the generation bump alone+ // catching it either, since that only tells this call a reset happened, not whether+ // stopAll()'s own sweep already passed the point where it could have found this+ // container. Registering first means stopAll() either sees it in the map and stops it,+ // or runs entirely before this put and gets caught by the generation check below;+ // there is no window where neither happens.+ containers.put(key, started);if (generation.get() != startedGeneration) {- // A stopAll() ran while this container was being built and could not have seen- // it: it was not in the map yet. Stop it now rather than register it, so a reset- // racing a first-use does not leave this one running behind.+ // A stopAll() ran while this container was being built and may have already swept+ // past this key (its forEach found nothing, since the put above hadn't happened+ // yet) before clearing the map out from under this registration. Remove and stop+ // it directly rather than trust the map state, which that same stopAll() may have+ // already cleared.LOG.warnv("Sidecar pool was reset while starting a container for key {0}; stopping it", key);+ containers.remove(key);lifecycleManager.stopAndRemove(started.containerId(), null);throw new IllegalStateException("Sidecar pool was reset while starting a container for key " + key);}- // Tracked before the health wait, not after: a timeout below must not stop- // stopContainer()/stopAll() from ever finding this already-running container again.- containers.put(key, started);SidecarHealthHelper.waitForHealth(started.baseUrl(), healthPath);return started.baseUrl();}diff --git a/src/main/java/io/github/hectorvent/floci/services/codeartifact/VerdaccioSidecarManager.java b/src/main/java/io/github/hectorvent/floci/services/codeartifact/VerdaccioSidecarManager.javaindex a80c3baf4d..4fbc109fbc 100644--- a/src/main/java/io/github/hectorvent/floci/services/codeartifact/VerdaccioSidecarManager.java+++ b/src/main/java/io/github/hectorvent/floci/services/codeartifact/VerdaccioSidecarManager.java@@ -38,8 +38,8 @@ import java.nio.charset.StandardCharsets;* <p>Implements {@link ContainerTeardown} rather than observing {@code ShutdownEvent} directly:* {@code ContainerTeardowns.stopAll} already runs every implementation both at process shutdown* and on {@code /state/reset}/{@code /state/nuke}, so this is what actually stops every- * repository's container on reset, not the generic {@code CodeArtifactService} repository-record- * walk, which only knows about repositories that still exist in storage.+ * repository's container on reset, driven by what {@link #pool} is really tracking rather than by+ * which repository records still exist in storage.*/@ApplicationScopedpublic class VerdaccioSidecarManager implements RepositorySidecarManager, ContainerTeardown {