Requête de recherche pour les bogues dans Apache Solr

Encore une fois, nous vérifions le produit Apache. Cette fois, nous avons choisi Solr, une plateforme de serveur de recherche open source. Solr vous permet de rechercher rapidement et efficacement des informations dans des bases de données et des ressources en ligne. Face à une tâche aussi complexe, il est facile de commettre des erreurs, même pour les développeurs Apache expérimentés. Dans cet article, nous examinerons ces types d'erreurs.

Qui es-tu, Solr ?

Il n'y a pas si longtemps, nous avons vérifié l'un des projets Apache les plus célèbres, l'EDI NetBeans. Nous avons trouvé de nombreux avertissements intéressants émis par notre analyseur lors du contrôle. D'ailleurs, les développeurs les ont rapidement remarqués et ont fait une pull request avant moi :) Cette fois-ci, nous avons décidé de jeter un œil à un autre de leurs gros produits, la plateforme de recherche en texte intégral Solr.

Introduit pour la première fois en 2006, Apache Solr offre un large éventail de fonctionnalités, du clustering dynamique et de l'intégration de bases de données au traitement de documents au format complexe. Ce moteur de recherche vous permet de rechercher et d'analyser des informations sur un site Web à grande vitesse et offre également la possibilité d'héberger des serveurs de recherche sur du matériel fonctionnant sous Linux. Solr a...

Pour faire court, je ne veux pas vous déranger avec trop de détails. Il suffit de savoir qu'il s'agit d'une plateforme logicielle pratique qui optimise le big data. Pourquoi ne pas examiner son code source et y rechercher quelque chose d'intéressant ou d'inhabituel ? C'est exactement ce que nous allons faire maintenant.

Ils ont mélangé quelque chose ici

Un programmeur n'est jamais à l'abri des fautes de frappe impliquant des opérateurs mélangés. En voici un :

public Map<String, List<String>> getIndexFilesPathForSnapshot(
    String collectionName, String snapshotName, String pathPrefix)
    throws SolrServerException, IOException {
  ....

  if (meta != null) {                  // <=
    throw new IllegalArgumentException(
                  "The snapshot named " + snapshotName +
                  " is not found for collection " + collectionName);
  }

  DocCollection collectionState = solrClient.getClusterState() 
                                            .getCollection(collectionName);
  for (Slice s : collectionState.getSlices()) {
    List<CoreSnapshotMetaData> replicaSnaps = 
                     meta.getReplicaSnapshotsForShard(s.getName());  // <=
    ....
  }
  return result;
}

Cette méthode utilise la variable meta pour stocker des informations sur les instantanés du système. Il y a aussi une vérification que meta != null dans le fragment de code ci-dessus. Si tel est le cas, alors IllegalArgumentException est levée. Cela seul semble bizarre. Bon, regardons maintenant le corps de la boucle. C'est ici que getReplicaSnapshotsForShard est appelé. Étant donné que meta ici est toujours nul, nous obtenons NullPointerException. Il semble que le développeur ait simplement fait une faute de frappe et confondu les opérateurs. Ainsi, une exception doit être levée si meta est égal à null.

L'analyseur PVS-Studio a fonctionné comme un correcteur d'épreuves et a signalé une faute de frappe détectée :

V6008 Déréférencement nul de 'méta'. SolrSnapshotsTool.java 262

Je voudrais ajouter que de telles erreurs avec des opérateurs mélangés sont plus courantes que vous ne le pensez. Par exemple, j'en ai déjà vu au moins deux similaires dans le projet NetBeans :

private SourcesModel getModel() {
  SourcesModel tm = model.get();
  if (tm == null) {
    tm.sourcePath.removePropertyChangeListener (this);
    tm.debugger.getSmartSteppingFilter ().
    removePropertyChangeListener (this);
  }
  return tm;
}

V6008 Déréférencement nul de 'tm'. SourcesModel.java 713

public void propertyChange(PropertyChangeEvent evt) {
  ....
  synchronized (this) {
    artifacts = null;
    if (listeners == null && listeners.isEmpty()) {
      return;
    }
    ....
  }
}

V6008 Déréférencement nul des 'auditeurs'. MavenArtifactsImplementation.java 613

Nous avons peut-être trouvé un nouveau modèle de faute de frappe. Nous continuerons à surveiller.

Passons au fragment suivant. Dans ce cours, un développeur a mélangé ce qu'il fallait retourner dans l'une des méthodes get.

public class FunctionQParser extends QParser {
  ....
  boolean parseMultipleSources = true;
  boolean parseToEnd = true;
  ....
  public void setParseMultipleSources(boolean parseMultipleSources) {
    this.parseMultipleSources = parseMultipleSources;
  }

  /** parse multiple comma separated value sources */
  public boolean getParseMultipleSources() {
    return parseMultipleSources;
  }

  public void setParseToEnd(boolean parseToEnd) {
    this.parseToEnd = parseToEnd;
  }

  /** throw exception if there is extra 
      stuff at the end of the parsed valuesource(s). */
  public boolean getParseToEnd() {
    return parseMultipleSources;
  }
  ....
}

Le cours est conçu pour analyser une certaine fonction mathématique écrite. Il existe deux propriétés pour modifier le comportement de l'analyseur : parseMultipleSources analyse toutes les sources de valeurs numériques, et parseToEnd vérifie si la fonction avec une chaîne doit être analysée jusqu'à la fin.

Regardons maintenant les méthodes get et set pour ces champs. Le champ parseMultipleSources est renvoyé dans getParseToEnd. Le programmeur a confondu le champ à renvoyer ici.

L'analyseur détecte facilement les champs renvoyés qui ne correspondent pas :

V6091 Implémentation du getter suspect. Le champ 'parseToEnd' devrait probablement être renvoyé à la place. FunctionQParser.java 87, FunctionQParser.java 57

La faute de frappe dans le fragment de code suivant peut entraîner une NullPointerException.

public void stringField(FieldInfo fieldInfo, String value) throws IOException {
  // trim the value if needed
  int len = value != null ? 
             UnicodeUtil.calcUTF16toUTF8Length(value, 0, value.length()) : 0;
  if (value.length() > maxLength) {               // <=    
    value = value.substring(0, maxLength);
  }
  countItem(fieldInfo.name, value, len);
}

Regardons de plus près : value est comparée à null d'abord, puis, dans la ligne suivante, la méthode length() est appelée valeur. Mais la variable peut être nulle ! Très probablement, le développeur aurait dû utiliser la variable len au lieu d'appeler length().

Nous avons trouvé cette faute de frappe grâce au message de l'analyseur :

V6008 Déréférencement nul potentiel de la « valeur ». IndexSizeEstimator.java 735, IndexSizeEstimator.java 736

Regardons un autre fragment de code avec une faute de frappe :

public Object doWork(Object value) throws IOException {
  ....
  List<?> list = (List<?>) value;
  // Validate all of same type and are comparable
  Object checkingObject = list.get(0);
  for (int idx = 0; idx < list.size(); ++idx) {
    Object item = list.get(0);                         // <=

    if (null == item) {
      throw new IOException(....);
    } else if (!(item instanceof Comparable<?>)) {
      throw new IOException(....);
    } else if (!item.getClass()
                    .getCanonicalName()
                    .equals(checkingObject.getClass()
                                          .getCanonicalName())) {
       throw new IOException(....);
    }
  }
  ....
}

Cela vaut la peine de prêter attention à la boucle for ici. Comme d'habitude, le programmeur déclare la boucle et la variable compteur idx, puis il obtient le nombre d'éléments dans la liste. Cependant, il y a un problème : chaque itération de la boucle ne prend que l'élément à l'index 0 : list.get(0).

L'analyseur PVS-Studio a détecté cette erreur :

V6016 Accès suspect à un élément de l'objet 'list' par un index constant à l'intérieur d'une boucle. AscEvaluator.java 56

The following example shows two methods with different names. However, they do the same thing.

private static List<Feature> makeFeatures(int[] featureIds) {
  final List<Feature> features = new ArrayList<>();
  for (final int i : featureIds) {
    Map<String, Object> params = new HashMap<String, Object>();
    params.put("value", i);
    final Feature f = Feature.getInstance(solrResourceLoader, 
                       ValueFeature.class.getName(), "f" + i, params);
    f.setIndex(i);
    features.add(f);
  }
  return features;
}

private static List<Feature> makeFilterFeatures(int[] featureIds) {
  final List<Feature> features = new ArrayList<>();
  for (final int i : featureIds) {
    Map<String, Object> params = new HashMap<String, Object>();
    params.put("value", i);
    final Feature f = Feature.getInstance(solrResourceLoader, 
                       ValueFeature.class.getName(), "f" + i, params);
    f.setIndex(i);
    features.add(f);
  }
  return features;
}

The first one creates a list of the Feature class objects. The second one, based on the name, should return a different type or filter these Features. If the FilterFeature type existed in the source code, we could assume that the developers simply made a typo. However, there's no such type. Maybe the method was copied and the developers forgot about it after copying it.

Anyway, this snippet looks very suspicious. And the analyzer proves this:

V6032 It is odd that the body of method 'makeFeatures' is fully equivalent to the body of another method 'makeFilterFeatures'. TestLTRScoringQuery.java 66, TestLTRScoringQuery.java 79

Forgot to check? Got error on track

If your gut tells you that null checks are unnecessary, don't trust it. In the code below, the "extra" check could've prevented NullPointerException.

public static Map<String, Object> postProcessCollectionJSON(
                                            Map<String, Object> collection) {
  final Map<String, Map<String, Object>> shards = collection != null   // <=
         ? (Map<String, Map<String, Object>>)
           collection.getOrDefault("shards", Collections.emptyMap())
         : Collections.emptyMap();
  final List<Health> healthStates = new ArrayList<>(shards.size());
  shards.forEach(
  ....
  );
  collection.put("health", Health.combine(healthStates).toString());   // <=
  return collection;
}

In the beginning of the method, the programmer checks if the collection reference is empty. If that's the case, then shards are derived from the collection. The most interesting thing is that healthStates is added to the collection at the end, regardless of whether the collection reference is empty or not.

Here's the analyzer warning for this code fragment:

V6008 Potential null dereference of 'collection'. ClusterStatus.java 303, ClusterStatus.java 335

And in the next example, the developers made an obvious mistake in the class constructor to support parallel distribution of thread work.

public class ParallelStream extends CloudSolrStream 
                            implements Expressible {
  ....
  private transient StreamFactory streamFactory;

  public ParallelStream(String zkHost, 
                        String collection, 
                        String expressionString, 
                        int workers, 
                        StreamComparator comp
) throws IOException {
    TupleStream tStream = this.streamFactory
                              .constructStream(expressionString);  // <=
    init(zkHost, collection, tStream, workers, comp);
  }  
  ....
}

The error lies in the first line of the constructor body. The streamFactory field is accessed here, but the field isn't initialized. The developers may have forgotten to add some logic in the constructor, or accidently may have written this line.

The PVS-Studio warning:

V6090 Field 'streamFactory' is being used before it was initialized. ParallelStream.java 61

However, they didn't forget to add a check in this method. Although, I think they put it in the wrong place.

private void createNewCollection(final String collection)
 throws InterruptedException {
  ....
  pending.add(completionService.submit(call));
  while (pending != null && pending.size() > 0) {
    Future<Object> future = completionService.take();
    if (future == null) return;
    pending.remove(future);
  }
}

Let's look at the interaction with the pending field: first the programmer called add, then they decided to make a loop in which they gradually removed elements from the method. The most interesting thing is that they checked that pending isn't null in the loop condition. It looks very suspicious, considering that there's no variable zeroing in the loop body. Seems like they should've added a check before calling the add method as well.

The analyzer warning:

V6060 The 'pending' reference was utilized before it was verified against null. AbstractBasicDistributedZkTestBase.java 1664, AbstractBasicDistributedZkTestBase.java 1665

Lost exception

Like many modern languages, Java has exception handling feature. The most important thing is to not lose them, as it happened here.

private void doSplitShardWithRule(SolrIndexSplitter.SplitMethod splitMethod) 
 throws Exception {
  ....
  try {
    ZkStateReader.from(cloudClient)
                 .waitForState(collectionName, 30, 
                             TimeUnit.SECONDS,
                             SolrCloudTestCase.activeClusterShape(1, 2));
  } catch (TimeoutException e) {
    new RuntimeException("Timeout waiting for
                          1shards and 2 replicas.", e);     // <=
  }
  ....
}

The error lies in the catch block: the developers created the RuntimeException object there, even added a link to the current intercepted TimeoutException and a message. But they forgot to write the throw keyword. So, the exception is never thrown.

The Lost and Found Bureau, in the form of our analyzer, found the lost exception and notified us about it:

V6006 The object was created but it is not being used. The 'throw' keyword could be missing. ShardSplitTest.java 773

How arithmetic errors interfere with testing

Does testing make software less buggy? Well, humans are the ones who write tests, and they can't help but make mistakes. This is what happened in the following example.

Public class SpellCheckCollatorTest extends SolrTestCaseJ4 {
  private static final int NUM_DOCS_WITH_TERM_EVERYOTHER = 8;
  private static final int NUM_DOCS = 17;
  ....
  @Test
  public void testEstimatedHitCounts() {
    ....
    for (int val = 5; val <= 20; val++) {
      String hitsXPath = xpathPrefix + "long[@name='hits']"; 

      if (val <= NUM_DOCS_WITH_TERM_EVERYOTHER) {
        int max = NUM_DOCS;
        int min = (/* min collected */ val) / 
                  (/* max docs possibly scanned */ NUM_DOCS);
        hitsXPath += "[" + min + " <= . and . <= " + max + "]";
      } 
    ....
    }
  }
  ....
}

A string containing the min variable, which in turn is the result of dividing val by NUM_DOCS, is written to hitsXPath here. Looking closer, you can see that the maximum and minimum values of val in this fragment are 8 and 5. The NUM_DOCS value is always 17. In all cases, min is zero in integer division. Most likely, the programmer forgot to convert division arguments to real numbers and change the type of the min variable.

We found this error using a brand-new diagnostic rule in the PVS-Studio analyzer:

V6113 The '(val) / (NUM_DOCS)' expression evaluates to 0 because the absolute value of the left operand 'val' is less than the value of the right operand 'NUM_DOCS'. SpellCheckCollatorTest.java 683

Danger of checking objects by reference

The class bellow describes the equalsTo comparison method.

private static class RandomQuery extends Query {
  private final long seed;
  private float density;
  private final List<BytesRef> docValues;
  ....
  private boolean equalsTo(RandomQuery other) {
    return seed == other.seed && 
           docValues == other.docValues && 
           density == other.density;
  }
}

Comparisons of the seed and density fields almost don't cause any questions (except, perhaps, for the density field that is a real number), because the values directly written into them are considered. However, since this field has a reference type, the docValues comparison via '==' looks very dubious. This check considers only the address and not the internal state of the object.

With such defect you can miss the case when two different lists store the same values because the lists are the same, but the references are different. It seems that when the developers named the equalsTo method, they hardly meant that it should compare references rather than the internal state of objects.

The PVS-Studio analyzer warning:

V6013 Objects 'docValues' and 'other.docValues' are compared by reference. Possibly an equality comparison was intended. TestFieldCacheSortRandom.java 341

Suspicious lack of synchronization

You won't find an error in the next fragment, but it's still potentially there. How's that possible? Take a look at this code and find out why.

public abstract class CachingDirectoryFactory extends DirectoryFactory {
  ....
  private static final Logger log = LoggerFactory.getLogger(....);
  protected Map<String, CacheValue> byPathCache = new HashMap<>();
  protected IdentityHashMap<Directory, CacheValue> byDirectoryCache = 
                                                  new IdentityHashMap<>();
  ....

  private void removeFromCache(CacheValue v) {
    log.debug("Removing from cache: {}", v);
    byDirectoryCache.remove(v.directory);
    byPathCache.remove(v.path);
  }
}

We won't be able to understand what's wrong here until we look at all the uses of the byDirectoryCache variable. In all other methods, the interaction occurs in the synchronized block. However, in the removeFromCache method, the programmer removes the collection elements outside of the synchronized blocks.

The analyzer detected this suspicious fragment:

V6102 Inconsistent synchronization of the 'byDirectoryCache' field. Consider synchronizing the field on all usages. CachingDirectoryFactory.java 92, CachingDirectoryFactory.java 228

At this point, one could say there's an error here, and the race condition could happen. However, it turns out that all removeFromCache calls are also enclosed in synchronized blocks. So, this is mostly a false positive that could be suppressed.

Although, we still can enhance this code, because there's a potential issue here. For example, when you need to use this method again, you may simply forget to enclose it in a synchronized block. Even though other methods have additional checks that the object exists in the byDirectoryCache collection, an unsynchronized call may delete an element already after the check. As a result, unnecessary actions are performed in another thread with a non-existent element of the collection, which can lead to errors in the program logic.

To protect ourselves from this, we can simply add the synchronized keyword to the removeFromCache method. So, even though there's no real error here, the static analyzer still urges us to write cleaner code.

By the way, we just recently released an article on the pitfalls of using synchronization.

Is it possible to create a few classes named the same?

In this fragment, the programmer didn't consider that classes can be renamed or declared with the same name in different packages.

private static String getFieldFlags(IndexableField f) {
  IndexOptions opts = (f == null) ? null : f.fieldType().indexOptions();

  StringBuilder flags = new StringBuilder();
  ....
  flags.append((f != null && f.getClass()
                              .getSimpleName()
                              .equals("LazyField"))  // <=
                                   ? FieldFlag.LAZY.getAbbreviation(): '-');
  ....
  return flags.toString();
}

This is an obviously unnecessary operation that may result in an error. The f variable has the getClass method called, which returns the object type, then gets and checks the name without specifying packages. All in all, there's no error here right now. However, it can arise for two reasons.

The first one is that there may be classes with the same name in different packages. In this case, it's unclear what kind of LazyField is required, and the program will run in a different way than intended.

The second one is related to changing the class name. If the name is changed, the code won't run as intended at all. And searching for all such strings in a huge code base is very difficult. Even if you resort to searching, it's something you can just forget about.

It'd be much safer to use the instanceof operator:

flags.append(f instanceof LazyDocument.LazyField
             ? FieldFlag.LAZY.getAbbreviation(): '-');

In this case, we wouldn't need to check for null, and the code would be much shorter. The chance of an error would also decrease, if there are classes with the same name in different packages, or if the name of the class changes.

The analyzer detected a potential error and issued a warning:

V6054 Classes should not be compared by their name. LukeRequestHandler.java 247

What about documentation, though?

As a final fragment, we'll look at the following code:

@Override
public UpdateCommand clone() {
  try {
    return (UpdateCommand) super.clone();
  } catch (CloneNotSupportedException e) {
    return null;                         // <=
  }
}

Let's see what's wrong with it, because everything seems fine at first glance. The analyzer informs us that returning null in the clone method is a bad idea:

V6073 It is not recommended to return null from 'clone' method. UpdateCommand.java 97

Why is it not recommended to return null from clone? It's time to consult the Java documentation:

Returns:

a clone of this instance.

Throws:

CloneNotSupportedException - if the object's class does not support the Cloneable interface. Subclasses that override the clone method can also throw this exception to indicate that an instance cannot be cloned.

The exception here indicates that the object can't be cloned. The method should return only a copy of the current object and nothing else. But why the analyzer doesn't recommend returning null from clone? It's all about further use of the code. If you constantly deviate from the recommendations in the documentation, it's difficult to catch non-standard situations.

Let's imagine a scenario where we want to use the UpdateCommand class, but the source code is unavailable, and we can't decompile it. Or we're just lazy. We can only use the built-in library with this class and focus on the interface. Our program needs us to use the clone method, so we write the following code:

try {
  UpdateCommand localCopy = field.clone(); 
  System.out.println(localCopy.toString();
} catch (CloneNotSupportedException e) {
  System.out.println("Could not clone the field"); 
}

In this code, we try to catch the CloneNotSupportedException, but we can't because the exception is a NullPointerException that causes the program to crash when calling localCopy.ToString(). This comes as a complete surprise to the developer. Deviating from official recommendations can be annoying, so it's better to always follow them

Conclusion

Let's stop here and take another look at the errors we found. Most of them are the result of carelessness, but there are some that require additional thought. For example, comparing class names without considering packages, or returning null instead of throwing an exception in the clone method.

Without special development tools like static analyzers, such bugs are difficult to find, especially in projects as large as Apache Solr. If you'd like to search for such non-obvious errors in your project, you may try our static analyzer here.

By the way, Solr isn't the only Apache product we checked:

• 21 bugs in 21st version of Apache NetBeans
• Big / Bug Data: analyzing the Apache Flink source code
• Apache Hadoop code quality: production vs test

Ce qui précède est le contenu détaillé de. pour plus d'informations, suivez d'autres articles connexes sur le site Web de PHP en chinois!