pnfsmanager: inroduce limit on number of concurrent listing of the sa…

…me directory

Motivation
----------
When many jobs run listings of large directories in parallel
thread starvation occurs, the list queue gets filled, the message
queue gets filled and PnfsManager becomes unresponsive.

Modification:
-------------
Before dispatching list request to a queue count how many
list request are already in the list queue.
Add property:

pnfsmanager.limits.number-of-concurrent-dir-listings = infinity

that controls number of list requests in the PnfsManager list queue.
If this number is exceeded PnfsManager refuses to serve the request.

Result:
------
PnfsManager is more resilient against flood of concurrent list requests.

Patch: https://rb.dcache.org/r/14098/
Acked-by: Albert Rossi
Target: trunk
Request: 9.2
Require-notes: yes

modules/dcache-chimera/src/main/resources/diskCacheV111/namespace/pnfsmanager-chimera.xml

@@ -36,6 +36,8 @@
       <property name="logSlowThreshold" value="${pnfsmanager.limits.log-slow-threshold}"/>
       <property name="folding" value="${pnfsmanager.enable.folding}"/>
       <property name="useParallelListing" value="${pnfsmanager.enable.parallel-listing}"/>
+      <property name="maxListRequestsInQueue"
+                  value="#{T(org.dcache.util.Strings).parseInt('${pnfsmanager.limits.number-of-concurrent-dir-listings}')}"/>
       <property name="directoryListLimit" value="${pnfsmanager.limits.list-chunk-size}"/>
       <property name="permissionHandler" ref="permission-handler"/>
       <property name="queueMaxSize" value="${pnfsmanager.limits.queue-length}"/>

modules/dcache/src/main/java/diskCacheV111/namespace/PnfsManagerV3.java

@@ -227,6 +227,7 @@ public class PnfsManagerV3
 
     private boolean useParentHashOnCreate;
     private boolean useParallelListing;
+    private int maxListRequestsInQueue;
 
     /**
      * Whether to use folding.
@@ -343,6 +344,11 @@ public void setUseParallelListing(boolean useParallelListing) {
         this.useParallelListing = useParallelListing;
     }
 
+    @Required
+    public void setMaxListRequestsInQueue(int maxListRequestsInQueue) {
+        this.maxListRequestsInQueue = maxListRequestsInQueue;
+    }
+
     @Required
     public void setScheduledExecutor(ScheduledExecutorService executor) {
         scheduledExecutor = executor;
@@ -2810,21 +2816,40 @@ public void messageArrived(CellMessage envelope, PnfsListDirectoryMessage messag
           throws CacheException {
 
         String path = message.getPnfsPath();
+
         if (path == null) {
             throw new InvalidMessageCacheException("Missing PNFS id and path");
         }
 
+	/**
+	 * when useParallelListing is true, we only have 1 queue in the
+	 * list of queues below
+	 */
 	int index = 0;
 
 	if (!useParallelListing) {
 	    index = (int)(Math.abs((long)Objects.hashCode(path.toString())) % _listThreads);
 	}
+        BlockingQueue<CellMessage> queue = _listQueues[index];
 
-	/**
-	 * when useParallelListing is true, we only have 1 queue in the
-	 * list of queues below
-	 */
-        if (!_listQueues[index].offer(envelope)) {
+        /**
+         * Do counts only if maxListRequestsInQueue is enabled
+         */
+        if (maxListRequestsInQueue < Integer.MAX_VALUE) {
+            int counter = 0;
+            for (CellMessage i : queue) {
+                PnfsListDirectoryMessage msg = (PnfsListDirectoryMessage)i.getMessageObject();
+                if (msg.getPnfsPath().equals(path))  {
+                    if (counter > maxListRequestsInQueue) {
+                        LOGGER.warn("Too many list requests for the same directory {}  in PnfsManager queue", path);
+                        throw new MissingResourceCacheException("Too many list requests for the same directory in PndsManager queue");
+                    }
+                    counter += 1;
+                }
+            }
+        }
+
+        if (!queue.offer(envelope)) {
             throw new MissingResourceCacheException("PnfsManager queue limit exceeded");
         }
     }

skel/share/defaults/pnfsmanager.properties

@@ -75,14 +75,14 @@ pnfsmanager.limits.threads = ${pnfsmanager.limits.threads-per-group}
 #  ---- Thread displatch mechanisms
 #
 #   Experimental feature. Normally message processing is dispatched
-#   to the same thread in the thread pool associated with pnfsid (or path) 
-#   of the namespace entry contained in the messagee. On massive 
-#   uploads (create entries) to a single directory we observed 
+#   to the same thread in the thread pool associated with pnfsid (or path)
+#   of the namespace entry contained in the messagee. On massive
+#   uploads (create entries) to a single directory we observed
 #   performance degradation caused by undelrying db back-end
 #   synchronization when updating mtime and link count of the target
-#   directory. This leads to all available threads being busy/hanging 
-#   processing create entry messages denying other users from 
-#   accessing the namespace. The switch below, if enabled, would cause 
+#   directory. This leads to all available threads being busy/hanging
+#   processing create entry messages denying other users from
+#   accessing the namespace. The switch below, if enabled, would cause
 #   the create mesages to be  dispatched to a thread associated
 #   with that entry's parent (that is the target directory).
 #
@@ -126,7 +126,15 @@ pnfsmanager.limits.list-chunk-size = 100
 
 (one-of?true|false)pnfsmanager.enable.parallel-listing = false
 
-
+# ---- Determines how many simultaneous same directory listings to allow
+#
+#  If number of allowed simultaneous listings of the same directory
+#  reaches the limit defined below the PnfsManager will reject
+#  new listing requests for that directory until the number of already
+#  queued and active listing requests for that directory drops
+#  below this number.
+#
+pnfsmanager.limits.number-of-concurrent-dir-listings = infinity
 
 #  ---- Threshold for when to log slow requests
 #

skel/share/services/pnfsmanager.batch

@@ -10,6 +10,7 @@ check -strong pnfsmanager.enable.acl
 check -strong pnfsmanager.default-retention-policy
 check -strong pnfsmanager.default-access-latency
 check -strong pnfsmanager.enable.parallel-listing
+check -strong pnfsmanager.limits.number-of-concurrent-dir-listings
 check pnfsmanager.destination.flush-notification
 check pnfsmanager.destination.cache-notification
 check pnfsmanager.destination.cancel-upload-notification