Synchronize on process factory to inhibit ETXTBSY

Refocus synchronization mechanism to cope with file descriptor set fork-
induced races to more tightly constrain concurrent fork/exec pairs. This
problem has been observed in bazel proper repeatedly, exhibiting as the
iconic ETXTBSY - Text file busy in wide worker pool builds and tests.

Evidence that this was discovered by @buchgr is in the comment and change
to the embedded ExecutionService implementation, and the description of
the race and the need for the synchronization was lifted from that scope
to the JavaSubprocessFactory. This factory is a singleton and represents
the gateway to all worker process execution, and serves as the correct
lock primitive to ensure that file descriptor sets are not duplicated
across forks, which gave rise to this issue.

To test this, I demonstrated a reproducer presented at
https://bugs.java.com/view_bug.do?bug_id=8068370
with 2.4% of invocations in that pathological case exhibiting the issue.
With a functionally equivalent change - synchronizing around a
processBuilder.start() call - as the only modification to the reproducer,
no further failures of any kind were observed, over several hundred runs.

Closes #5556.

PiperOrigin-RevId: 203947224

1 files changed, 4 insertions, 22 deletions

diff --git a/src/tools/remote/src/main/java/com/google/devtools/build/remote/worker/ExecutionServer.java b/src/tools/remote/src/main/java/com/google/devtools/build/remote/worker/ExecutionServer.java
index f9506987f6..19556dc9bd 100644
--- a/src/tools/remote/src/main/java/com/google/devtools/build/remote/worker/ExecutionServer.java
+++ b/src/tools/remote/src/main/java/com/google/devtools/build/remote/worker/ExecutionServer.java
@@ -76,7 +76,6 @@ import java.util.logging.Logger;
 final class ExecutionServer extends ExecutionImplBase {
   private static final Logger logger = Logger.getLogger(ExecutionServer.class.getName());
 
-  private final Object lock = new Object();
 
   // The name of the container image entry in the Platform proto
   // (see third_party/googleapis/devtools/remoteexecution/*/remote_execution.proto and
@@ -215,27 +214,10 @@ final class ExecutionServer extends ExecutionImplBase {
     CommandResult cmdResult = null;
 
     FutureCommandResult futureCmdResult = null;
-    synchronized (lock) {
-      // Linux does not provide a safe API for a multi-threaded program to fork a subprocess.
-      // Consider the case where two threads both write an executable file and then try to execute
-      // it. It can happen that the first thread writes its executable file, with the file
-      // descriptor still being open when the second thread forks, with the fork inheriting a copy
-      // of the file descriptor. Then the first thread closes the original file descriptor, and
-      // proceeds to execute the file. At that point Linux sees an open file descriptor to the file
-      // and returns ETXTBSY (Text file busy) as an error. This race is inherent in the fork / exec
-      // duality, with fork always inheriting a copy of the file descriptor table; if there was a
-      // way to fork without copying the entire file descriptor table (e.g., only copy specific
-      // entries), we could avoid this race.
-      //
-      // I was able to reproduce this problem reliably by running significantly more threads than
-      // there are CPU cores on my workstation - the more threads the more likely it happens.
-      //
-      // As a workaround, we put a synchronized block around the fork.
-      try {
-        futureCmdResult = cmd.executeAsync();
-      } catch (CommandException e) {
-        Throwables.throwIfInstanceOf(e.getCause(), IOException.class);
-      }
+    try {
+      futureCmdResult = cmd.executeAsync();
+    } catch (CommandException e) {
+      Throwables.throwIfInstanceOf(e.getCause(), IOException.class);
     }
 
     if (futureCmdResult != null) {