sync: fix connection error on macOS

With a large number of sync workers, the sync process may fail on
macOS due to connection errors. The root cause is that multiple
workers may attempt to connect to the multiprocessing manager server
at the same time when handling the first job. This can lead to
connection failures if there are too many pending connections, exceeding
the socket listening backlog.

Bug: 377538810
Change-Id: I1924d318d076ca3be61d75daa37bfa8d7dc23ed7
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/441541
Tested-by: Josip Sokcevic <sokcevic@google.com>
Commit-Queue: Josip Sokcevic <sokcevic@google.com>
Reviewed-by: Josip Sokcevic <sokcevic@google.com>

project.py

@@ -3375,24 +3375,29 @@ class Project:
             setting = fp.read()
             assert setting.startswith("gitdir:")
             git_worktree_path = setting.split(":", 1)[1].strip()
-        # Some platforms (e.g. Windows) won't let us update dotgit in situ
-        # because of file permissions.  Delete it and recreate it from scratch
-        # to avoid.
-        platform_utils.remove(dotgit)
-        # Use relative path from checkout->worktree & maintain Unix line endings
-        # on all OS's to match git behavior.
-        with open(dotgit, "w", newline="\n") as fp:
-            print(
-                "gitdir:",
-                os.path.relpath(git_worktree_path, self.worktree),
-                file=fp,
-            )
-        # Use relative path from worktree->checkout & maintain Unix line endings
-        # on all OS's to match git behavior.
-        with open(
-            os.path.join(git_worktree_path, "gitdir"), "w", newline="\n"
-        ) as fp:
-            print(os.path.relpath(dotgit, git_worktree_path), file=fp)
+
+        # `gitdir` maybe be either relative or absolute depending on the
+        # behavior of the local copy of git, so only convert the path to
+        # relative if it needs to be converted.
+        if os.path.isabs(git_worktree_path):
+            # Some platforms (e.g. Windows) won't let us update dotgit in situ
+            # because of file permissions.  Delete it and recreate it from
+            # scratch to avoid.
+            platform_utils.remove(dotgit)
+            # Use relative path from checkout->worktree & maintain Unix line
+            # endings on all OS's to match git behavior.
+            with open(dotgit, "w", newline="\n") as fp:
+                print(
+                    "gitdir:",
+                    os.path.relpath(git_worktree_path, self.worktree),
+                    file=fp,
+                )
+            # Use relative path from worktree->checkout & maintain Unix line
+            # endings on all OS's to match git behavior.
+            with open(
+                os.path.join(git_worktree_path, "gitdir"), "w", newline="\n"
+            ) as fp:
+                print(os.path.relpath(dotgit, git_worktree_path), file=fp)
 
         self._InitMRef()
 

subcmds/forall.py

@@ -279,7 +279,7 @@ without iterating through the remaining projects.
                 rc = rc or r
                 if r != 0 and opt.abort_on_errors:
                     raise Exception("Aborting due to previous error")
-                return rc
+            return rc
 
         try:
             config = self.manifest.manifestProject.config
@@ -298,7 +298,7 @@ without iterating through the remaining projects.
                 )
         except (KeyboardInterrupt, WorkerKeyboardInterrupt):
             # Catch KeyboardInterrupt raised inside and outside of workers
-            rc = rc or errno.EINTR
+            rc = errno.EINTR
         except Exception as e:
             # Catch any other exceptions raised
             logger.error(
@@ -306,7 +306,7 @@ without iterating through the remaining projects.
                 type(e).__name__,
                 e,
             )
-            rc = rc or getattr(e, "errno", 1)
+            rc = getattr(e, "errno", 1)
         if rc != 0:
             sys.exit(rc)
 

subcmds/sync.py

@@ -821,6 +821,16 @@ later is required to fix a server side protocol bug.
         jobs = jobs_str(len(items))
         return f"{jobs} | {elapsed_str(elapsed)} {earliest_proj}"
 
+    @classmethod
+    def InitWorker(cls):
+        # Force connect to the manager server now.
+        # This is good because workers are initialized one by one. Without this,
+        # multiple workers may connect to the manager when handling the first
+        # job at the same time. Then the connection may fail if too many
+        # connections are pending and execeeded the socket listening backlog,
+        # especially on MacOS.
+        len(cls.get_parallel_context()["sync_dict"])
+
     def _Fetch(self, projects, opt, err_event, ssh_proxy, errors):
         ret = True
 
@@ -913,6 +923,7 @@ later is required to fix a server side protocol bug.
                     # idle while other workers still have more than one job in
                     # their chunk queue.
                     chunksize=1,
+                    initializer=self.InitWorker,
                 )
             finally:
                 sync_event.set()

subcmds/upload.py

@@ -737,8 +737,8 @@ Gerrit Code Review:  https://www.gerritcodereview.com/
             pending = []
             for result in results:
                 project_idx, avail = result
+                project = projects[project_idx]
                 if avail is None:
-                    project = projects[project_idx]
                     logger.error(
                         'repo: error: %s: Unable to upload branch "%s". '
                         "You might be able to fix the branch by running:\n"
@@ -748,7 +748,7 @@ Gerrit Code Review:  https://www.gerritcodereview.com/
                         project.manifest.branch,
                     )
                 elif avail:
-                    pending.append(result)
+                    pending.append((project, avail))
             return pending
 
         with self.ParallelContext():