sync: switch network fetch to multiprocessing

This avoids GIL limitations with using threads for parallel processing.

This reworks the fetch logic to return results for processing in the
main thread instead of leaving every thread to do its own processing.

We have to tweak the chunking logic a little here because multiprocessing
favors batching over returning immediate results when using a larger value
for chunksize.  When a single job can be quite slow, this tradeoff is not
good UX.

Bug: https://crbug.com/gerrit/12389
Change-Id: I0f0512d15ad7332d1eb28aff52c29d378acc9e1d
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/298642
Reviewed-by: Chris Mcdonald <cjmcdonald@google.com>
Tested-by: Mike Frysinger <vapier@google.com>

1 files changed, 91 insertions, 124 deletions

diff --git a/subcmds/sync.py b/subcmds/sync.py
index b14ad246..bf1369c0 100644
--- a/subcmds/sync.py
+++ b/subcmds/sync.py
@@ -45,11 +45,6 @@ except ImportError:
   def _rlimit_nofile():
     return (256, 256)
 
-try:
-  import multiprocessing
-except ImportError:
-  multiprocessing = None
-
 import event_log
 from git_command import GIT, git_require
 from git_config import GetUrlCookieFile
@@ -69,10 +64,6 @@ from manifest_xml import GitcManifest
 _ONE_DAY_S = 24 * 60 * 60
 
 
-class _FetchError(Exception):
-  """Internal error thrown in _FetchHelper() when we don't want stack trace."""
-
-
 class Sync(Command, MirrorSafeCommand):
   jobs = 1
   common = True
@@ -315,148 +306,119 @@ later is required to fix a server side protocol bug.
     self._ReloadManifest(manifest_path)
     return manifest_path
 
-  def _FetchProjectList(self, opt, projects, sem, *args, **kwargs):
-    """Main function of the fetch threads.
+  def _FetchProjectList(self, opt, projects):
+    """Main function of the fetch worker.
+
+    The projects we're given share the same underlying git object store, so we
+    have to fetch them in serial.
 
     Delegates most of the work to _FetchHelper.
 
     Args:
       opt: Program options returned from optparse.  See _Options().
       projects: Projects to fetch.
-      sem: We'll release() this semaphore when we exit so that another thread
-          can be started up.
-      *args, **kwargs: Remaining arguments to pass to _FetchHelper. See the
-          _FetchHelper docstring for details.
     """
-    try:
-        for project in projects:
-          success = self._FetchHelper(opt, project, *args, **kwargs)
-          if not success and opt.fail_fast:
-            break
-    finally:
-        sem.release()
+    return [self._FetchOne(opt, x) for x in projects]
 
-  def _FetchHelper(self, opt, project, lock, fetched, pm, err_event,
-                   clone_filter):
+  def _FetchOne(self, opt, project):
     """Fetch git objects for a single project.
 
     Args:
       opt: Program options returned from optparse.  See _Options().
       project: Project object for the project to fetch.
-      lock: Lock for accessing objects that are shared amongst multiple
-          _FetchHelper() threads.
-      fetched: set object that we will add project.gitdir to when we're done
-          (with our lock held).
-      pm: Instance of a Project object.  We will call pm.update() (with our
-          lock held).
-      err_event: We'll set this event in the case of an error (after printing
-          out info about the error).
-      clone_filter: Filter for use in a partial clone.
 
     Returns:
       Whether the fetch was successful.
     """
-    # We'll set to true once we've locked the lock.
-    did_lock = False
-
-    # Encapsulate everything in a try/except/finally so that:
-    # - We always set err_event in the case of an exception.
-    # - We always make sure we unlock the lock if we locked it.
     start = time.time()
     success = False
     buf = io.StringIO()
-    with lock:
-      pm.start(project.name)
     try:
-      try:
-        success = project.Sync_NetworkHalf(
-            quiet=opt.quiet,
-            verbose=opt.verbose,
-            output_redir=buf,
-            current_branch_only=self._GetCurrentBranchOnly(opt),
-            force_sync=opt.force_sync,
-            clone_bundle=opt.clone_bundle,
-            tags=opt.tags, archive=self.manifest.IsArchive,
-            optimized_fetch=opt.optimized_fetch,
-            retry_fetches=opt.retry_fetches,
-            prune=opt.prune,
-            clone_filter=clone_filter)
-        self._fetch_times.Set(project, time.time() - start)
-
-        # Lock around all the rest of the code, since printing, updating a set
-        # and Progress.update() are not thread safe.
-        lock.acquire()
-        did_lock = True
-
-        output = buf.getvalue()
-        if opt.verbose and output:
-          pm.update(inc=0, msg=output.rstrip())
-
-        if not success:
-          err_event.set()
-          print('error: Cannot fetch %s from %s'
-                % (project.name, project.remote.url),
-                file=sys.stderr)
-          if opt.fail_fast:
-            raise _FetchError()
-
-        fetched.add(project.gitdir)
-      except _FetchError:
-        pass
-      except Exception as e:
-        print('error: Cannot fetch %s (%s: %s)'
-              % (project.name, type(e).__name__, str(e)), file=sys.stderr)
-        err_event.set()
-        raise
-    finally:
-      if not did_lock:
-        lock.acquire()
-      pm.finish(project.name)
-      lock.release()
-      finish = time.time()
-      self.event_log.AddSync(project, event_log.TASK_SYNC_NETWORK,
-                             start, finish, success)
+      success = project.Sync_NetworkHalf(
+          quiet=opt.quiet,
+          verbose=opt.verbose,
+          output_redir=buf,
+          current_branch_only=self._GetCurrentBranchOnly(opt),
+          force_sync=opt.force_sync,
+          clone_bundle=opt.clone_bundle,
+          tags=opt.tags, archive=self.manifest.IsArchive,
+          optimized_fetch=opt.optimized_fetch,
+          retry_fetches=opt.retry_fetches,
+          prune=opt.prune,
+          clone_filter=self.manifest.CloneFilter)
+
+      output = buf.getvalue()
+      if opt.verbose and output:
+        print('\n' + output.rstrip())
+
+      if not success:
+        print('error: Cannot fetch %s from %s'
+              % (project.name, project.remote.url),
+              file=sys.stderr)
+    except Exception as e:
+      print('error: Cannot fetch %s (%s: %s)'
+            % (project.name, type(e).__name__, str(e)), file=sys.stderr)
+      raise
 
-    return success
+    finish = time.time()
+    return (success, project, start, finish)
 
   def _Fetch(self, projects, opt, err_event):
+    ret = True
+
     fetched = set()
-    lock = _threading.Lock()
-    pm = Progress('Fetching', len(projects))
+    pm = Progress('Fetching', len(projects), delay=False)
 
     objdir_project_map = dict()
     for project in projects:
       objdir_project_map.setdefault(project.objdir, []).append(project)
+    projects_list = list(objdir_project_map.values())
+
+    def _ProcessResults(results_sets):
+      ret = True
+      for results in results_sets:
+        for (success, project, start, finish) in results:
+          self._fetch_times.Set(project, finish - start)
+          self.event_log.AddSync(project, event_log.TASK_SYNC_NETWORK,
+                                 start, finish, success)
+          # Check for any errors before running any more tasks.
+          # ...we'll let existing jobs finish, though.
+          if not success:
+            ret = False
+          else:
+            fetched.add(project.gitdir)
+          pm.update(msg=project.name)
+        if not ret and opt.fail_fast:
+          break
+      return ret
 
-    threads = set()
-    sem = _threading.Semaphore(self.jobs)
-    for project_list in objdir_project_map.values():
-      # Check for any errors before running any more tasks.
-      # ...we'll let existing threads finish, though.
-      if err_event.is_set() and opt.fail_fast:
-        break
-
-      sem.acquire()
-      kwargs = dict(opt=opt,
-                    projects=project_list,
-                    sem=sem,
-                    lock=lock,
-                    fetched=fetched,
-                    pm=pm,
-                    err_event=err_event,
-                    clone_filter=self.manifest.CloneFilter)
-      if self.jobs > 1:
-        t = _threading.Thread(target=self._FetchProjectList,
-                              kwargs=kwargs)
-        # Ensure that Ctrl-C will not freeze the repo process.
-        t.daemon = True
-        threads.add(t)
-        t.start()
+    # NB: Multiprocessing is heavy, so don't spin it up for one job.
+    if len(projects_list) == 1 or opt.jobs == 1:
+      if not _ProcessResults(self._FetchProjectList(opt, x) for x in projects_list):
+        ret = False
+    else:
+      # Favor throughput over responsiveness when quiet.  It seems that imap()
+      # will yield results in batches relative to chunksize, so even as the
+      # children finish a sync, we won't see the result until one child finishes
+      # ~chunksize jobs.  When using a large --jobs with large chunksize, this
+      # can be jarring as there will be a large initial delay where repo looks
+      # like it isn't doing anything and sits at 0%, but then suddenly completes
+      # a lot of jobs all at once.  Since this code is more network bound, we
+      # can accept a bit more CPU overhead with a smaller chunksize so that the
+      # user sees more immediate & continuous feedback.
+      if opt.quiet:
+        chunksize = WORKER_BATCH_SIZE
       else:
-        self._FetchProjectList(**kwargs)
-
-    for t in threads:
-      t.join()
+        pm.update(inc=0, msg='warming up')
+        chunksize = 4
+      with multiprocessing.Pool(opt.jobs) as pool:
+        results = pool.imap_unordered(
+            functools.partial(self._FetchProjectList, opt),
+            projects_list,
+            chunksize=chunksize)
+        if not _ProcessResults(results):
+          ret = False
+          pool.close()
 
     pm.end()
     self._fetch_times.Save()
@@ -464,7 +426,7 @@ later is required to fix a server side protocol bug.
     if not self.manifest.IsArchive:
       self._GCProjects(projects, opt, err_event)
 
-    return fetched
+    return (ret, fetched)
 
   def _CheckoutOne(self, opt, project):
     """Checkout work tree for one project
@@ -514,7 +476,7 @@ later is required to fix a server side protocol bug.
         self.event_log.AddSync(project, event_log.TASK_SYNC_LOCAL,
                                start, finish, success)
         # Check for any errors before running any more tasks.
-        # ...we'll let existing threads finish, though.
+        # ...we'll let existing jobs finish, though.
         if not success:
           err_results.append(project.relpath)
           if opt.fail_fast:
@@ -894,7 +856,9 @@ later is required to fix a server side protocol bug.
       to_fetch.extend(all_projects)
       to_fetch.sort(key=self._fetch_times.Get, reverse=True)
 
-      fetched = self._Fetch(to_fetch, opt, err_event)
+      success, fetched = self._Fetch(to_fetch, opt, err_event)
+      if not success:
+        err_event.set()
 
       _PostRepoFetch(rp, opt.repo_verify)
       if opt.network_only:
@@ -923,7 +887,10 @@ later is required to fix a server side protocol bug.
         if previously_missing_set == missing_set:
           break
         previously_missing_set = missing_set
-        fetched.update(self._Fetch(missing, opt, err_event))
+        success, new_fetched = self._Fetch(to_fetch, opt, err_event)
+        if not success:
+          err_event.set()
+        fetched.update(new_fetched)
 
       # If we saw an error, exit with code 1 so that other scripts can check.
       if err_event.is_set():