git-repo/subcmds/sync.py

2251 lines
80 KiB
Python
Raw Normal View History

2008-10-21 14:00:00 +00:00
# Copyright (C) 2008 The Android Open Source Project
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import collections
import functools
import http.cookiejar as cookielib
import io
import json
import multiprocessing
import netrc
import optparse
2008-10-21 14:00:00 +00:00
import os
from pathlib import Path
import sys
import tempfile
import time
from typing import List, NamedTuple, Set, Union
import urllib.error
import urllib.parse
import urllib.request
import xml.parsers.expat
import xmlrpc.client
2008-10-21 14:00:00 +00:00
try:
import threading as _threading
except ImportError:
import dummy_threading as _threading
try:
import resource
def _rlimit_nofile():
return resource.getrlimit(resource.RLIMIT_NOFILE)
except ImportError:
def _rlimit_nofile():
return (256, 256)
from command import Command
from command import DEFAULT_LOCAL_JOBS
from command import MirrorSafeCommand
from command import WORKER_BATCH_SIZE
from error import GitError
from error import RepoChangedException
from error import RepoError
from error import RepoExitError
from error import RepoUnhandledExceptionError
from error import SyncError
from error import UpdateManifestError
import event_log
sync: rework selfupdate logic The current logic has a downside in that it doesn't sync to the latest signed version available if the latest commit itself is unsigned. This can come up when using the "main" branch as it is sometimes signed, but often not as it's holding the latest merged commits. When people use the main branch, it's to get early testing on versions tagged but not yet released, and we don't want them to get stuck indefinitely on that old version of repo. For example, this series of events: * "stable" is at v2.12. * "main" is tagged with v2.13. * early testers use --repo-rev main to get v2.13. * new commits are merged to "main". * "main" is tagged with v2.14. * new commits are merged to "main". * devs who had synced in the past to test v2.13 are stuck on v2.13. repo sees "main" is unsigned and so doesn't try to upgrade at all. The only way to get unwedged is to re-run `repo init --repo-rev main`, or to manually sync once with repo verification disabled, or for us to leave "main" signed for a while and hope devs will sync in that window. The new logic is that whenever changes are available, we switch to the latest signed tag. We also replace some of the duplicated verification code in the sync command with the newer wrapper logic. This handles a couple of important scenarios inaddition to above: * rollback (e.g. v2.13.8 -> v2.13.7) * do not trash uncommitted changes (in case of ad-hoc testing) * switch tag histories (e.g. v2.13.8 -> v2.13.8-cr1) Change-Id: I5b45ba1dd26a7c582700ee3711f303dc7538579b Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/300122 Reviewed-by: Jonathan Nieder <jrn@google.com> Reviewed-by: Michael Mortensen <mmortensen@google.com> Tested-by: Mike Frysinger <vapier@google.com>
2021-03-15 18:58:52 +00:00
from git_command import git_require
from git_config import GetUrlCookieFile
from git_refs import HEAD
from git_refs import R_HEADS
import git_superproject
import platform_utils
from progress import elapsed_str
from progress import jobs_str
from progress import Progress
from project import DeleteWorktreeError
from project import Project
from project import RemoteSpec
from project import SyncBuffer
from repo_logging import RepoLogger
from repo_trace import Trace
import ssh
from wrapper import Wrapper
2008-10-21 14:00:00 +00:00
_ONE_DAY_S = 24 * 60 * 60
_REPO_ALLOW_SHALLOW = os.environ.get("REPO_ALLOW_SHALLOW")
logger = RepoLogger(__file__)
def _SafeCheckoutOrder(checkouts: List[Project]) -> List[List[Project]]:
"""Generate a sequence of checkouts that is safe to perform. The client
should checkout everything from n-th index before moving to n+1.
This is only useful if manifest contains nested projects.
E.g. if foo, foo/bar and foo/bar/baz are project paths, then foo needs to
finish before foo/bar can proceed, and foo/bar needs to finish before
foo/bar/baz."""
res = [[]]
current = res[0]
# depth_stack contains a current stack of parent paths.
depth_stack = []
# Checkouts are iterated in the hierarchical order. That way, it can easily
# be determined if the previous checkout is parent of the current checkout.
# We are splitting by the path separator so the final result is
# hierarchical, and not just lexicographical. For example, if the projects
# are: foo, foo/bar, foo-bar, lexicographical order produces foo, foo-bar
# and foo/bar, which doesn't work.
for checkout in sorted(checkouts, key=lambda x: x.relpath.split("/")):
checkout_path = Path(checkout.relpath)
while depth_stack:
try:
checkout_path.relative_to(depth_stack[-1])
except ValueError:
# Path.relative_to returns ValueError if paths are not relative.
# TODO(sokcevic): Switch to is_relative_to once min supported
# version is py3.9.
depth_stack.pop()
else:
if len(depth_stack) >= len(res):
# Another depth created.
res.append([])
break
current = res[len(depth_stack)]
current.append(checkout)
depth_stack.append(checkout_path)
return res
def _chunksize(projects: int, jobs: int) -> int:
"""Calculate chunk size for the given number of projects and jobs."""
return min(max(1, projects // jobs), WORKER_BATCH_SIZE)
class _FetchOneResult(NamedTuple):
"""_FetchOne return value.
Attributes:
success (bool): True if successful.
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
project_idx (int): The fetched project index.
start (float): The starting time.time().
finish (float): The ending time.time().
remote_fetched (bool): True if the remote was actually queried.
"""
success: bool
errors: List[Exception]
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
project_idx: int
start: float
finish: float
remote_fetched: bool
class _FetchResult(NamedTuple):
"""_Fetch return value.
Attributes:
success (bool): True if successful.
projects (Set[str]): The names of the git directories of fetched projects.
"""
success: bool
projects: Set[str]
class _FetchMainResult(NamedTuple):
"""_FetchMain return value.
Attributes:
all_projects (List[Project]): The fetched projects.
"""
all_projects: List[Project]
class _CheckoutOneResult(NamedTuple):
"""_CheckoutOne return value.
Attributes:
success (bool): True if successful.
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
project_idx (int): The project index.
start (float): The starting time.time().
finish (float): The ending time.time().
"""
success: bool
errors: List[Exception]
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
project_idx: int
start: float
finish: float
class SuperprojectError(SyncError):
"""Superproject sync repo."""
class SyncFailFastError(SyncError):
"""Sync exit error when --fail-fast set."""
class SmartSyncError(SyncError):
"""Smart sync exit error."""
class ManifestInterruptError(RepoError):
"""Aggregate Error to be logged when a user interrupts a manifest update."""
def __init__(self, output, **kwargs):
super().__init__(output, **kwargs)
self.output = output
def __str__(self):
error_type = type(self).__name__
return f"{error_type}:{self.output}"
class TeeStringIO(io.StringIO):
"""StringIO class that can write to an additional destination."""
def __init__(
self, io: Union[io.TextIOWrapper, None], *args, **kwargs
) -> None:
super().__init__(*args, **kwargs)
self.io = io
def write(self, s: str) -> int:
"""Write to additional destination."""
ret = super().write(s)
if self.io is not None:
self.io.write(s)
return ret
class Sync(Command, MirrorSafeCommand):
COMMON = True
MULTI_MANIFEST_SUPPORT = True
helpSummary = "Update working tree to the latest revision"
helpUsage = """
2008-10-21 14:00:00 +00:00
%prog [<project>...]
"""
helpDescription = """
2008-10-21 14:00:00 +00:00
The '%prog' command synchronizes local project directories
with the remote repositories specified in the manifest. If a local
project does not yet exist, it will clone a new local directory from
the remote repository and set up tracking branches as specified in
the manifest. If the local project already exists, '%prog'
will update the remote branches and rebase any new local changes
on top of the new remote changes.
'%prog' will synchronize all projects listed at the command
line. Projects can be specified either by name, or by a relative
or absolute path to the project's local directory. If no projects
are specified, '%prog' will synchronize all projects listed in
the manifest.
The -d/--detach option can be used to switch specified projects
back to the manifest revision. This option is especially helpful
if the project is currently on a topic branch, but the manifest
revision is temporarily needed.
The -s/--smart-sync option can be used to sync to a known good
build as specified by the manifest-server element in the current
manifest. The -t/--smart-tag option is similar and allows you to
specify a custom tag/label.
The -u/--manifest-server-username and -p/--manifest-server-password
options can be used to specify a username and password to authenticate
with the manifest server when using the -s or -t option.
If -u and -p are not specified when using the -s or -t option, '%prog'
will attempt to read authentication credentials for the manifest server
from the user's .netrc file.
'%prog' will not use authentication credentials from -u/-p or .netrc
if the manifest server specified in the manifest file already includes
credentials.
By default, all projects will be synced. The --fail-fast option can be used
to halt syncing as soon as possible when the first project fails to sync.
The --force-sync option can be used to overwrite existing git
directories if they have previously been linked to a different
object directory. WARNING: This may cause data to be lost since
refs may be removed when overwriting.
The --force-checkout option can be used to force git to switch revs even if the
index or the working tree differs from HEAD, and if there are untracked files.
WARNING: This may cause data to be lost since uncommitted changes may be
removed.
The --force-remove-dirty option can be used to remove previously used
projects with uncommitted changes. WARNING: This may cause data to be
lost since uncommitted changes may be removed with projects that no longer
exist in the manifest.
The --no-clone-bundle option disables any attempt to use
$URL/clone.bundle to bootstrap a new Git repository from a
resumeable bundle file on a content delivery network. This
may be necessary if there are problems with the local Python
HTTP client or proxy configuration, but the Git binary works.
Represent git-submodule as nested projects, take 2 (Previous submission of this change broke Android buildbot due to incorrect regular expression for parsing git-config output. During investigation, we also found that Android, which pulls Chromium, has a workaround for Chromium's submodules; its manifest includes Chromium's submodules. This new change, in addition to fixing the regex, also take this type of workarounds into consideration; it adds a new attribute that makes repo not fetch submodules unless submodules have a project element defined in the manifest, or this attribute is overridden by a parent project element or by the default element.) We need a representation of git-submodule in repo; otherwise repo will not sync submodules, and leave workspace in a broken state. Of course this will not be a problem if all projects are owned by the owner of the manifest file, who may simply choose not to use git-submodule in all projects. However, this is not possible in practice because manifest file owner is unlikely to own all upstream projects. As git submodules are simply git repositories, it is natural to treat them as plain repo projects that live inside a repo project. That is, we could use recursively declared projects to denote the is-submodule relation of git repositories. The behavior of repo remains the same to projects that do not have a sub-project within. As for parent projects, repo fetches them and their sub-projects as normal projects, and then checks out subprojects at the commit specified in parent's commit object. The sub-project is fetched at a path relative to parent project's working directory; so the path specified in manifest file should match that of .gitmodules file. If a submodule is not registered in repo manifest, repo will derive its properties from itself and its parent project, which might not always be correct. In such cases, the subproject is called a derived subproject. To a user, a sub-project is merely a git-submodule; so all tips of working with a git-submodule apply here, too. For example, you should not run `repo sync` in a parent repository if its submodule is dirty. Change-Id: I4b8344c1b9ccad2f58ad304573133e5d52e1faef
2012-01-11 03:28:42 +00:00
The --fetch-submodules option enables fetching Git submodules
of a project from server.
The -c/--current-branch option can be used to only fetch objects that
are on the branch specified by a project's revision.
The --optimized-fetch option can be used to only fetch projects that
are fixed to a sha1 revision if the sha1 revision does not already
exist locally.
The --prune option can be used to remove any refs that no longer
exist on the remote.
The --auto-gc option can be used to trigger garbage collection on all
projects. By default, repo does not run garbage collection.
# SSH Connections
If at least one project remote URL uses an SSH connection (ssh://,
git+ssh://, or user@host:path syntax) repo will automatically
enable the SSH ControlMaster option when connecting to that host.
This feature permits other projects in the same '%prog' session to
reuse the same SSH tunnel, saving connection setup overheads.
To disable this behavior on UNIX platforms, set the GIT_SSH
environment variable to 'ssh'. For example:
export GIT_SSH=ssh
%prog
# Compatibility
This feature is automatically disabled on Windows, due to the lack
of UNIX domain socket support.
This feature is not compatible with url.insteadof rewrites in the
user's ~/.gitconfig. '%prog' is currently not able to perform the
rewrite early enough to establish the ControlMaster tunnel.
If the remote SSH daemon is Gerrit Code Review, version 2.0.10 or
later is required to fix a server side protocol bug.
2008-10-21 14:00:00 +00:00
"""
# A value of 0 means we want parallel jobs, but we'll determine the default
# value later on.
PARALLEL_JOBS = 0
def _Options(self, p, show_smart=True):
p.add_option(
"--jobs-network",
default=None,
type=int,
metavar="JOBS",
help="number of network jobs to run in parallel (defaults to "
"--jobs or 1)",
)
p.add_option(
"--jobs-checkout",
default=None,
type=int,
metavar="JOBS",
help="number of local checkout jobs to run in parallel (defaults "
f"to --jobs or {DEFAULT_LOCAL_JOBS})",
)
p.add_option(
"-f",
"--force-broken",
dest="force_broken",
action="store_true",
help="obsolete option (to be deleted in the future)",
)
p.add_option(
"--fail-fast",
dest="fail_fast",
action="store_true",
help="stop syncing after first error is hit",
)
p.add_option(
"--force-sync",
dest="force_sync",
action="store_true",
help="overwrite an existing git directory if it needs to "
"point to a different object directory. WARNING: this "
"may cause loss of data",
)
p.add_option(
"--force-checkout",
dest="force_checkout",
action="store_true",
help="force checkout even if it results in throwing away "
"uncommitted modifications. "
"WARNING: this may cause loss of data",
)
p.add_option(
"--force-remove-dirty",
dest="force_remove_dirty",
action="store_true",
help="force remove projects with uncommitted modifications if "
"projects no longer exist in the manifest. "
"WARNING: this may cause loss of data",
)
p.add_option(
"--rebase",
dest="rebase",
action="store_true",
help="rebase local commits regardless of whether they are "
"published",
)
p.add_option(
"-l",
"--local-only",
dest="local_only",
action="store_true",
help="only update working tree, don't fetch",
)
p.add_option(
"--no-manifest-update",
"--nmu",
dest="mp_update",
action="store_false",
default="true",
help="use the existing manifest checkout as-is. "
"(do not update to the latest revision)",
)
p.add_option(
"-n",
"--network-only",
dest="network_only",
action="store_true",
help="fetch only, don't update working tree",
)
p.add_option(
"-d",
"--detach",
dest="detach_head",
action="store_true",
help="detach projects back to manifest revision",
)
p.add_option(
"-c",
"--current-branch",
dest="current_branch_only",
action="store_true",
help="fetch only current branch from server",
)
p.add_option(
"--no-current-branch",
dest="current_branch_only",
action="store_false",
help="fetch all branches from server",
)
p.add_option(
"-m",
"--manifest-name",
dest="manifest_name",
help="temporary manifest to use for this sync",
metavar="NAME.xml",
)
p.add_option(
"--clone-bundle",
action="store_true",
help="enable use of /clone.bundle on HTTP/HTTPS",
)
p.add_option(
"--no-clone-bundle",
dest="clone_bundle",
action="store_false",
help="disable use of /clone.bundle on HTTP/HTTPS",
)
p.add_option(
"-u",
"--manifest-server-username",
action="store",
dest="manifest_server_username",
help="username to authenticate with the manifest server",
)
p.add_option(
"-p",
"--manifest-server-password",
action="store",
dest="manifest_server_password",
help="password to authenticate with the manifest server",
)
p.add_option(
"--fetch-submodules",
dest="fetch_submodules",
action="store_true",
help="fetch submodules from server",
)
p.add_option(
"--use-superproject",
action="store_true",
help="use the manifest superproject to sync projects; implies -c",
)
p.add_option(
"--no-use-superproject",
action="store_false",
dest="use_superproject",
help="disable use of manifest superprojects",
)
p.add_option("--tags", action="store_true", help="fetch tags")
p.add_option(
"--no-tags",
dest="tags",
action="store_false",
help="don't fetch tags (default)",
)
p.add_option(
"--optimized-fetch",
dest="optimized_fetch",
action="store_true",
help="only fetch projects fixed to sha1 if revision does not exist "
"locally",
)
p.add_option(
"--retry-fetches",
default=0,
action="store",
type="int",
help="number of times to retry fetches on transient errors",
)
p.add_option(
"--prune",
action="store_true",
help="delete refs that no longer exist on the remote (default)",
)
p.add_option(
"--no-prune",
dest="prune",
action="store_false",
help="do not delete refs that no longer exist on the remote",
)
p.add_option(
"--auto-gc",
action="store_true",
default=None,
help="run garbage collection on all synced projects",
)
p.add_option(
"--no-auto-gc",
dest="auto_gc",
action="store_false",
help="do not run garbage collection on any projects (default)",
)
if show_smart:
p.add_option(
"-s",
"--smart-sync",
dest="smart_sync",
action="store_true",
help="smart sync using manifest from the latest known good "
"build",
)
p.add_option(
"-t",
"--smart-tag",
dest="smart_tag",
action="store",
help="smart sync using manifest from a known tag",
)
g = p.add_option_group("repo Version options")
g.add_option(
"--no-repo-verify",
dest="repo_verify",
default=True,
action="store_false",
help="do not verify repo source code",
)
g.add_option(
"--repo-upgraded",
dest="repo_upgraded",
action="store_true",
help=optparse.SUPPRESS_HELP,
)
def _GetBranch(self, manifest_project):
"""Returns the branch name for getting the approved smartsync manifest.
Args:
manifest_project: The manifestProject to query.
"""
b = manifest_project.GetBranch(manifest_project.CurrentBranch)
branch = b.merge
if branch.startswith(R_HEADS):
branch = branch[len(R_HEADS) :]
return branch
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
@classmethod
def _GetCurrentBranchOnly(cls, opt, manifest):
"""Returns whether current-branch or use-superproject options are
enabled.
Args:
opt: Program options returned from optparse. See _Options().
manifest: The manifest to use.
Returns:
True if a superproject is requested, otherwise the value of the
current_branch option (True, False or None).
"""
return (
git_superproject.UseSuperproject(opt.use_superproject, manifest)
or opt.current_branch_only
)
def _UpdateProjectsRevisionId(
self, opt, args, superproject_logging_data, manifest
):
"""Update revisionId of projects with the commit from the superproject.
This function updates each project's revisionId with the commit hash
from the superproject. It writes the updated manifest into a file and
reloads the manifest from it. When appropriate, sub manifests are also
processed.
Args:
opt: Program options returned from optparse. See _Options().
args: Arguments to pass to GetProjects. See the GetProjects
docstring for details.
superproject_logging_data: A dictionary of superproject data to log.
manifest: The manifest to use.
"""
have_superproject = manifest.superproject or any(
m.superproject for m in manifest.all_children
)
if not have_superproject:
return
if opt.local_only and manifest.superproject:
manifest_path = manifest.superproject.manifest_path
if manifest_path:
self._ReloadManifest(manifest_path, manifest)
return
all_projects = self.GetProjects(
args,
missing_ok=True,
submodules_ok=opt.fetch_submodules,
manifest=manifest,
all_manifests=not opt.this_manifest_only,
)
per_manifest = collections.defaultdict(list)
if opt.this_manifest_only:
per_manifest[manifest.path_prefix] = all_projects
else:
for p in all_projects:
per_manifest[p.manifest.path_prefix].append(p)
superproject_logging_data = {}
need_unload = False
for m in self.ManifestList(opt):
if m.path_prefix not in per_manifest:
continue
use_super = git_superproject.UseSuperproject(
opt.use_superproject, m
)
if superproject_logging_data:
superproject_logging_data["multimanifest"] = True
superproject_logging_data.update(
superproject=use_super,
haslocalmanifests=bool(m.HasLocalManifests),
hassuperprojecttag=bool(m.superproject),
)
if use_super and (m.IsMirror or m.IsArchive):
# Don't use superproject, because we have no working tree.
use_super = False
superproject_logging_data["superproject"] = False
superproject_logging_data["noworktree"] = True
if opt.use_superproject is not False:
logger.warning(
"%s: not using superproject because there is no "
"working tree.",
m.path_prefix,
)
if not use_super:
continue
m.superproject.SetQuiet(not opt.verbose)
print_messages = git_superproject.PrintMessages(
opt.use_superproject, m
)
m.superproject.SetPrintMessages(print_messages)
update_result = m.superproject.UpdateProjectsRevisionId(
per_manifest[m.path_prefix], git_event_log=self.git_event_log
)
manifest_path = update_result.manifest_path
superproject_logging_data["updatedrevisionid"] = bool(manifest_path)
if manifest_path:
m.SetManifestOverride(manifest_path)
need_unload = True
else:
if print_messages:
logger.warning(
"%s: warning: Update of revisionId from superproject "
"has failed, repo sync will not use superproject to "
"fetch the source. Please resync with the "
"--no-use-superproject option to avoid this repo "
"warning.",
m.path_prefix,
)
if update_result.fatal and opt.use_superproject is not None:
raise SuperprojectError()
if need_unload:
m.outer_client.manifest.Unload()
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
@classmethod
def _FetchProjectList(cls, 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 _FetchOne.
Args:
opt: Program options returned from optparse. See _Options().
projects: Projects to fetch.
"""
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
return [cls._FetchOne(opt, x) for x in projects]
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
@classmethod
def _FetchOne(cls, opt, project_idx):
"""Fetch git objects for a single project.
Args:
opt: Program options returned from optparse. See _Options().
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
project_idx: Project index for the project to fetch.
Returns:
Whether the fetch was successful.
"""
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
project = cls.get_parallel_context()["projects"][project_idx]
start = time.time()
k = f"{project.name} @ {project.relpath}"
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
cls.get_parallel_context()["sync_dict"][k] = start
success = False
remote_fetched = False
errors = []
buf = TeeStringIO(sys.stdout if opt.verbose else None)
try:
sync_result = project.Sync_NetworkHalf(
quiet=opt.quiet,
verbose=opt.verbose,
output_redir=buf,
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
current_branch_only=cls._GetCurrentBranchOnly(
opt, project.manifest
),
force_sync=opt.force_sync,
clone_bundle=opt.clone_bundle,
tags=opt.tags,
archive=project.manifest.IsArchive,
optimized_fetch=opt.optimized_fetch,
retry_fetches=opt.retry_fetches,
prune=opt.prune,
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
ssh_proxy=cls.get_parallel_context()["ssh_proxy"],
clone_filter=project.manifest.CloneFilter,
partial_clone_exclude=project.manifest.PartialCloneExclude,
clone_filter_for_depth=project.manifest.CloneFilterForDepth,
)
success = sync_result.success
remote_fetched = sync_result.remote_fetched
if sync_result.error:
errors.append(sync_result.error)
output = buf.getvalue()
if output and buf.io is None and not success:
print("\n" + output.rstrip())
if not success:
logger.error(
"error: Cannot fetch %s from %s",
project.name,
project.remote.url,
)
except KeyboardInterrupt:
logger.error("Keyboard interrupt while processing %s", project.name)
except GitError as e:
logger.error("error.GitError: Cannot fetch %s", e)
errors.append(e)
except Exception as e:
logger.error(
"error: Cannot fetch %s (%s: %s)",
project.name,
type(e).__name__,
e,
)
errors.append(e)
raise
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
finally:
del cls.get_parallel_context()["sync_dict"][k]
finish = time.time()
return _FetchOneResult(
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
success, errors, project_idx, start, finish, remote_fetched
)
def _GetSyncProgressMessage(self):
earliest_time = float("inf")
earliest_proj = None
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
items = self.get_parallel_context()["sync_dict"].items()
for project, t in items:
if t < earliest_time:
earliest_time = t
earliest_proj = project
if not earliest_proj:
# This function is called when sync is still running but in some
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
# cases (by chance), sync_dict can contain no entries. Return some
# text to indicate that sync is still working.
return "..working.."
elapsed = time.time() - earliest_time
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
fetched = set()
remote_fetched = set()
pm = Progress(
"Fetching",
len(projects),
delay=False,
quiet=opt.quiet,
show_elapsed=True,
elide=True,
)
sync_event = _threading.Event()
def _MonitorSyncLoop():
while True:
pm.update(inc=0, msg=self._GetSyncProgressMessage())
if sync_event.wait(timeout=1):
return
sync_progress_thread = _threading.Thread(target=_MonitorSyncLoop)
sync_progress_thread.daemon = True
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
def _ProcessResults(pool, pm, results_sets):
ret = True
for results in results_sets:
for result in results:
success = result.success
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
project = projects[result.project_idx]
start = result.start
finish = result.finish
self._fetch_times.Set(project, finish - start)
self._local_sync_state.SetFetchTime(project)
self.event_log.AddSync(
project,
event_log.TASK_SYNC_NETWORK,
start,
finish,
success,
)
if result.errors:
errors.extend(result.errors)
if result.remote_fetched:
remote_fetched.add(project)
# 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()
if not ret and opt.fail_fast:
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
if pool:
pool.close()
break
return ret
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
with self.ParallelContext():
self.get_parallel_context()["projects"] = projects
self.get_parallel_context()[
"sync_dict"
] = multiprocessing.Manager().dict()
objdir_project_map = dict()
for index, project in enumerate(projects):
objdir_project_map.setdefault(project.objdir, []).append(index)
projects_list = list(objdir_project_map.values())
jobs = max(1, min(opt.jobs_network, len(projects_list)))
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
# We pass the ssh proxy settings via the class. This allows
# multiprocessing to pickle it up when spawning children. We can't
# pass it as an argument to _FetchProjectList below as
# multiprocessing is unable to pickle those.
self.get_parallel_context()["ssh_proxy"] = ssh_proxy
sync_progress_thread.start()
if not opt.quiet:
pm.update(inc=0, msg="warming up")
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
try:
ret = self.ExecuteInParallel(
jobs,
functools.partial(self._FetchProjectList, opt),
projects_list,
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
callback=_ProcessResults,
output=pm,
# Use chunksize=1 to avoid the chance that some workers are
# idle while other workers still have more than one job in
# their chunk queue.
chunksize=1,
initializer=self.InitWorker,
)
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
finally:
sync_event.set()
sync_progress_thread.join()
self._fetch_times.Save()
self._local_sync_state.Save()
if not self.outer_client.manifest.IsArchive:
self._GCProjects(projects, opt, err_event)
return _FetchResult(ret, fetched)
def _FetchMain(
self, opt, args, all_projects, err_event, ssh_proxy, manifest, errors
):
"""The main network fetch loop.
Args:
opt: Program options returned from optparse. See _Options().
args: Command line args used to filter out projects.
all_projects: List of all projects that should be fetched.
err_event: Whether an error was hit while processing.
ssh_proxy: SSH manager for clients & masters.
manifest: The manifest to use.
Returns:
List of all projects that should be checked out.
"""
rp = manifest.repoProject
to_fetch = []
now = time.time()
if _ONE_DAY_S <= (now - rp.LastFetch):
to_fetch.append(rp)
to_fetch.extend(all_projects)
to_fetch.sort(key=self._fetch_times.Get, reverse=True)
result = self._Fetch(to_fetch, opt, err_event, ssh_proxy, errors)
success = result.success
fetched = result.projects
if not success:
err_event.set()
# Call self update, unless requested not to
if os.environ.get("REPO_SKIP_SELF_UPDATE", "0") == "0":
_PostRepoFetch(rp, opt.repo_verify)
if opt.network_only:
# Bail out now; the rest touches the working tree.
if err_event.is_set():
e = SyncError(
"error: Exited sync due to fetch errors.",
aggregate_errors=errors,
)
logger.error(e)
raise e
return _FetchMainResult([])
# Iteratively fetch missing and/or nested unregistered submodules.
previously_missing_set = set()
while True:
self._ReloadManifest(None, manifest)
all_projects = self.GetProjects(
args,
missing_ok=True,
submodules_ok=opt.fetch_submodules,
manifest=manifest,
all_manifests=not opt.this_manifest_only,
)
missing = []
for project in all_projects:
if project.gitdir not in fetched:
missing.append(project)
if not missing:
break
# Stop us from non-stopped fetching actually-missing repos: If set
# of missing repos has not been changed from last fetch, we break.
missing_set = {p.name for p in missing}
if previously_missing_set == missing_set:
break
previously_missing_set = missing_set
result = self._Fetch(missing, opt, err_event, ssh_proxy, errors)
success = result.success
new_fetched = result.projects
if not success:
err_event.set()
fetched.update(new_fetched)
return _FetchMainResult(all_projects)
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
@classmethod
def _CheckoutOne(
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
cls,
detach_head,
force_sync,
force_checkout,
force_rebase,
verbose,
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
project_idx,
):
"""Checkout work tree for one project
Args:
detach_head: Whether to leave a detached HEAD.
force_sync: Force checking out of .git directory (e.g. overwrite
existing git directory that was previously linked to a different
object directory).
force_checkout: Force checking out of the repo content.
force_rebase: Force rebase.
verbose: Whether to show verbose messages.
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
project_idx: Project index for the project to checkout.
Returns:
Whether the fetch was successful.
"""
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
project = cls.get_parallel_context()["projects"][project_idx]
start = time.time()
syncbuf = SyncBuffer(
project.manifest.manifestProject.config, detach_head=detach_head
)
success = False
errors = []
try:
project.Sync_LocalHalf(
syncbuf,
force_sync=force_sync,
force_checkout=force_checkout,
force_rebase=force_rebase,
errors=errors,
verbose=verbose,
)
success = syncbuf.Finish()
except GitError as e:
logger.error(
"error.GitError: Cannot checkout %s: %s", project.name, e
)
errors.append(e)
except Exception as e:
logger.error(
"error: Cannot checkout %s: %s: %s",
project.name,
type(e).__name__,
e,
)
raise
if not success:
logger.error("error: Cannot checkout %s", project.name)
finish = time.time()
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
return _CheckoutOneResult(success, errors, project_idx, start, finish)
def _Checkout(self, all_projects, opt, err_results, checkout_errors):
"""Checkout projects listed in all_projects
Args:
all_projects: List of all projects that should be checked out.
opt: Program options returned from optparse. See _Options().
err_results: A list of strings, paths to git repos where checkout
failed.
"""
# Only checkout projects with worktrees.
all_projects = [x for x in all_projects if x.worktree]
def _ProcessResults(pool, pm, results):
ret = True
for result in results:
success = result.success
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
project = self.get_parallel_context()["projects"][
result.project_idx
]
start = result.start
finish = result.finish
self.event_log.AddSync(
project, event_log.TASK_SYNC_LOCAL, start, finish, success
)
if result.errors:
checkout_errors.extend(result.errors)
# Check for any errors before running any more tasks.
# ...we'll let existing jobs finish, though.
if success:
self._local_sync_state.SetCheckoutTime(project)
else:
ret = False
err_results.append(
project.RelPath(local=opt.this_manifest_only)
)
if opt.fail_fast:
if pool:
pool.close()
return ret
pm.update(msg=project.name)
return ret
for projects in _SafeCheckoutOrder(all_projects):
sync: reduce multiprocessing serialization overhead Background: - Manifest object is large (for projects like Android) in terms of serialization cost and size (more than 1mb). - Lots of Project objects usually share only a few manifest objects. Before this CL, Project objects were passed to workers via function parameters. Function parameters are pickled separately (in chunk). In other words, manifests are serialized again and again. The major serialization overhead of repo sync was O(manifest_size * projects / chunksize) This CL uses following tricks to reduce serialization overhead. - All projects are pickled in one invocation. Because Project objects share manifests, pickle library remembers which objects are already seen and avoid the serialization cost. - Pass the Project objects to workers at worker intialization time. And pass project index as function parameters instead. The number of workers is much smaller than the number of projects. - Worker init state are shared on Linux (fork based). So it requires zero serialization for Project objects. On Linux (fork based), the serialization overhead is O(projects) --- one int per project On Windows (spawn based), the serialization overhead is O(manifest_size * min(workers, projects)) Moreover, use chunksize=1 to avoid the chance that some workers are idle while other workers still have more than one job in their chunk queue. Using 2.7k projects as the baseline, originally "repo sync" no-op sync takes 31s for fetch and 25s for checkout on my Linux workstation. With this CL, it takes 12s for fetch and 1s for checkout. Bug: b/371638995 Change-Id: Ifa22072ea54eacb4a5c525c050d84de371e87caa Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/439921 Tested-by: Kuang-che Wu <kcwu@google.com> Reviewed-by: Josip Sokcevic <sokcevic@google.com> Commit-Queue: Kuang-che Wu <kcwu@google.com>
2024-10-18 15:32:08 +00:00
with self.ParallelContext():
self.get_parallel_context()["projects"] = projects
proc_res = self.ExecuteInParallel(
opt.jobs_checkout,
functools.partial(
self._CheckoutOne,
opt.detach_head,
opt.force_sync,
opt.force_checkout,
opt.rebase,
opt.verbose,
),
range(len(projects)),
callback=_ProcessResults,
output=Progress(
"Checking out", len(all_projects), quiet=opt.quiet
),
# Use chunksize=1 to avoid the chance that some workers are
# idle while other workers still have more than one job in
# their chunk queue.
chunksize=1,
)
self._local_sync_state.Save()
return proc_res and not err_results
@staticmethod
def _GetPreciousObjectsState(project: Project, opt):
"""Get the preciousObjects state for the project.
Args:
project (Project): the project to examine, and possibly correct.
opt (optparse.Values): options given to sync.
Returns:
Expected state of extensions.preciousObjects:
False: Should be disabled. (not present)
True: Should be enabled.
"""
if project.use_git_worktrees:
return False
projects = project.manifest.GetProjectsWithName(
project.name, all_manifests=True
)
if len(projects) == 1:
return False
if len(projects) > 1:
# Objects are potentially shared with another project.
# See the logic in Project.Sync_NetworkHalf regarding UseAlternates.
# - When False, shared projects share (via symlink)
# .repo/project-objects/{PROJECT_NAME}.git as the one-and-only
# objects directory. All objects are precious, since there is no
# project with a complete set of refs.
# - When True, shared projects share (via info/alternates)
# .repo/project-objects/{PROJECT_NAME}.git as an alternate object
# store, which is written only on the first clone of the project,
# and is not written subsequently. (When Sync_NetworkHalf sees
# that it exists, it makes sure that the alternates file points
# there, and uses a project-local .git/objects directory for all
# syncs going forward.
# We do not support switching between the options. The environment
# variable is present for testing and migration only.
return not project.UseAlternates
return False
def _SetPreciousObjectsState(self, project: Project, opt):
"""Correct the preciousObjects state for the project.
Args:
project: the project to examine, and possibly correct.
opt: options given to sync.
"""
expected = self._GetPreciousObjectsState(project, opt)
actual = (
project.config.GetBoolean("extensions.preciousObjects") or False
)
relpath = project.RelPath(local=opt.this_manifest_only)
if expected != actual:
# If this is unexpected, log it and repair.
Trace(
f"{relpath} expected preciousObjects={expected}, got {actual}"
)
if expected:
if not opt.quiet:
print(
"\r%s: Shared project %s found, disabling pruning."
% (relpath, project.name)
)
if git_require((2, 7, 0)):
project.EnableRepositoryExtension("preciousObjects")
else:
# This isn't perfect, but it's the best we can do with old
# git.
logger.warning(
"%s: WARNING: shared projects are unreliable when "
"using old versions of git; please upgrade to "
"git-2.7.0+.",
relpath,
)
project.config.SetString("gc.pruneExpire", "never")
else:
project.config.SetString("extensions.preciousObjects", None)
project.config.SetString("gc.pruneExpire", None)
def _GCProjects(self, projects, opt, err_event):
"""Perform garbage collection.
If We are skipping garbage collection (opt.auto_gc not set), we still
want to potentially mark objects precious, so that `git gc` does not
discard shared objects.
"""
if not opt.auto_gc:
# Just repair preciousObjects state, and return.
for project in projects:
self._SetPreciousObjectsState(project, opt)
return
pm = Progress(
"Garbage collecting", len(projects), delay=False, quiet=opt.quiet
)
pm.update(inc=0, msg="prescan")
tidy_dirs = {}
for project in projects:
self._SetPreciousObjectsState(project, opt)
project.config.SetString("gc.autoDetach", "false")
# Only call git gc once per objdir, but call pack-refs for the
# remainder.
if project.objdir not in tidy_dirs:
tidy_dirs[project.objdir] = (
True, # Run a full gc.
project.bare_git,
)
elif project.gitdir not in tidy_dirs:
tidy_dirs[project.gitdir] = (
False, # Do not run a full gc; just run pack-refs.
project.bare_git,
)
jobs = opt.jobs
if jobs < 2:
for run_gc, bare_git in tidy_dirs.values():
pm.update(msg=bare_git._project.name)
if run_gc:
bare_git.gc("--auto")
else:
bare_git.pack_refs()
pm.end()
return
cpu_count = os.cpu_count()
config = {"pack.threads": cpu_count // jobs if cpu_count > jobs else 1}
threads = set()
sem = _threading.Semaphore(jobs)
def tidy_up(run_gc, bare_git):
pm.start(bare_git._project.name)
try:
try:
if run_gc:
bare_git.gc("--auto", config=config)
else:
bare_git.pack_refs(config=config)
except GitError:
err_event.set()
except Exception:
err_event.set()
raise
finally:
pm.finish(bare_git._project.name)
sem.release()
for run_gc, bare_git in tidy_dirs.values():
if err_event.is_set() and opt.fail_fast:
break
sem.acquire()
t = _threading.Thread(
target=tidy_up,
args=(
run_gc,
bare_git,
),
)
t.daemon = True
threads.add(t)
t.start()
for t in threads:
t.join()
pm.end()
def _ReloadManifest(self, manifest_name, manifest):
"""Reload the manfiest from the file specified by the |manifest_name|.
It unloads the manifest if |manifest_name| is None.
Args:
manifest_name: Manifest file to be reloaded.
manifest: The manifest to use.
"""
if manifest_name:
# Override calls Unload already.
manifest.Override(manifest_name)
else:
manifest.Unload()
def UpdateProjectList(self, opt, manifest):
"""Update the cached projects list for |manifest|
In a multi-manifest checkout, each manifest has its own project.list.
Args:
opt: Program options returned from optparse. See _Options().
manifest: The manifest to use.
Returns:
0: success
1: failure
"""
new_project_paths = []
for project in self.GetProjects(
None, missing_ok=True, manifest=manifest, all_manifests=False
):
if project.relpath:
new_project_paths.append(project.relpath)
file_name = "project.list"
file_path = os.path.join(manifest.subdir, file_name)
old_project_paths = []
if os.path.exists(file_path):
with open(file_path) as fd:
old_project_paths = fd.read().split("\n")
# In reversed order, so subfolders are deleted before parent folder.
for path in sorted(old_project_paths, reverse=True):
if not path:
continue
if path not in new_project_paths:
# If the path has already been deleted, we don't need to do
# it.
gitdir = os.path.join(manifest.topdir, path, ".git")
if os.path.exists(gitdir):
project = Project(
manifest=manifest,
name=path,
remote=RemoteSpec("origin"),
gitdir=gitdir,
objdir=gitdir,
use_git_worktrees=os.path.isfile(gitdir),
worktree=os.path.join(manifest.topdir, path),
relpath=path,
revisionExpr="HEAD",
revisionId=None,
groups=None,
)
project.DeleteWorktree(
verbose=opt.verbose, force=opt.force_remove_dirty
)
new_project_paths.sort()
with open(file_path, "w") as fd:
fd.write("\n".join(new_project_paths))
fd.write("\n")
return 0
def UpdateCopyLinkfileList(self, manifest):
"""Save all dests of copyfile and linkfile, and update them if needed.
Returns:
Whether update was successful.
"""
new_paths = {}
new_linkfile_paths = []
new_copyfile_paths = []
for project in self.GetProjects(
None, missing_ok=True, manifest=manifest, all_manifests=False
):
new_linkfile_paths.extend(x.dest for x in project.linkfiles)
new_copyfile_paths.extend(x.dest for x in project.copyfiles)
new_paths = {
"linkfile": new_linkfile_paths,
"copyfile": new_copyfile_paths,
}
copylinkfile_name = "copy-link-files.json"
copylinkfile_path = os.path.join(manifest.subdir, copylinkfile_name)
old_copylinkfile_paths = {}
if os.path.exists(copylinkfile_path):
with open(copylinkfile_path, "rb") as fp:
try:
old_copylinkfile_paths = json.load(fp)
except Exception:
logger.error(
"error: %s is not a json formatted file.",
copylinkfile_path,
)
platform_utils.remove(copylinkfile_path)
raise
need_remove_files = []
need_remove_files.extend(
set(old_copylinkfile_paths.get("linkfile", []))
- set(new_linkfile_paths)
)
need_remove_files.extend(
set(old_copylinkfile_paths.get("copyfile", []))
- set(new_copyfile_paths)
)
for need_remove_file in need_remove_files:
# Try to remove the updated copyfile or linkfile.
# So, if the file is not exist, nothing need to do.
platform_utils.remove(
os.path.join(self.client.topdir, need_remove_file),
missing_ok=True,
)
# Create copy-link-files.json, save dest path of "copyfile" and
# "linkfile".
with open(copylinkfile_path, "w", encoding="utf-8") as fp:
json.dump(new_paths, fp)
return True
def _SmartSyncSetup(self, opt, smart_sync_manifest_path, manifest):
if not manifest.manifest_server:
raise SmartSyncError(
"error: cannot smart sync: no manifest server defined in "
"manifest"
)
manifest_server = manifest.manifest_server
if not opt.quiet:
print("Using manifest server %s" % manifest_server)
if "@" not in manifest_server:
username = None
password = None
if opt.manifest_server_username and opt.manifest_server_password:
username = opt.manifest_server_username
password = opt.manifest_server_password
else:
try:
info = netrc.netrc()
except OSError:
# .netrc file does not exist or could not be opened.
pass
else:
try:
parse_result = urllib.parse.urlparse(manifest_server)
if parse_result.hostname:
auth = info.authenticators(parse_result.hostname)
if auth:
username, _account, password = auth
else:
logger.error(
"No credentials found for %s in .netrc",
parse_result.hostname,
)
except netrc.NetrcParseError as e:
logger.error("Error parsing .netrc file: %s", e)
if username and password:
manifest_server = manifest_server.replace(
"://", f"://{username}:{password}@", 1
)
transport = PersistentTransport(manifest_server)
if manifest_server.startswith("persistent-"):
manifest_server = manifest_server[len("persistent-") :]
try:
server = xmlrpc.client.Server(manifest_server, transport=transport)
if opt.smart_sync:
branch = self._GetBranch(manifest.manifestProject)
if "SYNC_TARGET" in os.environ:
target = os.environ["SYNC_TARGET"]
[success, manifest_str] = server.GetApprovedManifest(
branch, target
)
elif (
"TARGET_PRODUCT" in os.environ
and "TARGET_BUILD_VARIANT" in os.environ
and "TARGET_RELEASE" in os.environ
):
target = "%s-%s-%s" % (
os.environ["TARGET_PRODUCT"],
os.environ["TARGET_RELEASE"],
os.environ["TARGET_BUILD_VARIANT"],
)
[success, manifest_str] = server.GetApprovedManifest(
branch, target
)
elif (
"TARGET_PRODUCT" in os.environ
and "TARGET_BUILD_VARIANT" in os.environ
):
target = "%s-%s" % (
os.environ["TARGET_PRODUCT"],
os.environ["TARGET_BUILD_VARIANT"],
)
[success, manifest_str] = server.GetApprovedManifest(
branch, target
)
else:
[success, manifest_str] = server.GetApprovedManifest(branch)
else:
assert opt.smart_tag
[success, manifest_str] = server.GetManifest(opt.smart_tag)
if success:
manifest_name = os.path.basename(smart_sync_manifest_path)
try:
with open(smart_sync_manifest_path, "w") as f:
f.write(manifest_str)
except OSError as e:
raise SmartSyncError(
"error: cannot write manifest to %s:\n%s"
% (smart_sync_manifest_path, e),
aggregate_errors=[e],
)
self._ReloadManifest(manifest_name, manifest)
else:
raise SmartSyncError(
"error: manifest server RPC call failed: %s" % manifest_str
)
except (OSError, xmlrpc.client.Fault) as e:
raise SmartSyncError(
"error: cannot connect to manifest server %s:\n%s"
% (manifest.manifest_server, e),
aggregate_errors=[e],
)
except xmlrpc.client.ProtocolError as e:
raise SmartSyncError(
"error: cannot connect to manifest server %s:\n%d %s"
% (manifest.manifest_server, e.errcode, e.errmsg),
aggregate_errors=[e],
)
return manifest_name
def _UpdateAllManifestProjects(self, opt, mp, manifest_name, errors):
"""Fetch & update the local manifest project.
After syncing the manifest project, if the manifest has any sub
manifests, those are recursively processed.
Args:
opt: Program options returned from optparse. See _Options().
mp: the manifestProject to query.
manifest_name: Manifest file to be reloaded.
"""
if not mp.standalone_manifest_url:
self._UpdateManifestProject(opt, mp, manifest_name, errors)
if mp.manifest.submanifests:
for submanifest in mp.manifest.submanifests.values():
child = submanifest.repo_client.manifest
child.manifestProject.SyncWithPossibleInit(
submanifest,
current_branch_only=self._GetCurrentBranchOnly(opt, child),
verbose=opt.verbose,
tags=opt.tags,
git_event_log=self.git_event_log,
)
self._UpdateAllManifestProjects(
opt, child.manifestProject, None, errors
)
def _UpdateManifestProject(self, opt, mp, manifest_name, errors):
"""Fetch & update the local manifest project.
Args:
opt: Program options returned from optparse. See _Options().
mp: the manifestProject to query.
manifest_name: Manifest file to be reloaded.
"""
if not opt.local_only:
start = time.time()
buf = TeeStringIO(sys.stdout)
try:
result = mp.Sync_NetworkHalf(
quiet=not opt.verbose,
output_redir=buf,
verbose=opt.verbose,
current_branch_only=self._GetCurrentBranchOnly(
opt, mp.manifest
),
force_sync=opt.force_sync,
tags=opt.tags,
optimized_fetch=opt.optimized_fetch,
retry_fetches=opt.retry_fetches,
submodules=mp.manifest.HasSubmodules,
clone_filter=mp.manifest.CloneFilter,
partial_clone_exclude=mp.manifest.PartialCloneExclude,
clone_filter_for_depth=mp.manifest.CloneFilterForDepth,
)
if result.error:
errors.append(result.error)
except KeyboardInterrupt:
errors.append(
ManifestInterruptError(buf.getvalue(), project=mp.name)
)
raise
finish = time.time()
self.event_log.AddSync(
mp, event_log.TASK_SYNC_NETWORK, start, finish, result.success
)
if mp.HasChanges:
errors = []
syncbuf = SyncBuffer(mp.config)
start = time.time()
mp.Sync_LocalHalf(
syncbuf,
submodules=mp.manifest.HasSubmodules,
errors=errors,
verbose=opt.verbose,
)
clean = syncbuf.Finish()
self.event_log.AddSync(
mp, event_log.TASK_SYNC_LOCAL, start, time.time(), clean
)
if not clean:
raise UpdateManifestError(aggregate_errors=errors)
self._ReloadManifest(manifest_name, mp.manifest)
def ValidateOptions(self, opt, args):
if opt.force_broken:
logger.warning(
"warning: -f/--force-broken is now the default behavior, and "
"the options are deprecated"
)
if opt.network_only and opt.detach_head:
self.OptionParser.error("cannot combine -n and -d")
if opt.network_only and opt.local_only:
self.OptionParser.error("cannot combine -n and -l")
if opt.manifest_name and opt.smart_sync:
self.OptionParser.error("cannot combine -m and -s")
if opt.manifest_name and opt.smart_tag:
self.OptionParser.error("cannot combine -m and -t")
if opt.manifest_server_username or opt.manifest_server_password:
if not (opt.smart_sync or opt.smart_tag):
self.OptionParser.error(
"-u and -p may only be combined with -s or -t"
)
if None in [
opt.manifest_server_username,
opt.manifest_server_password,
]:
self.OptionParser.error("both -u and -p must be given")
if opt.prune is None:
opt.prune = True
def _ValidateOptionsWithManifest(self, opt, mp):
"""Like ValidateOptions, but after we've updated the manifest.
Needed to handle sync-xxx option defaults in the manifest.
Args:
opt: The options to process.
mp: The manifest project to pull defaults from.
"""
if not opt.jobs:
# If the user hasn't made a choice, use the manifest value.
opt.jobs = mp.manifest.default.sync_j
if opt.jobs:
# If --jobs has a non-default value, propagate it as the default for
# --jobs-xxx flags too.
if not opt.jobs_network:
opt.jobs_network = opt.jobs
if not opt.jobs_checkout:
opt.jobs_checkout = opt.jobs
else:
# Neither user nor manifest have made a choice, so setup defaults.
if not opt.jobs_network:
opt.jobs_network = 1
if not opt.jobs_checkout:
opt.jobs_checkout = DEFAULT_LOCAL_JOBS
opt.jobs = os.cpu_count()
# Try to stay under user rlimit settings.
#
# Since each worker requires at 3 file descriptors to run `git fetch`,
# use that to scale down the number of jobs. Unfortunately there isn't
# an easy way to determine this reliably as systems change, but it was
# last measured by hand in 2011.
soft_limit, _ = _rlimit_nofile()
jobs_soft_limit = max(1, (soft_limit - 5) // 3)
opt.jobs = min(opt.jobs, jobs_soft_limit)
opt.jobs_network = min(opt.jobs_network, jobs_soft_limit)
opt.jobs_checkout = min(opt.jobs_checkout, jobs_soft_limit)
def Execute(self, opt, args):
errors = []
try:
self._ExecuteHelper(opt, args, errors)
except (RepoExitError, RepoChangedException):
raise
except (KeyboardInterrupt, Exception) as e:
raise RepoUnhandledExceptionError(e, aggregate_errors=errors)
def _ExecuteHelper(self, opt, args, errors):
manifest = self.outer_manifest
if not opt.outer_manifest:
manifest = self.manifest
if opt.manifest_name:
manifest.Override(opt.manifest_name)
manifest_name = opt.manifest_name
smart_sync_manifest_path = os.path.join(
manifest.manifestProject.worktree, "smart_sync_override.xml"
)
if opt.clone_bundle is None:
opt.clone_bundle = manifest.CloneBundle
if opt.smart_sync or opt.smart_tag:
manifest_name = self._SmartSyncSetup(
opt, smart_sync_manifest_path, manifest
)
else:
if os.path.isfile(smart_sync_manifest_path):
try:
platform_utils.remove(smart_sync_manifest_path)
except OSError as e:
logger.error(
"error: failed to remove existing smart sync override "
"manifest: %s",
e,
)
err_event = multiprocessing.Event()
rp = manifest.repoProject
rp.PreSync()
cb = rp.CurrentBranch
if cb:
base = rp.GetBranch(cb).merge
if not base or not base.startswith("refs/heads/"):
logger.warning(
"warning: repo is not tracking a remote branch, so it will "
"not receive updates; run `repo init --repo-rev=stable` to "
"fix."
)
for m in self.ManifestList(opt):
if not m.manifestProject.standalone_manifest_url:
m.manifestProject.PreSync()
if opt.repo_upgraded:
_PostRepoUpgrade(manifest, quiet=opt.quiet)
mp = manifest.manifestProject
if _REPO_ALLOW_SHALLOW is not None:
if _REPO_ALLOW_SHALLOW == "1":
mp.ConfigureCloneFilterForDepth(None)
elif (
_REPO_ALLOW_SHALLOW == "0" and mp.clone_filter_for_depth is None
):
mp.ConfigureCloneFilterForDepth("blob:none")
if opt.mp_update:
self._UpdateAllManifestProjects(opt, mp, manifest_name, errors)
else:
print("Skipping update of local manifest project.")
# Now that the manifests are up-to-date, setup options whose defaults
# might be in the manifest.
self._ValidateOptionsWithManifest(opt, mp)
superproject_logging_data = {}
self._UpdateProjectsRevisionId(
opt, args, superproject_logging_data, manifest
)
all_projects = self.GetProjects(
args,
missing_ok=True,
submodules_ok=opt.fetch_submodules,
manifest=manifest,
all_manifests=not opt.this_manifest_only,
)
err_network_sync = False
err_update_projects = False
err_update_linkfiles = False
# Log the repo projects by existing and new.
existing = [x for x in all_projects if x.Exists]
mp.config.SetString("repo.existingprojectcount", str(len(existing)))
mp.config.SetString(
"repo.newprojectcount", str(len(all_projects) - len(existing))
)
self._fetch_times = _FetchTimes(manifest)
self._local_sync_state = LocalSyncState(manifest)
if not opt.local_only and not opt.repo_upgraded:
with multiprocessing.Manager() as manager:
with ssh.ProxyManager(manager) as ssh_proxy:
# Initialize the socket dir once in the parent.
ssh_proxy.sock()
result = self._FetchMain(
opt,
args,
all_projects,
err_event,
ssh_proxy,
manifest,
errors,
)
all_projects = result.all_projects
if opt.network_only:
return
# If we saw an error, exit with code 1 so that other scripts can
# check.
if err_event.is_set():
err_network_sync = True
if opt.fail_fast:
logger.error(
"error: Exited sync due to fetch errors.\n"
"Local checkouts *not* updated. Resolve network issues "
"& retry.\n"
"`repo sync -l` will update some local checkouts."
)
raise SyncFailFastError(aggregate_errors=errors)
for m in self.ManifestList(opt):
if m.IsMirror or m.IsArchive:
# Bail out now, we have no working tree.
continue
try:
self.UpdateProjectList(opt, m)
except Exception as e:
err_event.set()
err_update_projects = True
errors.append(e)
if isinstance(e, DeleteWorktreeError):
errors.extend(e.aggregate_errors)
if opt.fail_fast:
logger.error("error: Local checkouts *not* updated.")
raise SyncFailFastError(aggregate_errors=errors)
try:
self.UpdateCopyLinkfileList(m)
except Exception as e:
err_update_linkfiles = True
errors.append(e)
err_event.set()
if opt.fail_fast:
logger.error(
"error: Local update copyfile or linkfile failed."
)
raise SyncFailFastError(aggregate_errors=errors)
err_results = []
# NB: We don't exit here because this is the last step.
err_checkout = not self._Checkout(
all_projects, opt, err_results, errors
)
if err_checkout:
err_event.set()
printed_notices = set()
# If there's a notice that's supposed to print at the end of the sync,
# print it now... But avoid printing duplicate messages, and preserve
# order.
for m in sorted(self.ManifestList(opt), key=lambda x: x.path_prefix):
if m.notice and m.notice not in printed_notices:
print(m.notice)
printed_notices.add(m.notice)
# If we saw an error, exit with code 1 so that other scripts can check.
if err_event.is_set():
def print_and_log(err_msg):
self.git_event_log.ErrorEvent(err_msg)
logger.error("%s", err_msg)
print_and_log("error: Unable to fully sync the tree")
if err_network_sync:
print_and_log("error: Downloading network changes failed.")
if err_update_projects:
print_and_log("error: Updating local project lists failed.")
if err_update_linkfiles:
print_and_log("error: Updating copyfiles or linkfiles failed.")
if err_checkout:
print_and_log("error: Checking out local projects failed.")
if err_results:
# Don't log repositories, as it may contain sensitive info.
logger.error("Failing repos:\n%s", "\n".join(err_results))
# Not useful to log.
logger.error(
'Try re-running with "-j1 --fail-fast" to exit at the first '
"error."
)
raise SyncError(aggregate_errors=errors)
# Log the previous sync analysis state from the config.
self.git_event_log.LogDataConfigEvents(
mp.config.GetSyncAnalysisStateData(), "previous_sync_state"
)
# Update and log with the new sync analysis state.
mp.config.UpdateSyncAnalysisState(opt, superproject_logging_data)
self.git_event_log.LogDataConfigEvents(
mp.config.GetSyncAnalysisStateData(), "current_sync_state"
)
self._local_sync_state.PruneRemovedProjects()
if self._local_sync_state.IsPartiallySynced():
logger.warning(
"warning: Partial syncs are not supported. For the best "
"experience, sync the entire tree."
)
if not opt.quiet:
print("repo sync has finished successfully.")
def _PostRepoUpgrade(manifest, quiet=False):
# Link the docs for the internal .repo/ layout for people.
link = os.path.join(manifest.repodir, "internal-fs-layout.md")
if not platform_utils.islink(link):
target = os.path.join("repo", "docs", "internal-fs-layout.md")
try:
platform_utils.symlink(target, link)
except Exception:
pass
wrapper = Wrapper()
if wrapper.NeedSetupGnuPG():
wrapper.SetupGnuPG(quiet)
for project in manifest.projects:
if project.Exists:
project.PostRepoUpgrade()
def _PostRepoFetch(rp, repo_verify=True, verbose=False):
if rp.HasChanges:
logger.warning("info: A new version of repo is available")
wrapper = Wrapper()
try:
rev = rp.bare_git.describe(rp.GetRevisionId())
except GitError:
rev = None
_, new_rev = wrapper.check_repo_rev(
rp.gitdir, rev, repo_verify=repo_verify
)
# See if we're held back due to missing signed tag.
current_revid = rp.bare_git.rev_parse("HEAD")
new_revid = rp.bare_git.rev_parse("--verify", new_rev)
if current_revid != new_revid:
# We want to switch to the new rev, but also not trash any
# uncommitted changes. This helps with local testing/hacking.
# If a local change has been made, we will throw that away.
# We also have to make sure this will switch to an older commit if
# that's the latest tag in order to support release rollback.
try:
rp.work_git.reset("--keep", new_rev)
except GitError as e:
raise RepoUnhandledExceptionError(e)
print("info: Restarting repo with latest version")
raise RepoChangedException(["--repo-upgraded"])
else:
logger.warning("warning: Skipped upgrade to unverified version")
else:
if verbose:
print("repo version %s is current" % rp.work_git.describe(HEAD))
class _FetchTimes:
_ALPHA = 0.5
def __init__(self, manifest):
self._path = os.path.join(manifest.repodir, ".repo_fetchtimes.json")
self._saved = None
self._seen = {}
def Get(self, project):
self._Load()
return self._saved.get(project.name, _ONE_DAY_S)
def Set(self, project, t):
name = project.name
# For shared projects, save the longest time.
self._seen[name] = max(self._seen.get(name, 0), t)
def _Load(self):
if self._saved is None:
try:
with open(self._path) as f:
self._saved = json.load(f)
except (OSError, ValueError):
platform_utils.remove(self._path, missing_ok=True)
self._saved = {}
def Save(self):
if self._saved is None:
return
for name, t in self._seen.items():
# Keep a moving average across the previous/current sync runs.
old = self._saved.get(name, t)
self._seen[name] = (self._ALPHA * t) + ((1 - self._ALPHA) * old)
try:
with open(self._path, "w") as f:
json.dump(self._seen, f, indent=2)
except (OSError, TypeError):
platform_utils.remove(self._path, missing_ok=True)
class LocalSyncState:
_LAST_FETCH = "last_fetch"
_LAST_CHECKOUT = "last_checkout"
def __init__(self, manifest):
self._manifest = manifest
self._path = os.path.join(
self._manifest.repodir, ".repo_localsyncstate.json"
)
self._time = time.time()
self._state = None
self._Load()
def SetFetchTime(self, project):
self._Set(project, self._LAST_FETCH)
def SetCheckoutTime(self, project):
self._Set(project, self._LAST_CHECKOUT)
def GetFetchTime(self, project):
return self._Get(project, self._LAST_FETCH)
def GetCheckoutTime(self, project):
return self._Get(project, self._LAST_CHECKOUT)
def _Get(self, project, key):
self._Load()
p = project.relpath
if p not in self._state:
return
return self._state[p].get(key)
def _Set(self, project, key):
p = project.relpath
if p not in self._state:
self._state[p] = {}
self._state[p][key] = self._time
def _Load(self):
if self._state is None:
try:
with open(self._path) as f:
self._state = json.load(f)
except (OSError, ValueError):
platform_utils.remove(self._path, missing_ok=True)
self._state = {}
def Save(self):
if not self._state:
return
try:
with open(self._path, "w") as f:
json.dump(self._state, f, indent=2)
except (OSError, TypeError):
platform_utils.remove(self._path, missing_ok=True)
def PruneRemovedProjects(self):
"""Remove entries don't exist on disk and save."""
if not self._state:
return
delete = set()
for path in self._state:
gitdir = os.path.join(self._manifest.topdir, path, ".git")
if not os.path.exists(gitdir) or os.path.islink(gitdir):
delete.add(path)
if not delete:
return
for path in delete:
del self._state[path]
self.Save()
def IsPartiallySynced(self):
"""Return whether a partial sync state is detected."""
self._Load()
prev_checkout_t = None
for path, data in self._state.items():
if path == self._manifest.repoProject.relpath:
# The repo project isn't included in most syncs so we should
# ignore it here.
continue
checkout_t = data.get(self._LAST_CHECKOUT)
if not checkout_t:
return True
prev_checkout_t = prev_checkout_t or checkout_t
if prev_checkout_t != checkout_t:
return True
return False
# This is a replacement for xmlrpc.client.Transport using urllib2
# and supporting persistent-http[s]. It cannot change hosts from
# request to request like the normal transport, the real url
# is passed during initialization.
class PersistentTransport(xmlrpc.client.Transport):
def __init__(self, orig_host):
super().__init__()
self.orig_host = orig_host
def request(self, host, handler, request_body, verbose=False):
with GetUrlCookieFile(self.orig_host, not verbose) as (
cookiefile,
proxy,
):
# Python doesn't understand cookies with the #HttpOnly_ prefix
# Since we're only using them for HTTP, copy the file temporarily,
# stripping those prefixes away.
if cookiefile:
tmpcookiefile = tempfile.NamedTemporaryFile(mode="w")
tmpcookiefile.write("# HTTP Cookie File")
try:
with open(cookiefile) as f:
for line in f:
if line.startswith("#HttpOnly_"):
line = line[len("#HttpOnly_") :]
tmpcookiefile.write(line)
tmpcookiefile.flush()
cookiejar = cookielib.MozillaCookieJar(tmpcookiefile.name)
try:
cookiejar.load()
except cookielib.LoadError:
cookiejar = cookielib.CookieJar()
finally:
tmpcookiefile.close()
else:
cookiejar = cookielib.CookieJar()
proxyhandler = urllib.request.ProxyHandler
if proxy:
proxyhandler = urllib.request.ProxyHandler(
{"http": proxy, "https": proxy}
)
opener = urllib.request.build_opener(
urllib.request.HTTPCookieProcessor(cookiejar), proxyhandler
)
url = urllib.parse.urljoin(self.orig_host, handler)
parse_results = urllib.parse.urlparse(url)
scheme = parse_results.scheme
if scheme == "persistent-http":
scheme = "http"
if scheme == "persistent-https":
# If we're proxying through persistent-https, use http. The
# proxy itself will do the https.
if proxy:
scheme = "http"
else:
scheme = "https"
# Parse out any authentication information using the base class.
host, extra_headers, _ = self.get_host_info(parse_results.netloc)
url = urllib.parse.urlunparse(
(
scheme,
host,
parse_results.path,
parse_results.params,
parse_results.query,
parse_results.fragment,
)
)
request = urllib.request.Request(url, request_body)
if extra_headers is not None:
for name, header in extra_headers:
request.add_header(name, header)
request.add_header("Content-Type", "text/xml")
try:
response = opener.open(request)
except urllib.error.HTTPError as e:
if e.code == 501:
# We may have been redirected through a login process
# but our POST turned into a GET. Retry.
response = opener.open(request)
else:
raise
p, u = xmlrpc.client.getparser()
# Response should be fairly small, so read it all at once.
# This way we can show it to the user in case of error (e.g. HTML).
data = response.read()
try:
p.feed(data)
except xml.parsers.expat.ExpatError as e:
raise OSError(
f"Parsing the manifest failed: {e}\n"
f"Please report this to your manifest server admin.\n"
f'Here is the full response:\n{data.decode("utf-8")}'
)
p.close()
return u.close()
def close(self):
pass