git-repo/command.py

554 lines
19 KiB
Python
Raw Permalink 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.
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
import contextlib
import multiprocessing
2008-10-21 14:00:00 +00:00
import optparse
import os
import re
2008-10-21 14:00:00 +00:00
from error import InvalidProjectGroupsError
from error import NoSuchProjectError
from error import RepoExitError
from event_log import EventLog
import progress
2008-10-21 14:00:00 +00:00
# Are we generating man-pages?
GENERATE_MANPAGES = os.environ.get("_REPO_GENERATE_MANPAGES_") == " indeed! "
# Number of projects to submit to a single worker process at a time.
# This number represents a tradeoff between the overhead of IPC and finer
# grained opportunity for parallelism. This particular value was chosen by
# iterating through powers of two until the overall performance no longer
# improved. The performance of this batch size is not a function of the
# number of cores on the system.
WORKER_BATCH_SIZE = 32
# How many jobs to run in parallel by default? This assumes the jobs are
# largely I/O bound and do not hit the network.
DEFAULT_LOCAL_JOBS = min(os.cpu_count(), 8)
class UsageError(RepoExitError):
"""Exception thrown with invalid command usage."""
class Command:
"""Base class for any command line action in repo."""
# Singleton for all commands to track overall repo command execution and
# provide event summary to callers. Only used by sync subcommand currently.
#
# NB: This is being replaced by git trace2 events. See git_trace2_event_log.
event_log = EventLog()
# Whether this command is a "common" one, i.e. whether the user would
# commonly use it or it's a more uncommon command. This is used by the help
# command to show short-vs-full summaries.
COMMON = False
# Whether this command supports running in parallel. If greater than 0,
# it is the number of parallel jobs to default to.
PARALLEL_JOBS = None
# Whether this command supports Multi-manifest. If False, then main.py will
# iterate over the manifests and invoke the command once per (sub)manifest.
# This is only checked after calling ValidateOptions, so that partially
# migrated subcommands can set it to False.
MULTI_MANIFEST_SUPPORT = 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
# Shared data across parallel execution workers.
_parallel_context = None
@classmethod
def get_parallel_context(cls):
assert cls._parallel_context is not None
return cls._parallel_context
def __init__(
self,
repodir=None,
client=None,
manifest=None,
git_event_log=None,
outer_client=None,
outer_manifest=None,
):
self.repodir = repodir
self.client = client
self.outer_client = outer_client or client
self.manifest = manifest
self.git_event_log = git_event_log
self.outer_manifest = outer_manifest
# Cache for the OptionParser property.
self._optparse = None
def WantPager(self, _opt):
return False
def ReadEnvironmentOptions(self, opts):
"""Set options from environment variables."""
env_options = self._RegisteredEnvironmentOptions()
for env_key, opt_key in env_options.items():
# Get the user-set option value if any
opt_value = getattr(opts, opt_key)
# If the value is set, it means the user has passed it as a command
# line option, and we should use that. Otherwise we can try to set
# it with the value from the corresponding environment variable.
if opt_value is not None:
continue
env_value = os.environ.get(env_key)
if env_value is not None:
setattr(opts, opt_key, env_value)
return opts
@property
def OptionParser(self):
if self._optparse is None:
try:
me = "repo %s" % self.NAME
usage = self.helpUsage.strip().replace("%prog", me)
except AttributeError:
usage = "repo %s" % self.NAME
epilog = (
"Run `repo help %s` to view the detailed manual." % self.NAME
)
self._optparse = optparse.OptionParser(usage=usage, epilog=epilog)
self._CommonOptions(self._optparse)
self._Options(self._optparse)
return self._optparse
def _CommonOptions(self, p, opt_v=True):
"""Initialize the option parser with common options.
These will show up for *all* subcommands, so use sparingly.
NB: Keep in sync with repo:InitParser().
"""
g = p.add_option_group("Logging options")
opts = ["-v"] if opt_v else []
g.add_option(
*opts,
"--verbose",
dest="output_mode",
action="store_true",
help="show all output",
)
g.add_option(
"-q",
"--quiet",
dest="output_mode",
action="store_false",
help="only show errors",
)
if self.PARALLEL_JOBS is not None:
default = "based on number of CPU cores"
if not GENERATE_MANPAGES:
# Only include active cpu count if we aren't generating man
# pages.
default = f"%default; {default}"
p.add_option(
"-j",
"--jobs",
type=int,
default=self.PARALLEL_JOBS,
help=f"number of jobs to run in parallel (default: {default})",
)
m = p.add_option_group("Multi-manifest options")
m.add_option(
"--outer-manifest",
action="store_true",
default=None,
help="operate starting at the outermost manifest",
)
m.add_option(
"--no-outer-manifest",
dest="outer_manifest",
action="store_false",
help="do not operate on outer manifests",
)
m.add_option(
"--this-manifest-only",
action="store_true",
default=None,
help="only operate on this (sub)manifest",
)
m.add_option(
"--no-this-manifest-only",
"--all-manifests",
dest="this_manifest_only",
action="store_false",
help="operate on this manifest and its submanifests",
)
def _Options(self, p):
"""Initialize the option parser with subcommand-specific options."""
def _RegisteredEnvironmentOptions(self):
"""Get options that can be set from environment variables.
Return a dictionary mapping environment variable name
to option key name that it can override.
Example: {'REPO_MY_OPTION': 'my_option'}
Will allow the option with key value 'my_option' to be set
from the value in the environment variable named 'REPO_MY_OPTION'.
Note: This does not work properly for options that are explicitly
set to None by the user, or options that are defined with a
default value other than None.
"""
return {}
def Usage(self):
"""Display usage and terminate."""
self.OptionParser.print_usage()
raise UsageError()
def CommonValidateOptions(self, opt, args):
"""Validate common options."""
opt.quiet = opt.output_mode is False
opt.verbose = opt.output_mode is True
if opt.outer_manifest is None:
# By default, treat multi-manifest instances as a single manifest
# from the user's perspective.
opt.outer_manifest = True
def ValidateOptions(self, opt, args):
"""Validate the user options & arguments before executing.
This is meant to help break the code up into logical steps. Some tips:
* Use self.OptionParser.error to display CLI related errors.
* Adjust opt member defaults as makes sense.
* Adjust the args list, but do so inplace so the caller sees updates.
* Try to avoid updating self state. Leave that to Execute.
"""
def Execute(self, opt, args):
"""Perform the action, after option parsing is complete."""
raise NotImplementedError
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
@contextlib.contextmanager
def ParallelContext(cls):
"""Obtains the context, which is shared to ExecuteInParallel workers.
Callers can store data in the context dict before invocation of
ExecuteInParallel. The dict will then be shared to child workers of
ExecuteInParallel.
"""
assert cls._parallel_context is None
cls._parallel_context = {}
try:
yield
finally:
cls._parallel_context = None
@classmethod
def _InitParallelWorker(cls, context, initializer):
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._parallel_context = context
if initializer:
initializer()
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 ExecuteInParallel(
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,
jobs,
func,
inputs,
callback,
output=None,
ordered=False,
chunksize=WORKER_BATCH_SIZE,
initializer=None,
):
"""Helper for managing parallel execution boiler plate.
For subcommands that can easily split their work up.
Args:
jobs: How many parallel processes to use.
func: The function to apply to each of the |inputs|. Usually a
functools.partial for wrapping additional arguments. It will be
run in a separate process, so it must be pickalable, so nested
functions won't work. Methods on the subcommand Command class
should work.
inputs: The list of items to process. Must be a list.
callback: The function to pass the results to for processing. It
will be executed in the main thread and process the results of
|func| as they become available. Thus it may be a local nested
function. Its return value is passed back directly. It takes
three arguments:
- The processing pool (or None with one job).
- The |output| argument.
- An iterator for the results.
output: An output manager. May be progress.Progess or
color.Coloring.
ordered: Whether the jobs should be processed in order.
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
chunksize: The number of jobs processed in batch by parallel
workers.
initializer: Worker initializer.
Returns:
The |callback| function's results are returned.
"""
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
try:
# NB: Multiprocessing is heavy, so don't spin it up for one job.
if len(inputs) == 1 or jobs == 1:
return callback(None, output, (func(x) for x in inputs))
else:
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 multiprocessing.Pool(
jobs,
initializer=cls._InitParallelWorker,
initargs=(cls._parallel_context, initializer),
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
) as pool:
submit = pool.imap if ordered else pool.imap_unordered
return callback(
pool,
output,
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
submit(func, inputs, chunksize=chunksize),
)
finally:
if isinstance(output, progress.Progress):
output.end()
def _ResetPathToProjectMap(self, projects):
self._by_path = {p.worktree: p for p in projects}
def _UpdatePathToProjectMap(self, project):
self._by_path[project.worktree] = project
def _GetProjectByPath(self, manifest, path):
project = None
if os.path.exists(path):
oldpath = None
while path and path != oldpath and path != manifest.topdir:
try:
project = self._by_path[path]
break
except KeyError:
oldpath = path
path = os.path.dirname(path)
if not project and path == manifest.topdir:
try:
project = self._by_path[path]
except KeyError:
pass
else:
try:
project = self._by_path[path]
except KeyError:
pass
return project
def GetProjects(
self,
args,
manifest=None,
groups="",
missing_ok=False,
submodules_ok=False,
all_manifests=False,
):
"""A list of projects that match the arguments.
Args:
args: a list of (case-insensitive) strings, projects to search for.
manifest: an XmlManifest, the manifest to use, or None for default.
groups: a string, the manifest groups in use.
missing_ok: a boolean, whether to allow missing projects.
submodules_ok: a boolean, whether to allow submodules.
all_manifests: a boolean, if True then all manifests and
submanifests are used. If False, then only the local
(sub)manifest is used.
Returns:
A list of matching Project instances.
"""
if all_manifests:
if not manifest:
manifest = self.manifest.outer_client
all_projects_list = manifest.all_projects
else:
if not manifest:
manifest = self.manifest
all_projects_list = manifest.projects
result = []
if not groups:
groups = manifest.GetGroupsStr()
groups = [x for x in re.split(r"[,\s]+", groups) if x]
if not args:
derived_projects = {}
for project in all_projects_list:
if submodules_ok or project.sync_s:
derived_projects.update(
(p.name, p) for p in project.GetDerivedSubprojects()
)
all_projects_list.extend(derived_projects.values())
for project in all_projects_list:
if (missing_ok or project.Exists) and project.MatchesGroups(
groups
):
result.append(project)
else:
self._ResetPathToProjectMap(all_projects_list)
for arg in args:
# We have to filter by manifest groups in case the requested
# project is checked out multiple times or differently based on
# them.
projects = [
project
for project in manifest.GetProjectsWithName(
arg, all_manifests=all_manifests
)
if project.MatchesGroups(groups)
]
if not projects:
path = os.path.abspath(arg).replace("\\", "/")
tree = manifest
if all_manifests:
# Look for the deepest matching submanifest.
for tree in reversed(list(manifest.all_manifests)):
if path.startswith(tree.topdir):
break
project = self._GetProjectByPath(tree, path)
# If it's not a derived project, update path->project
# mapping and search again, as arg might actually point to
# a derived subproject.
if (
project
and not project.Derived
and (submodules_ok or project.sync_s)
):
search_again = False
for subproject in project.GetDerivedSubprojects():
self._UpdatePathToProjectMap(subproject)
search_again = True
if search_again:
project = (
self._GetProjectByPath(manifest, path)
or project
)
if project:
projects = [project]
if not projects:
raise NoSuchProjectError(arg)
for project in projects:
if not missing_ok and not project.Exists:
raise NoSuchProjectError(
"%s (%s)"
% (arg, project.RelPath(local=not all_manifests))
)
if not project.MatchesGroups(groups):
raise InvalidProjectGroupsError(arg)
result.extend(projects)
def _getpath(x):
return x.relpath
result.sort(key=_getpath)
return result
def FindProjects(self, args, inverse=False, all_manifests=False):
"""Find projects from command line arguments.
Args:
args: a list of (case-insensitive) strings, projects to search for.
inverse: a boolean, if True, then projects not matching any |args|
are returned.
all_manifests: a boolean, if True then all manifests and
submanifests are used. If False, then only the local
(sub)manifest is used.
"""
result = []
patterns = [re.compile(r"%s" % a, re.IGNORECASE) for a in args]
for project in self.GetProjects("", all_manifests=all_manifests):
paths = [project.name, project.RelPath(local=not all_manifests)]
for pattern in patterns:
match = any(pattern.search(x) for x in paths)
if not inverse and match:
result.append(project)
break
if inverse and match:
break
else:
if inverse:
result.append(project)
result.sort(
key=lambda project: (project.manifest.path_prefix, project.relpath)
)
return result
def ManifestList(self, opt):
"""Yields all of the manifests to traverse.
Args:
opt: The command options.
"""
top = self.outer_manifest
if not opt.outer_manifest or opt.this_manifest_only:
top = self.manifest
yield top
if not opt.this_manifest_only:
yield from top.all_children
2008-10-21 14:00:00 +00:00
class InteractiveCommand(Command):
"""Command which requires user interaction on the tty and must not run
within a pager, even if the user asks to.
"""
def WantPager(self, _opt):
return False
2008-10-21 14:00:00 +00:00
2008-10-21 14:00:00 +00:00
class PagedCommand(Command):
"""Command which defaults to output in a pager, as its display tends to be
larger than one screen full.
"""
def WantPager(self, _opt):
return True
class MirrorSafeCommand:
"""Command permits itself to run within a mirror, and does not require a
working directory.
"""
class GitcClientCommand:
"""Command that requires the local client to be a GITC client."""