Source code for ramble.stage
# Copyright 2022-2025 The Ramble Authors
#
# Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
# https://www.apache.org/licenses/LICENSE-2.0> or the MIT license
# <LICENSE-MIT or https://opensource.org/licenses/MIT>, at your
# option. This file may not be copied, modified, or distributed
# except according to those terms.
import errno
import getpass
import glob
import hashlib
import os
import shutil
import stat
import sys
import llnl.util.lang
import llnl.util.tty as tty
from llnl.util.filesystem import mkdirp, can_access, install, install_tree
from llnl.util.filesystem import partition_path, remove_linked_tree
import spack.paths
import spack.config
import spack.util.pattern as pattern
import spack.util.path as sup
import spack.util.url as url_util
from spack.util.crypto import prefix_bits, bit_length
import ramble.caches
import ramble.fetch_strategy as fs
import ramble.util.lock
import ramble.error
import ramble.mirror
from ramble.util.logger import logger
# The well-known stage source subdirectory name.
_input_subdir = "input"
[docs]
def create_stage_root(path):
"""Create the stage root directory and ensure appropriate access perms."""
assert path.startswith(os.path.sep) and len(path.strip()) > 1
err_msg = "Cannot create stage root {0}: Access to {1} is denied"
# TODO: (dwjacobsen) Remove when owner_uid is removed below
# user_uid = os.getuid()
# Obtain lists of ancestor and descendant paths of the $user node, if any.
group_paths, user_node, user_paths = partition_path(path, getpass.getuser())
for p in group_paths:
if not os.path.exists(p):
# Ensure access controls of subdirs created above `$user` inherit
# from the parent and share the group.
par_stat = os.stat(os.path.dirname(p))
mkdirp(p, group=par_stat.st_gid, mode=par_stat.st_mode)
p_stat = os.stat(p)
if par_stat.st_gid != p_stat.st_gid:
logger.warn(
f"Expected {p} to have group " f"{par_stat.st_gid}, but it is {p_stat.st_gid}"
)
if par_stat.st_mode & p_stat.st_mode != par_stat.st_mode:
logger.warn(
f"Expected {p} to support mode "
f"{par_stat.st_mode}, but it is {p_stat.st_mode}"
)
if not can_access(p):
raise OSError(errno.EACCES, err_msg.format(path, p))
# Add the path ending with the $user node to the user paths to ensure paths
# from $user (on down) meet the ownership and permission requirements.
if user_node:
user_paths.insert(0, user_node)
for p in user_paths:
# Ensure access controls of subdirs from `$user` on down are
# restricted to the user.
if not os.path.exists(p):
mkdirp(p, mode=stat.S_IRWXU)
p_stat = os.stat(p)
if p_stat.st_mode & stat.S_IRWXU != stat.S_IRWXU:
logger.error(
f"Expected {p} to support mode " f"{stat.S_IRWXU}, but it is {p_stat.st_mode}"
)
raise OSError(errno.EACCES, err_msg.format(path, p))
else:
p_stat = os.stat(p)
# TODO: (dwjacobsen) Remove at some point
# if user_uid != p_stat.st_uid:
# tty.warn("Expected user {0} to own {1}, but it is owned by {2}"
# .format(user_uid, p, owner_uid))
input_subdir = os.path.join(path, _input_subdir)
# When staging into a user-specified directory we need to ensure the
# `input` subdirectory exists, as we can't rely on it being created
# automatically by ramble.
if not os.path.isdir(input_subdir):
mkdirp(input_subdir, mode=stat.S_IRWXU)
def _first_accessible_path(paths):
"""Find the first path that is accessible, creating it if necessary."""
for path in paths:
try:
# Ensure the user has access, creating the directory if necessary.
if os.path.exists(path):
if can_access(path):
return path
else:
# Now create the stage root with the proper group/perms.
create_stage_root(path)
return path
except OSError as e:
logger.debug(f"OSError while checking stage path {path}: {e}")
return None
def _resolve_paths(candidates):
"""
Resolve candidate paths and make user-related adjustments.
Adjustments involve removing extra $user from $tempdir if $tempdir includes
$user and appending $user if it is not present in the path.
"""
temp_path = sup.canonicalize_path("$tempdir")
user = getpass.getuser()
tmp_has_usr = user in temp_path.split(os.path.sep)
paths = []
for path in candidates:
# Remove the extra `$user` node from a `$tempdir/$user` entry for
# hosts that automatically append `$user` to `$tempdir`.
if path.startswith(os.path.join("$tempdir", "$user")) and tmp_has_usr:
path = path.replace("/$user", "", 1)
# Ensure the path is unique per user.
can_path = sup.canonicalize_path(path)
if user not in can_path:
can_path = os.path.join(can_path, user)
paths.append(can_path)
return paths
# Cached stage path root
_stage_root = None
# TODO (dwj): If we want to support multiple mirrors, we'll need to
# figure out how to pass them to the stage.
def _mirror_roots():
mirrors = ramble.config.get("mirrors")
return [
(
sup.substitute_path_variables(root)
if root.endswith(os.sep)
else sup.substitute_path_variables(root) + os.sep
)
for root in mirrors.values()
]
[docs]
class InputStage:
"""Manages a stage directory for containing input files.
A Stage object is a context manager that handles a directory where
some workload inputs are downloaded.
It handles fetching inputs, either as an archive to be
expanded or by checking it out of a repository. A stage's
lifecycle looks like this::
with InputStage() as stage: # Context manager creates the stage
# directory
stage.fetch() # Fetch an input archive into the stage.
stage.expand_archive() # Expand the archive into input_path.
Because the input files are needed to execute various experiments,
the stage is retained by default.
You can also use the stage's create/destroy functions manually,
like this::
stage = InputStage()
try:
stage.create() # Explicitly create the stage directory.
stage.fetch() # Fetch a source archive into the stage.
stage.expand_archive() # Expand the archive into source_path.
<install> # Build and install the archive.
# (handled by user of Stage)
finally:
stage.destroy() # Explicitly destroy the stage directory.
InputStages are required to be named by default. The name should match
a workload namespace.
"""
"""Shared dict of all stage locks."""
stage_locks = {}
"""Most staging is managed by Ramble. DIYStage is one exception."""
managed_by_ramble = True
def __init__(
self,
url_or_fetch_strategy,
name=None,
path=None,
mirror_paths=None,
keep=True,
lock=True,
search_fn=None,
):
"""Create a stage object.
Parameters:
url_or_fetch_strategy
URL of the archive to be downloaded into this stage, OR
a valid FetchStrategy.
name
If a name is provided, then this stage is a named stage
and will persist between runs (or if you construct another
stage object later). If name is not provided, then this
stage will be given a unique name automatically.
mirror_paths
If provided, Stage will search Rambles's mirrors for
this archive at each of the provided relative mirror paths
before using the default fetch strategy.
keep
By default, when used as a context manager, the Stage
is deleted on exit when no exceptions are raised.
Pass True to keep the stage intact even if no
exceptions are raised.
path
If provided, the stage path to use for associated builds.
lock
True if the stage directory file lock is to be used, False
otherwise.
search_fn
The search function that provides the fetch strategy
instance.
"""
# TODO: fetch/stage coupling needs to be reworked -- the logic
# TODO: here is convoluted and not modular enough.
if isinstance(url_or_fetch_strategy, str):
self.fetcher = fs.from_url_scheme(url_or_fetch_strategy)
elif isinstance(url_or_fetch_strategy, fs.FetchStrategy):
self.fetcher = url_or_fetch_strategy
else:
raise ValueError("Can't construct Stage without url or fetch strategy")
self.fetcher.stage = self
# self.fetcher can change with mirrors.
self.default_fetcher = self.fetcher
self.search_fn = search_fn
# used for mirrored archives of repositories.
self.skip_checksum_for_mirror = True
self.srcdir = None
self.input_subdir = _input_subdir
self.name = name
if name is None:
raise StageError("InputStage requires a name")
self.mirror_paths = mirror_paths
# Use the provided path or construct an optionally named stage path.
self.path = path
# Flag to decide whether to delete the stage folder on exit or not
self.keep = keep
# File lock for the stage directory. We use one file for all
# stage locks. See spack.database.Database.prefix_lock for
# details on this approach.
self._lock = None
if lock:
if self.name not in InputStage.stage_locks:
sha1 = hashlib.sha1(self.name.encode("utf-8")).digest()
lock_id = prefix_bits(sha1, bit_length(sys.maxsize))
stage_lock_path = os.path.join(self.path, ".lock")
logger.debug(f"Creating stage lock {self.name}")
InputStage.stage_locks[self.name] = ramble.util.lock.Lock(
stage_lock_path, lock_id, 1, desc=self.name
)
self._lock = InputStage.stage_locks[self.name]
# When stages are reused, we need to know whether to re-create
# it. This marks whether it has been created/destroyed.
self.created = False
def __enter__(self):
"""
Entering a stage context will create the stage directory
Returns:
self
"""
if self._lock is not None:
self._lock.acquire_write(timeout=60)
self.create()
return self
def __exit__(self, exc_type, exc_val, exc_tb):
"""
Exiting from a stage context will unlock the stage directory.
Args:
exc_type: exception type
exc_val: exception value
exc_tb: exception traceback
Returns:
Boolean
"""
if self._lock is not None:
self._lock.release_write()
# Delete when there are no exceptions, unless asked to keep.
if exc_type is None and not self.keep:
self.destroy()
@property
def expected_archive_files(self):
"""Possible archive file paths."""
paths = []
fnames = []
expanded = True
if isinstance(self.default_fetcher, fs.URLFetchStrategy):
expanded = self.default_fetcher.expand_archive
fnames.append(os.path.basename(self.default_fetcher.url))
if self.mirror_paths:
fnames.extend(os.path.basename(x) for x in self.mirror_paths)
paths.extend(os.path.join(self.path, f) for f in fnames)
if not expanded:
# If the download file is not compressed, the "archive" is a
# single file placed in Stage.source_path
paths.extend(os.path.join(self.source_path, f) for f in fnames)
return paths
@property
def save_filename(self):
possible_filenames = self.expected_archive_files
if possible_filenames:
# This prefers using the URL associated with the default fetcher if
# available, so that the fetched resource name matches the remote
# name
return possible_filenames[0]
@property
def archive_file(self):
"""Path to the source archive within this stage directory."""
for path in self.expected_archive_files:
if os.path.exists(path):
return path
else:
return None
@property
def expanded(self):
"""Returns True if source path expanded; else False."""
return os.path.exists(self.source_path)
@property
def source_path(self):
"""Returns the well-known source directory path."""
return os.path.join(self.path, self.input_subdir)
[docs]
def fetch(self, mirror_only=False, err_msg=None):
"""Retrieves the code or archive
Args:
mirror_only (bool): only fetch from a mirror
err_msg (str or None): the error message to display if all fetchers
fail or ``None`` for the default fetch failure message
"""
fetchers = []
if not mirror_only:
fetchers.append(self.default_fetcher)
# TODO: move mirror logic out of here and clean it up!
# TODO: Or @alalazo may have some ideas about how to use a
# TODO: CompositeFetchStrategy here.
self.skip_checksum_for_mirror = True
if self.mirror_paths:
# Join URLs of mirror roots with mirror paths. Because
# urljoin() will strip everything past the final '/' in
# the root, so we add a '/' if it is not present.
mirror_urls = []
for mirror in ramble.mirror.MirrorCollection().values():
for rel_path in self.mirror_paths:
mirror_urls.append(url_util.join(mirror.fetch_url, rel_path))
# If this archive is normally fetched from a tarball URL,
# then use the same digest. `spack mirror` ensures that
# the checksum will be the same.
digest = None
expand = True
extension = None
if isinstance(self.default_fetcher, fs.URLFetchStrategy):
digest = self.default_fetcher.digest
expand = self.default_fetcher.expand_archive
extension = self.default_fetcher.extension
# Have to skip the checksum for things archived from
# repositories. How can this be made safer?
self.skip_checksum_for_mirror = not bool(digest)
# Add URL strategies for all the mirrors with the digest
# Insert fetchers in the order that the URLs are provided.
for url in reversed(mirror_urls):
fetchers.insert(
0, fs.from_url_scheme(url, digest, expand=expand, extension=extension)
)
if self.default_fetcher.cachable:
for rel_path in reversed(list(self.mirror_paths)):
cache_fetcher = ramble.caches.fetch_cache.fetcher(
rel_path, digest, expand=expand, extension=extension
)
fetchers.insert(0, cache_fetcher)
def generate_fetchers():
yield from fetchers
# The search function may be expensive, so wait until now to
# call it so the user can stop if a prior fetcher succeeded
if self.search_fn and not mirror_only:
dynamic_fetchers = self.search_fn()
yield from dynamic_fetchers
def print_errors(errors):
for msg in errors:
logger.debug(msg)
errors = []
for fetcher in generate_fetchers():
try:
fetcher.stage = self
self.fetcher = fetcher
self.fetcher.fetch()
break
except spack.fetch_strategy.NoCacheError:
# Don't bother reporting when something is not cached.
continue
except ramble.error.RambleError as e:
errors.append(f"Fetching from {fetcher} failed.")
logger.debug(e)
continue
except spack.util.web.SpackWebError as e:
errors.append(f"Fetching from {fetcher} failed.")
logger.debug(e)
continue
else:
print_errors(errors)
self.fetcher = self.default_fetcher
raise fs.FetchError(err_msg or "All fetchers failed", None)
print_errors(errors)
[docs]
def steal_source(self, dest):
"""Copy the source_path directory in its entirety to directory dest
This operation creates/fetches/expands the stage if it is not already,
and destroys the stage when it is done."""
if not self.created:
self.create()
if not self.expanded and not self.archive_file:
self.fetch()
if not self.expanded:
self.expand_archive()
if not os.path.isdir(dest):
mkdirp(dest)
# glob all files and directories in the source path
hidden_entries = glob.glob(os.path.join(self.source_path, ".*"))
entries = glob.glob(os.path.join(self.source_path, "*"))
# Move all files from stage to destination directory
# Include hidden files for VCS repo history
for entry in hidden_entries + entries:
if os.path.isdir(entry):
d = os.path.join(dest, os.path.basename(entry))
shutil.copytree(entry, d)
else:
shutil.copy2(entry, dest)
# copy archive file if we downloaded from url -- replaces for vcs
if self.archive_file and os.path.exists(self.archive_file):
shutil.copy2(self.archive_file, dest)
# remove leftover stage
self.destroy()
[docs]
def check(self):
"""Check the downloaded archive against a checksum digest.
No-op if this stage checks code out of a repository."""
if self.fetcher is not self.default_fetcher and self.skip_checksum_for_mirror:
logger.warn(
"Fetching from mirror without a checksum!",
"This input is normally checked out from a version "
"control system, but it has been archived on a "
"mirror. This means we cannot know a checksum for the "
"tarball in advance. Be sure that your connection to "
"this mirror is secure!",
)
elif ramble.config.get("config:checksum"):
self.fetcher.check()
[docs]
def cache_local(self):
ramble.caches.fetch_cache.store(self.fetcher, self.mirror_paths.storage_path)
[docs]
def cache_mirror(self, mirror, stats):
"""Perform a fetch if the resource is not already cached
Arguments:
mirror (ramble.caches.MirrorCache): the mirror to cache this Stage's resource in
stats (ramble.mirror.MirrorStats): this is updated depending on whether the
caching operation succeeded or failed
"""
if isinstance(self.default_fetcher, fs.BundleFetchStrategy):
# BundleFetchStrategy has no source to fetch. The associated
# fetcher does nothing but the associated stage may still exist.
# There is currently no method available on the fetcher to
# distinguish this ('cachable' refers to whether the fetcher
# refers to a resource with a fixed ID, which is not the same
# concept as whether there is anything to fetch at all) so we
# must examine the type of the fetcher.
return
if not fs.stable_target(self.default_fetcher):
return
absolute_storage_path = os.path.join(mirror.root, self.mirror_paths.storage_path)
if os.path.exists(absolute_storage_path):
logger.debug(f"Already existed: {absolute_storage_path}")
stats.already_existed(absolute_storage_path)
logger.debug(f" Stats? {stats.present}")
else:
self.fetch()
self.check()
mirror.store(self.fetcher, self.mirror_paths.storage_path)
logger.debug(f"Added: {absolute_storage_path}")
stats.added(absolute_storage_path)
if not os.path.exists(absolute_storage_path):
stats.error(absolute_storage_path)
mirror.symlink(self.mirror_paths)
[docs]
def expand_archive(self):
"""Changes to the stage directory and attempt to expand the downloaded
archive. Fail if the stage is not set up or if the archive is not yet
downloaded."""
if not self.expanded:
self.fetcher.expand()
logger.debug(f"Created stage in {self.path}")
else:
logger.debug(f"Already staged {self.name} in {self.path}")
[docs]
def restage(self):
"""Removes the expanded archive path if it exists, then re-expands
the archive.
"""
self.fetcher.reset()
[docs]
def create(self):
"""
Ensures the top-level (config:build_stage) directory exists.
"""
# Emulate file permissions for tempfile.mkdtemp.
if not os.path.exists(self.path):
mkdirp(self.path, mode=stat.S_IRWXU)
elif not os.path.isdir(self.path):
os.remove(self.path)
mkdirp(self.path, mode=stat.S_IRWXU)
# Make sure we can actually do something with the stage we made.
ensure_access(self.path)
self.created = True
[docs]
def destroy(self):
"""Removes this stage directory."""
remove_linked_tree(self.path)
# Make sure we don't end up in a removed directory
try:
os.getcwd()
except OSError as e:
logger.debug(e)
os.chdir(os.path.dirname(self.path))
# mark as destroyed
self.created = False
[docs]
class ResourceStage(InputStage):
def __init__(self, url_or_fetch_strategy, root, resource, **kwargs):
super().__init__(url_or_fetch_strategy, **kwargs)
self.root_stage = root
self.resource = resource
def _add_to_root_stage(self):
"""
Move the extracted resource to the root stage (according to placement).
"""
root_stage = self.root_stage
resource = self.resource
if resource.placement:
placement = resource.placement
elif self.srcdir:
placement = self.srcdir
else:
placement = self.source_path
if not isinstance(placement, dict):
placement = {"": placement}
target_path = os.path.join(root_stage.source_path, resource.destination)
try:
os.makedirs(target_path)
except OSError as err:
logger.debug(err)
if err.errno == errno.EEXIST and os.path.isdir(target_path):
pass
else:
raise
for key, value in placement.items():
destination_path = os.path.join(target_path, value)
source_path = os.path.join(self.source_path, key)
if not os.path.exists(destination_path):
logger.info(
"Moving resource stage\n\tsource : "
"{stage}\n\tdestination : {destination}".format(
stage=source_path, destination=destination_path
)
)
src = os.path.realpath(source_path)
if os.path.isdir(src):
install_tree(src, destination_path)
else:
install(src, destination_path)
[docs]
class StageComposite(pattern.Composite):
"""Composite for Stage type objects. The first item in this composite is
considered to be the root workload, and operations that return a value are
forwarded to it."""
#
# __enter__ and __exit__ delegate to all stages in the composite.
#
def __init__(self):
super().__init__(
[
"fetch",
"create",
"created",
"check",
"expand_archive",
"restage",
"destroy",
"cache_local",
"cache_mirror",
"steal_source",
"managed_by_ramble",
]
)
def __enter__(self):
for item in self:
item.__enter__()
return self
def __exit__(self, exc_type, exc_val, exc_tb):
for item in reversed(self):
item.keep = getattr(self, "keep", False)
item.__exit__(exc_type, exc_val, exc_tb)
#
# Below functions act only on the *first* stage in the composite.
#
@property
def source_path(self):
return self[0].source_path
@property
def expanded(self):
return self[0].expanded
@property
def path(self):
return self[0].path
@property
def archive_file(self):
return self[0].archive_file
[docs]
class DIYStage:
"""
Simple class that allows any directory to be a ramble input stage.
Consequently, it does not expect or require that the source path adhere to
the standard directory naming convention.
"""
"""DIY staging is, by definition, not managed by Ramble."""
managed_by_ramble = False
def __init__(self, path):
if path is None:
raise ValueError("Cannot construct DIYStage without a path.")
elif not os.path.isdir(path):
raise StagePathError("The stage path directory does not exist:", path)
self.archive_file = None
self.path = path
self.source_path = path
self.created = True
# DIY stages do nothing as context managers.
def __enter__(self):
pass
def __exit__(self, exc_type, exc_val, exc_tb):
pass
@property
def expanded(self):
"""Returns True since the source_path must exist."""
return True
[docs]
def ensure_access(file):
"""Ensure we can access a directory and die with an error if we can't."""
if not can_access(file):
logger.die(f"Insufficient permissions for {file}")
# TODO (dwj): Need to add checksums for inputs.
[docs]
def get_checksums_for_versions(
url_dict, name, first_stage_function=None, keep_stage=False, fetch_options=None, batch=False
):
"""Fetches and checksums archives from URLs.
This function is called by both ``ramble checksum`` and ``ramble
create``. The ``first_stage_function`` argument allows the caller to
inspect the first downloaded archive, e.g., to determine the build
system.
Args:
url_dict (dict): A dictionary of the form: version -> URL
name (str): The name of the input
first_stage_function (Callable): function that takes a Stage and a URL;
this is run on the stage of the first URL downloaded
keep_stage (bool): whether to keep staging area when command completes
batch (bool): whether to ask user how many versions to fetch (false)
or fetch all versions (true)
fetch_options (dict): Options used for the fetcher (such as timeout
or cookies)
Returns:
(str): A multi-line string containing versions and corresponding hashes
"""
sorted_versions = sorted(url_dict.keys(), reverse=True)
# Find length of longest string in the list for padding
max_len = max(len(str(v)) for v in sorted_versions)
num_ver = len(sorted_versions)
logger.msg(
"Found {} version{} of {}:".format(num_ver, "" if num_ver == 1 else "s", name),
"",
*llnl.util.lang.elide_list(
["{0:{1}} {2}".format(str(v), max_len, url_dict[v]) for v in sorted_versions]
),
)
print()
if batch:
archives_to_fetch = len(sorted_versions)
else:
archives_to_fetch = tty.get_number(
"How many would you like to checksum?", default=1, abort="q"
)
if not archives_to_fetch:
logger.die("Aborted.")
versions = sorted_versions[:archives_to_fetch]
urls = [url_dict[v] for v in versions]
logger.debug("Downloading...")
version_hashes = []
i = 0
errors = []
for url, version in zip(urls, versions):
try:
if fetch_options:
url_or_fs = fs.URLFetchStrategy(url, fetch_options=fetch_options)
else:
url_or_fs = url
with InputStage(url_or_fs, keep=keep_stage) as stage:
# Fetch the archive
stage.fetch()
if i == 0 and first_stage_function:
# Only run first_stage_function the first time,
# no need to run it every time
first_stage_function(stage, url)
# Checksum the archive and add it to the list
version_hashes.append(
(version, spack.util.crypto.checksum(hashlib.sha256, stage.archive_file))
)
i += 1
except FailedDownloadError:
errors.append(f"Failed to fetch {url}")
except Exception as e:
logger.msg(f"Something failed on {url}, skipping. ({e})")
for msg in errors:
logger.debug(msg)
if not version_hashes:
logger.die(f"Could not fetch any versions for {name}")
# Find length of longest string in the list for padding
max_len = max(len(str(v)) for v, h in version_hashes)
# Generate the version directives to put in a package.py
version_lines = "\n".join(
[
" version('{}', {}sha256='{}')".format(v, " " * (max_len - len(str(v))), h)
for v, h in version_hashes
]
)
num_hash = len(version_hashes)
logger.debug(
"Checksummed {} version{} of {}:".format(num_hash, "" if num_hash == 1 else "s", name)
)
return version_lines
[docs]
class StageError(ramble.error.RambleError):
""" "Superclass for all errors encountered during staging."""
[docs]
class VersionFetchError(StageError):
"""Raised when we can't determine a URL to fetch an input."""
# Keep this in namespace for convenience
FailedDownloadError = fs.FailedDownloadError