dnf-json

#!/usr/bin/python3

import datetime
import hashlib
import json
import tempfile
import os
import socket
import socketserver
import logging
import sys
from http.server import BaseHTTPRequestHandler
import pathlib
from multiprocessing import Lock
import queue
import shutil
from datetime import datetime, timedelta

import dnf
import hawkey
import pickle

# Logging setup (to systemd if available)
formatter = logging.Formatter(
        fmt="%(asctime)s %(name)s.%(levelname)s: %(message)s",
        datefmt="%Y.%m.%d %H:%M:%S"
        )
handler = logging.StreamHandler(stream=sys.stdout)
handler.setFormatter(formatter)
log = logging.getLogger('dnf-json')
log.addHandler(handler)
log.setLevel(logging.INFO)

# Synchronisation necessary for the multiprocess request handling.
process_lock = Lock()

class CacheState():
    """
    A CacheState keeps track of the cache folders.
    CacheState objects can update the list of used cache folders, which
    will associate a date object to the name of the folder.
    CacheState objects can ask to clean the unused cache folders.
    Meaning that the folders having reach a certain timeout will be deleted.
    """

    def __init__(self, cache_dir, cache_timeout, folder_dict={}):
        """
        cache_dir the place where the state is stored
        folder_dict a dict containing the existing list of cache folders
        cache_timeout a timedelta before a cache folder can be deleted
        """
        self.cache_dir = cache_dir
        self.folder_dict = folder_dict
        self.cache_timeout = cache_timeout

    def update_used(self, folder):
        """
        Associate a datetime.now() to the folders given as parameters
        """
        log.debug("Folder %s was used", folder)
        self.folder_dict[folder] = datetime.now()

    def clean_unused(self):
        """
        Delete the folders having reach the timeout
        """
        log.info("clean unused folders")
        now = datetime.now()
        list_folder_to_delete = []
        for folder, then in self.folder_dict.items():
            delta = now - then
            log.debug("delete %s if %s > than %s", folder, delta, self.cache_timeout)
            if delta > self.cache_timeout:
                list_folder_to_delete.append(folder)
        for folder in list_folder_to_delete:
            del self.folder_dict[folder]
            shutil.rmtree(folder)

    @staticmethod
    def load_cache_state_from_disk(cache_dir):
        try:
            with open(os.path.join(cache_dir,"cache_state.pkl"), "rb") as inp:
                return pickle.load(inp)
        except FileNotFoundError:
            return CacheState(cache_dir, timedelta(hours=24))

    def store_on_disk(self):
        with open(os.path.join(self.cache_dir, "cache_state.pkl"), "wb") as outp:
            return pickle.dump(self, outp)

class Solver():

    def __init__(self, repos, module_platform_id, persistdir, cachedir, arch):
        self.base = dnf.Base()

        # Enable fastestmirror to ensure we choose the fastest mirrors for
        # downloading metadata (when depsolving) and downloading packages.
        self.base.conf.fastestmirror = True

        # We use the same cachedir for multiple architectures. Unfortunately,
        # this is something that doesn't work well in certain situations
        # with zchunk:
        # Imagine that we already have cache for arch1. Then, we use dnf-json
        # to depsolve for arch2. If ZChunk is enabled and available (that's
        # the case for Fedora), dnf will try to download only differences
        # between arch1 and arch2 metadata. But, as these are completely
        # different, dnf must basically redownload everything.
        # For downloding deltas, zchunk uses HTTP range requests. Unfortunately,
        # if the mirror doesn't support multi range requests, then zchunk will
        # download one small segment per a request. Because we need to update
        # the whole metadata (10s of MB), this can be extremely slow in some cases.
        # I think that we can come up with a better fix but let's just disable
        # zchunk for now. As we are already downloading a lot of data when
        # building images, I don't care if we download even more.
        self.base.conf.zchunk = False

        # Try another mirror if it takes longer than 5 seconds to connect.
        self.base.conf.timeout = 5

        # Set the rest of the dnf configuration.
        self.base.conf.module_platform_id = module_platform_id
        self.base.conf.config_file_path = "/dev/null"
        self.base.conf.persistdir = persistdir
        self.base.conf.cachedir = cachedir
        self.base.conf.substitutions['arch'] = arch
        self.base.conf.substitutions['basearch'] = dnf.rpm.basearch(arch)

        for repo in repos:
            self.base.repos.add(self._dnfrepo(repo, self.base.conf))
        self.base.fill_sack(load_system_repo=False)

    def _dnfrepo(self, desc, parent_conf=None):
        """Makes a dnf.repo.Repo out of a JSON repository description"""

        repo = dnf.repo.Repo(desc["id"], parent_conf)

        if "baseurl" in desc:
            repo.baseurl = desc["baseurl"]
        elif "metalink" in desc:
            repo.metalink = desc["metalink"]
        elif "mirrorlist" in desc:
            repo.mirrorlist = desc["mirrorlist"]
        else:
            assert False

        if desc.get("ignoressl", False):
            repo.sslverify = False
        if "sslcacert" in desc:
            repo.sslcacert = desc["sslcacert"]
        if "sslclientkey" in desc:
            repo.sslclientkey = desc["sslclientkey"]
        if "sslclientcert" in desc:
            repo.sslclientcert = desc["sslclientcert"]

        # In dnf, the default metadata expiration time is 48 hours. However,
        # some repositories never expire the metadata, and others expire it much
        # sooner than that. We therefore allow this to be configured. If nothing
        # is provided we error on the side of checking if we should invalidate
        # the cache. If cache invalidation is not necessary, the overhead of
        # checking is in the hundreds of milliseconds. In order to avoid this
        # overhead accumulating for API calls that consist of several dnf calls,
        # we set the expiration to a short time period, rather than 0.
        repo.metadata_expire = desc.get("metadata_expire", "20s")

        return repo

    def _repo_checksums(self):
        checksums = {}
        for repo in self.base.repos.iter_enabled():
            # Uses the same algorithm as libdnf to find cache dir:
            #   https://github.com/rpm-software-management/libdnf/blob/master/libdnf/repo/Repo.cpp#L1288
            if repo.metalink:
                url = repo.metalink
            elif repo.mirrorlist:
                url = repo.mirrorlist
            elif repo.baseurl:
                url = repo.baseurl[0]
            else:
                assert False

            digest = hashlib.sha256(url.encode()).hexdigest()[:16]

            repomd_file = f"{repo.id}-{digest}/repodata/repomd.xml"
            with open(f"{self.base.conf.cachedir}/{repomd_file}", "rb") as f:
                repomd = f.read()

            checksums[repo.id] = "sha256:" + hashlib.sha256(repomd).hexdigest()

        return checksums

    def _timestamp_to_rfc3339(self, timestamp):
        d = datetime.utcfromtimestamp(timestamp)
        return d.strftime('%Y-%m-%dT%H:%M:%SZ')

    def dump(self):
        packages = []
        for package in self.base.sack.query().available():
            packages.append({
                "name": package.name,
                "summary": package.summary,
                "description": package.description,
                "url": package.url,
                "epoch": package.epoch,
                "version": package.version,
                "release": package.release,
                "arch": package.arch,
                "buildtime": self._timestamp_to_rfc3339(package.buildtime),
                "license": package.license
            })
        return {
            "checksums": self._repo_checksums(),
            "packages": packages
        }

    def depsolve(self, package_spec, exclude_spec):
        self.base.install_specs(package_spec, exclude_spec)
        self.base.resolve()
        dependencies = []
        for tsi in self.base.transaction:
            # Avoid using the install_set() helper, as it does not guarantee
            # a stable order
            if tsi.action not in dnf.transaction.FORWARD_ACTIONS:
                continue
            package = tsi.pkg

            dependencies.append({
                "name": package.name,
                "epoch": package.epoch,
                "version": package.version,
                "release": package.release,
                "arch": package.arch,
                "repo_id": package.reponame,
                "path": package.relativepath,
                "remote_location": package.remote_location(),
                "checksum": (
                    f"{hawkey.chksum_name(package.chksum[0])}:"
                    f"{package.chksum[1].hex()}"
                )
            })
        return {
            "checksums": self._repo_checksums(),
            "dependencies": dependencies
        }

class DnfJsonRequestHandler(BaseHTTPRequestHandler):
    """
    Answers Http requests to depsolve or dump packages.
    """


    def init_cache_folder_list(self, repos):
        cache_folders = []
        for repo in repos:
            if "baseurl" in repo:
                url = repo["baseurl"]
            elif "metalink" in repo:
                url = repo["metalink"]
            elif "mirrorlist" in repo:
                url = repo["mirrorlist"]
            else:
                assert False
            digest = hashlib.sha256(url.encode()).hexdigest()[:16]
            repoid = repo["id"]
            cache_folders.append(f"{self.cache_dir}/{repoid}-{digest}")
        return cache_folders

    def _send(self):
        self.client_address=('',)

    def response_with_dnf_error(self, kind: str, reason: str):
        self._send()
        self.send_response(500)
        self.send_header("Content-Type", "application/json")
        self.end_headers()
        self.wfile.write(json.dumps({"kind": kind, "reason":
            reason}).encode("utf-8"))

    def response_success(self, json_object):
        self._send()
        self.send_response(200)
        self.send_header("Content-Type", "application/json")
        self.end_headers()
        self.wfile.write(json.dumps(json_object).encode("utf-8"))

    def do_POST(self):
        """
        do_Post answer the request of a depsolving/dump.
        Depsolving and dump require synchronizing caches on disk to perform their
        operations. Caches folders are created for each remote repository. Since
        the user request contains the list of repository the space taken can grow
        indefinitely.

        This solution implements a cache cleaning mechanism. It works by deleting
        the directories on a timeout based rule and by keeping the last used date
        in a synced file on disks.

        This method is multiprocess safe.
        """
        with process_lock:
            try:
                content_len = int(self.headers.get('Content-Length'))
                data = self.rfile.read(content_len)
                call = json.loads(data.decode("utf-8"))
                command = call["command"]
                arguments = call["arguments"]
                repos = arguments.get("repos", {})
                arch = arguments["arch"]
                self.cache_dir = arguments["cachedir"]
                cache_state = CacheState.load_cache_state_from_disk(self.cache_dir)
                module_platform_id = arguments["module_platform_id"]

                with tempfile.TemporaryDirectory() as persistdir:
                    try:
                        solver = Solver(
                            repos,
                            module_platform_id,
                            persistdir,
                            self.cache_dir,
                            arch
                        )
                        if command == "dump":
                            self.response_success(solver.dump())
                            log.info("dump success")
                        elif command == "depsolve":
                            self.response_success(
                                    solver.depsolve(
                                        arguments["package-specs"],
                                        arguments.get("exclude-specs", [])
                                        )
                                    )
                            log.info("depsolve success")

                    except dnf.exceptions.MarkingErrors as e:
                        log.info("error install_specs")
                        self.response_with_dnf_error(
                            "MarkingErrors",
                            f"Error occurred when marking packages for installation: {e}"
                        )
                    except dnf.exceptions.DepsolveError as e:
                        log.info("error depsolve")
                        self.response_with_dnf_error(
                            "DepsolveError",
                            (
                                "There was a problem depsolving "
                                f"{arguments['package-specs']}: {e}"
                            )
                        )
                    except dnf.exceptions.Error as e:
                        self.response_with_dnf_error(
                            type(e).__name__,
                            f"Error occurred when setting up repo: {e}")
            finally:
                for cache_folder in self.init_cache_folder_list(repos):
                    cache_state.update_used(cache_folder)
                cache_state.clean_unused()
                cache_state.store_on_disk()

log.info("Starting the dnf-json server")

LISTEN_FDS = int(os.environ.get("LISTEN_FDS", 0))
SOCK_PATH = "/run/osbuild-dnf-json/"
SOCK_NAME = "api.sock"

# The dnf-json web server has to use forks to serve the requests. Because the
# dnf library is leaking memory in its Cpp side.
class SystemDActivationSocketServer(socketserver.ForkingMixIn, socketserver.UnixStreamServer):
    def server_bind(self):
        log.debug("service bind")
        if LISTEN_FDS == 0:
            log.debug("create new socket")
            socketserver.UnixStreamServer.server_bind(self)
        else:
            log.debug("rebind socket")
            log.debug("address_family: %d ", self.address_family)
            log.debug("socket_type: %d ", self.socket_type)
            self.socket = socket.fromfd(3, self.address_family, self.socket_type)

# start the web server
pathlib.Path(SOCK_PATH).mkdir(parents=True, exist_ok=True)
server = SystemDActivationSocketServer(f"{SOCK_PATH}{SOCK_NAME}", DnfJsonRequestHandler)
server.serve_forever()