mercurial/commandserver.py

wireproto: add streams to frame-based protocol

Previously, the frame-based protocol was just a series of frames,
with each frame associated with a request ID.

In order to scale the protocol, we'll want to enable the use of
compression. While it is possible to enable compression at the
socket/pipe level, this has its disadvantages. The big one is it
undermines the point of frames being standalone, atomic units that
can be read and written: if you add compression above the framing
protocol, you are back to having a stream-based protocol as opposed
to something frame-based.

So in order to preserve frames, compression needs to occur at
the frame payload level.

Compressing each frame's payload individually will limit compression
ratios because the window size of the compressor will be limited
by the max frame size, which is 32-64kb as currently defined. It
will also add CPU overhead, as it is more efficient for compressors
to operate on fewer, larger blocks of data than more, smaller blocks.

So compressing each frame independently is out.

This means we need to compress each frame's payload as if it is part
of a larger stream.

The simplest approach is to have 1 stream per connection. This
could certainly work. However, it has disadvantages (documented below).

We could also have 1 stream per RPC/command invocation. (This is the
model HTTP/2 goes with.) This also has disadvantages.

The main disadvantage to one global stream is that it has the very
real potential to create CPU bottlenecks doing compression. Networks
are only getting faster and the performance of single CPU cores has
been relatively flat. Newer compression formats like zstandard offer
better CPU cycle efficiency than predecessors like zlib. But it still
all too common to saturate your CPU with compression overhead long
before you saturate the network pipe.

The main disadvantage with streams per request is that you can't
reap the benefits of the compression context for multiple requests.
For example, if you send 1000 RPC requests (or HTTP/2 requests for
that matter), the response to each would have its own compression
context. The overall size of the raw responses would be larger because
compression contexts wouldn't be able to reference data from another
request or response.

The approach for streams as implemented in this commit is to support
N streams per connection and for streams to potentially span requests
and responses. As explained by the added internals docs, this
facilitates servers and clients delegating independent streams and
compression to independent threads / CPU cores. This helps alleviate
the CPU bottleneck of compression. This design also allows compression
contexts to be reused across requests/responses. This can result in
improved compression ratios and less overhead for compressors and
decompressors having to build new contexts.

Another feature that was defined was the ability for individual frames
within a stream to declare whether that individual frame's payload
uses the content encoding (read: compression) defined by the stream.
The idea here is that some servers may serve data from a combination
of caches and dynamic resolution. Data coming from caches may be
pre-compressed. We want to facilitate servers being able to essentially
stream bytes from caches to the wire with minimal overhead. Being
able to mix and match with frames are compressed within a stream
enables these types of advanced server functionality.

This commit defines the new streams mechanism. Basic code for
supporting streams in frames has been added. But that code is
seriously lacking and doesn't fully conform to the defined protocol.
For example, we don't close any streams. And support for content
encoding within streams is not yet implemented. The change was
rather invasive and I didn't think it would be reasonable to implement
the entire feature in a single commit.

For the record, I would have loved to reuse an existing multiplexing
protocol to build the new wire protocol on top of. However, I couldn't
find a protocol that offers the performance and scaling characteristics
that I desired. Namely, it should support multiple compression
contexts to facilitate scaling out to multiple CPU cores and
compression contexts should be able to live longer than single RPC
requests. HTTP/2 *almost* fits the bill. But the semantics of HTTP
message exchange state that streams can only live for a single
request-response. We /could/ tunnel on top of HTTP/2 streams and
frames with HEADER and DATA frames. But there's no guarantee that
HTTP/2 libraries and proxies would allow us to use HTTP/2 streams
and frames without the HTTP message exchange semantics defined in
RFC 7540 Section 8. Other RPC protocols like gRPC tunnel are built
on top of HTTP/2 and thus preserve its semantics of stream per
RPC invocation. Even QUIC does this. We could attempt to invent a
higher-level stream that spans HTTP/2 streams. But this would be
violating HTTP/2 because there is no guarantee that HTTP/2 streams
are routed to the same server. The best we can do - which is what
this protocol does - is shoehorn all request and response data into
a single HTTP message and create streams within. At that point, we've
defined a Content-Type in HTTP parlance. It just so happens our
media type can also work as a standalone, stream-based protocol,
without leaning on HTTP or similar protocol.

Differential Revision: https://phab.mercurial-scm.org/D2907

# commandserver.py - communicate with Mercurial's API over a pipe
#
#  Copyright Matt Mackall <[e-mail]>
#
# This software may be used and distributed according to the terms of the
# GNU General Public License version 2 or any later version.

from __future__ import absolute_import

import errno
import gc
import os
import random
import signal
import socket
import struct
import traceback

try:
    import selectors
    selectors.BaseSelector
except ImportError:
    from .thirdparty import selectors2 as selectors

from .i18n import _
from . import (
    encoding,
    error,
    pycompat,
    util,
)
from .utils import (
    procutil,
)

logfile = None

def log(*args):
    if not logfile:
        return

    for a in args:
        logfile.write(str(a))

    logfile.flush()

class channeledoutput(object):
    """
    Write data to out in the following format:

    data length (unsigned int),
    data
    """
    def __init__(self, out, channel):
        self.out = out
        self.channel = channel

    @property
    def name(self):
        return '<%c-channel>' % self.channel

    def write(self, data):
        if not data:
            return
        # single write() to guarantee the same atomicity as the underlying file
        self.out.write(struct.pack('>cI', self.channel, len(data)) + data)
        self.out.flush()

    def __getattr__(self, attr):
        if attr in ('isatty', 'fileno', 'tell', 'seek'):
            raise AttributeError(attr)
        return getattr(self.out, attr)

class channeledinput(object):
    """
    Read data from in_.

    Requests for input are written to out in the following format:
    channel identifier - 'I' for plain input, 'L' line based (1 byte)
    how many bytes to send at most (unsigned int),

    The client replies with:
    data length (unsigned int), 0 meaning EOF
    data
    """

    maxchunksize = 4 * 1024

    def __init__(self, in_, out, channel):
        self.in_ = in_
        self.out = out
        self.channel = channel

    @property
    def name(self):
        return '<%c-channel>' % self.channel

    def read(self, size=-1):
        if size < 0:
            # if we need to consume all the clients input, ask for 4k chunks
            # so the pipe doesn't fill up risking a deadlock
            size = self.maxchunksize
            s = self._read(size, self.channel)
            buf = s
            while s:
                s = self._read(size, self.channel)
                buf += s

            return buf
        else:
            return self._read(size, self.channel)

    def _read(self, size, channel):
        if not size:
            return ''
        assert size > 0

        # tell the client we need at most size bytes
        self.out.write(struct.pack('>cI', channel, size))
        self.out.flush()

        length = self.in_.read(4)
        length = struct.unpack('>I', length)[0]
        if not length:
            return ''
        else:
            return self.in_.read(length)

    def readline(self, size=-1):
        if size < 0:
            size = self.maxchunksize
            s = self._read(size, 'L')
            buf = s
            # keep asking for more until there's either no more or
            # we got a full line
            while s and s[-1] != '\n':
                s = self._read(size, 'L')
                buf += s

            return buf
        else:
            return self._read(size, 'L')

    def __iter__(self):
        return self

    def next(self):
        l = self.readline()
        if not l:
            raise StopIteration
        return l

    def __getattr__(self, attr):
        if attr in ('isatty', 'fileno', 'tell', 'seek'):
            raise AttributeError(attr)
        return getattr(self.in_, attr)

class server(object):
    """
    Listens for commands on fin, runs them and writes the output on a channel
    based stream to fout.
    """
    def __init__(self, ui, repo, fin, fout):
        self.cwd = pycompat.getcwd()

        # developer config: cmdserver.log
        logpath = ui.config("cmdserver", "log")
        if logpath:
            global logfile
            if logpath == '-':
                # write log on a special 'd' (debug) channel
                logfile = channeledoutput(fout, 'd')
            else:
                logfile = open(logpath, 'a')

        if repo:
            # the ui here is really the repo ui so take its baseui so we don't
            # end up with its local configuration
            self.ui = repo.baseui
            self.repo = repo
            self.repoui = repo.ui
        else:
            self.ui = ui
            self.repo = self.repoui = None

        self.cerr = channeledoutput(fout, 'e')
        self.cout = channeledoutput(fout, 'o')
        self.cin = channeledinput(fin, fout, 'I')
        self.cresult = channeledoutput(fout, 'r')

        self.client = fin

    def cleanup(self):
        """release and restore resources taken during server session"""

    def _read(self, size):
        if not size:
            return ''

        data = self.client.read(size)

        # is the other end closed?
        if not data:
            raise EOFError

        return data

    def _readstr(self):
        """read a string from the channel

        format:
        data length (uint32), data
        """
        length = struct.unpack('>I', self._read(4))[0]
        if not length:
            return ''
        return self._read(length)

    def _readlist(self):
        """read a list of NULL separated strings from the channel"""
        s = self._readstr()
        if s:
            return s.split('\0')
        else:
            return []

    def runcommand(self):
        """ reads a list of \0 terminated arguments, executes
        and writes the return code to the result channel """
        from . import dispatch  # avoid cycle

        args = self._readlist()

        # copy the uis so changes (e.g. --config or --verbose) don't
        # persist between requests
        copiedui = self.ui.copy()
        uis = [copiedui]
        if self.repo:
            self.repo.baseui = copiedui
            # clone ui without using ui.copy because this is protected
            repoui = self.repoui.__class__(self.repoui)
            repoui.copy = copiedui.copy # redo copy protection
            uis.append(repoui)
            self.repo.ui = self.repo.dirstate._ui = repoui
            self.repo.invalidateall()

        for ui in uis:
            ui.resetstate()
            # any kind of interaction must use server channels, but chg may
            # replace channels by fully functional tty files. so nontty is
            # enforced only if cin is a channel.
            if not util.safehasattr(self.cin, 'fileno'):
                ui.setconfig('ui', 'nontty', 'true', 'commandserver')

        req = dispatch.request(args[:], copiedui, self.repo, self.cin,
                               self.cout, self.cerr)

        try:
            ret = (dispatch.dispatch(req) or 0) & 255 # might return None
            self.cresult.write(struct.pack('>i', int(ret)))
        finally:
            # restore old cwd
            if '--cwd' in args:
                os.chdir(self.cwd)

    def getencoding(self):
        """ writes the current encoding to the result channel """
        self.cresult.write(encoding.encoding)

    def serveone(self):
        cmd = self.client.readline()[:-1]
        if cmd:
            handler = self.capabilities.get(cmd)
            if handler:
                handler(self)
            else:
                # clients are expected to check what commands are supported by
                # looking at the servers capabilities
                raise error.Abort(_('unknown command %s') % cmd)

        return cmd != ''

    capabilities = {'runcommand': runcommand,
                    'getencoding': getencoding}

    def serve(self):
        hellomsg = 'capabilities: ' + ' '.join(sorted(self.capabilities))
        hellomsg += '\n'
        hellomsg += 'encoding: ' + encoding.encoding
        hellomsg += '\n'
        hellomsg += 'pid: %d' % procutil.getpid()
        if util.safehasattr(os, 'getpgid'):
            hellomsg += '\n'
            hellomsg += 'pgid: %d' % os.getpgid(0)

        # write the hello msg in -one- chunk
        self.cout.write(hellomsg)

        try:
            while self.serveone():
                pass
        except EOFError:
            # we'll get here if the client disconnected while we were reading
            # its request
            return 1

        return 0

class pipeservice(object):
    def __init__(self, ui, repo, opts):
        self.ui = ui
        self.repo = repo

    def init(self):
        pass

    def run(self):
        ui = self.ui
        # redirect stdio to null device so that broken extensions or in-process
        # hooks will never cause corruption of channel protocol.
        with procutil.protectedstdio(ui.fin, ui.fout) as (fin, fout):
            try:
                sv = server(ui, self.repo, fin, fout)
                return sv.serve()
            finally:
                sv.cleanup()

def _initworkerprocess():
    # use a different process group from the master process, in order to:
    # 1. make the current process group no longer "orphaned" (because the
    #    parent of this process is in a different process group while
    #    remains in a same session)
    #    according to POSIX 2.2.2.52, orphaned process group will ignore
    #    terminal-generated stop signals like SIGTSTP (Ctrl+Z), which will
    #    cause trouble for things like ncurses.
    # 2. the client can use kill(-pgid, sig) to simulate terminal-generated
    #    SIGINT (Ctrl+C) and process-exit-generated SIGHUP. our child
    #    processes like ssh will be killed properly, without affecting
    #    unrelated processes.
    os.setpgid(0, 0)
    # change random state otherwise forked request handlers would have a
    # same state inherited from parent.
    random.seed()

def _serverequest(ui, repo, conn, createcmdserver):
    fin = conn.makefile('rb')
    fout = conn.makefile('wb')
    sv = None
    try:
        sv = createcmdserver(repo, conn, fin, fout)
        try:
            sv.serve()
        # handle exceptions that may be raised by command server. most of
        # known exceptions are caught by dispatch.
        except error.Abort as inst:
            ui.warn(_('abort: %s\n') % inst)
        except IOError as inst:
            if inst.errno != errno.EPIPE:
                raise
        except KeyboardInterrupt:
            pass
        finally:
            sv.cleanup()
    except: # re-raises
        # also write traceback to error channel. otherwise client cannot
        # see it because it is written to server's stderr by default.
        if sv:
            cerr = sv.cerr
        else:
            cerr = channeledoutput(fout, 'e')
        traceback.print_exc(file=cerr)
        raise
    finally:
        fin.close()
        try:
            fout.close()  # implicit flush() may cause another EPIPE
        except IOError as inst:
            if inst.errno != errno.EPIPE:
                raise

class unixservicehandler(object):
    """Set of pluggable operations for unix-mode services

    Almost all methods except for createcmdserver() are called in the main
    process. You can't pass mutable resource back from createcmdserver().
    """

    pollinterval = None

    def __init__(self, ui):
        self.ui = ui

    def bindsocket(self, sock, address):
        util.bindunixsocket(sock, address)
        sock.listen(socket.SOMAXCONN)
        self.ui.status(_('listening at %s\n') % address)
        self.ui.flush()  # avoid buffering of status message

    def unlinksocket(self, address):
        os.unlink(address)

    def shouldexit(self):
        """True if server should shut down; checked per pollinterval"""
        return False

    def newconnection(self):
        """Called when main process notices new connection"""

    def createcmdserver(self, repo, conn, fin, fout):
        """Create new command server instance; called in the process that
        serves for the current connection"""
        return server(self.ui, repo, fin, fout)

class unixforkingservice(object):
    """
    Listens on unix domain socket and forks server per connection
    """

    def __init__(self, ui, repo, opts, handler=None):
        self.ui = ui
        self.repo = repo
        self.address = opts['address']
        if not util.safehasattr(socket, 'AF_UNIX'):
            raise error.Abort(_('unsupported platform'))
        if not self.address:
            raise error.Abort(_('no socket path specified with --address'))
        self._servicehandler = handler or unixservicehandler(ui)
        self._sock = None
        self._oldsigchldhandler = None
        self._workerpids = set()  # updated by signal handler; do not iterate
        self._socketunlinked = None

    def init(self):
        self._sock = socket.socket(socket.AF_UNIX)
        self._servicehandler.bindsocket(self._sock, self.address)
        if util.safehasattr(procutil, 'unblocksignal'):
            procutil.unblocksignal(signal.SIGCHLD)
        o = signal.signal(signal.SIGCHLD, self._sigchldhandler)
        self._oldsigchldhandler = o
        self._socketunlinked = False

    def _unlinksocket(self):
        if not self._socketunlinked:
            self._servicehandler.unlinksocket(self.address)
            self._socketunlinked = True

    def _cleanup(self):
        signal.signal(signal.SIGCHLD, self._oldsigchldhandler)
        self._sock.close()
        self._unlinksocket()
        # don't kill child processes as they have active clients, just wait
        self._reapworkers(0)

    def run(self):
        try:
            self._mainloop()
        finally:
            self._cleanup()

    def _mainloop(self):
        exiting = False
        h = self._servicehandler
        selector = selectors.DefaultSelector()
        selector.register(self._sock, selectors.EVENT_READ)
        while True:
            if not exiting and h.shouldexit():
                # clients can no longer connect() to the domain socket, so
                # we stop queuing new requests.
                # for requests that are queued (connect()-ed, but haven't been
                # accept()-ed), handle them before exit. otherwise, clients
                # waiting for recv() will receive ECONNRESET.
                self._unlinksocket()
                exiting = True
            ready = selector.select(timeout=h.pollinterval)
            if not ready:
                # only exit if we completed all queued requests
                if exiting:
                    break
                continue
            try:
                conn, _addr = self._sock.accept()
            except socket.error as inst:
                if inst.args[0] == errno.EINTR:
                    continue
                raise

            pid = os.fork()
            if pid:
                try:
                    self.ui.debug('forked worker process (pid=%d)\n' % pid)
                    self._workerpids.add(pid)
                    h.newconnection()
                finally:
                    conn.close()  # release handle in parent process
            else:
                try:
                    self._runworker(conn)
                    conn.close()
                    os._exit(0)
                except:  # never return, hence no re-raises
                    try:
                        self.ui.traceback(force=True)
                    finally:
                        os._exit(255)
        selector.close()

    def _sigchldhandler(self, signal, frame):
        self._reapworkers(os.WNOHANG)

    def _reapworkers(self, options):
        while self._workerpids:
            try:
                pid, _status = os.waitpid(-1, options)
            except OSError as inst:
                if inst.errno == errno.EINTR:
                    continue
                if inst.errno != errno.ECHILD:
                    raise
                # no child processes at all (reaped by other waitpid()?)
                self._workerpids.clear()
                return
            if pid == 0:
                # no waitable child processes
                return
            self.ui.debug('worker process exited (pid=%d)\n' % pid)
            self._workerpids.discard(pid)

    def _runworker(self, conn):
        signal.signal(signal.SIGCHLD, self._oldsigchldhandler)
        _initworkerprocess()
        h = self._servicehandler
        try:
            _serverequest(self.ui, self.repo, conn, h.createcmdserver)
        finally:
            gc.collect()  # trigger __del__ since worker process uses os._exit