wireproto: compress data from a generator
Currently, the "getbundle" wire protocol command obtains a generator of
data, converts it to a util.chunkbuffer, then converts it back to a
generator via the protocol's groupchunks() implementation. For the SSH
protocol, groupchunks() simply reads 4kb chunks then write()s the
data to a file descriptor. For the HTTP protocol, groupchunks() reads
32kb chunks, feeds those into a zlib compressor, emits compressed data
as it is available, and that is sent to the WSGI layer, where it is
likely turned into HTTP chunked transfer chunks as is or further
buffered and turned into a larger chunk.
For both the SSH and HTTP protocols, there is inefficiency from using
util.chunkbuffer.
For SSH, emitting consistent 4kb chunks sounds nice. However, the file
descriptor it is writing to is almost certainly buffered. That means
that a Python .write() probably doesn't translate into exactly what is
written to the I/O layer.
For HTTP, we're going through an intermediate layer to zlib compress
data. So all util.chunkbuffer is doing is ensuring that the chunks we
feed into the zlib compressor are of uniform size. This means more CPU
time in Python buffering and emitting chunks in util.chunkbuffer but
fewer function calls to zlib.
This patch introduces and implements a new wire protocol abstract
method: compresschunks(). It is like groupchunks() except it operates
on a generator instead of something with a .read(). The SSH
implementation simply proxies chunks. The HTTP implementation uses
zlib compression.
To avoid duplicate code, the HTTP groupchunks() has been reimplemented
in terms of compresschunks().
To prove this all works, the "getbundle" wire protocol command has been
switched to compresschunks(). This removes the util.chunkbuffer from
that command. Now, data essentially streams straight from the
changegroup emitter to the wire, possibly through a zlib compressor.
Generators all the way, baby.
There were slim to no performance changes on the server as measured
with the mozilla-central repository. This is likely because CPU
time is dominated by reading revlogs, producing the changegroup, and
zlib compressing the output stream. Still, this brings us a little
closer to our ideal of using generators everywhere.
# Minimal support for git commands on an hg repository
#
# Copyright 2005, 2006 Chris Mason <mason@suse.com>
#
# This software may be used and distributed according to the terms of the
# GNU General Public License version 2 or any later version.
'''browse the repository in a graphical way
The hgk extension allows browsing the history of a repository in a
graphical way. It requires Tcl/Tk version 8.4 or later. (Tcl/Tk is not
distributed with Mercurial.)
hgk consists of two parts: a Tcl script that does the displaying and
querying of information, and an extension to Mercurial named hgk.py,
which provides hooks for hgk to get information. hgk can be found in
the contrib directory, and the extension is shipped in the hgext
repository, and needs to be enabled.
The :hg:`view` command will launch the hgk Tcl script. For this command
to work, hgk must be in your search path. Alternately, you can specify
the path to hgk in your configuration file::
[hgk]
path = /location/of/hgk
hgk can make use of the extdiff extension to visualize revisions.
Assuming you had already configured extdiff vdiff command, just add::
[hgk]
vdiff=vdiff
Revisions context menu will now display additional entries to fire
vdiff on hovered and selected revisions.
'''
from __future__ import absolute_import
import os
from mercurial.i18n import _
from mercurial.node import (
nullid,
nullrev,
short,
)
from mercurial import (
cmdutil,
commands,
obsolete,
patch,
scmutil,
)
cmdtable = {}
command = cmdutil.command(cmdtable)
# Note for extension authors: ONLY specify testedwith = 'ships-with-hg-core' for
# extensions which SHIP WITH MERCURIAL. Non-mainline extensions should
# be specifying the version(s) of Mercurial they are tested with, or
# leave the attribute unspecified.
testedwith = 'ships-with-hg-core'
@command('debug-diff-tree',
[('p', 'patch', None, _('generate patch')),
('r', 'recursive', None, _('recursive')),
('P', 'pretty', None, _('pretty')),
('s', 'stdin', None, _('stdin')),
('C', 'copy', None, _('detect copies')),
('S', 'search', "", _('search'))],
('[OPTION]... NODE1 NODE2 [FILE]...'),
inferrepo=True)
def difftree(ui, repo, node1=None, node2=None, *files, **opts):
"""diff trees from two commits"""
def __difftree(repo, node1, node2, files=[]):
assert node2 is not None
mmap = repo[node1].manifest()
mmap2 = repo[node2].manifest()
m = scmutil.match(repo[node1], files)
modified, added, removed = repo.status(node1, node2, m)[:3]
empty = short(nullid)
for f in modified:
# TODO get file permissions
ui.write((":100664 100664 %s %s M\t%s\t%s\n") %
(short(mmap[f]), short(mmap2[f]), f, f))
for f in added:
ui.write((":000000 100664 %s %s N\t%s\t%s\n") %
(empty, short(mmap2[f]), f, f))
for f in removed:
ui.write((":100664 000000 %s %s D\t%s\t%s\n") %
(short(mmap[f]), empty, f, f))
##
while True:
if opts['stdin']:
try:
line = raw_input().split(' ')
node1 = line[0]
if len(line) > 1:
node2 = line[1]
else:
node2 = None
except EOFError:
break
node1 = repo.lookup(node1)
if node2:
node2 = repo.lookup(node2)
else:
node2 = node1
node1 = repo.changelog.parents(node1)[0]
if opts['patch']:
if opts['pretty']:
catcommit(ui, repo, node2, "")
m = scmutil.match(repo[node1], files)
diffopts = patch.difffeatureopts(ui)
diffopts.git = True
chunks = patch.diff(repo, node1, node2, match=m,
opts=diffopts)
for chunk in chunks:
ui.write(chunk)
else:
__difftree(repo, node1, node2, files=files)
if not opts['stdin']:
break
def catcommit(ui, repo, n, prefix, ctx=None):
nlprefix = '\n' + prefix
if ctx is None:
ctx = repo[n]
# use ctx.node() instead ??
ui.write(("tree %s\n" % short(ctx.changeset()[0])))
for p in ctx.parents():
ui.write(("parent %s\n" % p))
date = ctx.date()
description = ctx.description().replace("\0", "")
ui.write(("author %s %s %s\n" % (ctx.user(), int(date[0]), date[1])))
if 'committer' in ctx.extra():
ui.write(("committer %s\n" % ctx.extra()['committer']))
ui.write(("revision %d\n" % ctx.rev()))
ui.write(("branch %s\n" % ctx.branch()))
if obsolete.isenabled(repo, obsolete.createmarkersopt):
if ctx.obsolete():
ui.write(("obsolete\n"))
ui.write(("phase %s\n\n" % ctx.phasestr()))
if prefix != "":
ui.write("%s%s\n" % (prefix,
description.replace('\n', nlprefix).strip()))
else:
ui.write(description + "\n")
if prefix:
ui.write('\0')
@command('debug-merge-base', [], _('REV REV'))
def base(ui, repo, node1, node2):
"""output common ancestor information"""
node1 = repo.lookup(node1)
node2 = repo.lookup(node2)
n = repo.changelog.ancestor(node1, node2)
ui.write(short(n) + "\n")
@command('debug-cat-file',
[('s', 'stdin', None, _('stdin'))],
_('[OPTION]... TYPE FILE'),
inferrepo=True)
def catfile(ui, repo, type=None, r=None, **opts):
"""cat a specific revision"""
# in stdin mode, every line except the commit is prefixed with two
# spaces. This way the our caller can find the commit without magic
# strings
#
prefix = ""
if opts['stdin']:
try:
(type, r) = raw_input().split(' ')
prefix = " "
except EOFError:
return
else:
if not type or not r:
ui.warn(_("cat-file: type or revision not supplied\n"))
commands.help_(ui, 'cat-file')
while r:
if type != "commit":
ui.warn(_("aborting hg cat-file only understands commits\n"))
return 1
n = repo.lookup(r)
catcommit(ui, repo, n, prefix)
if opts['stdin']:
try:
(type, r) = raw_input().split(' ')
except EOFError:
break
else:
break
# git rev-tree is a confusing thing. You can supply a number of
# commit sha1s on the command line, and it walks the commit history
# telling you which commits are reachable from the supplied ones via
# a bitmask based on arg position.
# you can specify a commit to stop at by starting the sha1 with ^
def revtree(ui, args, repo, full="tree", maxnr=0, parents=False):
def chlogwalk():
count = len(repo)
i = count
l = [0] * 100
chunk = 100
while True:
if chunk > i:
chunk = i
i = 0
else:
i -= chunk
for x in xrange(chunk):
if i + x >= count:
l[chunk - x:] = [0] * (chunk - x)
break
if full is not None:
if (i + x) in repo:
l[x] = repo[i + x]
l[x].changeset() # force reading
else:
if (i + x) in repo:
l[x] = 1
for x in xrange(chunk - 1, -1, -1):
if l[x] != 0:
yield (i + x, full is not None and l[x] or None)
if i == 0:
break
# calculate and return the reachability bitmask for sha
def is_reachable(ar, reachable, sha):
if len(ar) == 0:
return 1
mask = 0
for i in xrange(len(ar)):
if sha in reachable[i]:
mask |= 1 << i
return mask
reachable = []
stop_sha1 = []
want_sha1 = []
count = 0
# figure out which commits they are asking for and which ones they
# want us to stop on
for i, arg in enumerate(args):
if arg.startswith('^'):
s = repo.lookup(arg[1:])
stop_sha1.append(s)
want_sha1.append(s)
elif arg != 'HEAD':
want_sha1.append(repo.lookup(arg))
# calculate the graph for the supplied commits
for i, n in enumerate(want_sha1):
reachable.append(set())
visit = [n]
reachable[i].add(n)
while visit:
n = visit.pop(0)
if n in stop_sha1:
continue
for p in repo.changelog.parents(n):
if p not in reachable[i]:
reachable[i].add(p)
visit.append(p)
if p in stop_sha1:
continue
# walk the repository looking for commits that are in our
# reachability graph
for i, ctx in chlogwalk():
if i not in repo:
continue
n = repo.changelog.node(i)
mask = is_reachable(want_sha1, reachable, n)
if mask:
parentstr = ""
if parents:
pp = repo.changelog.parents(n)
if pp[0] != nullid:
parentstr += " " + short(pp[0])
if pp[1] != nullid:
parentstr += " " + short(pp[1])
if not full:
ui.write("%s%s\n" % (short(n), parentstr))
elif full == "commit":
ui.write("%s%s\n" % (short(n), parentstr))
catcommit(ui, repo, n, ' ', ctx)
else:
(p1, p2) = repo.changelog.parents(n)
(h, h1, h2) = map(short, (n, p1, p2))
(i1, i2) = map(repo.changelog.rev, (p1, p2))
date = ctx.date()[0]
ui.write("%s %s:%s" % (date, h, mask))
mask = is_reachable(want_sha1, reachable, p1)
if i1 != nullrev and mask > 0:
ui.write("%s:%s " % (h1, mask)),
mask = is_reachable(want_sha1, reachable, p2)
if i2 != nullrev and mask > 0:
ui.write("%s:%s " % (h2, mask))
ui.write("\n")
if maxnr and count >= maxnr:
break
count += 1
# git rev-list tries to order things by date, and has the ability to stop
# at a given commit without walking the whole repo. TODO add the stop
# parameter
@command('debug-rev-list',
[('H', 'header', None, _('header')),
('t', 'topo-order', None, _('topo-order')),
('p', 'parents', None, _('parents')),
('n', 'max-count', 0, _('max-count'))],
('[OPTION]... REV...'))
def revlist(ui, repo, *revs, **opts):
"""print revisions"""
if opts['header']:
full = "commit"
else:
full = None
copy = [x for x in revs]
revtree(ui, copy, repo, full, opts['max_count'], opts['parents'])
@command('view',
[('l', 'limit', '',
_('limit number of changes displayed'), _('NUM'))],
_('[-l LIMIT] [REVRANGE]'))
def view(ui, repo, *etc, **opts):
"start interactive history viewer"
os.chdir(repo.root)
optstr = ' '.join(['--%s %s' % (k, v) for k, v in opts.iteritems() if v])
if repo.filtername is None:
optstr += '--hidden'
cmd = ui.config("hgk", "path", "hgk") + " %s %s" % (optstr, " ".join(etc))
ui.debug("running %s\n" % cmd)
ui.system(cmd)