sapling/hgsql.py
Durham Goode d8a2f1ef92 Allow lazily taking the lock in bundle2 pushes
Summary:
In order to increase commit throughput on the server side, we need to be able to
run hooks and do caching outside of the lock. Upstream Mercurial and the
pushrebase extension are doing this by making the unbundle lock possible to
acquire lazily.

For hgsql, this means we can't just wrap the entire unbundle command in a lock.
Instead we need to take the lock only when it's requested.

Test Plan:
Wrote a test, and changed the existing bundle2 test to use the lazy
locking

Reviewers: #sourcecontrol

Differential Revision: https://phabricator.fb.com/D2525414
2015-10-09 13:13:18 -07:00

1385 lines
51 KiB
Python

# hgsql.py
#
# Copyright 2013 Facebook, Inc.
#
# This software may be used and distributed according to the terms of the
# GNU General Public License version 2 or any later version.
"""
CREATE TABLE revisions(
repo CHAR(64) BINARY NOT NULL,
path VARCHAR(512) BINARY NOT NULL,
chunk INT UNSIGNED NOT NULL,
chunkcount INT UNSIGNED NOT NULL,
linkrev INT UNSIGNED NOT NULL,
rev INT UNSIGNED NOT NULL,
node CHAR(40) BINARY NOT NULL,
entry BINARY(64) NOT NULL,
data0 CHAR(1) NOT NULL,
data1 LONGBLOB NOT NULL,
createdtime DATETIME NOT NULL,
INDEX linkrevs (repo, linkrev),
PRIMARY KEY (repo, path, rev, chunk)
);
CREATE TABLE revision_references(
autoid INT UNSIGNED NOT NULL AUTO_INCREMENT PRIMARY KEY,
repo CHAR(32) BINARY NOT NULL,
namespace CHAR(32) BINARY NOT NULL,
name VARCHAR(256) BINARY,
value char(40) BINARY NOT NULL,
UNIQUE KEY bookmarkindex (repo, namespace, name)
);
"""
from mercurial.node import bin, hex, nullid, nullrev
from mercurial.i18n import _
from mercurial.extensions import wrapfunction, wrapcommand
from mercurial import changelog, error, cmdutil, revlog, localrepo, transaction
from mercurial import wireproto, bookmarks, repair, commands, hg, mdiff, phases
from mercurial import util, changegroup, exchange, bundle2
import MySQLdb, struct, time, Queue, threading, _mysql_exceptions
from MySQLdb import cursors
import warnings
import sys
cmdtable = {}
command = cmdutil.command(cmdtable)
testedwith = 'internal'
writelock = 'write_lock'
INITIAL_SYNC_NORMAL = 'normal'
INITIAL_SYNC_DISABLE = 'disabled'
INITIAL_SYNC_FORCE = 'force'
initialsync = INITIAL_SYNC_NORMAL
class CorruptionException(Exception):
pass
def uisetup(ui):
# Enable SQL for local commands that write to the repository.
wrapcommand(commands.table, 'pull', pull)
wrapcommand(commands.table, 'commit', commit)
wrapcommand(commands.table, 'bookmark', bookmarkcommand)
wrapfunction(exchange, '_localphasemove', _localphasemove)
wrapfunction(exchange, 'push', push)
# Enable SQL for remote commands that write to the repository
wireproto.commands['unbundle'] = (wireproto.unbundle, 'heads')
wrapfunction(exchange, 'unbundle', unbundle)
wrapfunction(wireproto, 'pushkey', pushkey)
wireproto.commands['pushkey'] = (wireproto.pushkey, 'namespace key old new')
wrapfunction(bookmarks, 'updatefromremote', updatefromremote)
wrapfunction(changegroup, 'addchangegroup', addchangegroup)
# Record revlog writes
def writeentry(orig, self, transaction, ifh, dfh, entry, data, link, offset):
"""records each revlog write to the repo's pendingrev list"""
if not util.safehasattr(transaction, "repo"):
return orig(self, transaction, ifh, dfh, entry, data, link, offset)
e = struct.unpack(revlog.indexformatng, entry)
node = hex(e[7])
data0 = data[0] or ''
transaction.repo.pendingrevs.append((self.indexfile, link,
len(self) - 1, node, entry, data0, data[1]))
return orig(self, transaction, ifh, dfh, entry, data, link, offset)
wrapfunction(revlog.revlog, '_writeentry', writeentry)
# Reorder incoming revs to be in linkrev order
wrapfunction(revlog.revlog, 'addgroup', addgroup)
# Write SQL bookmarks at the same time as local bookmarks
wrapfunction(bookmarks.bmstore, 'write', bookmarkwrite)
def extsetup(ui):
if ui.configbool('hgsql', 'enabled'):
commands.globalopts.append(
('', 'forcesync', False,
_('force hgsql sync even on read-only commands'),
_('TYPE')))
# Directly examining argv seems like a terrible idea, but it seems
# neccesary unless we refactor mercurial dispatch code. This is because
# the first place we have access to parsed options is in the same function
# (dispatch.dispatch) that created the repo and the repo creation initiates
# the sync operation in which the lock is elided unless we set this.
if '--forcesync' in sys.argv:
ui.debug('forcesync enabled\n')
global initialsync
initialsync = INITIAL_SYNC_FORCE
def reposetup(ui, repo):
if repo.ui.configbool("hgsql", "enabled"):
wraprepo(repo)
if initialsync != INITIAL_SYNC_DISABLE:
# Use a noop to force a sync
def noop():
pass
waitforlock = (initialsync == INITIAL_SYNC_FORCE)
executewithsql(repo, noop, waitforlock=waitforlock)
# Incoming commits are only allowed via push and pull
def unbundle(orig, repo, cg, *args, **kwargs):
if not repo.ui.configbool("hgsql", "enabled"):
return orig(repo, cg, *args, **kwargs)
isbundle2 = util.safehasattr(cg, 'params')
islazylocking = repo.ui.configbool('experimental', 'bundle2lazylocking')
if isbundle2 and islazylocking:
# lazy locked bundle2
exception = None
oldopclass = None
context = None
try:
context = sqlcontext(repo, takelock=True, waitforlock=True)
# Temporarily replace bundleoperation so we can hook into it's
# locking mechanism.
oldopclass = bundle2.bundleoperation
class sqllockedoperation(bundle2.bundleoperation):
def __init__(self, repo, transactiongetter, *args, **kwargs):
def sqllocktr():
if not context.active():
context.__enter__()
return transactiongetter()
super(sqllockedoperation, self).__init__(repo, sqllocktr,
*args, **kwargs)
# undo our temporary wrapping
bundle2.bundleoperation = oldopclass
bundle2.bundleoperation = sqllockedoperation
return orig(repo, cg, *args, **kwargs)
except Exception as ex:
exception = ex
raise
finally:
# Be extra sure to undo our wrapping
if oldopclass:
bundle2.bundleoperation = oldopclass
# release sqllock here in exit
if context:
type = value = traceback = None
if exception:
type = exception.__class__
value = exception
traceback = [] # This isn't really important
context.__exit__(type, value, traceback)
else:
# bundle1 or non-lazy locked
return executewithsql(repo, orig, True, repo, cg, *args, **kwargs)
def pull(orig, *args, **kwargs):
repo = args[1]
if repo.ui.configbool("hgsql", "enabled"):
return executewithsql(repo, orig, True, *args, **kwargs)
else:
return orig(*args, **kwargs)
def push(orig, *args, **kwargs):
repo = args[0]
if repo.ui.configbool("hgsql", "enabled"):
# A push locks the local repo in order to update phase data, so we need
# to take the lock for the local repo during a push.
return executewithsql(repo, orig, True, *args, **kwargs)
else:
return orig(*args, **kwargs)
def commit(orig, *args, **kwargs):
repo = args[1]
if repo.ui.configbool("hgsql", "enabled"):
return executewithsql(repo, orig, True, *args, **kwargs)
else:
return orig(*args, **kwargs)
def updatefromremote(orig, *args, **kwargs):
repo = args[1]
if repo.ui.configbool("hgsql", "enabled"):
return executewithsql(repo, orig, True, *args, **kwargs)
else:
return orig(*args, **kwargs)
def addchangegroup(orig, *args, **kwargs):
repo = args[0]
if repo.ui.configbool("hgsql", "enabled"):
return executewithsql(repo, orig, True, *args, **kwargs)
else:
return orig(*args, **kwargs)
def _localphasemove(orig, pushop, *args, **kwargs):
repo = pushop.repo
if repo.ui.configbool("hgsql", "enabled"):
return executewithsql(repo, orig, True, pushop, *args, **kwargs)
else:
return orig(pushop, *args, **kwargs)
class sqlcontext(object):
def __init__(self, repo, takelock=False, waitforlock=False):
self.repo = repo
self.takelock = takelock
self.waitforlock = waitforlock
self._connected = False
self._locked = False
self._used = False
self._startlocktime = 0
self._active = False
def active(self):
return self._active
def __enter__(self):
if self._used:
raise Exception("error: using sqlcontext twice")
self._used = True
self._active = True
repo = self.repo
if not repo.sqlconn:
repo.sqlconnect()
self._connected = True
if self.takelock and not writelock in repo.heldlocks:
startwait = time.time()
repo.sqllock(writelock)
self._locked = True
repo.ui.log("sqllock", "waited for sql lock for %s seconds\n",
time.time() - startwait)
self._startlocktime = time.time()
if self._connected:
repo.syncdb(waitforlock=self.waitforlock)
def __exit__(self, type, value, traceback):
try:
repo = self.repo
if self._locked:
repo.ui.log("sqllock", "held sql lock for %s seconds\n",
time.time() - self._startlocktime)
repo.sqlunlock(writelock)
if self._connected:
repo.sqlclose()
self._active = False
except (_mysql_exceptions.ProgrammingError,
_mysql_exceptions.OperationalError):
# Only raise sql exceptions if the wrapped code threw no exception
if type is None:
raise
def executewithsql(repo, action, sqllock=False, *args, **kwargs):
"""Executes the given action while having a SQL connection open.
If a locks are specified, those locks are held for the duration of the
action.
"""
# executewithsql can be executed in a nested scenario (ex: writing
# bookmarks during a pull), so track whether this call performed
# the connect.
waitforlock = sqllock
if 'waitforlock' in kwargs:
if not waitforlock:
waitforlock = kwargs['waitforlock']
del kwargs['waitforlock']
with sqlcontext(repo, takelock=sqllock, waitforlock=waitforlock):
return action(*args, **kwargs)
def wraprepo(repo):
class sqllocalrepo(repo.__class__):
def sqlconnect(self):
if self.sqlconn:
raise Exception("SQL connection already open")
if self.sqlcursor:
raise Exception("SQL cursor already open without connection")
retry = 3
while True:
try:
self.sqlconn = MySQLdb.connect(**self.sqlargs)
break
except (_mysql_exceptions.ProgrammingError,
_mysql_exceptions.OperationalError):
# mysql can be flakey occasionally, so do some minimal
# retrying.
retry -= 1
if retry == 0:
raise
time.sleep(0.2)
self.sqlconn.autocommit(False)
waittimeout = self.ui.config('hgsql', 'waittimeout', '300')
waittimeout = self.sqlconn.escape_string("%s" % (waittimeout,))
self.locktimeout = self.ui.config('hgsql', 'locktimeout', '60')
self.locktimeout = self.sqlconn.escape_string("%s" %
(self.locktimeout,))
self.sqlconn.query("SET wait_timeout=%s" % waittimeout)
self.sqlconn.query("SET innodb_lock_wait_timeout=%s" %
self.locktimeout)
self.sqlcursor = self.sqlconn.cursor()
def sqlclose(self):
with warnings.catch_warnings():
warnings.simplefilter("ignore")
self.sqlcursor.close()
self.sqlconn.close()
self.sqlcursor = None
self.sqlconn = None
def sqllock(self, name):
escapedname = self.sqlconn.escape_string("%s_%s" %
(name, self.sqlreponame))
# cast to int to prevent passing bad sql data
self.sqlconn.query("SELECT GET_LOCK('%s', %s)" %
(escapedname, self.locktimeout))
result = self.sqlconn.store_result().fetch_row()[0][0]
if result != 1:
raise util.Abort("timed out waiting for mysql repo lock (%s)" % escapedname)
self.heldlocks.add(name)
def hassqllock(self, name):
if not name in self.heldlocks:
return False
escapedname = self.sqlconn.escape_string("%s_%s" % (name, self.sqlreponame))
self.sqlconn.query("SELECT IS_USED_LOCK('%s')" % (escapedname,))
lockheldby = self.sqlconn.store_result().fetch_row()[0][0]
if lockheldby == None:
raise Exception("unable to check %s lock" % escapedname)
self.sqlconn.query("SELECT CONNECTION_ID()")
myconnectid = self.sqlconn.store_result().fetch_row()[0][0]
if myconnectid == None:
raise Exception("unable to read connection id")
return lockheldby == myconnectid
def sqlunlock(self, name):
escapedname = self.sqlconn.escape_string("%s_%s" % (name, self.sqlreponame))
self.sqlconn.query("SELECT RELEASE_LOCK('%s')" % (escapedname,))
self.sqlconn.store_result().fetch_row()
self.heldlocks.discard(name)
def transaction(self, *args, **kwargs):
tr = super(sqllocalrepo, self).transaction(*args, **kwargs)
if tr.count > 1:
return tr
validator = tr.validator
def pretxnclose(tr):
validator(tr)
self.committodb(tr)
del self.pendingrevs[:]
tr.validator = pretxnclose
def transactionabort(orig):
del self.pendingrevs[:]
return orig()
wrapfunction(tr, "_abort", transactionabort)
tr.repo = self
return tr
def needsync(self):
"""Returns True if the local repo is not in sync with the database.
If it returns False, the heads and bookmarks match the database.
"""
self.sqlcursor.execute("""SELECT namespace, name, value
FROM revision_references WHERE repo = %s""", (self.sqlreponame))
sqlheads = set()
sqlbookmarks = {}
tip = -1
for namespace, name, value in self.sqlcursor:
if namespace == "heads":
sqlheads.add(bin(value))
elif namespace == "bookmarks":
sqlbookmarks[name] = bin(value)
elif namespace == "tip":
tip = int(value)
heads = set(self.heads())
bookmarks = self._bookmarks
outofsync = heads != sqlheads or bookmarks != sqlbookmarks or tip != len(self) - 1
return outofsync, sqlheads, sqlbookmarks, tip
def syncdb(self, waitforlock=False):
ui = self.ui
if not self.needsync()[0]:
ui.debug("syncing not needed\n")
return
ui.debug("syncing with mysql\n")
# Save a copy of the old manifest cache so we can put it back
# afterwards.
oldmancache = self.manifest._mancache
try:
lock = self.lock(wait=waitforlock)
except error.LockHeld:
# Oh well. Don't block this non-critical read-only operation.
ui.debug("skipping sync for current operation\n")
return
configbackups = []
try:
# Disable all pretxnclose hooks, since these revisions are
# technically already commited.
for name, value in ui.configitems("hooks"):
if name.startswith("pretxnclose"):
configbackups.append(ui.backupconfig("hooks", name))
ui.setconfig("hooks", name, None)
# The hg-ssh wrapper installs a hook to block all writes. We need to
# circumvent this when we sync from the server.
configbackups.append(ui.backupconfig("hooks", "pretxnopen.hg-ssh"))
self.ui.setconfig("hooks", "pretxnopen.hg-ssh", None)
# Someone else may have synced us while we were waiting.
# Restart the transaction so we have access to the latest rows.
self.sqlconn.rollback()
outofsync, sqlheads, sqlbookmarks, fetchend = self.needsync()
if not outofsync:
return
transaction = self.transaction("syncdb")
self.hook('presyncdb', throw=True)
try:
# Inspect the changelog now that we have the lock
fetchstart = len(self.changelog)
queue = Queue.Queue()
abort = threading.Event()
t = threading.Thread(target=self.fetchthread,
args=(queue, abort, fetchstart, fetchend))
t.setDaemon(True)
try:
t.start()
addentries(self, queue, transaction)
finally:
abort.set()
phases.advanceboundary(self, transaction, phases.public,
self.heads())
transaction.close()
finally:
transaction.release()
# We circumvent the changelog and manifest when we add entries to
# the revlogs. So clear all the caches.
self.invalidate()
self._filecache.pop('changelog', None)
self._filecache.pop('manifest', None)
self._filecache.pop('_phasecache', None)
# Refill the cache
self.manifest._mancache = oldmancache
heads = set(self.heads())
heads.discard(nullid)
if heads != sqlheads:
raise CorruptionException("heads don't match after sync")
if len(self) - 1 != fetchend:
raise CorruptionException("tip doesn't match after sync")
self.disablesync = True
try:
bm = self._bookmarks
bm.clear()
self.sqlcursor.execute("""SELECT name, value FROM revision_references
WHERE namespace = 'bookmarks' AND repo = %s""",
(self.sqlreponame))
for name, node in self.sqlcursor:
node = bin(node)
if node in self:
bm[name] = node
bm.write()
finally:
self.disablesync = False
if bm != sqlbookmarks:
raise CorruptionException("bookmarks don't match after sync")
finally:
for backup in configbackups:
ui.restoreconfig(backup)
lock.release()
def fetchthread(self, queue, abort, fetchstart, fetchend):
"""Fetches every revision from fetchstart to fetchend (inclusive)
and places them on the queue. This function is meant to run on a
background thread and listens to the abort event to abort early.
"""
ui = self.ui
clrev = fetchstart
chunksize = 1000
try:
while True:
if abort.isSet():
break
maxrev = min(clrev + chunksize, fetchend + 1)
self.sqlcursor.execute("""SELECT path, chunk, chunkcount,
linkrev, entry, data0, data1 FROM revisions WHERE repo = %s
AND linkrev > %s AND linkrev < %s ORDER BY linkrev ASC""",
(self.sqlreponame, clrev - 1, maxrev))
# Put split chunks back together into a single revision
groupedrevdata = {}
for revdata in self.sqlcursor:
name = revdata[0]
chunk = revdata[1]
linkrev = revdata[3]
groupedrevdata.setdefault((name, linkrev), {})[chunk] = revdata
if not groupedrevdata:
break
fullrevisions = []
for chunks in groupedrevdata.itervalues():
chunkcount = chunks[0][2]
if chunkcount == 1:
fullrevisions.append(chunks[0])
elif chunkcount == len(chunks):
fullchunk = list(chunks[0])
data1 = ""
for i in range(0, chunkcount):
data1 += chunks[i][6]
fullchunk[6] = data1
fullrevisions.append(tuple(fullchunk))
else:
raise Exception("missing revision chunk - expected %s got %s" %
(chunkcount, len(chunks)))
fullrevisions = sorted(fullrevisions, key=lambda revdata: revdata[3])
for revdata in fullrevisions:
queue.put(revdata)
clrev += chunksize
except Exception as ex:
queue.put(ex)
return
queue.put(False)
def pushkey(self, namespace, key, old, new):
def _pushkey():
return super(sqllocalrepo, self).pushkey(namespace, key, old, new)
return executewithsql(self, _pushkey, namespace == 'bookmarks')
def committodb(self, tr):
"""Commits all pending revisions to the database
"""
if self.sqlconn == None:
raise util.Abort("invalid repo change - only hg push and pull" +
" are allowed")
if not self.pendingrevs and not 'bookmark_moved' in tr.hookargs:
return
reponame = self.sqlreponame
cursor = self.sqlcursor
maxcommitsize = self.maxcommitsize
maxinsertsize = self.maxinsertsize
maxrowsize = self.maxrowsize
if self.pendingrevs:
self._validatependingrevs()
def insert(args, values):
argstring = ','.join(args)
cursor.execute("INSERT INTO revisions(repo, path, "
"chunk, chunkcount, linkrev, rev, node, entry, "
"data0, data1, createdtime) VALUES %s" %
argstring, values)
try:
commitsize = 0
insertsize = 0
args = []
values = []
now = time.strftime('%Y-%m-%d %H:%M:%S')
for revision in self.pendingrevs:
path, linkrev, rev, node, entry, data0, data1 = revision
start = 0
chunk = 0
datalen = len(data1)
chunkcount = datalen / maxrowsize
if datalen % maxrowsize != 0 or datalen == 0:
chunkcount += 1
# We keep row size down by breaking large revisions down into
# smaller chunks.
while chunk == 0 or start < len(data1):
end = min(len(data1), start + maxrowsize)
datachunk = data1[start:end]
args.append('(%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s)')
values.extend((reponame, path, chunk, chunkcount, linkrev, rev,
node, entry, data0, datachunk, now))
size = len(datachunk)
commitsize += size
insertsize += size
chunk += 1
start = end
# Minimize roundtrips by doing bulk inserts
if insertsize > maxinsertsize:
insert(args, values)
del args[:]
del values[:]
insertsize = 0
# MySQL transactions can only reach a certain size, so we commit
# every so often. As long as we don't update the tip pushkey,
# this is ok.
if commitsize > maxcommitsize:
self.sqlconn.commit()
commitsize = 0
if args:
insert(args, values)
# commit at the end just to make sure we're clean
self.sqlconn.commit()
# Compute new heads, and delete old heads
newheads = set(self.heads())
oldheads = []
cursor.execute(
"SELECT value FROM revision_references "
"WHERE repo = %s AND namespace='heads'",
(reponame,)
)
for head in cursor:
head = head[0]
if head in newheads:
newheads.discard(head)
else:
oldheads.append(head)
if oldheads:
headargs = ','.join(['%s'] * len(oldheads))
cursor.execute(
"DELETE revision_references FROM revision_references " +
"FORCE INDEX (bookmarkindex) " +
"WHERE namespace = 'heads' " +
"AND repo = %s AND value IN (" + headargs + ")",
(reponame,) + tuple(oldheads)
)
# Compute new bookmarks, and delete old bookmarks
newbookmarks = self._bookmarks.copy()
oldbookmarks = []
cursor.execute(
"SELECT name, value FROM revision_references "
"WHERE namespace = 'bookmarks' AND repo = %s",
(reponame,)
)
for k, v in cursor:
if newbookmarks.get(k) == v:
del newbookmarks[k]
else:
oldbookmarks.append(k)
if oldbookmarks:
bookargs = ','.join(['%s'] * len(oldbookmarks))
cursor.execute(
"DELETE revision_references FROM revision_references " +
"FORCE INDEX (bookmarkindex) " +
"WHERE namespace = 'bookmarks' AND repo = %s " +
"AND name IN (" + bookargs + ")",
(repo.sqlreponame,) + tuple(oldbookmarks)
)
tmpl = []
values = []
for head in newheads:
tmpl.append("(%s, 'heads', NULL, %s)")
values.append(reponame)
values.append(hex(head))
for k, v in newbookmarks.iteritems():
tmpl.append("(%s, 'bookmarks', %s, %s)")
values.append(repo.sqlreponame)
values.append(k)
values.append(hex(v))
if tmpl:
cursor.execute("INSERT INTO revision_references(repo, namespace, name, value) " +
"VALUES %s" % ','.join(tmpl), tuple(values))
# revision_references has multiple keys (primary key, and a unique index), so
# mysql gives a warning when using ON DUPLICATE KEY since it would only update one
# row despite multiple key duplicates. This doesn't matter for us, since we know
# there is only one row that will share the same key. So suppress the warning.
cursor.execute("""INSERT INTO revision_references(repo, namespace, name, value)
VALUES(%s, 'tip', 'tip', %s) ON DUPLICATE KEY UPDATE value=%s""",
(reponame, len(self) - 1, len(self) - 1))
# Just to be super sure, check the write lock before doing the final commit
if not self.hassqllock(writelock):
raise Exception("attempting to write to sql without holding %s (precommit)"
% writelock)
self.sqlconn.commit()
except:
self.sqlconn.rollback()
raise
finally:
del self.pendingrevs[:]
def _validatependingrevs(self):
"""Validates that the current pending revisions will be valid when
written to the database.
"""
reponame = self.sqlreponame
cursor = self.sqlcursor
# Ensure we hold the write lock
if not self.hassqllock(writelock):
raise Exception("attempting to write to sql without holding %s (prevalidate)"
% writelock)
# Validate that we are appending to the correct linkrev
cursor.execute("""SELECT value FROM revision_references WHERE repo = %s AND
namespace = 'tip'""", reponame)
tipresults = cursor.fetchall()
if len(tipresults) == 0:
maxlinkrev = -1
elif len(tipresults) == 1:
maxlinkrev = int(tipresults[0][0])
else:
raise CorruptionException(("multiple tips for %s in " +
" the database") % reponame)
minlinkrev = min(self.pendingrevs, key= lambda x: x[1])[1]
if maxlinkrev == None or maxlinkrev != minlinkrev - 1:
raise CorruptionException("attempting to write non-sequential " +
"linkrev %s, expected %s" % (minlinkrev, maxlinkrev + 1))
# Clean up excess revisions left from interrupted commits.
# Since MySQL can only hold so much data in a transaction, we allow
# committing across multiple db transactions. That means if
# the commit is interrupted, the next transaction needs to clean
# up bad revisions.
cursor.execute("""DELETE FROM revisions WHERE repo = %s AND
linkrev > %s""", (reponame, maxlinkrev))
# Validate that all rev dependencies (base, p1, p2) have the same
# node in the database
pending = set([(path, rev) for path, _, rev, _, _, _, _ in self.pendingrevs])
expectedrevs = set()
for revision in self.pendingrevs:
path, linkrev, rev, node, entry, data0, data1 = revision
e = struct.unpack(revlog.indexformatng, entry)
_, _, _, base, _, p1r, p2r, _ = e
if p1r != nullrev and not (path, p1r) in pending:
expectedrevs.add((path, p1r))
if p2r != nullrev and not (path, p2r) in pending:
expectedrevs.add((path, p2r))
if (base != nullrev and base != rev and
not (path, base) in pending):
expectedrevs.add((path, base))
if not expectedrevs:
return
missingrevs = []
expectedlist = list(expectedrevs)
expectedcount = len(expectedrevs)
batchsize = self.ui.configint('hgsql', 'verifybatchsize', 1000)
i = 0
while i < expectedcount:
checkrevs = set(expectedlist[i:i + batchsize])
i += batchsize
whereclauses = []
args = []
args.append(reponame)
for path, rev in checkrevs:
whereclauses.append("(path, rev, chunk) = (%s, %s, 0)")
args.append(path)
args.append(rev)
whereclause = ' OR '.join(whereclauses)
cursor.execute("""SELECT path, rev, node FROM revisions WHERE
repo = %s AND (""" + whereclause + ")",
args)
for path, rev, node in cursor:
rev = int(rev)
checkrevs.remove((path, rev))
rl = None
if path == '00changelog.i':
rl = self.changelog
elif path == '00manifest.i':
rl = self.manifest
else:
rl = revlog.revlog(self.svfs, path)
localnode = hex(rl.node(rev))
if localnode != node:
raise CorruptionException(("expected node %s at rev %d of "
"%s but found %s") % (node, rev, path, localnode))
if len(checkrevs) > 0:
missingrevs.extend(checkrevs)
if missingrevs:
raise CorruptionException(("unable to verify %d dependent " +
"revisions before adding a commit") % (len(missingrevs)))
ui = repo.ui
sqlargs = {}
sqlargs['host'] = ui.config("hgsql", "host")
sqlargs['db'] = ui.config("hgsql", "database")
sqlargs['user'] = ui.config("hgsql", "user")
sqlargs['port'] = ui.configint("hgsql", "port")
password = ui.config("hgsql", "password", "")
if password:
sqlargs['passwd'] = password
sqlargs['cursorclass'] = cursors.SSCursor
repo.sqlargs = sqlargs
repo.sqlreponame = ui.config("hgsql", "reponame")
if not repo.sqlreponame:
raise Exception("missing hgsql.reponame")
repo.maxcommitsize = ui.configbytes("hgsql", "maxcommitsize", 52428800) # 50MB
repo.maxinsertsize = ui.configbytes("hgsql", "maxinsertsize", 1048576) # 1MB
repo.maxrowsize = ui.configbytes("hgsql", "maxrowsize", 1048576)
repo.sqlconn = None
repo.sqlcursor = None
repo.disablesync = False
repo.pendingrevs = []
repo.heldlocks = set()
repo.__class__ = sqllocalrepo
class bufferedopener(object):
"""Opener implementation that buffers all writes in memory until
flush or close is called.
"""
def __init__(self, opener, path, mode):
self.opener = opener
self.path = path
self.mode = mode
self.buffer = []
self.closed = False
def write(self, value):
if self.closed:
raise Exception("attempted to write to a closed bufferedopener")
self.buffer.append(value)
def flush(self):
buffer = self.buffer
self.buffer = []
if buffer:
fp = self.opener(self.path, self.mode)
fp.write(''.join(buffer))
fp.close()
def close(self):
self.flush()
self.closed = True
def addentries(repo, queue, transaction):
"""Reads new rev entries from a queue and writes them to a buffered
revlog. At the end it flushes all the new data to disk.
"""
opener = repo.svfs
revlogs = {}
def writeentry(revdata):
# Instantiates pseudo-revlogs for us to write data directly to
path, chunk, chunkcount, link, entry, data0, data1 = revdata
revlog = revlogs.get(path)
if not revlog:
revlog = EntryRevlog(opener, path)
revlogs[path] = revlog
# Replace the existing openers with buffered ones so we can
# perform the flush to disk all at once at the end.
if not hasattr(revlog, 'ifh') or revlog.ifh.closed:
dfh = None
if not revlog._inline:
dfh = bufferedopener(opener, revlog.datafile, "a")
ifh = bufferedopener(opener, revlog.indexfile, "a+")
revlog.ifh = ifh
revlog.dfh = dfh
revlog.addentry(transaction, revlog.ifh, revlog.dfh, entry,
data0, data1)
leftover = None
exit = False
# Read one linkrev at a time from the queue
while not exit:
currentlinkrev = -1
revisions = []
if leftover:
revisions.append(leftover)
leftover = None
# Read everything from the current linkrev
while True:
revdata = queue.get()
if not revdata:
exit = True
break
# The background thread had an exception, rethrow from the
# foreground thread.
if isinstance(revdata, Exception):
raise revdata
linkrev = revdata[3]
if currentlinkrev == -1:
currentlinkrev = linkrev
if linkrev == currentlinkrev:
revisions.append(revdata)
elif linkrev < currentlinkrev:
raise CorruptionException("SQL data is not in linkrev order")
else:
leftover = revdata
currentlinkrev = linkrev
break
if not revisions:
continue
for revdata in revisions:
writeentry(revdata)
# Flush filelogs, then manifest, then changelog
changelog = revlogs.pop("00changelog.i", None)
manifest = revlogs.pop("00manifest.i", None)
def flushrevlog(revlog):
if not revlog.ifh.closed:
revlog.ifh.flush()
if revlog.dfh and not revlog.dfh.closed:
revlog.dfh.flush()
for filelog in revlogs.itervalues():
flushrevlog(filelog)
if manifest:
flushrevlog(manifest)
if changelog:
flushrevlog(changelog)
class EntryRevlog(revlog.revlog):
"""Pseudo-revlog implementation that allows applying data directly to
the end of the revlog.
"""
def addentry(self, transaction, ifh, dfh, entry, data0, data1):
curr = len(self)
offset = self.end(curr - 1)
e = struct.unpack(revlog.indexformatng, entry)
offsettype, datalen, textlen, base, link, p1r, p2r, node = e
# The first rev has special metadata encoded in it that should be
# stripped before being added to the index.
if curr == 0:
elist = list(e)
type = revlog.gettype(offsettype)
offsettype = revlog.offset_type(0, type)
elist[0] = offsettype
e = tuple(elist)
# Verify that the rev's parents and base appear earlier in the revlog
if p1r >= curr or p2r >= curr:
raise CorruptionException("parent revision is not in revlog: %s" %
self.indexfile)
if base > curr:
raise CorruptionException("base revision is not in revlog: %s" %
self.indexfile)
expectedoffset = revlog.getoffset(offsettype)
if expectedoffset != 0 and expectedoffset != offset:
raise CorruptionException("revision offset doesn't match prior length " +
"(%s offset vs %s length): %s" %
(expectedoffset, offset, self.indexfile))
if node not in self.nodemap:
self.index.insert(-1, e)
self.nodemap[node] = len(self) - 1
if not self._inline:
transaction.add(self.datafile, offset)
transaction.add(self.indexfile, curr * len(entry))
if data0:
dfh.write(data0)
dfh.write(data1)
ifh.write(entry)
else:
offset += curr * self._io.size
transaction.add(self.indexfile, offset, curr)
ifh.write(entry)
ifh.write(data0)
ifh.write(data1)
self.checkinlinesize(transaction, ifh)
def addgroup(orig, self, bundle, linkmapper, transaction, addrevisioncb=None):
"""Copy paste of revlog.addgroup, but we ensure that the revisions are
added in linkrev order.
"""
if not util.safehasattr(transaction, "repo"):
return orig(self, bundle, linkmapper, transaction,
addrevisioncb=addrevisioncb)
# track the base of the current delta log
content = []
node = None
r = len(self)
end = 0
if r:
end = self.end(r - 1)
ifh = self.opener(self.indexfile, "a+")
isize = r * self._io.size
if self._inline:
transaction.add(self.indexfile, end + isize, r)
dfh = None
else:
transaction.add(self.indexfile, isize, r)
transaction.add(self.datafile, end)
dfh = self.opener(self.datafile, "a")
try:
# loop through our set of deltas
chunkdatas = []
chunkmap = {}
lastlinkrev = -1
reorder = False
# Read all of the data from the stream
chain = None
while True:
chunkdata = bundle.deltachunk(chain)
if not chunkdata:
break
node = chunkdata['node']
cs = chunkdata['cs']
link = linkmapper(cs)
if link < lastlinkrev:
reorder = True
lastlinkrev = link
chunkdatas.append((link, chunkdata))
chunkmap[node] = chunkdata
chain = node
# If we noticed a incoming rev was not in linkrev order
# we reorder all the revs appropriately.
if reorder:
chunkdatas = sorted(chunkdatas)
fulltexts = {}
def getfulltext(node):
if node in fulltexts:
return fulltexts[node]
if node in self.nodemap:
return self.revision(node)
chunkdata = chunkmap[node]
deltabase = chunkdata['deltabase']
delta = chunkdata['delta']
deltachain = []
currentbase = deltabase
while True:
if currentbase in fulltexts:
deltachain.append(fulltexts[currentbase])
break
elif currentbase in self.nodemap:
deltachain.append(self.revision(currentbase))
break
elif currentbase == nullid:
break
else:
deltachunk = chunkmap[currentbase]
currentbase = deltachunk['deltabase']
deltachain.append(deltachunk['delta'])
prevtext = deltachain.pop()
while deltachain:
prevtext = mdiff.patch(prevtext, deltachain.pop())
fulltext = mdiff.patch(prevtext, delta)
fulltexts[node] = fulltext
return fulltext
visited = set()
prevnode = self.node(len(self) - 1)
for link, chunkdata in chunkdatas:
node = chunkdata['node']
deltabase = chunkdata['deltabase']
if (not deltabase in self.nodemap and
not deltabase in visited):
fulltext = getfulltext(node)
ptext = getfulltext(prevnode)
delta = mdiff.textdiff(ptext, fulltext)
chunkdata['delta'] = delta
chunkdata['deltabase'] = prevnode
prevnode = node
visited.add(node)
# Apply the reordered revs to the revlog
for link, chunkdata in chunkdatas:
node = chunkdata['node']
p1 = chunkdata['p1']
p2 = chunkdata['p2']
cs = chunkdata['cs']
deltabase = chunkdata['deltabase']
delta = chunkdata['delta']
content.append(node)
link = linkmapper(cs)
if node in self.nodemap:
# this can happen if two branches make the same change
continue
for p in (p1, p2):
if p not in self.nodemap:
raise LookupError(p, self.indexfile,
_('unknown parent'))
if deltabase not in self.nodemap:
raise LookupError(deltabase, self.indexfile,
_('unknown delta base'))
baserev = self.rev(deltabase)
chain = self._addrevision(node, None, transaction, link, p1, p2,
revlog.REVIDX_DEFAULT_FLAGS, (baserev, delta),
ifh, dfh)
if addrevisioncb:
# Data for added revision can't be read unless flushed
# because _loadchunk always opensa new file handle and
# there is no guarantee data was actually written yet.
if dfh:
dfh.flush()
ifh.flush()
addrevisioncb(self, chain)
if not dfh and not self._inline:
# addrevision switched from inline to conventional
# reopen the index
ifh.close()
dfh = self.opener(self.datafile, "a")
ifh = self.opener(self.indexfile, "a")
finally:
if dfh:
dfh.close()
ifh.close()
return content
def bookmarkcommand(orig, ui, repo, *names, **opts):
if not repo.ui.configbool("hgsql", "enabled"):
return orig(ui, repo, *names, **opts)
write = (opts.get('delete') or opts.get('rename')
or opts.get('inactive') or names)
def _bookmarkcommand():
return orig(ui, repo, *names, **opts)
if write:
return executewithsql(repo, _bookmarkcommand, True)
else:
return _bookmarkcommand()
def bookmarkwrite(orig, self):
repo = self._repo
if not repo.ui.configbool("hgsql", "enabled") or repo.disablesync:
return orig(self)
if not repo.sqlconn:
raise util.Abort("attempted bookmark write without sql connection")
elif not repo.hassqllock(writelock):
raise util.Abort("attempted bookmark write without write lock")
try:
cursor = repo.sqlcursor
cursor.execute("""DELETE FROM revision_references WHERE repo = %s AND
namespace = 'bookmarks'""", (repo.sqlreponame))
for k, v in self.iteritems():
cursor.execute("""INSERT INTO revision_references(repo, namespace, name, value)
VALUES(%s, 'bookmarks', %s, %s)""",
(repo.sqlreponame, k, hex(v)))
repo.sqlconn.commit()
return orig(self)
except:
repo.sqlconn.rollback()
raise
def pushkey(orig, repo, proto, namespace, key, old, new):
if repo.ui.configbool("hgsql", "enabled"):
def commitpushkey():
return orig(repo, proto, namespace, key, old, new)
return executewithsql(repo, commitpushkey, namespace == 'bookmarks')
else:
return orig(repo, proto, namespace, key, old, new)
# recover must be a norepo command because loading the repo fails
commands.norepo += " sqlrecover sqlstrip"
@command('^sqlrecover', [
('f', 'force', '', _('strips as far back as necessary'), ''),
('', 'no-backup', '', _('does not produce backup bundles for strips'), ''),
], _('hg sqlrecover'))
def sqlrecover(ui, *args, **opts):
"""
Strips commits from the local repo until it is back in sync with the SQL
server.
"""
global initialsync
initialsync = INITIAL_SYNC_DISABLE
repo = hg.repository(ui, ui.environ['PWD'])
repo.disablesync = True
if repo.recover():
ui.status("recovered from incomplete transaction")
def iscorrupt():
repo.sqlconnect()
try:
repo.syncdb()
except:
return True
finally:
repo.sqlclose()
return False
reposize = len(repo)
stripsize = 10
while iscorrupt() and len(repo) > 0:
if not opts.get('force') and reposize > len(repo) + 10000:
ui.warn("unable to fix repo after stripping 10000 commits " +
"(use -f to strip more)")
striprev = max(0, len(repo) - stripsize)
nodelist = [repo[striprev].node()]
ui.status("stripping back to %s commits" % (striprev))
backup = "none" if opts.get("no-backup") else "all"
repair.strip(ui, repo, nodelist, backup=backup, topic="sqlrecover")
stripsize = min(stripsize * 5, 10000)
if len(repo) == 0:
ui.warn(_("unable to fix repo corruption\n"))
elif len(repo) == reposize:
ui.status(_("local repo was not corrupt - no action taken\n"))
else:
ui.status(_("local repo now matches SQL\n"))
@command('^sqlstrip', [
('', 'i-know-what-i-am-doing', None, _('only run sqlstrip if you know ' +
'exactly what you\'re doing')),
], _('hg sqlstrip [OPTIONS] REV'))
def sqlstrip(ui, rev, *args, **opts):
"""strips all revisions greater than or equal to the given revision from the sql database
Deletes all revisions with linkrev >= the given revision from the local
repo and from the sql database. This is permanent and cannot be undone.
Once the revisions are deleted from the database, you will need to run
this command on each master server before proceeding to write new revisions.
"""
if not opts.get('i_know_what_i_am_doing'):
raise util.Abort("You must pass --i-know-what-i-am-doing to run this " +
"command. If you have multiple servers using the database, this " +
"command will break your servers until you run it on each one. " +
"Only the Mercurial server admins should ever run this.")
global initialsync
initialsync = INITIAL_SYNC_DISABLE
repo = hg.repository(ui, ui.environ['PWD'])
repo.disablesync = True
try:
rev = int(rev)
except ValueError:
raise util.Abort("specified rev must be an integer: '%s'" % rev)
lock = repo.lock()
try:
repo.sqlconnect()
try:
repo.sqllock(writelock)
if rev not in repo:
raise util.Abort("revision %s is not in the repo" % rev)
reponame = repo.sqlreponame
cursor = repo.sqlcursor
changelog = repo.changelog
revs = repo.revs('%s:' % rev)
# strip locally
ui.status("stripping locally\n")
repair.strip(ui, repo, [changelog.node(r) for r in revs], "all")
ui.status("stripping from the database\n")
ui.status("deleting old references\n")
cursor.execute("""DELETE FROM revision_references WHERE repo = %s""", (reponame,))
ui.status("adding new head references\n")
for head in repo.heads():
cursor.execute("""INSERT INTO revision_references(repo, namespace, value)
VALUES(%s, 'heads', %s)""",
(reponame, hex(head)))
ui.status("adding new tip reference\n")
cursor.execute("""INSERT INTO revision_references(repo, namespace, name, value)
VALUES(%s, 'tip', 'tip', %s)""",
(reponame, len(repo) - 1))
ui.status("adding new bookmark references\n")
for k, v in repo._bookmarks.iteritems():
cursor.execute("""INSERT INTO revision_references(repo, namespace, name, value)
VALUES(%s, 'bookmarks', %s, %s)""",
(reponame, k, hex(v)))
ui.status("deleting revision data\n")
cursor.execute("""DELETE FROM revisions WHERE repo = %s and linkrev > %s""",
(reponame, rev))
repo.sqlconn.commit()
finally:
repo.sqlunlock(writelock)
repo.sqlclose()
finally:
lock.release()