##// END OF EJS Templates
lfs: fix the stall and corruption issue when concurrently uploading blobs...
lfs: fix the stall and corruption issue when concurrently uploading blobs We've avoided the issue up to this point by gating worker usage with an experimental config. See 10e62d5efa73, and the thread linked there for some of the initial diagnosis, but essentially some data was being read from the blob before an error occurred and `keepalive` retried, but didn't rewind the file pointer. So the leading data was lost from the blob on the server, and the connection stalled, trying to send more data than available. In trying to recreate this, I was unable to do so uploading from Windows to CentOS 7. But it reproduced every time going from CentOS 7 to another CentOS 7 over https. I found recent fixes in the FaceBook repo to address this[1][2]. The commit message for the first is: The KeepAlive HTTP implementation is bugged in it's retry logic, it supports reading from a file pointer, but doesn't support rewinding of the seek cursor when it performs a retry. So it can happen that an upload fails for whatever reason and will then 'hang' on the retry event. The sequence of events that get triggered are: - Upload file A, goes OK. Keep-Alive caches connection. - Upload file B, fails due to (for example) failing Keep-Alive, but LFS file pointer has been consumed for the upload and fd has been closed. - Retry for file B starts, sets the Content-Length properly to the expected file size, but since file pointer has been consumed no data will be uploaded, causing the server to wait for the uploaded data until either client or server reaches a timeout, making it seem as our mercurial process hangs. This is just a stop-gap measure to prevent this behavior from blocking Mercurial (LFS has retry logic). A proper solutions need to be build on top of this stop-gap measure: for upload from file pointers, we should support fseek() on the interface. Since we expect to consume the whole file always anyways, this should be safe. This way we can seek back to the beginning on a retry. I ported those two patches, and it works. But I see that `url._sendfile()` does a rewind on `httpsendfile` objects[3], so maybe it's better to keep this all in one place and avoid a second seek. We may still want the first FaceBook patch as extra protection for this problem in general. The other two uses of `httpsendfile` are in the wire protocol to upload bundles, and to upload largefiles. Neither of these appear to use a worker, and I'm not sure why workers seem to trigger this, or if this could have happened without a worker. Since `httpsendfile` already has a `close()` method, that is dropped. That class also explicitly says there's no `__len__` attribute, so that is removed too. The override for `read()` is necessary to avoid the progressbar usage per file. [1] https://github.com/facebookexperimental/eden/commit/c350d6536d90c044c837abdd3675185644481469 [2] https://github.com/facebookexperimental/eden/commit/77f0d3fd0415e81b63e317e457af9c55c46103ee [3] https://www.mercurial-scm.org/repo/hg/file/5.2.2/mercurial/url.py#l176 Differential Revision: https://phab.mercurial-scm.org/D7962

File last commit:

r44530:b9e174d4 default
r44746:43eea17a default
Show More
wrapper.py
546 lines | 16.7 KiB | text/x-python | PythonLexer
# wrapper.py - methods wrapping core mercurial logic
#
# Copyright 2017 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.
from __future__ import absolute_import
import hashlib
from mercurial.i18n import _
from mercurial.node import bin, hex, nullid, short
from mercurial.pycompat import (
getattr,
setattr,
)
from mercurial import (
bundle2,
changegroup,
cmdutil,
context,
error,
exchange,
exthelper,
localrepo,
pycompat,
revlog,
scmutil,
upgrade,
util,
vfs as vfsmod,
wireprotov1server,
)
from mercurial.interfaces import repository
from mercurial.utils import (
storageutil,
stringutil,
)
from ..largefiles import lfutil
from . import (
blobstore,
pointer,
)
eh = exthelper.exthelper()
@eh.wrapfunction(localrepo, b'makefilestorage')
def localrepomakefilestorage(orig, requirements, features, **kwargs):
if b'lfs' in requirements:
features.add(repository.REPO_FEATURE_LFS)
return orig(requirements=requirements, features=features, **kwargs)
@eh.wrapfunction(changegroup, b'allsupportedversions')
def allsupportedversions(orig, ui):
versions = orig(ui)
versions.add(b'03')
return versions
@eh.wrapfunction(wireprotov1server, b'_capabilities')
def _capabilities(orig, repo, proto):
'''Wrap server command to announce lfs server capability'''
caps = orig(repo, proto)
if util.safehasattr(repo.svfs, b'lfslocalblobstore'):
# Advertise a slightly different capability when lfs is *required*, so
# that the client knows it MUST load the extension. If lfs is not
# required on the server, there's no reason to autoload the extension
# on the client.
if b'lfs' in repo.requirements:
caps.append(b'lfs-serve')
caps.append(b'lfs')
return caps
def bypasscheckhash(self, text):
return False
def readfromstore(self, text):
"""Read filelog content from local blobstore transform for flagprocessor.
Default tranform for flagprocessor, returning contents from blobstore.
Returns a 2-typle (text, validatehash) where validatehash is True as the
contents of the blobstore should be checked using checkhash.
"""
p = pointer.deserialize(text)
oid = p.oid()
store = self.opener.lfslocalblobstore
if not store.has(oid):
p.filename = self.filename
self.opener.lfsremoteblobstore.readbatch([p], store)
# The caller will validate the content
text = store.read(oid, verify=False)
# pack hg filelog metadata
hgmeta = {}
for k in p.keys():
if k.startswith(b'x-hg-'):
name = k[len(b'x-hg-') :]
hgmeta[name] = p[k]
if hgmeta or text.startswith(b'\1\n'):
text = storageutil.packmeta(hgmeta, text)
return (text, True, {})
def writetostore(self, text, sidedata):
# hg filelog metadata (includes rename, etc)
hgmeta, offset = storageutil.parsemeta(text)
if offset and offset > 0:
# lfs blob does not contain hg filelog metadata
text = text[offset:]
# git-lfs only supports sha256
oid = hex(hashlib.sha256(text).digest())
self.opener.lfslocalblobstore.write(oid, text)
# replace contents with metadata
longoid = b'sha256:%s' % oid
metadata = pointer.gitlfspointer(oid=longoid, size=b'%d' % len(text))
# by default, we expect the content to be binary. however, LFS could also
# be used for non-binary content. add a special entry for non-binary data.
# this will be used by filectx.isbinary().
if not stringutil.binary(text):
# not hg filelog metadata (affecting commit hash), no "x-hg-" prefix
metadata[b'x-is-binary'] = b'0'
# translate hg filelog metadata to lfs metadata with "x-hg-" prefix
if hgmeta is not None:
for k, v in pycompat.iteritems(hgmeta):
metadata[b'x-hg-%s' % k] = v
rawtext = metadata.serialize()
return (rawtext, False)
def _islfs(rlog, node=None, rev=None):
if rev is None:
if node is None:
# both None - likely working copy content where node is not ready
return False
rev = rlog.rev(node)
else:
node = rlog.node(rev)
if node == nullid:
return False
flags = rlog.flags(rev)
return bool(flags & revlog.REVIDX_EXTSTORED)
# Wrapping may also be applied by remotefilelog
def filelogaddrevision(
orig,
self,
text,
transaction,
link,
p1,
p2,
cachedelta=None,
node=None,
flags=revlog.REVIDX_DEFAULT_FLAGS,
**kwds
):
# The matcher isn't available if reposetup() wasn't called.
lfstrack = self._revlog.opener.options.get(b'lfstrack')
if lfstrack:
textlen = len(text)
# exclude hg rename meta from file size
meta, offset = storageutil.parsemeta(text)
if offset:
textlen -= offset
if lfstrack(self._revlog.filename, textlen):
flags |= revlog.REVIDX_EXTSTORED
return orig(
self,
text,
transaction,
link,
p1,
p2,
cachedelta=cachedelta,
node=node,
flags=flags,
**kwds
)
# Wrapping may also be applied by remotefilelog
def filelogrenamed(orig, self, node):
if _islfs(self._revlog, node):
rawtext = self._revlog.rawdata(node)
if not rawtext:
return False
metadata = pointer.deserialize(rawtext)
if b'x-hg-copy' in metadata and b'x-hg-copyrev' in metadata:
return metadata[b'x-hg-copy'], bin(metadata[b'x-hg-copyrev'])
else:
return False
return orig(self, node)
# Wrapping may also be applied by remotefilelog
def filelogsize(orig, self, rev):
if _islfs(self._revlog, rev=rev):
# fast path: use lfs metadata to answer size
rawtext = self._revlog.rawdata(rev)
metadata = pointer.deserialize(rawtext)
return int(metadata[b'size'])
return orig(self, rev)
@eh.wrapfunction(revlog, b'_verify_revision')
def _verify_revision(orig, rl, skipflags, state, node):
if _islfs(rl, node=node):
rawtext = rl.rawdata(node)
metadata = pointer.deserialize(rawtext)
# Don't skip blobs that are stored locally, as local verification is
# relatively cheap and there's no other way to verify the raw data in
# the revlog.
if rl.opener.lfslocalblobstore.has(metadata.oid()):
skipflags &= ~revlog.REVIDX_EXTSTORED
elif skipflags & revlog.REVIDX_EXTSTORED:
# The wrapped method will set `skipread`, but there's enough local
# info to check renames.
state[b'safe_renamed'].add(node)
orig(rl, skipflags, state, node)
@eh.wrapfunction(context.basefilectx, b'cmp')
def filectxcmp(orig, self, fctx):
"""returns True if text is different than fctx"""
# some fctx (ex. hg-git) is not based on basefilectx and do not have islfs
if self.islfs() and getattr(fctx, 'islfs', lambda: False)():
# fast path: check LFS oid
p1 = pointer.deserialize(self.rawdata())
p2 = pointer.deserialize(fctx.rawdata())
return p1.oid() != p2.oid()
return orig(self, fctx)
@eh.wrapfunction(context.basefilectx, b'isbinary')
def filectxisbinary(orig, self):
if self.islfs():
# fast path: use lfs metadata to answer isbinary
metadata = pointer.deserialize(self.rawdata())
# if lfs metadata says nothing, assume it's binary by default
return bool(int(metadata.get(b'x-is-binary', 1)))
return orig(self)
def filectxislfs(self):
return _islfs(self.filelog()._revlog, self.filenode())
@eh.wrapfunction(cmdutil, b'_updatecatformatter')
def _updatecatformatter(orig, fm, ctx, matcher, path, decode):
orig(fm, ctx, matcher, path, decode)
fm.data(rawdata=ctx[path].rawdata())
@eh.wrapfunction(scmutil, b'wrapconvertsink')
def convertsink(orig, sink):
sink = orig(sink)
if sink.repotype == b'hg':
class lfssink(sink.__class__):
def putcommit(
self,
files,
copies,
parents,
commit,
source,
revmap,
full,
cleanp2,
):
pc = super(lfssink, self).putcommit
node = pc(
files,
copies,
parents,
commit,
source,
revmap,
full,
cleanp2,
)
if b'lfs' not in self.repo.requirements:
ctx = self.repo[node]
# The file list may contain removed files, so check for
# membership before assuming it is in the context.
if any(f in ctx and ctx[f].islfs() for f, n in files):
self.repo.requirements.add(b'lfs')
self.repo._writerequirements()
return node
sink.__class__ = lfssink
return sink
# bundlerepo uses "vfsmod.readonlyvfs(othervfs)", we need to make sure lfs
# options and blob stores are passed from othervfs to the new readonlyvfs.
@eh.wrapfunction(vfsmod.readonlyvfs, b'__init__')
def vfsinit(orig, self, othervfs):
orig(self, othervfs)
# copy lfs related options
for k, v in othervfs.options.items():
if k.startswith(b'lfs'):
self.options[k] = v
# also copy lfs blobstores. note: this can run before reposetup, so lfs
# blobstore attributes are not always ready at this time.
for name in [b'lfslocalblobstore', b'lfsremoteblobstore']:
if util.safehasattr(othervfs, name):
setattr(self, name, getattr(othervfs, name))
def _prefetchfiles(repo, revs, match):
"""Ensure that required LFS blobs are present, fetching them as a group if
needed."""
if not util.safehasattr(repo.svfs, b'lfslocalblobstore'):
return
pointers = []
oids = set()
localstore = repo.svfs.lfslocalblobstore
for rev in revs:
ctx = repo[rev]
for f in ctx.walk(match):
p = pointerfromctx(ctx, f)
if p and p.oid() not in oids and not localstore.has(p.oid()):
p.filename = f
pointers.append(p)
oids.add(p.oid())
if pointers:
# Recalculating the repo store here allows 'paths.default' that is set
# on the repo by a clone command to be used for the update.
blobstore.remote(repo).readbatch(pointers, localstore)
def _canskipupload(repo):
# Skip if this hasn't been passed to reposetup()
if not util.safehasattr(repo.svfs, b'lfsremoteblobstore'):
return True
# if remotestore is a null store, upload is a no-op and can be skipped
return isinstance(repo.svfs.lfsremoteblobstore, blobstore._nullremote)
def candownload(repo):
# Skip if this hasn't been passed to reposetup()
if not util.safehasattr(repo.svfs, b'lfsremoteblobstore'):
return False
# if remotestore is a null store, downloads will lead to nothing
return not isinstance(repo.svfs.lfsremoteblobstore, blobstore._nullremote)
def uploadblobsfromrevs(repo, revs):
'''upload lfs blobs introduced by revs
Note: also used by other extensions e. g. infinitepush. avoid renaming.
'''
if _canskipupload(repo):
return
pointers = extractpointers(repo, revs)
uploadblobs(repo, pointers)
def prepush(pushop):
"""Prepush hook.
Read through the revisions to push, looking for filelog entries that can be
deserialized into metadata so that we can block the push on their upload to
the remote blobstore.
"""
return uploadblobsfromrevs(pushop.repo, pushop.outgoing.missing)
@eh.wrapfunction(exchange, b'push')
def push(orig, repo, remote, *args, **kwargs):
"""bail on push if the extension isn't enabled on remote when needed, and
update the remote store based on the destination path."""
if b'lfs' in repo.requirements:
# If the remote peer is for a local repo, the requirement tests in the
# base class method enforce lfs support. Otherwise, some revisions in
# this repo use lfs, and the remote repo needs the extension loaded.
if not remote.local() and not remote.capable(b'lfs'):
# This is a copy of the message in exchange.push() when requirements
# are missing between local repos.
m = _(b"required features are not supported in the destination: %s")
raise error.Abort(
m % b'lfs', hint=_(b'enable the lfs extension on the server')
)
# Repositories where this extension is disabled won't have the field.
# But if there's a requirement, then the extension must be loaded AND
# there may be blobs to push.
remotestore = repo.svfs.lfsremoteblobstore
try:
repo.svfs.lfsremoteblobstore = blobstore.remote(repo, remote.url())
return orig(repo, remote, *args, **kwargs)
finally:
repo.svfs.lfsremoteblobstore = remotestore
else:
return orig(repo, remote, *args, **kwargs)
# when writing a bundle via "hg bundle" command, upload related LFS blobs
@eh.wrapfunction(bundle2, b'writenewbundle')
def writenewbundle(
orig, ui, repo, source, filename, bundletype, outgoing, *args, **kwargs
):
"""upload LFS blobs added by outgoing revisions on 'hg bundle'"""
uploadblobsfromrevs(repo, outgoing.missing)
return orig(
ui, repo, source, filename, bundletype, outgoing, *args, **kwargs
)
def extractpointers(repo, revs):
"""return a list of lfs pointers added by given revs"""
repo.ui.debug(b'lfs: computing set of blobs to upload\n')
pointers = {}
makeprogress = repo.ui.makeprogress
with makeprogress(
_(b'lfs search'), _(b'changesets'), len(revs)
) as progress:
for r in revs:
ctx = repo[r]
for p in pointersfromctx(ctx).values():
pointers[p.oid()] = p
progress.increment()
return sorted(pointers.values(), key=lambda p: p.oid())
def pointerfromctx(ctx, f, removed=False):
"""return a pointer for the named file from the given changectx, or None if
the file isn't LFS.
Optionally, the pointer for a file deleted from the context can be returned.
Since no such pointer is actually stored, and to distinguish from a non LFS
file, this pointer is represented by an empty dict.
"""
_ctx = ctx
if f not in ctx:
if not removed:
return None
if f in ctx.p1():
_ctx = ctx.p1()
elif f in ctx.p2():
_ctx = ctx.p2()
else:
return None
fctx = _ctx[f]
if not _islfs(fctx.filelog()._revlog, fctx.filenode()):
return None
try:
p = pointer.deserialize(fctx.rawdata())
if ctx == _ctx:
return p
return {}
except pointer.InvalidPointer as ex:
raise error.Abort(
_(b'lfs: corrupted pointer (%s@%s): %s\n')
% (f, short(_ctx.node()), ex)
)
def pointersfromctx(ctx, removed=False):
"""return a dict {path: pointer} for given single changectx.
If ``removed`` == True and the LFS file was removed from ``ctx``, the value
stored for the path is an empty dict.
"""
result = {}
m = ctx.repo().narrowmatch()
# TODO: consider manifest.fastread() instead
for f in ctx.files():
if not m(f):
continue
p = pointerfromctx(ctx, f, removed=removed)
if p is not None:
result[f] = p
return result
def uploadblobs(repo, pointers):
"""upload given pointers from local blobstore"""
if not pointers:
return
remoteblob = repo.svfs.lfsremoteblobstore
remoteblob.writebatch(pointers, repo.svfs.lfslocalblobstore)
@eh.wrapfunction(upgrade, b'_finishdatamigration')
def upgradefinishdatamigration(orig, ui, srcrepo, dstrepo, requirements):
orig(ui, srcrepo, dstrepo, requirements)
# Skip if this hasn't been passed to reposetup()
if util.safehasattr(
srcrepo.svfs, b'lfslocalblobstore'
) and util.safehasattr(dstrepo.svfs, b'lfslocalblobstore'):
srclfsvfs = srcrepo.svfs.lfslocalblobstore.vfs
dstlfsvfs = dstrepo.svfs.lfslocalblobstore.vfs
for dirpath, dirs, files in srclfsvfs.walk():
for oid in files:
ui.write(_(b'copying lfs blob %s\n') % oid)
lfutil.link(srclfsvfs.join(oid), dstlfsvfs.join(oid))
@eh.wrapfunction(upgrade, b'preservedrequirements')
@eh.wrapfunction(upgrade, b'supporteddestrequirements')
def upgraderequirements(orig, repo):
reqs = orig(repo)
if b'lfs' in repo.requirements:
reqs.add(b'lfs')
return reqs