fix race in localrepo.addchangegroup.
localrepo.addchangegroup writes to changelog, then manifest, then normal
files. this breaks access ordering. if reader reads changelog while
manifest is being written, can find pointers into places in manifest
that are not yet written. same can happen for manifest and normal files.
fix is to make almost no change to localrepo.addchangegroup. it must
to write changelog and manifest data early because it has to read them
while writing other files. instead, write changelog and manifest data
to temp file that reader cannot see, then append temp data to manifest
after all normal files written, finally append temp data to changelog.
temp file code is in new appendfile module. can be used in other places
with small changes.
much smaller race still left. we write all new data in one write call,
but reader can maybe see partial update because python or os or filesystem
cannot always make write really atomic. file locking no help: slow, not
portable, not reliable over nfs. only real safe other plan is write to
temp file every time and rename, but performance bad when manifest or
changelog is big.
#!/bin/sh
# No local source
hg clone a b
echo $?
# No remote source
hg clone http://127.0.0.1:3121/a b
echo $?
rm -rf b # work around bug with http clone
# Inaccessible source
mkdir a
chmod 000 a
hg clone a b
echo $?
# Inaccessible destination
mkdir b
cd b
hg init
hg clone . ../a
echo $?
cd ..
chmod 700 a
rm -rf a b
# Source of wrong type
mkfifo a
hg clone a b
echo $?
rm a
# Default destination, same directory
mkdir q
cd q
hg init
cd ..
hg clone q
true