Addurl can fail due to an apparent race condition when watch or assistant is running and the repository is in direct mode. The following stress test script encounters the bug consistently on my system. I am running git-annex 5.20140320 on on Ubuntu 13.10.
#!/bin/sh
set -eu
cleanup() {
local dir
dir="$1"; shift
if [ -d "$dir" ]; then
(
set -x
fuser -k -w "$dir/annex/.git/annex/daemon.log" || :
find "$dir" -type d -exec chmod 700 '{}' '+'
find "$dir" -type f -exec chmod 600 '{}' '+'
rm -fr "$dir"
)
fi
}
go() {
local dir
dir="$(mktemp -d "${TMP:-/tmp}/stress-annex.XXXXXXXXXX")"
trap "cleanup '$dir'" 0 1 2 13 15
(
cd "$dir"
mkdir annex
cd annex
set -x
git init
git annex init
git annex direct
git annex watch
for n in $(seq 100); do
git annex addurl --file=foo http://heh.fi/robots.txt
git annex sync
rm -f foo
git annex sync
done
git annex watch --stop
git annex uninit
)
cleanup "$dir"
trap - 0 1 2 13 14
}
go
Script output:
% ./stress-annex
+ git init
Initialized empty Git repository in /tmp/stress-annex.OKj6D8kVmV/annex/.git/
+ git annex init
init ok
(Recording state in git...)
+ git annex direct
commit
On branch master
Initial commit
nothing to commit
ok
direct ok
+ git annex watch
+ seq 100
+ git annex addurl --file=foo http://heh.fi/robots.txt
addurl foo (downloading http://heh.fi/robots.txt ...)
--2014-03-27 03:14:29-- http://heh.fi/robots.txt
Resolving heh.fi (heh.fi)... 83.145.237.222
Connecting to heh.fi (heh.fi)|83.145.237.222|:80... connected.
HTTP request sent, awaiting response... 200 OK
Length: 0 [text/plain]
Saving to: ‘/tmp/stress-annex.OKj6D8kVmV/annex/.git/annex/tmp/URL--http&c%%heh.fi%robots.txt’
[ <=> ] 0 --.-K/s in 0s
2014-03-27 03:14:29 (0.00 B/s) - ‘/tmp/stress-annex.OKj6D8kVmV/annex/.git/annex/tmp/URL--http&c%%heh.fi%robots.txt’ saved [0/0]
(Recording state in git...)
ok
(Recording state in git...)
+ git annex sync
commit ok
+ rm -f foo
+ git annex sync
commit (Recording state in git...)
ok
(Recording state in git...)
+ git annex addurl --file=foo http://heh.fi/robots.txt
addurl foo (downloading http://heh.fi/robots.txt ...)
--2014-03-27 03:14:29-- http://heh.fi/robots.txt
Resolving heh.fi (heh.fi)... 83.145.237.222
Connecting to heh.fi (heh.fi)|83.145.237.222|:80... connected.
HTTP request sent, awaiting response... 200 OK
Length: 0 [text/plain]
Saving to: ‘/tmp/stress-annex.OKj6D8kVmV/annex/.git/annex/tmp/URL--http&c%%heh.fi%robots.txt’
[ <=> ] 0 --.-K/s in 0s
2014-03-27 03:14:29 (0.00 B/s) - ‘/tmp/stress-annex.OKj6D8kVmV/annex/.git/annex/tmp/URL--http&c%%heh.fi%robots.txt’ saved [0/0]
(Recording state in git...)
ok
(Recording state in git...)
+ git annex sync
commit ok
+ rm -f foo
+ git annex sync
commit (Recording state in git...)
ok
(Recording state in git...)
+ git annex addurl --file=foo http://heh.fi/robots.txt
addurl foo (downloading http://heh.fi/robots.txt ...)
--2014-03-27 03:14:29-- http://heh.fi/robots.txt
Resolving heh.fi (heh.fi)... 83.145.237.222
Connecting to heh.fi (heh.fi)|83.145.237.222|:80... connected.
HTTP request sent, awaiting response... 200 OK
Length: 0 [text/plain]
Saving to: ‘/tmp/stress-annex.OKj6D8kVmV/annex/.git/annex/tmp/URL--http&c%%heh.fi%robots.txt’
[ <=> ] 0 --.-K/s in 0s
2014-03-27 03:14:29 (0.00 B/s) - ‘/tmp/stress-annex.OKj6D8kVmV/annex/.git/annex/tmp/URL--http&c%%heh.fi%robots.txt’ saved [0/0]
git-annex: /tmp/stress-annex.OKj6D8kVmV/annex/.git/annex/objects/pX/ZJ/SHA256E-s0--e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855/: openTempFile: permission denied (Permission denied)
failed
git-annex: addurl: 1 failed
+ fuser -k -w /tmp/stress-annex.OKj6D8kVmV/annex/.git/annex/daemon.log
/tmp/stress-annex.OKj6D8kVmV/annex/.git/annex/daemon.log: 30704 30709 30735 30738 30778
+ find /tmp/stress-annex.OKj6D8kVmV -type d -exec chmod 700 {} +
+ find /tmp/stress-annex.OKj6D8kVmV -type f -exec chmod 600 {} +
+ rm -fr /tmp/stress-annex.OKj6D8kVmV
The script also seems to encounter another issue. The output when seq 100 is changed to seq 1 and addurl happens to succeed:
+ git init
Initialized empty Git repository in /tmp/stress-annex.QEs0pNyS9z/annex/.git/
+ git annex init
init ok
(Recording state in git...)
+ git annex direct
commit
On branch master
Initial commit
nothing to commit
ok
direct ok
+ git annex watch
+ seq 1
+ git annex addurl --file=foo http://heh.fi/robots.txt
addurl foo (downloading http://heh.fi/robots.txt ...)
--2014-03-27 03:17:20-- http://heh.fi/robots.txt
Resolving heh.fi (heh.fi)... 83.145.237.222
Connecting to heh.fi (heh.fi)|83.145.237.222|:80... connected.
HTTP request sent, awaiting response... 200 OK
Length: 0 [text/plain]
Saving to: ‘/tmp/stress-annex.QEs0pNyS9z/annex/.git/annex/tmp/URL--http&c%%heh.fi%robots.txt’
[ <=> ] 0 --.-K/s in 0s
2014-03-27 03:17:20 (0.00 B/s) - ‘/tmp/stress-annex.QEs0pNyS9z/annex/.git/annex/tmp/URL--http&c%%heh.fi%robots.txt’ saved [0/0]
(Recording state in git...)
ok
(Recording state in git...)
+ git annex sync
commit ok
+ rm -f foo
+ git annex sync
commit (Recording state in git...)
ok
(Recording state in git...)
+ git annex watch --stop
+ git annex uninit
git-annex: /tmp/stress-annex.QEs0pNyS9z/annex/.git/annex/objects/pX/ZJ/SHA256E-s0--e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855/SHA256E-s0--e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855.map: removeLink: permission denied (Permission denied)
+ fuser -k -w /tmp/stress-annex.QEs0pNyS9z/annex/.git/annex/daemon.log
+ :
+ find /tmp/stress-annex.QEs0pNyS9z -type d -exec chmod 700 {} +
+ find /tmp/stress-annex.QEs0pNyS9z -type f -exec chmod 600 {} +
+ rm -fr /tmp/stress-annex.QEs0pNyS9z
Here’s another race condition which seems related:
The only lines in daemon.log from that time:
These races look beniegn, as far as I can see it doesn't cause any data to be lost, or indeed anything to happen that wouldn't have happened if addurl had been run without the assistant running.
The first race probably has addurl and the assistant both trying to move the file object into the annex at the same time. One wins and moves it; the other loses and sulks.
The second race has addurl trying to
git addthe file, while the assistant has already noticed the file appeared,git added it, and committed the add.The only way to really avoid these races would be to add a lot of lock checking. Or just make
git annex addurland presumably alsogit annex addand maybe several other commands refuse to run when the assistant is running.