Hi Kim,
On Sun, 15 Apr 2018, Kim Gybels wrote:
> On (13/04/18 15:03), Johannes Schindelin wrote:
> > I wonder whether you found a reliable way to trigger this? It would be
> > nice to have a regression test for this.
>
> On my system, it reproduced reliably using Oleg's example [1], below is
> my bash version of it.
Okay.
> Script to generate repository with some history:
>
> $ cat example.sh
> #!/bin/bash
>
> git init example
> cd example
>
> git --help > foo.txt
>
> for i in $(seq 1 12); do
> cat foo.txt foo.txt > bar.txt
> mv bar.txt foo.txt
> git add foo.txt
> git commit -m v$i
> done
Okay, so this sets up a minimal repository with a moderate size, inflating
foo.txt from the initial Git help text by factor 2**12 = 4096. The help
text is around 2kB, so we end up with an ~8MB large file in the end that
grew exponentially.
> $ ./example.sh
> Initialized empty Git repository in C:/git/bug/example/.git/
> [master (root-commit) 2e44b4a] v1
> 1 file changed, 84 insertions(+)
> create mode 100644 foo.txt
> [master 9791332] v2
> 1 file changed, 84 insertions(+)
> [master 524e672] v3
> 1 file changed, 168 insertions(+)
> [master afec6ef] v4
> 1 file changed, 336 insertions(+)
> [master 1bcd9cc] v5
> 1 file changed, 672 insertions(+)
> [master 2f38a8e] v6
> 1 file changed, 1344 insertions(+)
> [master 33382fe] v7
> 1 file changed, 2688 insertions(+)
> [master 6c2cbd6] v8
> 1 file changed, 5376 insertions(+)
> [master 8d0770f] v9
> 1 file changed, 10752 insertions(+)
> [master 517d650] v10
> 1 file changed, 21504 insertions(+)
> [master 9e12406] v11
> 1 file changed, 43008 insertions(+)
> [master 4c4f600] v12
> 1 file changed, 86016 insertions(+)
>
> Server side:
>
> $ git daemon --verbose --reuseaddr --base-path=$(pwd) --export-all
> [4760] Ready to rumble
> [696] Connection from 127.0.0.1:2054
> [696] unable to set SO_KEEPALIVE on socket: No such file or directory
> [696] Extended attribute "host": 127.0.0.1
> [696] Request upload-pack for '/example'
I guess apart from the generated repo size (which would make this an
expensive test in and of itself), the fact that we have to run the daemon
(and that lib-git-daemon.sh requires the PIPE prerequisite which we
disabled on Windows) makes it hard to turn this into a regular regression
test in t/.
Although we *could* of course introduce a test that does not use
lib-git-daemon.sh and is hidden behind the EXPENSIVE prerequisite or some
such.
BTW I just tested this, and it indeed fixes the problem, so I am eager to
ship it to the Git for Windows users. Let's see whether we can convince
Junio that your quite small and easy-to-review patches are not too much of
a maintenance burden, and the fact that they fix a long-standing bug
should help.
BTW I had to apply this to build with DEVELOPER=1:
diff --git a/daemon.c b/daemon.c
index 97fadd62d10..1cc901e9739 100644
--- a/daemon.c
+++ b/daemon.c
@@ -1162,13 +1162,13 @@ static int service_loop(struct socketlist
*socklist)
signal(SIGCHLD, child_handler);
for (;;) {
- int i;
+ int i, ret;
check_dead_children();
#ifdef NO_POLL
poll_timeout = live_children ? 100 : -1;
#endif
- int ret = poll(pfd, socklist->nr, poll_timeout);
+ ret = poll(pfd, socklist->nr, poll_timeout);
if (ret == 0) {
continue;
} else if (ret < 0) {
Would you mind squashing that in?
Thanks,
Dscho
