Hi Petter and Steve,
On Monday 18 August 2008 16:46:08 Petter Reinholdtsen wrote:
> [Steve Langasek]
> > In this case, NFS...
>
> Right. So ext3 and NFS should be ok. :)
>
> > I can try against a different filesystem tomorrow if you like.
>
> If you happen to know what file system is used on the autobuilder, or
> if ext2/ext3 is not the default file system on alpha, it would be
> useful to have a test with those file systems as well. If not, it
> will be a shot in the dark, and I have no idea what file system to
> suggest.
I think insserv can handle this situation without disabling optimisations, the
problem seems to concern the use of an underterministic directory sequence,
as returned by a while readdir() loop, when scanning the init.d/ directory to
load LSB init info.
The fact that the problem was discovered with a symlink test, is I think
interesting, but not neccessarily an indication of some lower level problem
with kernel / filesystem interaction. The problem can be reproduced with a more
simple test too:
---
test_deterministic_order() {
echo
echo "info: test two or more initscripts providing same facility, making sure"
echo "info: that the first script can be registered with insserv, but others
fail."
echo
initdir_purge
addscript abc <<'EOF' || true
### BEGIN INIT INFO
# Provides: service
# Required-Start:
# Required-Stop:
# Default-Start: S
# Default-Stop:
### END INIT INFO
EOF
addscript xyz <<'EOF' || true
### BEGIN INIT INFO
# Provides: service
# Required-Start:
# Required-Stop:
# Default-Start: S
# Default-Stop:
### END INIT INFO
EOF
addscript hjk <<'EOF' || true
### BEGIN INIT INFO
# Provides: service
# Required-Start:
# Required-Stop:
# Default-Start: S
# Default-Stop:
### END INIT INFO
EOF
insserv_reg xyz || true
insserv_reg hjk || true
insserv_reg abc || true
list_rclinks
check_script_present S xyz
check_script_not_present S abc
check_script_not_present S hjk
}
---
The above test will always fail, but differently when run on different
filesystems
(eg. ext3 vs tmpfs). It seems the information that readdir(3) returns from a
filesystem is not always the same.
insserv could defend against this behaviour, by ensuring that the script being
registered is parsed first, followed by all other scripts. I've put together a
hack to do just that, it is attached. WIll commit it to the packaging SVN with
test case for further discussions.
Thanks, Kel.
---
--- a/insserv.c
+++ b/insserv.c
@@ -2238,6 +2238,7 @@
boolean del = false;
boolean defaults = false;
boolean ignore = false;
+ boolean loadarg = argc;
myname = basename(*argv);
@@ -2490,17 +2491,45 @@
/*
* Scan now all scripts found in the init.d/ directory
*/
- while ((d = readdir(initdir)) != (struct dirent*)0) {
- const boolean isarg = chkfor(d->d_name, argv, argc);
+ for (;;) {
service_t * service = (service_t*)0;
char * token;
char * begin = (char*)0; /* hold start pointer of strings
handled by strsep() */
boolean hard = false;
+ boolean isarg = false;
uchar lsb = 0;
#if defined(DEBUG) && (DEBUG > 0)
int nobug = 0;
#endif
+ if ((d = readdir(initdir)) == (struct dirent*)0) {
+ /*
+ * If first script in argument list was loaded in advance, then
+ * rewind the init.d/ directory stream and attempt to load all
+ * other scripts.
+ */
+ if (loadarg) {
+ loadarg = false;
+ rewinddir(initdir);
+ continue;
+ }
+ break;
+ }
+
+ isarg = chkfor(d->d_name, argv, argc);
+
+ /*
+ * Load first script in argument list before all other scripts. This
+ * avoids problems with loading services in underterministic sequence
+ * returned by readdir(3).
+ */
+ if (loadarg && !isarg)
+ continue;
+ else if (loadarg && isarg && (curr_argc != 0))
+ continue;
+ else if (!loadarg && isarg && (curr_argc == 0))
+ continue;
+
if (*d->d_name == '.')
continue;
errno = 0;
---
--
To UNSUBSCRIBE, email to [EMAIL PROTECTED]
with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
