Created attachment 1042096 [details]
Patch - fix direct map expire not set for initial empty map

Hi, Ian,

Thanks for discovering and reporting this bug, and I think I understand the root cause and the purpose of your patch.

Based on more code reading, I think maybe it's better for us to put the supplement of "alarm_add" as following: 

--- ../autofs-5.0.7/daemon/state.c  2015-01-23 17:10:12.740443598 +0800
+++ daemon/state.c  2015-07-01 15:48:26.235034830 +0800
@@ -488,6 +486,7 @@
        status = lookup_ghost(ap, ap->path);
    } else {
        struct mapent *me;
+       int append_alarm = !ap->exp_runfreq;

        mnts = tree_make_mnt_tree(_PROC_MOUNTS, "/");
        pthread_cleanup_push(tree_mnts_cleanup, mnts);
@@ -516,6 +515,13 @@
            map->stale = 0;
            map = map->next;
        }
+
+       /* If the direct mount map was empty at startup no expire alarm will
+        * have been added. So add it here if there are now map entries.
+        */
+        if (append_alarm && ap->exp_runfreq)
+               alarm_add(ap, ap->exp_runfreq + rand() % ap->exp_runfreq);
+
        pthread_cleanup_pop(1);
        pthread_cleanup_pop(1);
        pthread_cleanup_pop(1);

Because in this way, we use the final ap->exp_runfreq as the timeout. And this behavor for alarm_add is consistent with the handle_mounts function.
I am not sure, just proposed it, you decide it.


Another question, what's the supposed behavor for expire if the map files(direct and indirect) become empty and reload called?

Thanks.

This problem is introduced by the following patch, so it can only be reproduced since autofs-5.0.7+

commit f3dedc7bf56a8b226ee9bb7bade8d58954c0a7de
Author: Leonardo Chiquitto <leonardo.lists>
Date:   Tue Jan 15 09:26:11 2013 +0800

    autofs-5.0.7 - don't schedule new alarms after readmap

(In reply to XuWang from comment #2)
> Hi, Ian,
> 
> Thanks for discovering and reporting this bug, and I think I understand the
> root cause and the purpose of your patch.
> 
> Based on more code reading, I think maybe it's better for us to put the
> supplement of "alarm_add" as following: 
> 
> --- ../autofs-5.0.7/daemon/state.c  2015-01-23 17:10:12.740443598 +0800
> +++ daemon/state.c  2015-07-01 15:48:26.235034830 +0800
> @@ -488,6 +486,7 @@
>         status = lookup_ghost(ap, ap->path);
>     } else {
>         struct mapent *me;
> +       int append_alarm = !ap->exp_runfreq;
> 
>         mnts = tree_make_mnt_tree(_PROC_MOUNTS, "/");
>         pthread_cleanup_push(tree_mnts_cleanup, mnts);
> @@ -516,6 +515,13 @@
>             map->stale = 0;
>             map = map->next;
>         }
> +
> +       /* If the direct mount map was empty at startup no expire alarm will
> +        * have been added. So add it here if there are now map entries.
> +        */
> +        if (append_alarm && ap->exp_runfreq)
> +               alarm_add(ap, ap->exp_runfreq + rand() % ap->exp_runfreq);
> +
>         pthread_cleanup_pop(1);
>         pthread_cleanup_pop(1);
>         pthread_cleanup_pop(1);
> 
> Because in this way, we use the final ap->exp_runfreq as the timeout. And
> this behavor for alarm_add is consistent with the handle_mounts function.
> I am not sure, just proposed it, you decide it.

I like this patch better than mine, I'll use it.

It has lower overhead because my change would cause an additional
comparison to be done for every direct mount map entry whereas this
will do a single comparison at the end.

> 
> 
> Another question, what's the supposed behavor for expire if the map
> files(direct and indirect) become empty and reload called?

The alarm will continue to trigger.
That was the expected behaviour before commit f3dedc7b.

An improvement could be made to remove the alarm when the map
becomes empty but isn't a high priority since maps usually aren't
empty and the alarm overhead is small.

Ian

(In reply to XuWang from comment #3)
> This problem is introduced by the following patch, so it can only be
> reproduced since autofs-5.0.7+
> 
> commit f3dedc7bf56a8b226ee9bb7bade8d58954c0a7de
> Author: Leonardo Chiquitto <leonardo.lists>
> Date:   Tue Jan 15 09:26:11 2013 +0800
> 
>     autofs-5.0.7 - don't schedule new alarms after readmap

Yes, it looks like that at first sight but alarm_add(ap, 0)
won't add an alarm (have a look at lib/alarm.c:alarm_add()).

This patch isn't in RHEL-6 autofs but the problem is still
present because of it. In fact that's were I first saw it.

Ian

Created attachment 1045341 [details]
Patch - fix direct map expire not set for initial empty map (updated)

run test case /autofs/sanity/reload for autofs-5.0.7-53.el7, seems good.

autofs previous version:

:: [  BEGIN   ] :: Running 'tests/expire.sh'
posix          	on
Redirecting to /bin/systemctl stop  autofs.service
Redirecting to /bin/systemctl restart  autofs.service
:: [  BEGIN   ] :: Running 'check_indirect_reload'
Redirecting to /bin/systemctl reload  autofs.service
:: [   PASS   ] :: Command 'check_indirect_reload' (Expected 0, got 0)
:: [  BEGIN   ] :: Running 'check_direct_reload'
Redirecting to /bin/systemctl reload  autofs.service
:: [   PASS   ] :: Command 'check_direct_reload' (Expected 0, got 0)
:: [  BEGIN   ] :: Running 'check_already_umount /indirect/plus'
:: [   PASS   ] :: Command 'check_already_umount /indirect/plus' (Expected 0, got 0)
:: [  BEGIN   ] :: Running 'check_already_umount /direct/plus'
:: [   FAIL   ] :: Command 'check_already_umount /direct/plus' (Expected 0, got 1)
Redirecting to /bin/systemctl stop  autofs.service
:: [   PASS   ] :: Command 'tests/expire.sh' (Expected 0, got 0)



autofs-5.0.7-53.el7:
[04:22:07 root@ ~~]# tests/expire.sh
posix          	on
Redirecting to /bin/systemctl stop  autofs.service
Redirecting to /bin/systemctl restart  autofs.service
:: [  BEGIN   ] :: Running 'check_indirect_reload'
Redirecting to /bin/systemctl reload  autofs.service
:: [   PASS   ] :: Command 'check_indirect_reload' (Expected 0, got 0)
:: [  BEGIN   ] :: Running 'check_direct_reload'
Redirecting to /bin/systemctl reload  autofs.service
:: [   PASS   ] :: Command 'check_direct_reload' (Expected 0, got 0)
:: [  BEGIN   ] :: Running 'check_already_umount /indirect/plus'
:: [   PASS   ] :: Command 'check_already_umount /indirect/plus' (Expected 0, got 0)
:: [  BEGIN   ] :: Running 'check_already_umount /direct/plus'
:: [   PASS   ] :: Command 'check_already_umount /direct/plus' (Expected 0, got 0)
Redirecting to /bin/systemctl stop  autofs.service
:: [   PASS   ] :: Running 'tests/expire.sh' (Expected 0, got 0)
-------------------------------------------------------------------------------


also run regression, connectathon, stress, bugzillas, sanity test cases for autofs-5.0.7-53.el7, covers x86_64, s390x, ppc64, ppc64le, no new issues.

so I will change this bug status to be verified.

Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.

For information on the advisory, and where to find the updated
files, follow the link below.

If the solution does not work for you, open a new bug report.

https://rhn.redhat.com/errata/RHSA-2015-2417.html