statd calls atexit(statd_unregister) to unregister statd service on exit,
which actually has a side-effect that ha_callout() unregisters statd
service even when the child callout process exits on execl() failure.
Certain clustering software's deployment script adds -H option with its
specified file non-existent, when it is configured not to use callout.
In other words, -H seems to be used no matter if callout is needed or not,
but when callout is unnecessary, the specified callout program is not
deployed.
This causes statd not to work once a lock is requested by its NFS client,
as execl() in ha_callout() results in ENOENT and exit() of the child
process calls exit-handler statd_unregister(). Eventually, the NFS client
gets stuck with messages "lockd: cannot monitor xxx" on the NFS server.
Also, execl() could fail for other reasons like ENFILE or EIO as well.
A forked child must not unregister the statd RPC server, so use
_exit(), which does not call any exit-handlers, instead of exit().
Signed-off-by: Toshiaki Makita <[email protected]>
Reviewed-by: Chuck Lever <[email protected]>
---
v2:
- Simplified changelog.
support/include/ha-callout.h | 2 +-
1 files changed, 1 insertions(+), 1 deletions(-)
diff --git a/support/include/ha-callout.h b/support/include/ha-callout.h
index 1164336..a454bdb 100644
--- a/support/include/ha-callout.h
+++ b/support/include/ha-callout.h
@@ -47,7 +47,7 @@ ha_callout(char *event, char *arg1, char *arg2, int arg3)
arg3 < 0 ? NULL : buf,
NULL);
perror("execl");
- exit(2);
+ _exit(2);
case -1: perror("fork");
break;
default: pid = waitpid(pid, &ret, 0);
--
1.7.1
Hi Steve,
On 2016/02/16 9:36, Toshiaki Makita wrote:
> statd calls atexit(statd_unregister) to unregister statd service on exit,
> which actually has a side-effect that ha_callout() unregisters statd
> service even when the child callout process exits on execl() failure.
>
> Certain clustering software's deployment script adds -H option with its
> specified file non-existent, when it is configured not to use callout.
> In other words, -H seems to be used no matter if callout is needed or not,
> but when callout is unnecessary, the specified callout program is not
> deployed.
> This causes statd not to work once a lock is requested by its NFS client,
> as execl() in ha_callout() results in ENOENT and exit() of the child
> process calls exit-handler statd_unregister(). Eventually, the NFS client
> gets stuck with messages "lockd: cannot monitor xxx" on the NFS server.
>
> Also, execl() could fail for other reasons like ENFILE or EIO as well.
>
> A forked child must not unregister the statd RPC server, so use
> _exit(), which does not call any exit-handlers, instead of exit().
>
> Signed-off-by: Toshiaki Makita <[email protected]>
> Reviewed-by: Chuck Lever <[email protected]>
Would you tell me the status of this patch?
Regards,
Toshiaki Makita
Hey,
On 03/02/2016 07:20 PM, Toshiaki Makita wrote:
> Hi Steve,
>
> On 2016/02/16 9:36, Toshiaki Makita wrote:
>> statd calls atexit(statd_unregister) to unregister statd service on exit,
>> which actually has a side-effect that ha_callout() unregisters statd
>> service even when the child callout process exits on execl() failure.
>>
>> Certain clustering software's deployment script adds -H option with its
>> specified file non-existent, when it is configured not to use callout.
>> In other words, -H seems to be used no matter if callout is needed or not,
>> but when callout is unnecessary, the specified callout program is not
>> deployed.
>> This causes statd not to work once a lock is requested by its NFS client,
>> as execl() in ha_callout() results in ENOENT and exit() of the child
>> process calls exit-handler statd_unregister(). Eventually, the NFS client
>> gets stuck with messages "lockd: cannot monitor xxx" on the NFS server.
>>
>> Also, execl() could fail for other reasons like ENFILE or EIO as well.
>>
>> A forked child must not unregister the statd RPC server, so use
>> _exit(), which does not call any exit-handlers, instead of exit().
>>
>> Signed-off-by: Toshiaki Makita <[email protected]>
>> Reviewed-by: Chuck Lever <[email protected]>
>
> Would you tell me the status of this patch?
Its on my too do list.... I've been traveling but have every
intention on catching up asap...
steved.
>
> Regards,
> Toshiaki Makita
>
>
On 2016/03/03 9:27, Steve Dickson wrote:
> Hey,
>
> On 03/02/2016 07:20 PM, Toshiaki Makita wrote:
>> Hi Steve,
>>
>> On 2016/02/16 9:36, Toshiaki Makita wrote:
>>> statd calls atexit(statd_unregister) to unregister statd service on exit,
>>> which actually has a side-effect that ha_callout() unregisters statd
>>> service even when the child callout process exits on execl() failure.
>>>
>>> Certain clustering software's deployment script adds -H option with its
>>> specified file non-existent, when it is configured not to use callout.
>>> In other words, -H seems to be used no matter if callout is needed or not,
>>> but when callout is unnecessary, the specified callout program is not
>>> deployed.
>>> This causes statd not to work once a lock is requested by its NFS client,
>>> as execl() in ha_callout() results in ENOENT and exit() of the child
>>> process calls exit-handler statd_unregister(). Eventually, the NFS client
>>> gets stuck with messages "lockd: cannot monitor xxx" on the NFS server.
>>>
>>> Also, execl() could fail for other reasons like ENFILE or EIO as well.
>>>
>>> A forked child must not unregister the statd RPC server, so use
>>> _exit(), which does not call any exit-handlers, instead of exit().
>>>
>>> Signed-off-by: Toshiaki Makita <[email protected]>
>>> Reviewed-by: Chuck Lever <[email protected]>
>>
>> Would you tell me the status of this patch?
> Its on my too do list.... I've been traveling but have every
> intention on catching up asap...
I just wanted to know if it is being processed and not in hurry ;)
Thank you.
Regards,
Toshiaki Makita
On 02/15/2016 07:36 PM, Toshiaki Makita wrote:
> statd calls atexit(statd_unregister) to unregister statd service on exit,
> which actually has a side-effect that ha_callout() unregisters statd
> service even when the child callout process exits on execl() failure.
>
> Certain clustering software's deployment script adds -H option with its
> specified file non-existent, when it is configured not to use callout.
> In other words, -H seems to be used no matter if callout is needed or not,
> but when callout is unnecessary, the specified callout program is not
> deployed.
> This causes statd not to work once a lock is requested by its NFS client,
> as execl() in ha_callout() results in ENOENT and exit() of the child
> process calls exit-handler statd_unregister(). Eventually, the NFS client
> gets stuck with messages "lockd: cannot monitor xxx" on the NFS server.
>
> Also, execl() could fail for other reasons like ENFILE or EIO as well.
>
> A forked child must not unregister the statd RPC server, so use
> _exit(), which does not call any exit-handlers, instead of exit().
>
> Signed-off-by: Toshiaki Makita <[email protected]>
> Reviewed-by: Chuck Lever <[email protected]>
Committed....
steved.
> ---
> v2:
> - Simplified changelog.
>
> support/include/ha-callout.h | 2 +-
> 1 files changed, 1 insertions(+), 1 deletions(-)
>
> diff --git a/support/include/ha-callout.h b/support/include/ha-callout.h
> index 1164336..a454bdb 100644
> --- a/support/include/ha-callout.h
> +++ b/support/include/ha-callout.h
> @@ -47,7 +47,7 @@ ha_callout(char *event, char *arg1, char *arg2, int arg3)
> arg3 < 0 ? NULL : buf,
> NULL);
> perror("execl");
> - exit(2);
> + _exit(2);
> case -1: perror("fork");
> break;
> default: pid = waitpid(pid, &ret, 0);
>