Hi Nikita,
[ Cc: Steve as user-space maintainer, also Neil and whole linux-nfs ]
> On systemd-based linux hosts, rpcbind service is typically started via
> socket activation, when the first client connects. If no client has
> connected before LTP rpc test starts, rpcbind process will not be
> running at the time of check_portmap_rpcbind() execution, causing
> check_portmap_rpcbind() to report TCONF error.
> Fix that by adding a quiet invocation of 'rpcinfo' before checking for
> rpcbind.
Looks reasonable, but I'd prefer to have confirmation from NFS experts.
> For portmap, similar step is likely not needed, because portmap is used
> only on old systemd and those don't use systemd.
> Signed-off-by: Nikita Yushchenko <[email protected]>
> ---
> testcases/network/rpc/basic_tests/rpc_lib.sh | 6 ++++++
> 1 file changed, 6 insertions(+)
> diff --git a/testcases/network/rpc/basic_tests/rpc_lib.sh b/testcases/network/rpc/basic_tests/rpc_lib.sh
> index c7c868709..e882e41b3 100644
> --- a/testcases/network/rpc/basic_tests/rpc_lib.sh
> +++ b/testcases/network/rpc/basic_tests/rpc_lib.sh
> @@ -8,6 +8,12 @@ check_portmap_rpcbind()
> if pgrep portmap > /dev/null; then
> PORTMAPPER="portmap"
> else
> + # In case of systemd socket activation, rpcbind could be
> + # not started until somebody tries to connect to it's socket.
> + #
> + # To handle that case properly, run a client now.
> + rpcinfo >/dev/null 2>&1
nit: Shouldn't we keep stderr? In LTP we put required commands into
$TST_NEEDS_CMDS. It'd be better not require rpcinfo (not a hard dependency),
and thus it'd be better to see "command not found" when rpcinfo missing and test
fails.
Kind regards,
Petr
> +
> pgrep rpcbind > /dev/null && PORTMAPPER="rpcbind" || \
> tst_brk TCONF "portmap or rpcbind is not running"
> fi
21.01.2022 00:01, Petr Vorel wrote:
> Hi Nikita,
>
> [ Cc: Steve as user-space maintainer, also Neil and whole linux-nfs ]
>
>> On systemd-based linux hosts, rpcbind service is typically started via
>> socket activation, when the first client connects. If no client has
>> connected before LTP rpc test starts, rpcbind process will not be
>> running at the time of check_portmap_rpcbind() execution, causing
>> check_portmap_rpcbind() to report TCONF error.
>
>> Fix that by adding a quiet invocation of 'rpcinfo' before checking for
>> rpcbind.
>
> Looks reasonable, but I'd prefer to have confirmation from NFS experts.
NFS is not involved here, this is about sunrpc tests.
I had to add this patch to make 'runltp -f net.rpc' pass just after container is started - that happens
in container autotests here.
Nikita
Hi Nikita,
> 21.01.2022 00:01, Petr Vorel wrote:
> > Hi Nikita,
> > [ Cc: Steve as user-space maintainer, also Neil and whole linux-nfs ]
> > > On systemd-based linux hosts, rpcbind service is typically started via
> > > socket activation, when the first client connects. If no client has
> > > connected before LTP rpc test starts, rpcbind process will not be
> > > running at the time of check_portmap_rpcbind() execution, causing
> > > check_portmap_rpcbind() to report TCONF error.
> > > Fix that by adding a quiet invocation of 'rpcinfo' before checking for
> > > rpcbind.
> > Looks reasonable, but I'd prefer to have confirmation from NFS experts.
> NFS is not involved here, this is about sunrpc tests.
Sure. Just tirpc (in libtirpc or the the old SUN-RPC already removed from glibc)
are used in NFS. Steve is the libtirpc maintainer.
> I had to add this patch to make 'runltp -f net.rpc' pass just after
> container is started - that happens in container autotests here.
Yep, I suspected this. Because on normal linux distro it's working right after
boot (tested on rpc01.sh). Can't this be a setup issue?
Kind regards,
Petr
> Nikita
>
>> I had to add this patch to make 'runltp -f net.rpc' pass just after
>> container is started - that happens in container autotests here.
> Yep, I suspected this. Because on normal linux distro it's working right after
> boot (tested on rpc01.sh). Can't this be a setup issue?
This depends on what is installed and how it is configured.
But definitely the state with rpcbind process not running and systemd is listening on rpcbind sockets -
is valid.
In the setup where the issue was caught, the test harness creates a container with minimal centos8 setup
inside, boots it, and starts ltp inside.
Just reproduced manually:
[root@vz8 ~]# vzctl start 1000
Starting Container ...
...
[root@vz8 ~]# vzctl enter 1000
entered into CT 1000
CT-1000 /# pidof rpcbind
CT-1000 /# rpcinfo > /dev/null 2>&1
CT-1000 /# pidof rpcbind
678
CT-1000 /#
On Fri, 21 Jan 2022, Petr Vorel wrote:
> Hi Nikita,
>
> [ Cc: Steve as user-space maintainer, also Neil and whole linux-nfs ]
>
> > On systemd-based linux hosts, rpcbind service is typically started via
> > socket activation, when the first client connects. If no client has
> > connected before LTP rpc test starts, rpcbind process will not be
> > running at the time of check_portmap_rpcbind() execution, causing
> > check_portmap_rpcbind() to report TCONF error.
>
> > Fix that by adding a quiet invocation of 'rpcinfo' before checking for
> > rpcbind.
>
> Looks reasonable, but I'd prefer to have confirmation from NFS experts.
>
> > For portmap, similar step is likely not needed, because portmap is used
> > only on old systemd and those don't use systemd.
>
> > Signed-off-by: Nikita Yushchenko <[email protected]>
> > ---
> > testcases/network/rpc/basic_tests/rpc_lib.sh | 6 ++++++
> > 1 file changed, 6 insertions(+)
>
> > diff --git a/testcases/network/rpc/basic_tests/rpc_lib.sh b/testcases/network/rpc/basic_tests/rpc_lib.sh
> > index c7c868709..e882e41b3 100644
> > --- a/testcases/network/rpc/basic_tests/rpc_lib.sh
> > +++ b/testcases/network/rpc/basic_tests/rpc_lib.sh
> > @@ -8,6 +8,12 @@ check_portmap_rpcbind()
> > if pgrep portmap > /dev/null; then
> > PORTMAPPER="portmap"
> > else
> > + # In case of systemd socket activation, rpcbind could be
> > + # not started until somebody tries to connect to it's socket.
> > + #
> > + # To handle that case properly, run a client now.
> > + rpcinfo >/dev/null 2>&1
If it were me, I would remove the 'pgrep's and just call "rpcbind -p"
and make sure something responds.
NeilBrown
> nit: Shouldn't we keep stderr? In LTP we put required commands into
> $TST_NEEDS_CMDS. It'd be better not require rpcinfo (not a hard dependency),
> and thus it'd be better to see "command not found" when rpcinfo missing and test
> fails.
>
> Kind regards,
> Petr
>
> > +
> > pgrep rpcbind > /dev/null && PORTMAPPER="rpcbind" || \
> > tst_brk TCONF "portmap or rpcbind is not running"
> > fi
>
>
> On Fri, 21 Jan 2022, Petr Vorel wrote:
> > Hi Nikita,
> > [ Cc: Steve as user-space maintainer, also Neil and whole linux-nfs ]
> > > On systemd-based linux hosts, rpcbind service is typically started via
> > > socket activation, when the first client connects. If no client has
> > > connected before LTP rpc test starts, rpcbind process will not be
> > > running at the time of check_portmap_rpcbind() execution, causing
> > > check_portmap_rpcbind() to report TCONF error.
> > > Fix that by adding a quiet invocation of 'rpcinfo' before checking for
> > > rpcbind.
> > Looks reasonable, but I'd prefer to have confirmation from NFS experts.
> > > For portmap, similar step is likely not needed, because portmap is used
> > > only on old systemd and those don't use systemd.
> > > Signed-off-by: Nikita Yushchenko <[email protected]>
> > > ---
> > > testcases/network/rpc/basic_tests/rpc_lib.sh | 6 ++++++
> > > 1 file changed, 6 insertions(+)
> > > diff --git a/testcases/network/rpc/basic_tests/rpc_lib.sh b/testcases/network/rpc/basic_tests/rpc_lib.sh
> > > index c7c868709..e882e41b3 100644
> > > --- a/testcases/network/rpc/basic_tests/rpc_lib.sh
> > > +++ b/testcases/network/rpc/basic_tests/rpc_lib.sh
> > > @@ -8,6 +8,12 @@ check_portmap_rpcbind()
> > > if pgrep portmap > /dev/null; then
> > > PORTMAPPER="portmap"
> > > else
> > > + # In case of systemd socket activation, rpcbind could be
> > > + # not started until somebody tries to connect to it's socket.
> > > + #
> > > + # To handle that case properly, run a client now.
> > > + rpcinfo >/dev/null 2>&1
> If it were me, I would remove the 'pgrep's and just call "rpcbind -p"
> and make sure something responds.
Hi Neil,
I guess you mean: rpcinfo -p
Good idea, thanks!
Kind regards,
Petr
> NeilBrown
> > nit: Shouldn't we keep stderr? In LTP we put required commands into
> > $TST_NEEDS_CMDS. It'd be better not require rpcinfo (not a hard dependency),
> > and thus it'd be better to see "command not found" when rpcinfo missing and test
> > fails.
> > Kind regards,
> > Petr
> > > +
> > > pgrep rpcbind > /dev/null && PORTMAPPER="rpcbind" || \
> > > tst_brk TCONF "portmap or rpcbind is not running"
> > > fi