[RFC,2/2] netstress: Update help for -m behavior

For udp, udp_lite and DCCP on client is timeout set 100.
For udp and udp_lite on server on low timeout it reaches "recv failed"
error (due no listen in udp and udp_lite).

Signed-off-by: Petr Vorel <pvorel@suse.cz>
---
I documented the current state, but I it looks to me that udp and
udp_lite on server is a bug. Or wanted behaviour?


Kind regards,
Petr
---
 testcases/network/netstress/netstress.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

On 19.07.2018 15:53, Petr Vorel wrote:
> For udp, udp_lite and DCCP on client is timeout set 100.
> For udp and udp_lite on server on low timeout it reaches "recv failed"
> error (due no listen in udp and udp_lite).
> 
> Signed-off-by: Petr Vorel <pvorel@suse.cz>
> ---
> I documented the current state, but I it looks to me that udp and
> udp_lite on server is a bug. Or wanted behaviour?
> 
> 

Hi Petr,

Since the server waits for requests from the client, the timeout
value for UDP/DCCP is the same as for the other protocols. Also the
server starts earlier than the client, so it should wait some time
to get the client requests.

I've changed the client side only because either request from the
client or reply from the server might be lost.


> Kind regards,
> Petr
> ---
>  testcases/network/netstress/netstress.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/testcases/network/netstress/netstress.c b/testcases/network/netstress/netstress.c
> index 48fb0aabd..62301ca9f 100644
> --- a/testcases/network/netstress/netstress.c
> +++ b/testcases/network/netstress/netstress.c
> @@ -999,7 +999,7 @@ static struct tst_option options[] = {
>  	{"r:", &rarg, "-r x     Number of client requests"},
>  	{"n:", &narg, "-n x     Client message size"},
>  	{"N:", &Narg, "-N x     Server message size"},
> -	{"m:", &Targ, "-m x     Reply timeout in microsec."},
> +	{"m:", &Targ, "-m x     Reply timeout in microsec (only for tcp and sctp, also affects server on udp and udp_lite)."},

What about dccp? May be "Receive timeout in milliseconds (not affects UDP/DCCP client)"?

Thanks,
Alexey

Hi Alexey,

> Hi Petr,

> Since the server waits for requests from the client, the timeout
> value for UDP/DCCP is the same as for the other protocols. Also the
> server starts earlier than the client, so it should wait some time
> to get the client requests.

> I've changed the client side only because either request from the
> client or reply from the server might be lost.

Thanks for your explanation.
I mean: UDP itself is the only protocol which don't support listen() so
netstress server using UDP timeouts after some time. The default is 100ms.
But due obvious return before listen() for UDP in server_init() -m parameter
affects behavior of the server in UDP (how quickly it timeouts):

$ date +"%T.%3N"; testcases/network/netstress/netstress -m 1 -T udp; date +"%T.%3N"
15:52:34.501
tst_test.c:1015: INFO: Timeout per run is 0h 05m 00s
netstress.c:917: INFO: max requests '3'
netstress.c:944: INFO: using UDP
netstress.c:676: INFO: assigning a name to the server socket...
netstress.c:683: INFO: bind to port 47728
netstress.c:575: FAIL: recv failed, sock '3'
netstress.c:642: BROK: Server closed
...
15:52:34.516

$ date +"%T.%3N"; testcases/network/netstress/netstress -m 1000 -T udp; date +"%T.%3N"
16:01:27.064
...
16:01:28.079

date +"%T.%3N"; testcases/network/netstress/netstress -m 10000 -T udp; date +"%T.%3N"
16:00:39.647
...
16:00:49.672

I just wonder if we want to document it (or at least don't state -m as "client only config").

> > -	{"m:", &Targ, "-m x     Reply timeout in microsec."},
> > +	{"m:", &Targ, "-m x     Reply timeout in microsec (only for tcp and sctp, also affects server on udp and udp_lite)."},
> What about dccp?
Yes, I left it, thanks!
> May be "Receive timeout in milliseconds (not affects UDP/DCCP client)"?
Yes, this is better.


Kind regards,
Petr

On 07/24/2018 05:14 PM, Petr Vorel wrote:
> Hi Alexey,
> 
>> Hi Petr,
> 
>> Since the server waits for requests from the client, the timeout
>> value for UDP/DCCP is the same as for the other protocols. Also the
>> server starts earlier than the client, so it should wait some time
>> to get the client requests.
> 
>> I've changed the client side only because either request from the
>> client or reply from the server might be lost.
> 
> Thanks for your explanation.
> I mean: UDP itself is the only protocol which don't support listen() so
> netstress server using UDP timeouts after some time. The default is 100ms.

Hmm, for the server the default should be 60 sec. Does it timeout earlier?

It is more than enough for the tests to start the client on the other end
and start sending requests. If it hasn't happened during this time period
there is no reason to have it running there.

> But due obvious return before listen() for UDP in server_init() -m parameter
> affects behavior of the server in UDP (how quickly it timeouts):
> 

And with listen() it won't wait forever either, it will be killed by LTP
timeout per test run...

> $ date +"%T.%3N"; testcases/network/netstress/netstress -m 1 -T udp; date +"%T.%3N"
> 15:52:34.501
 > tst_test.c:1015: INFO: Timeout per run is 0h 05m 00s
> netstress.c:917: INFO: max requests '3'
> netstress.c:944: INFO: using UDP
> netstress.c:676: INFO: assigning a name to the server socket...
> netstress.c:683: INFO: bind to port 47728
> netstress.c:575: FAIL: recv failed, sock '3'
> netstress.c:642: BROK: Server closed
> ...
> 15:52:34.516

Thanks,
Alexey