Re: [libmicrohttpd] Require info on setting TCP_NODELAY for MHD

[Badari Prasad]

Thank you Christian for the response.

Thanks

 badari 
 

On Wed, Nov 21, 2018 at 9:57 PM Christian Grothoff <address@hidden> wrote:

Dear Badari,

Thanks for your surprisingly timely post. I just this week initiated an
off-list discussion with Evgeny about some odd corking/nodelay behavior
one of my students observed on OS X.  So this is very much precisely on
my radar these days (but not yet a priority).

Based on my discussion with Evgeny, the *desired* behavior should be
that MHD uncorks after (1) a large HTTP header *unless* the response
fits with the header in ~1200 bytes [Q: change behavior for 200 vs.
other status codes?], (2) at the end of the response, and (3) if the
application did send data and then calls suspend (without completing the
response).

That said, this is not yet implemented correctly (suspend not at all,
the rest seems to be broken on some OSes and fine on others...), and
would need to be both documented / specified and tested.

Help with this would be welcome, in the meantime all I can say is that
it is on our radar...

Happy hacking!

Christian

On 11/21/18 4:37 PM, Badari Prasad wrote:
> Hi Christian, 
>  My application runs on fedora and is single threaded, for my testing
> application just receives the request and responds with 200OK with some
> json body (size <100 bytes )in the response.
>
> I did try running strace on MHD  this is what I got. (client I am using
> libcurl application which sends 33k of body in http post with out Expect
> 100 header) 
>
> 14:38:14.623432 epoll_wait(12, [{EPOLLIN|EPOLLOUT, {u32=21908544,
> u64=21908544}}], 128, 0) = 1
> *14:38:14.623509 recvfrom(13, "POST /hello/TestMsg/async"..., 16384, 0,
> NULL, NULL) = 16384*
> 14:38:14.623825 rt_sigprocmask(SIG_UNBLOCK, [HUP QUIT PIPE ALRM TERM
> CHLD CONT TSTP TTIN TTOU WINCH IO PWR], NULL, 8) = 0
> 14:38:14.623880 epoll_wait(5, [], 2, 50) = 0
> 14:38:14.674038 rt_sigprocmask(SIG_BLOCK, [HUP QUIT PIPE ALRM TERM CHLD
> CONT TSTP TTIN TTOU WINCH IO PWR], NULL, 8) = 0
> 14:38:14.674135 epoll_wait(12, [], 128, 0) = 0
> *14:38:14.674175 recvfrom(13, "ey\": \"4\"\n},{\n    \"key\": \"0\",\n 
>  "..., 15919, 0, NULL, NULL) = 15919*
> 14:38:14.674329 rt_sigprocmask(SIG_UNBLOCK, [HUP QUIT PIPE ALRM TERM
> CHLD CONT TSTP TTIN TTOU WINCH IO PWR], NULL, 8) = 0
> 14:38:14.674375 epoll_wait(5, [], 2, 50) = 0
> 14:38:14.724568 rt_sigprocmask(SIG_BLOCK, [HUP QUIT PIPE ALRM TERM CHLD
> CONT TSTP TTIN TTOU WINCH IO PWR], NULL, 8) = 0
> 14:38:14.724666 epoll_wait(12, [], 128, 0) = 0
> *14:38:14.724715 recvfrom(13, "key\": \"3\",\n    \"key\": \"4\"\n},{\n 
> "..., 15919, 0, NULL, NULL) = 1878*
> ....
> ....
>
> 14:38:14.743832 setsockopt(13, SOL_TCP, TCP_NODELAY, [0], 4) = 0
>
> 14:38:14.743908 setsockopt(13, SOL_TCP, TCP_CORK, [1], 4) = 0
>
> *14:38:14.743949 sendto(13, "HTTP/1.1 200 OK\r\nConnection: Kee"...,
> 162, MSG_NOSIGNAL, NULL, 0) = 162*
>
> 14:38:14.744022 setsockopt(13, SOL_TCP, TCP_CORK, [0], 4) = 0
>
> 14:38:14.744189 setsockopt(13, SOL_TCP, TCP_NODELAY, [1], 4) = 0
>
> 14:38:14.744234 setsockopt(13, SOL_TCP, TCP_NODELAY, [0], 4) = 0
>
> 14:38:14.744286 setsockopt(13, SOL_TCP, TCP_CORK, [1], 4) = 0    
>
> Time diff I see is about 120 milli secs between MHD receiving request
> and sending response to the client.
>
> With this I got to understand that TCP_CORK and TCP_NODELAY is kind off
> delaying the response being sent out on transport layer... 
> Is their a way to always enable TCP_NODELAY on MHD ? and check if we see
> this delay on transport layer ?
>
> Thanks in advance,
>  Badari 
>
>
> On Sat, Nov 10, 2018 at 11:53 PM Christian Grothoff <address@hidden
> <mailto:address@hidden>> wrote:
>
>     On 11/9/18 5:44 PM, Badari Prasad wrote:
>     > Hi,
>     >   I was going through old mailist archives, found one thread on
>     > TCP_NODELAY
>     >
>     : https://lists.gnu.org/archive/html/libmicrohttpd/2015-11/msg00077.html
>     > But couldn't know if the option provided is it at compile time flag or
>     > runtime flag per connection basis.
>     >
>     > Kindly let me know the information on TCP_NODELAY option at MHD.
>
>     Hi!
>
>     To the best of my knowledge (i.e. unless there is a bug), TCP_NODELAY
>     should be set/unset by MHD automatically whenever it makes sense to do
>     so. Your application doesn't need to do anything here.
>
>     If you find/suspect that corking is inappropriately used, please send a
>     packet trace and maybe a bit of explanation as to how your code works
>     (in particular, which platform you use and which threading mode is
>     in use).
>
>     Happy hacking!
>
>     Christian
>