Re: [libmicrohttpd] Select timeout being set to 0

[Christian Grothoff]

Dear Jon,

I've tried (and failed) to reproduce your problem.  I've attached my 
modified version of 'minimal_example.c' (modified to use a thread pool). 
 When I connect to the server with firefox, the socket stays open (will 
be re-used by the browser) but MHD does NOT use any CPU after the page 
is served initially.

I do recall a bugs that could cause similar behavior of MHD looping in 
older MHD versions (prior to 4.5) and there was also a bug with looping 
for up to 1s before 0.9.12.  Which MHD version are you using? (We're 
currently at 0.9.15, another release might happen later this weekend.)

Happy hacking!

Christian

On 10/27/2011 10:07 PM, Jon Nalley wrote:
> Hi,
>
> I am seeing some odd behavior when using MHD_USE_SELECT_INTERNALLY +
> MHD_OPTION_THREAD_POOL_SIZE.
>
> When a request is handled, my CPU usage goes to 100% (I am running on a
> PowerPC 405ex - Linux+uClibc).
>
> I was able to determine that MHD_select() in src/daemon/daemon.c is setting
> the select timeout to 0.
>
> The following check is true, and ltimeout is 0 which causes the timeout
> passed to select() to be 0.
>
> else if ( (0 == (daemon->options&  MHD_USE_THREAD_PER_CONNECTION))&&
>              (MHD_YES == MHD_get_timeout (daemon,&ltimeout)) )
>
> If I use MHD_USE_THREAD_PER_CONNECTION instead of MHD_USE_SELECT_INTERNALLY
> + MHD_OPTION_THREAD_POOL_SIZE I do *NOT* see the issue.
>
> I took a look at MHD_get_timeout() and noticed I am hitting the following:
>
> if (earliest_deadline<  now)
>      *timeout = 0;
>
> I also confirmed that if I close my browser, the connections are closed and
> the select() is once again called with a NULL timeout (blocking forever).
>
> I am not sure exactly what is going on here.  It seems like the browser (I
> tested with Chrome and Firefox) is keeping the connection open, which causes
> MHD_get_timeout() to return MHD_YES (and set ltimeout = 0 in MHD_select())
> which results in my process effectively entering a "busy loop" calling
> select with a timeout of 0 repeatedly.  The page has been fully loaded by
> the browser, so I am not sure why the connections remain "ESTABLISHED" (as
> reported by netstat -atn).  Perhaps the browser keeps the connections open
> in order to speed up subsequent requests?
>
> I would appreciate any feedback as to what I may be overlooking and how I
> can resolve this issue.
>
> Regards,
>
> Jon Nalley

minimal_example.c
Description: Text Data