Conference turris::digital_unix

8851.0. "xti osi select with threads" by VARESE::BIOTTI () Mon Feb 17 1997 11:26

 {cross posted in CMA, DNU_OSI and DIGITAL_UNIX conf.}

 I'm using select and xti osi with threads. 

 I've realised that I've a problem starting from OSF32 
 and DNA32.   All was fine with OSF30 DNA30.

 When I issue more than one t_snd in a very short time from a thread
 and I wait for events (answers) coming from the network 
 with a select call in another thread, I observe that the 
 select only gets the first 2 or 3 events in an acceptable
 time, then select fires with an event only at a distance
 of some (3 or 4) seconds. 
 I've done a decnet osi thace with ctf and I see the delays. 

 select is called with a timeout of 1 second. 
 xti calls are wrapped. 
 I've tried to put pthread_lock_global_np and a pthread_unlock_global_np
 around select and I get good response times, that is select doesn't 
 complete many times with timeout before getting the event. 
 If I do all in a syncronous way it's fine even without select wrap.

 I know I have a work around that would be to decrease the timeout,
 wrap select and may be call pthread_delay_np or/and pthread_yeld to allow 
 other threads to take the control but I would prefer to understand 
 if there is a problem somewhere.
 
 With OSF40 (plus an xti patch sharable I've got for a t_look problem)
 the behaviour is even worst in the sense that select doesn't fire 
 anymore even if there are some data that should still come in.

 I've the problem with OSF32C too.
 I've the problem with DNA32A too. 
 I haven't installed yet DNA32B MUP, I've only read the release 
 notes and they don't seem to mention anything about a problem like the
 above.  Anyway the installation of DNA32B  is my next step. 

 Thanks for any advice 
 ABiotti Basestar Open group
  

    The XTI library was not thread safe till V4.0 (to get something 
    workable you need V4.0 + patches).
    
    		/ron

>    The XTI library was not thread safe till V4.0 (to get something 
>    workable you need V4.0 + patches).
    
  
  I know that.  And all my XTI calls are between lock and unlock
  global of pthreads on 30 32 and 40 too. 
  Anyway good to know that with 4.0 I can get rid of them.
  I've always used select without lock unlock. 
  I've also generated a .lis of the file when compiling and I've 
  seen that the 

  #define select cma_select   

  doesn't take effect.    That is I have  

  X #define select cma_select

  Is that a problem?  If yes wasn't that a problem on OSF30?

  Thanks

You may need the tli/xti patches that QAR#48633 produced. 

Regards, 

Arnold

    
    I already have them (for 4.0) and with them I've the problem. 
    

.2> I've  seen that the 
.2> 
.2>   #define select cma_select   
.2> 
.2> doesn't take effect.

That's not a problem on Digital Unix (i.e., what you're calling "OSF").  That
is old moldly code from our Ulitrix support...  It's quite unnecessary on
Digtal Unix.


					Webb