Re: Defect in XSH pthread_sigmask (revised 1)

[Dave Butenhof <yyyyyyyyyyyyyy@xxxxxx>]

Loic,

First, thanks for the effort in trying to write example code. This is a 
great contribution, in an area that's been sadly overlooked. Please 
don't take any of this as a complaint about what you did!

At the teleconference today, we discussed ERN 114, which is your lengthy 
example for the use of pthread_sigmask() and sigwait(). I could make 
some minor stylistic comments (for example, that the thread start 
routine really should have the standard prototype 'void *name (void*)' 
rather than 'void name(void)'). But we talked about the separate issues 
of size and complexity.

There would be value in having large and fully worked examples that tie 
together various aspects of the standard. However, those would be best 
represented somewhere outside the document; where someone can download 
them and use them, rather than having to try to cut and paste from a pdf 
or html document, or even worse re-type from a printed document. We 
didn't try to work out how that would be done.

Examples in the main body of the standard should be small and concise, 
so they can easily be read and absorbed. They should present one or two 
concepts in less than a page of text so that it can be seen at the same 
time as the context it explains.

I've stripped your unfortunate program "to the bones", though as always 
there's plenty of room for discussion about what constitutes "bone". For 
example, your universal check for errors is academically laudable, but 
does increase the size. I've substituted a simple comment in most 
places. I removed error checking entirely for sigemptyset() and 
sigaddset() because when using standard signal number macros there 
should be no excuse for any implementation reporting an error; and it 
saves space. There's no need to show the header files, because they're 
well defined by the standard... and again, it saves space.

I've removed your nicely generalized signal mask and 'die' routines, 
again in the name of conserving space. There's no need, in the first 
place, for a generalized routine to manage signal masks here. You use 
the routine twice, and it's redundant because both calls construct the 
same mask. Instead, I've used the sig*() functions directly, once, on a 
static mask.

We'd actually agreed to remove the main() wrapper and show the sigmask 
and thread create code "disembodied". But as I sat down to do it I 
realized that's shaving the bones themselves rather thin. In this case 
is IS relevant that the pthread_sigmask() be done "at the top of" main() 
in order to ensure that the mask is inherited by all threads. Else the 
sigwait() protocol would be unreliable.

So, for the purpose of comment and discussion, here's the "stripped" 
version:

   static sigset_t   signal_mask;  /* signals to block         */

   int main (int argc, char *argv[])
   {
       pthread_t  sig_thr_id;      /* signal handler thread ID */
       int        rc;              /* return code              */

       sigemptyset (&signal_mask);
       sigaddset (&signal_mask, SIGINT);
       sigaddset (&signal_mask, SIGTERM);
       sigaddset (&signal_mask, SIGQUIT);
       rc = pthread_sigmask (SIG_BLOCK, &signal_mask, NULL);
       if (rc != 0) {
           /* handle error */
       }
       rc = pthread_create (&sig_thr_id, NULL, signal_thread, NULL);
       if (rc != 0) {
           /* handle error */
       }

       /* APPLICATION CODE */

       return 0;
   }

   void *signal_thread (void *arg)
   {
       int       sig_caught;    /* signal caught       */
       int       rc;            /* returned code       */

       rc = sigwait (&signal_mask, &sig_caught);
       if (rc != 0) {
           /* handle error */
       }
       switch (sig_caught)
       {
       case SIGINT:     /* process SIGINT  */
           ...
           break;
       case SIGQUIT:    /* process SIGQUIT */
           ...
           break;
       case SIGTERM:    /* process SIGTERM */
           ...
           break;   
       default:         /* should normally not happen */
           fprintf (stderr, "\nUnexpected signal %d\n", sig_caught);
           break;
       }
   }

I think that your die() routine would make a good example of the use of 
strerror_r(); and when compressed somewhat vertically it's not very 
large. Similarly, create_sigmask() could be used as an example of the 
sig*set() functions.

--
/--------------------[ yyyyyyyyyyyyyy@xxxxxx ]--------------------\
| Hewlett-Packard Company       Tru64 UNIX & VMS Thread Architect |
|     My book: http://www.awl.com/cseng/titles/0-201-63392-2/     |
\----[ http://homepage.mac.com/dbutenhof/Threads/Threads.html ]---/