ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/radiance/ray/src/rt/raypcalls.c
(Generate patch)

Comparing ray/src/rt/raypcalls.c (file contents):
Revision 2.5 by schorsch, Sun Jul 27 22:12:03 2003 UTC vs.
Revision 2.24 by greg, Sat Dec 12 23:08:13 2009 UTC

# Line 13 | Line 13 | static const char      RCSid[] = "$Id$";
13   *  These calls are designed similarly to the ones in raycalls.c,
14   *  but allow for multiple rendering processes on the same host
15   *  machine.  There is no sense in specifying more child processes
16 < *  than you have processors, but one child may help by allowing
16 > *  than you have processor cores, but one child may help by allowing
17   *  asynchronous ray computation in an interactive program, and
18   *  will protect the caller from fatal rendering errors.
19   *
20 < *  You should first read and undrstand the header in raycalls.c,
20 > *  You should first read and understand the header in raycalls.c,
21   *  as some things are explained there that are not repated here.
22   *
23   *  The first step is opening one or more rendering processes
24   *  with a call to ray_pinit(oct, nproc).  Before calling fork(),
25   *  ray_pinit() loads the octree and data structures into the
26 < *  caller's memory.  This permits all sorts of queries that
27 < *  wouldn't be possible otherwise, without causing any real
26 > *  caller's memory, and ray_popen() synchronizes the ambient
27 > *  file, if any.  Shared memory permits all sorts of queries
28 > *  that wouldn't be possible otherwise without causing any real
29   *  memory overhead, since all the static data are shared
30 < *  between processes.  Rays are then traced using a simple
30 > *  between processes.  Rays are traced using a simple
31   *  queuing mechanism, explained below.
32   *
33 < *  The ray queue holds as many rays as there are rendering
34 < *  processes.  Rays are queued and returned by a single
33 > *  The ray queue buffers RAYQLEN rays before sending to
34 > *  children, each of which may internally buffer RAYQLEN rays
35 > *  during evaluation.  Rays are not returned in the order
36 > *  they are sent when multiple processes are open.
37 > *
38 > *  Rays are queued and returned by a single
39   *  ray_pqueue() call.  A ray_pqueue() return
40   *  value of 0 indicates that no rays are ready
41   *  and the queue is not yet full.  A return value of 1
# Line 43 | Line 48 | static const char      RCSid[] = "$Id$";
48   *      myRay.rorg = ( ray origin point )
49   *      myRay.rdir = ( normalized ray direction )
50   *      myRay.rmax = ( maximum length, or zero for no limit )
51 < *      rayorigin(&myRay, NULL, PRIMARY, 1.0);
51 > *      rayorigin(&myRay, PRIMARY, NULL, NULL);
52   *      myRay.rno = ( my personal ray identifier )
53   *      if (ray_pqueue(&myRay) == 1)
54   *              { do something with results }
# Line 51 | Line 56 | static const char      RCSid[] = "$Id$";
56   *  Note the differences between this and the simpler ray_trace()
57   *  call.  In particular, the call may or may not return a value
58   *  in the passed ray structure.  Also, you need to call rayorigin()
59 < *  yourself, which is normally for you by ray_trace().  The
60 < *  great thing is that ray_pqueue() will trace rays faster in
59 > *  yourself, which is normally called for you by ray_trace().  The
60 > *  benefit is that ray_pqueue() will trace rays faster in
61   *  proportion to the number of CPUs you have available on your
62   *  system.  If the ray queue is full before the call, ray_pqueue()
63   *  will block until a result is ready so it can queue this one.
# Line 66 | Line 71 | static const char      RCSid[] = "$Id$";
71   *
72   *  If the second argument is 1, the call won't block when
73   *  results aren't ready, but will immediately return 0.
74 + *  (A special value of -1 returns 0 unless a ray is
75 + *  ready in the queue and no system calls are needed.)
76   *  If the second argument is 0, the call will block
77   *  until a value is available, returning 0 only if the
78   *  queue is completely empty.  A negative return value
79   *  indicates that a rendering process died.  If this
80 < *  happens, ray_close(0) is automatically called to close
80 > *  happens, ray_pclose(0) is automatically called to close
81   *  all child processes, and ray_pnprocs is set to zero.
82   *
83   *  If you just want to fill the ray queue without checking for
# Line 81 | Line 88 | static const char      RCSid[] = "$Id$";
88   *              ray_psend(&myRay);
89   *      }
90   *
91 < *  The ray_presult() and/or ray_pqueue() functions may then be
92 < *  called to read back the results.
91 > *  Note that it is a fatal error to call ra_psend() when
92 > *  ray_pnidle is zero.  The ray_presult() and/or ray_pqueue()
93 > *  functions may be called subsequently to read back the results.
94   *
95   *  When you are done, you may call ray_pdone(1) to close
96   *  all child processes and clean up memory used by Radiance.
97   *  Any queued ray calculations will be awaited and discarded.
98   *  As with ray_done(), ray_pdone(0) hangs onto data files
99   *  and fonts that are likely to be used in subsequent renderings.
100 < *  Whether you want to bother cleaning up memory or not, you
101 < *  should at least call ray_pclose(0) to clean the child processes.
100 > *  Whether you need to clean up memory or not, you should
101 > *  at least call ray_pclose(0) to await the child processes.
102 > *  The caller should define a quit() function that calls
103 > *  ray_pclose(0) if ray_pnprocs > 0.
104   *
105   *  Warning:  You cannot affect any of the rendering processes
106   *  by changing global parameter values onece ray_pinit() has
# Line 99 | Line 109 | static const char      RCSid[] = "$Id$";
109   *  If you just want to reap children so that you can alter the
110   *  rendering parameters without reloading the scene, use the
111   *  ray_pclose(0) and ray_popen(nproc) calls to close
112 < *  then restart the child processes.
112 > *  then restart the child processes after the changes are made.
113   *
114   *  Note:  These routines are written to coordinate with the
115   *  definitions in raycalls.c, and in fact depend on them.
116   *  If you want to trace a ray and get a result synchronously,
117   *  use the ray_trace() call to compute it in the parent process.
118 + *  This will not interfere with any subprocess calculations,
119 + *  but beware that a fatal error may end with a call to quit().
120   *
121   *  Note:  One of the advantages of using separate processes
122   *  is that it gives the calling program some immunity from
123   *  fatal rendering errors.  As discussed in raycalls.c,
124   *  Radiance tends to throw up its hands and exit at the
125   *  first sign of trouble, calling quit() to return control
126 < *  to the system.  Although you can avoid exit() with
126 > *  to the top level.  Although you can avoid exit() with
127   *  your own longjmp() in quit(), the cleanup afterwards
128   *  is always suspect.  Through the use of subprocesses,
129   *  we avoid this pitfall by closing the processes and
130   *  returning a negative value from ray_pqueue() or
131   *  ray_presult().  If you get a negative value from either
132   *  of these calls, you can assume that the processes have
133 < *  been cleaned up with a call to ray_close(), though you
133 > *  been cleaned up with a call to ray_pclose(), though you
134   *  will have to call ray_pdone() yourself if you want to
135 < *  free memory.  Obviously, you cannot continue rendering,
136 < *  but otherwise your process should not be compromised.
135 > *  free memory.  Obviously, you cannot continue rendering
136 > *  without risking further errors, but otherwise your
137 > *  process should not be compromised.
138   */
139  
140 + #include  "rtprocess.h"
141   #include  "ray.h"
142 <
142 > #include  "ambient.h"
143 > #include  <sys/types.h>
144 > #include  <sys/wait.h>
145   #include  "selcall.h"
146  
147   #ifndef RAYQLEN
148 < #define RAYQLEN         16              /* # rays to send at once */
148 > #define RAYQLEN         12              /* # rays to send at once */
149   #endif
150  
151   #ifndef MAX_RPROCS
# Line 150 | Line 166 | static struct child_proc {
166          int     fd_send;                        /* write to child here */
167          int     fd_recv;                        /* read from child here */
168          int     npending;                       /* # rays in process */
169 <        unsigned long  rno[RAYQLEN];            /* working on these rays */
169 >        RNUMBER rno[RAYQLEN];                   /* working on these rays */
170   } r_proc[MAX_NPROCS];                   /* our child processes */
171  
172   static RAY      r_queue[2*RAYQLEN];     /* ray i/o buffer */
173 < static int      r_send_next;            /* next send ray placement */
174 < static int      r_recv_first;           /* position of first unreported ray */
175 < static int      r_recv_next;            /* next receive ray placement */
173 > static int      r_send_next = 0;        /* next send ray placement */
174 > static int      r_recv_first = RAYQLEN; /* position of first unreported ray */
175 > static int      r_recv_next = RAYQLEN;  /* next received ray placement */
176  
177   #define sendq_full()    (r_send_next >= RAYQLEN)
178  
179 + static int ray_pflush(void);
180 + static void ray_pchild(int fd_in, int fd_out);
181  
182 +
183   void
184 < ray_pinit(otnm, nproc)          /* initialize ray-tracing processes */
185 < char    *otnm;
186 < int     nproc;
184 > ray_pinit(              /* initialize ray-tracing processes */
185 >        char    *otnm,
186 >        int     nproc
187 > )
188   {
189          if (nobjects > 0)               /* close old calculation */
190                  ray_pdone(0);
191  
192          ray_init(otnm);                 /* load the shared scene */
193  
174 –        preload_objs();                 /* preload auxiliary data */
175 –
176 –                                        /* set shared memory boundary */
177 –        shm_boundary = (char *)malloc(16);
178 –        strcpy(shm_boundary, "SHM_BOUNDARY");
179 –
180 –        r_send_next = 0;                /* set up queue */
181 –        r_recv_first = r_recv_next = RAYQLEN;
182 –
194          ray_popen(nproc);               /* fork children */
195   }
196  
197  
198   static int
199 < ray_pflush()                    /* send queued rays to idle children */
199 > ray_pflush(void)                        /* send queued rays to idle children */
200   {
201          int     nc, n, nw, i, sfirst;
202  
# Line 213 | Line 224 | ray_pflush()                   /* send queued rays to idle children */
224                  ray_pnidle--;           /* now she's busy */
225          }
226          if (sfirst != r_send_next)
227 <                error(CONSISTENCY, "code screwup in ray_pflush");
227 >                error(CONSISTENCY, "code screwup in ray_pflush()");
228          r_send_next = 0;
229          return(sfirst);                 /* return total # sent */
230   }
231  
232  
233   void
234 < ray_psend(r)                    /* add a ray to our send queue */
235 < RAY     *r;
234 > ray_psend(                      /* add a ray to our send queue */
235 >        RAY     *r
236 > )
237   {
238          if (r == NULL)
239                  return;
240                                          /* flush output if necessary */
241          if (sendq_full() && ray_pflush() <= 0)
242 <                error(INTERNAL, "ray_pflush failed in ray_psend");
242 >                error(INTERNAL, "ray_pflush failed in ray_psend()");
243  
244 <        r_queue[r_send_next] = *r;
233 <        r_send_next++;
244 >        r_queue[r_send_next++] = *r;
245   }
246  
247  
248   int
249 < ray_pqueue(r)                   /* queue a ray for computation */
250 < RAY     *r;
249 > ray_pqueue(                     /* queue a ray for computation */
250 >        RAY     *r
251 > )
252   {
253          if (r == NULL)
254                  return(0);
255                                          /* check for full send queue */
256          if (sendq_full()) {
257 <                RAY     mySend;
246 <                int     rval;
247 <                mySend = *r;
257 >                RAY     mySend = *r;
258                                          /* wait for a result */
259 <                rval = ray_presult(r, 0);
259 >                if (ray_presult(r, 0) <= 0)
260 >                        return(-1);
261                                          /* put new ray in queue */
262 <                r_queue[r_send_next] = mySend;
263 <                r_send_next++;
264 <                return(rval);           /* done */
262 >                r_queue[r_send_next++] = mySend;
263 >                                /* XXX r_send_next may now be > RAYQLEN */
264 >                return(1);
265          }
266 <                                        /* add ray to send queue */
267 <        r_queue[r_send_next] = *r;
257 <        r_send_next++;
266 >                                        /* else add ray to send queue */
267 >        r_queue[r_send_next++] = *r;
268                                          /* check for returned ray... */
269          if (r_recv_first >= r_recv_next)
270                  return(0);
271                                          /* ...one is sitting in queue */
272 <        *r = r_queue[r_recv_first];
263 <        r_recv_first++;
272 >        *r = r_queue[r_recv_first++];
273          return(1);
274   }
275  
276  
277   int
278 < ray_presult(r, poll)            /* check for a completed ray */
279 < RAY     *r;
280 < int     poll;
278 > ray_presult(            /* check for a completed ray */
279 >        RAY     *r,
280 >        int     poll
281 > )
282   {
283          static struct timeval   tpoll;  /* zero timeval struct */
284          static fd_set   readset, errset;
# Line 279 | Line 289 | int    poll;
289                  return(0);
290                                          /* check queued results first */
291          if (r_recv_first < r_recv_next) {
292 <                *r = r_queue[r_recv_first];
283 <                r_recv_first++;
292 >                *r = r_queue[r_recv_first++];
293                  return(1);
294          }
295 +        if (poll < 0)                   /* immediate polling mode? */
296 +                return(0);
297 +
298          n = ray_pnprocs - ray_pnidle;   /* pending before flush? */
299  
300          if (ray_pflush() < 0)           /* send new rays to process */
# Line 294 | Line 306 | int    poll;
306                  n = ray_pnprocs - ray_pnidle;
307          if (n <= 0)                     /* return if nothing to await */
308                  return(0);
309 +        if (!poll && ray_pnprocs == 1)  /* one process -> skip select() */
310 +                FD_SET(r_proc[0].fd_recv, &readset);
311 +
312   getready:                               /* any children waiting for us? */
313          for (pn = ray_pnprocs; pn--; )
314                  if (FD_ISSET(r_proc[pn].fd_recv, &readset) ||
315                                  FD_ISSET(r_proc[pn].fd_recv, &errset))
316                          break;
317 <                                        /* call select if we must */
317 >                                        /* call select() if we must */
318          if (pn < 0) {
319                  FD_ZERO(&readset); FD_ZERO(&errset); n = 0;
320                  for (pn = ray_pnprocs; pn--; ) {
# Line 314 | Line 329 | getready:                              /* any children waiting for us? */
329                                  poll ? &tpoll : (struct timeval *)NULL)) < 0)
330                          if (errno != EINTR) {
331                                  error(WARNING,
332 <                                        "select call failed in ray_presult");
332 >                                        "select call failed in ray_presult()");
333                                  ray_pclose(0);
334                                  return(-1);
335                          }
# Line 354 | Line 369 | getready:                              /* any children waiting for us? */
369                  rp->slights = NULL;
370          }
371                                          /* return first ray received */
372 <        *r = r_queue[r_recv_first];
358 <        r_recv_first++;
372 >        *r = r_queue[r_recv_first++];
373          return(1);
374   }
375  
376  
377   void
378 < ray_pdone(freall)               /* reap children and free data */
379 < int     freall;
378 > ray_pdone(              /* reap children and free data */
379 >        int     freall
380 > )
381   {
382          ray_pclose(0);                  /* close child processes */
383  
# Line 370 | Line 385 | int    freall;
385                  free((void *)shm_boundary);
386                  shm_boundary = NULL;
387          }
388 +
389          ray_done(freall);               /* free rendering data */
390   }
391  
392  
393   static void
394 < ray_pchild(fd_in, fd_out)       /* process rays (never returns) */
395 < int     fd_in;
396 < int     fd_out;
394 > ray_pchild(     /* process rays (never returns) */
395 >        int     fd_in,
396 >        int     fd_out
397 > )
398   {
399          int     n;
400          register int    i;
401 +                                        /* flag child process for quit() */
402 +        ray_pnprocs = -1;
403                                          /* read each ray request set */
404          while ((n = read(fd_in, (char *)r_queue, sizeof(r_queue))) > 0) {
405                  int     n2;
406 <                if (n % sizeof(RAY))
406 >                if (n < sizeof(RAY))
407                          break;
389 –                n /= sizeof(RAY);
408                                          /* get smuggled set length */
409 <                n2 = r_queue[0].crtype - n;
409 >                n2 = sizeof(RAY)*r_queue[0].crtype - n;
410                  if (n2 < 0)
411 <                        error(INTERNAL, "buffer over-read in ray_pchild");
411 >                        error(INTERNAL, "buffer over-read in ray_pchild()");
412                  if (n2 > 0) {           /* read the rest of the set */
413 <                        i = readbuf(fd_in, (char *)(r_queue+n),
414 <                                        sizeof(RAY)*n2);
397 <                        if (i != sizeof(RAY)*n2)
413 >                        i = readbuf(fd_in, (char *)r_queue + n, n2);
414 >                        if (i != n2)
415                                  break;
416                          n += n2;
417                  }
418 +                n /= sizeof(RAY);
419                                          /* evaluate rays */
420                  for (i = 0; i < n; i++) {
421                          r_queue[i].crtype = r_queue[i].rtype;
422                          r_queue[i].parent = NULL;
423                          r_queue[i].clipset = NULL;
424                          r_queue[i].slights = NULL;
425 <                        r_queue[i].revf = raytrace;
425 >                        r_queue[i].rlvl = 0;
426                          samplendx++;
427                          rayclear(&r_queue[i]);
428                          rayvalue(&r_queue[i]);
# Line 412 | Line 430 | int    fd_out;
430                                          /* write back our results */
431                  i = writebuf(fd_out, (char *)r_queue, sizeof(RAY)*n);
432                  if (i != sizeof(RAY)*n)
433 <                        error(SYSTEM, "write error in ray_pchild");
433 >                        error(SYSTEM, "write error in ray_pchild()");
434          }
435          if (n)
436 <                error(SYSTEM, "read error in ray_pchild");
436 >                error(SYSTEM, "read error in ray_pchild()");
437          ambsync();
438          quit(0);                        /* normal exit */
439   }
440  
441  
442   void
443 < ray_popen(nadd)                 /* open the specified # processes */
444 < int     nadd;
443 > ray_popen(                      /* open the specified # processes */
444 >        int     nadd
445 > )
446   {
447                                          /* check if our table has room */
448          if (ray_pnprocs + nadd > MAX_NPROCS)
449                  nadd = MAX_NPROCS - ray_pnprocs;
450          if (nadd <= 0)
451                  return;
452 <        fflush(stderr);                 /* clear pending output */
453 <        fflush(stdout);
452 >        ambsync();                      /* load any new ambient values */
453 >        if (shm_boundary == NULL) {     /* first child process? */
454 >                preload_objs();         /* preload auxiliary data */
455 >                                        /* set shared memory boundary */
456 >                shm_boundary = (char *)malloc(16);
457 >                strcpy(shm_boundary, "SHM_BOUNDARY");
458 >        }
459 >        fflush(NULL);                   /* clear pending output */
460          while (nadd--) {                /* fork each new process */
461                  int     p0[2], p1[2];
462                  if (pipe(p0) < 0 || pipe(p1) < 0)
# Line 443 | Line 468 | int    nadd;
468                                  close(r_proc[pn].fd_recv);
469                          }
470                          close(p0[0]); close(p1[1]);
471 +                        close(0);       /* don't share stdin */
472                                          /* following call never returns */
473                          ray_pchild(p1[0], p0[1]);
474                  }
475                  if (r_proc[ray_pnprocs].pid < 0)
476                          error(SYSTEM, "cannot fork child process");
477                  close(p1[0]); close(p0[1]);
478 +                /*
479 +                 * Close write stream on exec to avoid multiprocessing deadlock.
480 +                 * No use in read stream without it, so set flag there as well.
481 +                 */
482 +                fcntl(p1[1], F_SETFD, FD_CLOEXEC);
483 +                fcntl(p0[0], F_SETFD, FD_CLOEXEC);
484                  r_proc[ray_pnprocs].fd_send = p1[1];
485                  r_proc[ray_pnprocs].fd_recv = p0[0];
486                  r_proc[ray_pnprocs].npending = 0;
# Line 459 | Line 491 | int    nadd;
491  
492  
493   void
494 < ray_pclose(nsub)                /* close one or more child processes */
495 < int     nsub;
494 > ray_pclose(             /* close one or more child processes */
495 >        int     nsub
496 > )
497   {
498          static int      inclose = 0;
499          RAY     res;
# Line 474 | Line 507 | int    nsub;
507                                          /* clear our ray queue */
508          while (ray_presult(&res,0) > 0)
509                  ;
510 +        r_send_next = 0;                /* hard reset in case of error */
511 +        r_recv_first = r_recv_next = RAYQLEN;
512                                          /* clean up children */
513          while (nsub--) {
514                  int     status;
515                  ray_pnprocs--;
481 –                close(r_proc[ray_pnprocs].fd_recv);
516                  close(r_proc[ray_pnprocs].fd_send);
517 <                while (wait(&status) != r_proc[ray_pnprocs].pid)
518 <                        ;
517 >                if (waitpid(r_proc[ray_pnprocs].pid, &status, 0) < 0)
518 >                        status = 127<<8;
519 >                close(r_proc[ray_pnprocs].fd_recv);
520                  if (status) {
521                          sprintf(errmsg,
522                                  "rendering process %d exited with code %d",
# Line 491 | Line 526 | int    nsub;
526                  ray_pnidle--;
527          }
528          inclose--;
494 – }
495 –
496 –
497 – void
498 – quit(ec)                        /* make sure exit is called */
499 – int     ec;
500 – {
501 –        exit(ec);
529   }

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines