ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/radiance/ray/src/rt/raypcalls.c
Revision: 2.9
Committed: Mon Sep 20 16:26:58 2004 UTC (19 years, 7 months ago) by greg
Content type: text/plain
Branch: MAIN
CVS Tags: rad3R6, rad3R6P1
Changes since 2.8: +7 -1 lines
Log Message:
Added close-on-exec flag to pipes to prevent possible deadlocks

File Contents

# Content
1 #ifndef lint
2 static const char RCSid[] = "$Id: raypcalls.c,v 2.8 2004/09/17 21:43:50 greg Exp $";
3 #endif
4 /*
5 * raypcalls.c - interface for parallel rendering using Radiance
6 *
7 * External symbols declared in ray.h
8 */
9
10 #include "copyright.h"
11
12 /*
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
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,
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
28 * memory overhead, since all the static data are shared
29 * between processes. Rays are then traced using a simple
30 * queuing mechanism, explained below.
31 *
32 * The ray queue holds as many rays as there are rendering
33 * processes. Rays are queued and returned by a single
34 * ray_pqueue() call. A ray_pqueue() return
35 * value of 0 indicates that no rays are ready
36 * and the queue is not yet full. A return value of 1
37 * indicates that a ray was returned, though it is probably
38 * not the one you just requested. Rays may be identified by
39 * the rno member of the RAY struct, which is incremented
40 * by the rayorigin() call, or may be set explicitly by
41 * the caller. Below is an example call sequence:
42 *
43 * myRay.rorg = ( ray origin point )
44 * myRay.rdir = ( normalized ray direction )
45 * myRay.rmax = ( maximum length, or zero for no limit )
46 * rayorigin(&myRay, NULL, PRIMARY, 1.0);
47 * myRay.rno = ( my personal ray identifier )
48 * if (ray_pqueue(&myRay) == 1)
49 * { do something with results }
50 *
51 * Note the differences between this and the simpler ray_trace()
52 * call. In particular, the call may or may not return a value
53 * in the passed ray structure. Also, you need to call rayorigin()
54 * yourself, which is normally called for you by ray_trace(). The
55 * benefit is that ray_pqueue() will trace rays faster in
56 * proportion to the number of CPUs you have available on your
57 * system. If the ray queue is full before the call, ray_pqueue()
58 * will block until a result is ready so it can queue this one.
59 * The global int ray_pnidle indicates the number of currently idle
60 * children. If you want to check for completed rays without blocking,
61 * or get the results from rays that have been queued without
62 * queuing any new ones, the ray_presult() call is for you:
63 *
64 * if (ray_presult(&myRay, 1) == 1)
65 * { do something with results }
66 *
67 * If the second argument is 1, the call won't block when
68 * results aren't ready, but will immediately return 0.
69 * If the second argument is 0, the call will block
70 * until a value is available, returning 0 only if the
71 * queue is completely empty. A negative return value
72 * indicates that a rendering process died. If this
73 * happens, ray_close(0) is automatically called to close
74 * all child processes, and ray_pnprocs is set to zero.
75 *
76 * If you just want to fill the ray queue without checking for
77 * results, check ray_pnidle and call ray_psend():
78 *
79 * while (ray_pnidle) {
80 * ( set up ray )
81 * ray_psend(&myRay);
82 * }
83 *
84 * Note that it is a fatal error to call ra_psend() when
85 * ray_pnidle is zero. The ray_presult() and/or ray_pqueue()
86 * functions may be called subsequently to read back the results.
87 *
88 * When you are done, you may call ray_pdone(1) to close
89 * all child processes and clean up memory used by Radiance.
90 * Any queued ray calculations will be awaited and discarded.
91 * As with ray_done(), ray_pdone(0) hangs onto data files
92 * and fonts that are likely to be used in subsequent renderings.
93 * Whether you want to bother cleaning up memory or not, you
94 * should at least call ray_pclose(0) to clean the child processes.
95 *
96 * Warning: You cannot affect any of the rendering processes
97 * by changing global parameter values onece ray_pinit() has
98 * been called. Changing global parameters will have no effect
99 * until the next call to ray_pinit(), which restarts everything.
100 * If you just want to reap children so that you can alter the
101 * rendering parameters without reloading the scene, use the
102 * ray_pclose(0) and ray_popen(nproc) calls to close
103 * then restart the child processes after the changes are made.
104 *
105 * Note: These routines are written to coordinate with the
106 * definitions in raycalls.c, and in fact depend on them.
107 * If you want to trace a ray and get a result synchronously,
108 * use the ray_trace() call to compute it in the parent process
109 * This will not interfere with any subprocess calculations,
110 * but beware that a fatal error may end with a call to quit().
111 *
112 * Note: One of the advantages of using separate processes
113 * is that it gives the calling program some immunity from
114 * fatal rendering errors. As discussed in raycalls.c,
115 * Radiance tends to throw up its hands and exit at the
116 * first sign of trouble, calling quit() to return control
117 * to the top level. Although you can avoid exit() with
118 * your own longjmp() in quit(), the cleanup afterwards
119 * is always suspect. Through the use of subprocesses,
120 * we avoid this pitfall by closing the processes and
121 * returning a negative value from ray_pqueue() or
122 * ray_presult(). If you get a negative value from either
123 * of these calls, you can assume that the processes have
124 * been cleaned up with a call to ray_close(), though you
125 * will have to call ray_pdone() yourself if you want to
126 * free memory. Obviously, you cannot continue rendering
127 * without risking further errors, but otherwise your
128 * process should not be compromised.
129 */
130
131 #include <stdio.h>
132 #include <sys/types.h>
133 #include <sys/wait.h> /* XXX platform */
134
135 #include "rtprocess.h"
136 #include "ray.h"
137 #include "ambient.h"
138 #include "selcall.h"
139
140 #ifndef RAYQLEN
141 #define RAYQLEN 16 /* # rays to send at once */
142 #endif
143
144 #ifndef MAX_RPROCS
145 #if (FD_SETSIZE/2-4 < 64)
146 #define MAX_NPROCS (FD_SETSIZE/2-4)
147 #else
148 #define MAX_NPROCS 64 /* max. # rendering processes */
149 #endif
150 #endif
151
152 extern char *shm_boundary; /* boundary of shared memory */
153
154 int ray_pnprocs = 0; /* number of child processes */
155 int ray_pnidle = 0; /* number of idle children */
156
157 static struct child_proc {
158 int pid; /* child process id */
159 int fd_send; /* write to child here */
160 int fd_recv; /* read from child here */
161 int npending; /* # rays in process */
162 unsigned long rno[RAYQLEN]; /* working on these rays */
163 } r_proc[MAX_NPROCS]; /* our child processes */
164
165 static RAY r_queue[2*RAYQLEN]; /* ray i/o buffer */
166 static int r_send_next; /* next send ray placement */
167 static int r_recv_first; /* position of first unreported ray */
168 static int r_recv_next; /* next receive ray placement */
169
170 #define sendq_full() (r_send_next >= RAYQLEN)
171
172 static int ray_pflush(void);
173 static void ray_pchild(int fd_in, int fd_out);
174
175
176 extern void
177 ray_pinit( /* initialize ray-tracing processes */
178 char *otnm,
179 int nproc
180 )
181 {
182 if (nobjects > 0) /* close old calculation */
183 ray_pdone(0);
184
185 ray_init(otnm); /* load the shared scene */
186
187 preload_objs(); /* preload auxiliary data */
188
189 /* set shared memory boundary */
190 shm_boundary = (char *)malloc(16);
191 strcpy(shm_boundary, "SHM_BOUNDARY");
192
193 r_send_next = 0; /* set up queue */
194 r_recv_first = r_recv_next = RAYQLEN;
195
196 ray_popen(nproc); /* fork children */
197 }
198
199
200 static int
201 ray_pflush(void) /* send queued rays to idle children */
202 {
203 int nc, n, nw, i, sfirst;
204
205 if ((ray_pnidle <= 0) | (r_send_next <= 0))
206 return(0); /* nothing we can send */
207
208 sfirst = 0; /* divvy up labor */
209 nc = ray_pnidle;
210 for (i = ray_pnprocs; nc && i--; ) {
211 if (r_proc[i].npending > 0)
212 continue; /* child looks busy */
213 n = (r_send_next - sfirst)/nc--;
214 if (!n)
215 continue;
216 /* smuggle set size in crtype */
217 r_queue[sfirst].crtype = n;
218 nw = writebuf(r_proc[i].fd_send, (char *)&r_queue[sfirst],
219 sizeof(RAY)*n);
220 if (nw != sizeof(RAY)*n)
221 return(-1); /* write error */
222 r_proc[i].npending = n;
223 while (n--) /* record ray IDs */
224 r_proc[i].rno[n] = r_queue[sfirst+n].rno;
225 sfirst += r_proc[i].npending;
226 ray_pnidle--; /* now she's busy */
227 }
228 if (sfirst != r_send_next)
229 error(CONSISTENCY, "code screwup in ray_pflush");
230 r_send_next = 0;
231 return(sfirst); /* return total # sent */
232 }
233
234
235 extern void
236 ray_psend( /* add a ray to our send queue */
237 RAY *r
238 )
239 {
240 if (r == NULL)
241 return;
242 /* flush output if necessary */
243 if (sendq_full() && ray_pflush() <= 0)
244 error(INTERNAL, "ray_pflush failed in ray_psend");
245
246 r_queue[r_send_next] = *r;
247 r_send_next++;
248 }
249
250
251 extern int
252 ray_pqueue( /* queue a ray for computation */
253 RAY *r
254 )
255 {
256 if (r == NULL)
257 return(0);
258 /* check for full send queue */
259 if (sendq_full()) {
260 RAY mySend;
261 int rval;
262 mySend = *r;
263 /* wait for a result */
264 rval = ray_presult(r, 0);
265 /* put new ray in queue */
266 r_queue[r_send_next] = mySend;
267 r_send_next++;
268 return(rval); /* done */
269 }
270 /* add ray to send queue */
271 r_queue[r_send_next] = *r;
272 r_send_next++;
273 /* check for returned ray... */
274 if (r_recv_first >= r_recv_next)
275 return(0);
276 /* ...one is sitting in queue */
277 *r = r_queue[r_recv_first];
278 r_recv_first++;
279 return(1);
280 }
281
282
283 extern int
284 ray_presult( /* check for a completed ray */
285 RAY *r,
286 int poll
287 )
288 {
289 static struct timeval tpoll; /* zero timeval struct */
290 static fd_set readset, errset;
291 int n, ok;
292 register int pn;
293
294 if (r == NULL)
295 return(0);
296 /* check queued results first */
297 if (r_recv_first < r_recv_next) {
298 *r = r_queue[r_recv_first];
299 r_recv_first++;
300 return(1);
301 }
302 n = ray_pnprocs - ray_pnidle; /* pending before flush? */
303
304 if (ray_pflush() < 0) /* send new rays to process */
305 return(-1);
306 /* reset receive queue */
307 r_recv_first = r_recv_next = RAYQLEN;
308
309 if (!poll) /* count newly sent unless polling */
310 n = ray_pnprocs - ray_pnidle;
311 if (n <= 0) /* return if nothing to await */
312 return(0);
313 getready: /* any children waiting for us? */
314 for (pn = ray_pnprocs; pn--; )
315 if (FD_ISSET(r_proc[pn].fd_recv, &readset) ||
316 FD_ISSET(r_proc[pn].fd_recv, &errset))
317 break;
318 /* call select if we must */
319 if (pn < 0) {
320 FD_ZERO(&readset); FD_ZERO(&errset); n = 0;
321 for (pn = ray_pnprocs; pn--; ) {
322 if (r_proc[pn].npending > 0)
323 FD_SET(r_proc[pn].fd_recv, &readset);
324 FD_SET(r_proc[pn].fd_recv, &errset);
325 if (r_proc[pn].fd_recv >= n)
326 n = r_proc[pn].fd_recv + 1;
327 }
328 /* find out who is ready */
329 while ((n = select(n, &readset, (fd_set *)NULL, &errset,
330 poll ? &tpoll : (struct timeval *)NULL)) < 0)
331 if (errno != EINTR) {
332 error(WARNING,
333 "select call failed in ray_presult");
334 ray_pclose(0);
335 return(-1);
336 }
337 if (n > 0) /* go back and get it */
338 goto getready;
339 return(0); /* else poll came up empty */
340 }
341 if (r_recv_next + r_proc[pn].npending > sizeof(r_queue)/sizeof(RAY))
342 error(CONSISTENCY, "buffer shortage in ray_presult()");
343
344 /* read rendered ray data */
345 n = readbuf(r_proc[pn].fd_recv, (char *)&r_queue[r_recv_next],
346 sizeof(RAY)*r_proc[pn].npending);
347 if (n > 0) {
348 r_recv_next += n/sizeof(RAY);
349 ok = (n == sizeof(RAY)*r_proc[pn].npending);
350 } else
351 ok = 0;
352 /* reset child's status */
353 FD_CLR(r_proc[pn].fd_recv, &readset);
354 if (n <= 0)
355 FD_CLR(r_proc[pn].fd_recv, &errset);
356 r_proc[pn].npending = 0;
357 ray_pnidle++;
358 /* check for rendering errors */
359 if (!ok) {
360 ray_pclose(0); /* process died -- clean up */
361 return(-1);
362 }
363 /* preen returned rays */
364 for (n = r_recv_next - r_recv_first; n--; ) {
365 register RAY *rp = &r_queue[r_recv_first + n];
366 rp->rno = r_proc[pn].rno[n];
367 rp->parent = NULL;
368 rp->newcset = rp->clipset = NULL;
369 rp->rox = NULL;
370 rp->slights = NULL;
371 }
372 /* return first ray received */
373 *r = r_queue[r_recv_first];
374 r_recv_first++;
375 return(1);
376 }
377
378
379 extern void
380 ray_pdone( /* reap children and free data */
381 int freall
382 )
383 {
384 ray_pclose(0); /* close child processes */
385
386 if (shm_boundary != NULL) { /* clear shared memory boundary */
387 free((void *)shm_boundary);
388 shm_boundary = NULL;
389 }
390 ray_done(freall); /* free rendering data */
391 }
392
393
394 static void
395 ray_pchild( /* process rays (never returns) */
396 int fd_in,
397 int fd_out
398 )
399 {
400 int n;
401 register int i;
402 /* read each ray request set */
403 while ((n = read(fd_in, (char *)r_queue, sizeof(r_queue))) > 0) {
404 int n2;
405 if (n % sizeof(RAY))
406 break;
407 n /= sizeof(RAY);
408 /* get smuggled set length */
409 n2 = r_queue[0].crtype - n;
410 if (n2 < 0)
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);
415 if (i != sizeof(RAY)*n2)
416 break;
417 n += n2;
418 }
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;
426 samplendx++;
427 rayclear(&r_queue[i]);
428 rayvalue(&r_queue[i]);
429 }
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");
434 }
435 if (n)
436 error(SYSTEM, "read error in ray_pchild");
437 ambsync();
438 quit(0); /* normal exit */
439 }
440
441
442 extern void
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);
454 while (nadd--) { /* fork each new process */
455 int p0[2], p1[2];
456 if (pipe(p0) < 0 || pipe(p1) < 0)
457 error(SYSTEM, "cannot create pipe");
458 if ((r_proc[ray_pnprocs].pid = fork()) == 0) {
459 int pn; /* close others' descriptors */
460 for (pn = ray_pnprocs; pn--; ) {
461 close(r_proc[pn].fd_send);
462 close(r_proc[pn].fd_recv);
463 }
464 close(p0[0]); close(p1[1]);
465 /* following call never returns */
466 ray_pchild(p1[0], p0[1]);
467 }
468 if (r_proc[ray_pnprocs].pid < 0)
469 error(SYSTEM, "cannot fork child process");
470 close(p1[0]); close(p0[1]);
471 /*
472 * Close write stream on exec to avoid multiprocessing deadlock.
473 * No use in read stream without it, so set flag there as well.
474 */
475 fcntl(p1[1], F_SETFD, FD_CLOEXEC);
476 fcntl(p0[0], F_SETFD, FD_CLOEXEC);
477 r_proc[ray_pnprocs].fd_send = p1[1];
478 r_proc[ray_pnprocs].fd_recv = p0[0];
479 r_proc[ray_pnprocs].npending = 0;
480 ray_pnprocs++;
481 ray_pnidle++;
482 }
483 }
484
485
486 extern void
487 ray_pclose( /* close one or more child processes */
488 int nsub
489 )
490 {
491 static int inclose = 0;
492 RAY res;
493 /* check recursion */
494 if (inclose)
495 return;
496 inclose++;
497 /* check argument */
498 if ((nsub <= 0) | (nsub > ray_pnprocs))
499 nsub = ray_pnprocs;
500 /* clear our ray queue */
501 while (ray_presult(&res,0) > 0)
502 ;
503 /* clean up children */
504 while (nsub--) {
505 int status;
506 ray_pnprocs--;
507 close(r_proc[ray_pnprocs].fd_recv);
508 close(r_proc[ray_pnprocs].fd_send);
509 if (waitpid(r_proc[ray_pnprocs].pid, &status, 0) < 0)
510 status = 127<<8;
511 if (status) {
512 sprintf(errmsg,
513 "rendering process %d exited with code %d",
514 r_proc[ray_pnprocs].pid, status>>8);
515 error(WARNING, errmsg);
516 }
517 ray_pnidle--;
518 }
519 inclose--;
520 }
521
522
523 void
524 quit(ec) /* make sure exit is called */
525 int ec;
526 {
527 exit(ec);
528 }