1 |
#ifndef lint |
2 |
static const char RCSid[] = "$Id: raypcalls.c,v 2.15 2007/09/12 03:57:00 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, 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 |
31 |
* queuing mechanism, explained below. |
32 |
* |
33 |
* The ray queue buffers RAYQLEN rays before sending to |
34 |
* children, each of which may internally buffer RAYQLEN rays. |
35 |
* |
36 |
* Rays are queued and returned by a single |
37 |
* ray_pqueue() call. A ray_pqueue() return |
38 |
* value of 0 indicates that no rays are ready |
39 |
* and the queue is not yet full. A return value of 1 |
40 |
* indicates that a ray was returned, though it is probably |
41 |
* not the one you just requested. Rays may be identified by |
42 |
* the rno member of the RAY struct, which is incremented |
43 |
* by the rayorigin() call, or may be set explicitly by |
44 |
* the caller. Below is an example call sequence: |
45 |
* |
46 |
* myRay.rorg = ( ray origin point ) |
47 |
* myRay.rdir = ( normalized ray direction ) |
48 |
* myRay.rmax = ( maximum length, or zero for no limit ) |
49 |
* rayorigin(&myRay, PRIMARY, NULL, NULL); |
50 |
* myRay.rno = ( my personal ray identifier ) |
51 |
* if (ray_pqueue(&myRay) == 1) |
52 |
* { do something with results } |
53 |
* |
54 |
* Note the differences between this and the simpler ray_trace() |
55 |
* call. In particular, the call may or may not return a value |
56 |
* in the passed ray structure. Also, you need to call rayorigin() |
57 |
* yourself, which is normally called for you by ray_trace(). The |
58 |
* benefit is that ray_pqueue() will trace rays faster in |
59 |
* proportion to the number of CPUs you have available on your |
60 |
* system. If the ray queue is full before the call, ray_pqueue() |
61 |
* will block until a result is ready so it can queue this one. |
62 |
* The global int ray_pnidle indicates the number of currently idle |
63 |
* children. If you want to check for completed rays without blocking, |
64 |
* or get the results from rays that have been queued without |
65 |
* queuing any new ones, the ray_presult() call is for you: |
66 |
* |
67 |
* if (ray_presult(&myRay, 1) == 1) |
68 |
* { do something with results } |
69 |
* |
70 |
* If the second argument is 1, the call won't block when |
71 |
* results aren't ready, but will immediately return 0. |
72 |
* If the second argument is 0, the call will block |
73 |
* until a value is available, returning 0 only if the |
74 |
* queue is completely empty. A negative return value |
75 |
* indicates that a rendering process died. If this |
76 |
* happens, ray_close(0) is automatically called to close |
77 |
* all child processes, and ray_pnprocs is set to zero. |
78 |
* |
79 |
* If you just want to fill the ray queue without checking for |
80 |
* results, check ray_pnidle and call ray_psend(): |
81 |
* |
82 |
* while (ray_pnidle) { |
83 |
* ( set up ray ) |
84 |
* ray_psend(&myRay); |
85 |
* } |
86 |
* |
87 |
* Note that it is a fatal error to call ra_psend() when |
88 |
* ray_pnidle is zero. The ray_presult() and/or ray_pqueue() |
89 |
* functions may be called subsequently to read back the results. |
90 |
* |
91 |
* When you are done, you may call ray_pdone(1) to close |
92 |
* all child processes and clean up memory used by Radiance. |
93 |
* Any queued ray calculations will be awaited and discarded. |
94 |
* As with ray_done(), ray_pdone(0) hangs onto data files |
95 |
* and fonts that are likely to be used in subsequent renderings. |
96 |
* Whether you want to bother cleaning up memory or not, you |
97 |
* should at least call ray_pclose(0) to clean the child processes. |
98 |
* |
99 |
* Warning: You cannot affect any of the rendering processes |
100 |
* by changing global parameter values onece ray_pinit() has |
101 |
* been called. Changing global parameters will have no effect |
102 |
* until the next call to ray_pinit(), which restarts everything. |
103 |
* If you just want to reap children so that you can alter the |
104 |
* rendering parameters without reloading the scene, use the |
105 |
* ray_pclose(0) and ray_popen(nproc) calls to close |
106 |
* then restart the child processes after the changes are made. |
107 |
* |
108 |
* Note: These routines are written to coordinate with the |
109 |
* definitions in raycalls.c, and in fact depend on them. |
110 |
* If you want to trace a ray and get a result synchronously, |
111 |
* use the ray_trace() call to compute it in the parent process. |
112 |
* This will not interfere with any subprocess calculations, |
113 |
* but beware that a fatal error may end with a call to quit(). |
114 |
* |
115 |
* Note: One of the advantages of using separate processes |
116 |
* is that it gives the calling program some immunity from |
117 |
* fatal rendering errors. As discussed in raycalls.c, |
118 |
* Radiance tends to throw up its hands and exit at the |
119 |
* first sign of trouble, calling quit() to return control |
120 |
* to the top level. Although you can avoid exit() with |
121 |
* your own longjmp() in quit(), the cleanup afterwards |
122 |
* is always suspect. Through the use of subprocesses, |
123 |
* we avoid this pitfall by closing the processes and |
124 |
* returning a negative value from ray_pqueue() or |
125 |
* ray_presult(). If you get a negative value from either |
126 |
* of these calls, you can assume that the processes have |
127 |
* been cleaned up with a call to ray_close(), though you |
128 |
* will have to call ray_pdone() yourself if you want to |
129 |
* free memory. Obviously, you cannot continue rendering |
130 |
* without risking further errors, but otherwise your |
131 |
* process should not be compromised. |
132 |
*/ |
133 |
|
134 |
#include <stdio.h> |
135 |
#include <sys/types.h> |
136 |
#include <sys/wait.h> /* XXX platform */ |
137 |
|
138 |
#include "rtprocess.h" |
139 |
#include "ray.h" |
140 |
#include "ambient.h" |
141 |
#include "selcall.h" |
142 |
|
143 |
#ifndef RAYQLEN |
144 |
#define RAYQLEN 12 /* # rays to send at once */ |
145 |
#endif |
146 |
|
147 |
#ifndef MAX_RPROCS |
148 |
#if (FD_SETSIZE/2-4 < 64) |
149 |
#define MAX_NPROCS (FD_SETSIZE/2-4) |
150 |
#else |
151 |
#define MAX_NPROCS 64 /* max. # rendering processes */ |
152 |
#endif |
153 |
#endif |
154 |
|
155 |
extern char *shm_boundary; /* boundary of shared memory */ |
156 |
|
157 |
int ray_pnprocs = 0; /* number of child processes */ |
158 |
int ray_pnidle = 0; /* number of idle children */ |
159 |
|
160 |
static struct child_proc { |
161 |
int pid; /* child process id */ |
162 |
int fd_send; /* write to child here */ |
163 |
int fd_recv; /* read from child here */ |
164 |
int npending; /* # rays in process */ |
165 |
unsigned long rno[RAYQLEN]; /* working on these rays */ |
166 |
} r_proc[MAX_NPROCS]; /* our child processes */ |
167 |
|
168 |
static RAY r_queue[2*RAYQLEN]; /* ray i/o buffer */ |
169 |
static int r_send_next; /* next send ray placement */ |
170 |
static int r_recv_first; /* position of first unreported ray */ |
171 |
static int r_recv_next; /* next receive ray placement */ |
172 |
|
173 |
#define sendq_full() (r_send_next >= RAYQLEN) |
174 |
|
175 |
static int ray_pflush(void); |
176 |
static void ray_pchild(int fd_in, int fd_out); |
177 |
|
178 |
|
179 |
extern void |
180 |
ray_pinit( /* initialize ray-tracing processes */ |
181 |
char *otnm, |
182 |
int nproc |
183 |
) |
184 |
{ |
185 |
if (nobjects > 0) /* close old calculation */ |
186 |
ray_pdone(0); |
187 |
|
188 |
ray_init(otnm); /* load the shared scene */ |
189 |
|
190 |
preload_objs(); /* preload auxiliary data */ |
191 |
|
192 |
/* set shared memory boundary */ |
193 |
shm_boundary = (char *)malloc(16); |
194 |
strcpy(shm_boundary, "SHM_BOUNDARY"); |
195 |
|
196 |
r_send_next = 0; /* set up queue */ |
197 |
r_recv_first = r_recv_next = RAYQLEN; |
198 |
|
199 |
ray_popen(nproc); /* fork children */ |
200 |
} |
201 |
|
202 |
|
203 |
static int |
204 |
ray_pflush(void) /* send queued rays to idle children */ |
205 |
{ |
206 |
int nc, n, nw, i, sfirst; |
207 |
|
208 |
if ((ray_pnidle <= 0) | (r_send_next <= 0)) |
209 |
return(0); /* nothing we can send */ |
210 |
|
211 |
sfirst = 0; /* divvy up labor */ |
212 |
nc = ray_pnidle; |
213 |
for (i = ray_pnprocs; nc && i--; ) { |
214 |
if (r_proc[i].npending > 0) |
215 |
continue; /* child looks busy */ |
216 |
n = (r_send_next - sfirst)/nc--; |
217 |
if (!n) |
218 |
continue; |
219 |
/* smuggle set size in crtype */ |
220 |
r_queue[sfirst].crtype = n; |
221 |
nw = writebuf(r_proc[i].fd_send, (char *)&r_queue[sfirst], |
222 |
sizeof(RAY)*n); |
223 |
if (nw != sizeof(RAY)*n) |
224 |
return(-1); /* write error */ |
225 |
r_proc[i].npending = n; |
226 |
while (n--) /* record ray IDs */ |
227 |
r_proc[i].rno[n] = r_queue[sfirst+n].rno; |
228 |
sfirst += r_proc[i].npending; |
229 |
ray_pnidle--; /* now she's busy */ |
230 |
} |
231 |
if (sfirst != r_send_next) |
232 |
error(CONSISTENCY, "code screwup in ray_pflush"); |
233 |
r_send_next = 0; |
234 |
return(sfirst); /* return total # sent */ |
235 |
} |
236 |
|
237 |
|
238 |
extern void |
239 |
ray_psend( /* add a ray to our send queue */ |
240 |
RAY *r |
241 |
) |
242 |
{ |
243 |
if (r == NULL) |
244 |
return; |
245 |
/* flush output if necessary */ |
246 |
if (sendq_full() && ray_pflush() <= 0) |
247 |
error(INTERNAL, "ray_pflush failed in ray_psend"); |
248 |
|
249 |
r_queue[r_send_next++] = *r; |
250 |
} |
251 |
|
252 |
|
253 |
extern int |
254 |
ray_pqueue( /* queue a ray for computation */ |
255 |
RAY *r |
256 |
) |
257 |
{ |
258 |
if (r == NULL) |
259 |
return(0); |
260 |
/* check for full send queue */ |
261 |
if (sendq_full()) { |
262 |
RAY mySend; |
263 |
int rval; |
264 |
mySend = *r; |
265 |
/* wait for a result */ |
266 |
rval = ray_presult(r, 0); |
267 |
/* put new ray in queue */ |
268 |
r_queue[r_send_next++] = mySend; |
269 |
return(rval); /* done */ |
270 |
} |
271 |
/* else add ray to send queue */ |
272 |
r_queue[r_send_next++] = *r; |
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 |
return(1); |
279 |
} |
280 |
|
281 |
|
282 |
extern int |
283 |
ray_presult( /* check for a completed ray */ |
284 |
RAY *r, |
285 |
int poll |
286 |
) |
287 |
{ |
288 |
static struct timeval tpoll; /* zero timeval struct */ |
289 |
static fd_set readset, errset; |
290 |
int n, ok; |
291 |
register int pn; |
292 |
|
293 |
if (r == NULL) |
294 |
return(0); |
295 |
/* check queued results first */ |
296 |
if (r_recv_first < r_recv_next) { |
297 |
*r = r_queue[r_recv_first++]; |
298 |
return(1); |
299 |
} |
300 |
n = ray_pnprocs - ray_pnidle; /* pending before flush? */ |
301 |
|
302 |
if (ray_pflush() < 0) /* send new rays to process */ |
303 |
return(-1); |
304 |
/* reset receive queue */ |
305 |
r_recv_first = r_recv_next = RAYQLEN; |
306 |
|
307 |
if (!poll) /* count newly sent unless polling */ |
308 |
n = ray_pnprocs - ray_pnidle; |
309 |
if (n <= 0) /* return if nothing to await */ |
310 |
return(0); |
311 |
if (!poll && ray_pnprocs == 1) /* one process -> skip select() */ |
312 |
FD_SET(r_proc[0].fd_recv, &readset); |
313 |
|
314 |
getready: /* any children waiting for us? */ |
315 |
for (pn = ray_pnprocs; pn--; ) |
316 |
if (FD_ISSET(r_proc[pn].fd_recv, &readset) || |
317 |
FD_ISSET(r_proc[pn].fd_recv, &errset)) |
318 |
break; |
319 |
/* call select if we must */ |
320 |
if (pn < 0) { |
321 |
FD_ZERO(&readset); FD_ZERO(&errset); n = 0; |
322 |
for (pn = ray_pnprocs; pn--; ) { |
323 |
if (r_proc[pn].npending > 0) |
324 |
FD_SET(r_proc[pn].fd_recv, &readset); |
325 |
FD_SET(r_proc[pn].fd_recv, &errset); |
326 |
if (r_proc[pn].fd_recv >= n) |
327 |
n = r_proc[pn].fd_recv + 1; |
328 |
} |
329 |
/* find out who is ready */ |
330 |
while ((n = select(n, &readset, (fd_set *)NULL, &errset, |
331 |
poll ? &tpoll : (struct timeval *)NULL)) < 0) |
332 |
if (errno != EINTR) { |
333 |
error(WARNING, |
334 |
"select call failed in ray_presult"); |
335 |
ray_pclose(0); |
336 |
return(-1); |
337 |
} |
338 |
if (n > 0) /* go back and get it */ |
339 |
goto getready; |
340 |
return(0); /* else poll came up empty */ |
341 |
} |
342 |
if (r_recv_next + r_proc[pn].npending > sizeof(r_queue)/sizeof(RAY)) |
343 |
error(CONSISTENCY, "buffer shortage in ray_presult()"); |
344 |
|
345 |
/* read rendered ray data */ |
346 |
n = readbuf(r_proc[pn].fd_recv, (char *)&r_queue[r_recv_next], |
347 |
sizeof(RAY)*r_proc[pn].npending); |
348 |
if (n > 0) { |
349 |
r_recv_next += n/sizeof(RAY); |
350 |
ok = (n == sizeof(RAY)*r_proc[pn].npending); |
351 |
} else |
352 |
ok = 0; |
353 |
/* reset child's status */ |
354 |
FD_CLR(r_proc[pn].fd_recv, &readset); |
355 |
if (n <= 0) |
356 |
FD_CLR(r_proc[pn].fd_recv, &errset); |
357 |
r_proc[pn].npending = 0; |
358 |
ray_pnidle++; |
359 |
/* check for rendering errors */ |
360 |
if (!ok) { |
361 |
ray_pclose(0); /* process died -- clean up */ |
362 |
return(-1); |
363 |
} |
364 |
/* preen returned rays */ |
365 |
for (n = r_recv_next - r_recv_first; n--; ) { |
366 |
register RAY *rp = &r_queue[r_recv_first + n]; |
367 |
rp->rno = r_proc[pn].rno[n]; |
368 |
rp->parent = NULL; |
369 |
rp->newcset = rp->clipset = NULL; |
370 |
rp->rox = NULL; |
371 |
rp->slights = NULL; |
372 |
} |
373 |
/* return first ray received */ |
374 |
*r = r_queue[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 |
/* flag child process for quit() */ |
403 |
ray_pnprocs = -1; |
404 |
/* read each ray request set */ |
405 |
while ((n = read(fd_in, (char *)r_queue, sizeof(r_queue))) > 0) { |
406 |
int n2; |
407 |
if (n < sizeof(RAY)) |
408 |
break; |
409 |
/* get smuggled set length */ |
410 |
n2 = sizeof(RAY)*r_queue[0].crtype - n; |
411 |
if (n2 < 0) |
412 |
error(INTERNAL, "buffer over-read in ray_pchild"); |
413 |
if (n2 > 0) { /* read the rest of the set */ |
414 |
i = readbuf(fd_in, (char *)r_queue + n, n2); |
415 |
if (i != n2) |
416 |
break; |
417 |
n += n2; |
418 |
} |
419 |
n /= sizeof(RAY); |
420 |
/* evaluate rays */ |
421 |
for (i = 0; i < n; i++) { |
422 |
r_queue[i].crtype = r_queue[i].rtype; |
423 |
r_queue[i].parent = NULL; |
424 |
r_queue[i].clipset = NULL; |
425 |
r_queue[i].slights = NULL; |
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 |
ambsync(); /* load any new ambient values */ |
453 |
fflush(NULL); /* clear pending output */ |
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 |
if (ray_pnprocs > 0) /* close children if any */ |
528 |
ray_pclose(0); |
529 |
exit(ec); |
530 |
} |