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 |
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 |
> |
* |
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 |
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, NULL, PRIMARY, 1.0); |
49 |
> |
* rayorigin(&myRay, PRIMARY, NULL, NULL); |
50 |
|
* myRay.rno = ( my personal ray identifier ) |
51 |
|
* if (ray_pqueue(&myRay) == 1) |
52 |
|
* { do something with results } |
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 for you by ray_trace(). The |
58 |
< |
* great thing is that ray_pqueue() will trace rays faster in |
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_idle indicates the number of currently idle |
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: |
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_nprocs is set to zero. |
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_idle and call ray_psend(): |
80 |
> |
* results, check ray_pnidle and call ray_psend(): |
81 |
|
* |
82 |
< |
* while (ray_idle) { |
82 |
> |
* while (ray_pnidle) { |
83 |
|
* ( set up ray ) |
84 |
|
* ray_psend(&myRay); |
85 |
|
* } |
86 |
|
* |
87 |
< |
* The ray_presult() and/or ray_pqueue() functions may then be |
88 |
< |
* called to read back the results. |
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. |
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. |
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 system. Although you can avoid exit() with |
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 |
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 |
< |
* but otherwise your process should not be compromised. |
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 "rtprocess.h" |
135 |
|
#include "ray.h" |
136 |
< |
|
136 |
> |
#include "ambient.h" |
137 |
> |
#include <sys/types.h> |
138 |
> |
#include <sys/wait.h> |
139 |
|
#include "selcall.h" |
140 |
|
|
141 |
|
#ifndef RAYQLEN |
142 |
< |
#define RAYQLEN 16 /* # rays to send at once */ |
142 |
> |
#define RAYQLEN 12 /* # rays to send at once */ |
143 |
|
#endif |
144 |
|
|
145 |
|
#ifndef MAX_RPROCS |
152 |
|
|
153 |
|
extern char *shm_boundary; /* boundary of shared memory */ |
154 |
|
|
155 |
< |
int ray_nprocs = 0; /* number of child processes */ |
156 |
< |
int ray_idle = 0; /* number of idle children */ |
155 |
> |
int ray_pnprocs = 0; /* number of child processes */ |
156 |
> |
int ray_pnidle = 0; /* number of idle children */ |
157 |
|
|
158 |
|
static struct child_proc { |
159 |
|
int pid; /* child process id */ |
170 |
|
|
171 |
|
#define sendq_full() (r_send_next >= RAYQLEN) |
172 |
|
|
173 |
+ |
static int ray_pflush(void); |
174 |
+ |
static void ray_pchild(int fd_in, int fd_out); |
175 |
|
|
176 |
< |
void |
177 |
< |
ray_pinit(otnm, nproc) /* initialize ray-tracing processes */ |
178 |
< |
char *otnm; |
179 |
< |
int nproc; |
176 |
> |
|
177 |
> |
extern void |
178 |
> |
ray_pinit( /* initialize ray-tracing processes */ |
179 |
> |
char *otnm, |
180 |
> |
int nproc |
181 |
> |
) |
182 |
|
{ |
183 |
|
if (nobjects > 0) /* close old calculation */ |
184 |
|
ray_pdone(0); |
199 |
|
|
200 |
|
|
201 |
|
static int |
202 |
< |
ray_pflush() /* send queued rays to idle children */ |
202 |
> |
ray_pflush(void) /* send queued rays to idle children */ |
203 |
|
{ |
204 |
|
int nc, n, nw, i, sfirst; |
205 |
|
|
206 |
< |
if ((ray_idle <= 0 | r_send_next <= 0)) |
206 |
> |
if ((ray_pnidle <= 0) | (r_send_next <= 0)) |
207 |
|
return(0); /* nothing we can send */ |
208 |
|
|
209 |
|
sfirst = 0; /* divvy up labor */ |
210 |
< |
nc = ray_idle; |
211 |
< |
for (i = ray_nprocs; nc && i--; ) { |
210 |
> |
nc = ray_pnidle; |
211 |
> |
for (i = ray_pnprocs; nc && i--; ) { |
212 |
|
if (r_proc[i].npending > 0) |
213 |
|
continue; /* child looks busy */ |
214 |
|
n = (r_send_next - sfirst)/nc--; |
224 |
|
while (n--) /* record ray IDs */ |
225 |
|
r_proc[i].rno[n] = r_queue[sfirst+n].rno; |
226 |
|
sfirst += r_proc[i].npending; |
227 |
< |
ray_idle--; /* now she's busy */ |
227 |
> |
ray_pnidle--; /* now she's busy */ |
228 |
|
} |
229 |
|
if (sfirst != r_send_next) |
230 |
|
error(CONSISTENCY, "code screwup in ray_pflush"); |
233 |
|
} |
234 |
|
|
235 |
|
|
236 |
< |
void |
237 |
< |
ray_psend(r) /* add a ray to our send queue */ |
238 |
< |
RAY *r; |
236 |
> |
extern void |
237 |
> |
ray_psend( /* add a ray to our send queue */ |
238 |
> |
RAY *r |
239 |
> |
) |
240 |
|
{ |
241 |
|
if (r == NULL) |
242 |
|
return; |
244 |
|
if (sendq_full() && ray_pflush() <= 0) |
245 |
|
error(INTERNAL, "ray_pflush failed in ray_psend"); |
246 |
|
|
247 |
< |
copystruct(&r_queue[r_send_next], r); |
233 |
< |
r_send_next++; |
247 |
> |
r_queue[r_send_next++] = *r; |
248 |
|
} |
249 |
|
|
250 |
|
|
251 |
< |
int |
252 |
< |
ray_pqueue(r) /* queue a ray for computation */ |
253 |
< |
RAY *r; |
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; |
246 |
< |
int rval; |
247 |
< |
copystruct(&mySend, r); |
260 |
> |
RAY mySend = *r; |
261 |
|
/* wait for a result */ |
262 |
< |
rval = ray_presult(r, 0); |
262 |
> |
if (ray_presult(r, 0) <= 0) |
263 |
> |
return(-1); |
264 |
|
/* put new ray in queue */ |
265 |
< |
copystruct(&r_queue[r_send_next], &mySend); |
266 |
< |
r_send_next++; |
267 |
< |
return(rval); /* done */ |
265 |
> |
r_queue[r_send_next++] = mySend; |
266 |
> |
/* XXX r_send_next may now be > RAYQLEN */ |
267 |
> |
return(1); |
268 |
|
} |
269 |
< |
/* add ray to send queue */ |
270 |
< |
copystruct(&r_queue[r_send_next], r); |
257 |
< |
r_send_next++; |
269 |
> |
/* else add ray to send queue */ |
270 |
> |
r_queue[r_send_next++] = *r; |
271 |
|
/* check for returned ray... */ |
272 |
|
if (r_recv_first >= r_recv_next) |
273 |
|
return(0); |
274 |
|
/* ...one is sitting in queue */ |
275 |
< |
copystruct(r, &r_queue[r_recv_first]); |
263 |
< |
r_recv_first++; |
275 |
> |
*r = r_queue[r_recv_first++]; |
276 |
|
return(1); |
277 |
|
} |
278 |
|
|
279 |
|
|
280 |
< |
int |
281 |
< |
ray_presult(r, poll) /* check for a completed ray */ |
282 |
< |
RAY *r; |
283 |
< |
int poll; |
280 |
> |
extern int |
281 |
> |
ray_presult( /* check for a completed ray */ |
282 |
> |
RAY *r, |
283 |
> |
int poll |
284 |
> |
) |
285 |
|
{ |
286 |
|
static struct timeval tpoll; /* zero timeval struct */ |
287 |
|
static fd_set readset, errset; |
292 |
|
return(0); |
293 |
|
/* check queued results first */ |
294 |
|
if (r_recv_first < r_recv_next) { |
295 |
< |
copystruct(r, &r_queue[r_recv_first]); |
283 |
< |
r_recv_first++; |
295 |
> |
*r = r_queue[r_recv_first++]; |
296 |
|
return(1); |
297 |
|
} |
298 |
< |
n = ray_nprocs - ray_idle; /* pending before flush? */ |
298 |
> |
n = ray_pnprocs - ray_pnidle; /* pending before flush? */ |
299 |
|
|
300 |
|
if (ray_pflush() < 0) /* send new rays to process */ |
301 |
|
return(-1); |
303 |
|
r_recv_first = r_recv_next = RAYQLEN; |
304 |
|
|
305 |
|
if (!poll) /* count newly sent unless polling */ |
306 |
< |
n = ray_nprocs - ray_idle; |
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_nprocs; pn--; ) |
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 */ |
318 |
|
if (pn < 0) { |
319 |
|
FD_ZERO(&readset); FD_ZERO(&errset); n = 0; |
320 |
< |
for (pn = ray_nprocs; pn--; ) { |
320 |
> |
for (pn = ray_pnprocs; pn--; ) { |
321 |
|
if (r_proc[pn].npending > 0) |
322 |
|
FD_SET(r_proc[pn].fd_recv, &readset); |
323 |
|
FD_SET(r_proc[pn].fd_recv, &errset); |
353 |
|
if (n <= 0) |
354 |
|
FD_CLR(r_proc[pn].fd_recv, &errset); |
355 |
|
r_proc[pn].npending = 0; |
356 |
< |
ray_idle++; |
356 |
> |
ray_pnidle++; |
357 |
|
/* check for rendering errors */ |
358 |
|
if (!ok) { |
359 |
|
ray_pclose(0); /* process died -- clean up */ |
369 |
|
rp->slights = NULL; |
370 |
|
} |
371 |
|
/* return first ray received */ |
372 |
< |
copystruct(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; |
377 |
> |
extern void |
378 |
> |
ray_pdone( /* reap children and free data */ |
379 |
> |
int freall |
380 |
> |
) |
381 |
|
{ |
382 |
|
ray_pclose(0); /* close child processes */ |
383 |
|
|
390 |
|
|
391 |
|
|
392 |
|
static void |
393 |
< |
ray_pchild(fd_in, fd_out) /* process rays (never returns) */ |
394 |
< |
int fd_in; |
395 |
< |
int fd_out; |
393 |
> |
ray_pchild( /* process rays (never returns) */ |
394 |
> |
int fd_in, |
395 |
> |
int fd_out |
396 |
> |
) |
397 |
|
{ |
398 |
|
int n; |
399 |
|
register int i; |
400 |
+ |
/* flag child process for quit() */ |
401 |
+ |
ray_pnprocs = -1; |
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)) |
405 |
> |
if (n < sizeof(RAY)) |
406 |
|
break; |
389 |
– |
n /= sizeof(RAY); |
407 |
|
/* get smuggled set length */ |
408 |
< |
n2 = r_queue[0].crtype - n; |
408 |
> |
n2 = sizeof(RAY)*r_queue[0].crtype - n; |
409 |
|
if (n2 < 0) |
410 |
|
error(INTERNAL, "buffer over-read in ray_pchild"); |
411 |
|
if (n2 > 0) { /* read the rest of the set */ |
412 |
< |
i = readbuf(fd_in, (char *)(r_queue+n), |
413 |
< |
sizeof(RAY)*n2); |
397 |
< |
if (i != sizeof(RAY)*n2) |
412 |
> |
i = readbuf(fd_in, (char *)r_queue + n, n2); |
413 |
> |
if (i != n2) |
414 |
|
break; |
415 |
|
n += n2; |
416 |
|
} |
417 |
+ |
n /= sizeof(RAY); |
418 |
|
/* evaluate rays */ |
419 |
|
for (i = 0; i < n; i++) { |
420 |
|
r_queue[i].crtype = r_queue[i].rtype; |
421 |
|
r_queue[i].parent = NULL; |
422 |
|
r_queue[i].clipset = NULL; |
423 |
|
r_queue[i].slights = NULL; |
407 |
– |
r_queue[i].revf = raytrace; |
424 |
|
samplendx++; |
425 |
|
rayclear(&r_queue[i]); |
426 |
|
rayvalue(&r_queue[i]); |
437 |
|
} |
438 |
|
|
439 |
|
|
440 |
< |
void |
441 |
< |
ray_popen(nadd) /* open the specified # processes */ |
442 |
< |
int nadd; |
440 |
> |
extern void |
441 |
> |
ray_popen( /* open the specified # processes */ |
442 |
> |
int nadd |
443 |
> |
) |
444 |
|
{ |
445 |
|
/* check if our table has room */ |
446 |
< |
if (ray_nprocs + nadd > MAX_NPROCS) |
447 |
< |
nadd = MAX_NPROCS - ray_nprocs; |
446 |
> |
if (ray_pnprocs + nadd > MAX_NPROCS) |
447 |
> |
nadd = MAX_NPROCS - ray_pnprocs; |
448 |
|
if (nadd <= 0) |
449 |
|
return; |
450 |
< |
fflush(stderr); /* clear pending output */ |
451 |
< |
fflush(stdout); |
450 |
> |
ambsync(); /* load any new ambient values */ |
451 |
> |
fflush(NULL); /* clear pending output */ |
452 |
|
while (nadd--) { /* fork each new process */ |
453 |
|
int p0[2], p1[2]; |
454 |
|
if (pipe(p0) < 0 || pipe(p1) < 0) |
455 |
|
error(SYSTEM, "cannot create pipe"); |
456 |
< |
if ((r_proc[ray_nprocs].pid = fork()) == 0) { |
456 |
> |
if ((r_proc[ray_pnprocs].pid = fork()) == 0) { |
457 |
|
int pn; /* close others' descriptors */ |
458 |
< |
for (pn = ray_nprocs; pn--; ) { |
458 |
> |
for (pn = ray_pnprocs; pn--; ) { |
459 |
|
close(r_proc[pn].fd_send); |
460 |
|
close(r_proc[pn].fd_recv); |
461 |
|
} |
463 |
|
/* following call never returns */ |
464 |
|
ray_pchild(p1[0], p0[1]); |
465 |
|
} |
466 |
< |
if (r_proc[ray_nprocs].pid < 0) |
466 |
> |
if (r_proc[ray_pnprocs].pid < 0) |
467 |
|
error(SYSTEM, "cannot fork child process"); |
468 |
|
close(p1[0]); close(p0[1]); |
469 |
< |
r_proc[ray_nprocs].fd_send = p1[1]; |
470 |
< |
r_proc[ray_nprocs].fd_recv = p0[0]; |
471 |
< |
r_proc[ray_nprocs].npending = 0; |
472 |
< |
ray_nprocs++; |
473 |
< |
ray_idle++; |
469 |
> |
/* |
470 |
> |
* Close write stream on exec to avoid multiprocessing deadlock. |
471 |
> |
* No use in read stream without it, so set flag there as well. |
472 |
> |
*/ |
473 |
> |
fcntl(p1[1], F_SETFD, FD_CLOEXEC); |
474 |
> |
fcntl(p0[0], F_SETFD, FD_CLOEXEC); |
475 |
> |
r_proc[ray_pnprocs].fd_send = p1[1]; |
476 |
> |
r_proc[ray_pnprocs].fd_recv = p0[0]; |
477 |
> |
r_proc[ray_pnprocs].npending = 0; |
478 |
> |
ray_pnprocs++; |
479 |
> |
ray_pnidle++; |
480 |
|
} |
481 |
|
} |
482 |
|
|
483 |
|
|
484 |
< |
void |
485 |
< |
ray_pclose(nsub) /* close one or more child processes */ |
486 |
< |
int nsub; |
484 |
> |
extern void |
485 |
> |
ray_pclose( /* close one or more child processes */ |
486 |
> |
int nsub |
487 |
> |
) |
488 |
|
{ |
489 |
|
static int inclose = 0; |
490 |
|
RAY res; |
493 |
|
return; |
494 |
|
inclose++; |
495 |
|
/* check argument */ |
496 |
< |
if ((nsub <= 0 | nsub > ray_nprocs)) |
497 |
< |
nsub = ray_nprocs; |
496 |
> |
if ((nsub <= 0) | (nsub > ray_pnprocs)) |
497 |
> |
nsub = ray_pnprocs; |
498 |
|
/* clear our ray queue */ |
499 |
|
while (ray_presult(&res,0) > 0) |
500 |
|
; |
501 |
|
/* clean up children */ |
502 |
|
while (nsub--) { |
503 |
|
int status; |
504 |
< |
ray_nprocs--; |
505 |
< |
close(r_proc[ray_nprocs].fd_recv); |
506 |
< |
close(r_proc[ray_nprocs].fd_send); |
507 |
< |
while (wait(&status) != r_proc[ray_nprocs].pid) |
508 |
< |
; |
504 |
> |
ray_pnprocs--; |
505 |
> |
close(r_proc[ray_pnprocs].fd_recv); |
506 |
> |
close(r_proc[ray_pnprocs].fd_send); |
507 |
> |
if (waitpid(r_proc[ray_pnprocs].pid, &status, 0) < 0) |
508 |
> |
status = 127<<8; |
509 |
|
if (status) { |
510 |
|
sprintf(errmsg, |
511 |
|
"rendering process %d exited with code %d", |
512 |
< |
r_proc[ray_nprocs].pid, status>>8); |
512 |
> |
r_proc[ray_pnprocs].pid, status>>8); |
513 |
|
error(WARNING, errmsg); |
514 |
|
} |
515 |
< |
ray_idle--; |
515 |
> |
ray_pnidle--; |
516 |
|
} |
517 |
|
inclose--; |
518 |
|
} |
522 |
|
quit(ec) /* make sure exit is called */ |
523 |
|
int ec; |
524 |
|
{ |
525 |
+ |
if (ray_pnprocs > 0) /* close children if any */ |
526 |
+ |
ray_pclose(0); |
527 |
|
exit(ec); |
528 |
|
} |