| 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 |
| 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 } |
| 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. |
| 75 |
|
* until a value is available, returning 0 only if the |
| 76 |
|
* queue is completely empty. A negative return value |
| 77 |
|
* indicates that a rendering process died. If this |
| 78 |
< |
* happens, ray_close(0) is automatically called to close |
| 78 |
> |
* happens, ray_pclose(0) is automatically called to close |
| 79 |
|
* all child processes, and ray_pnprocs is set to zero. |
| 80 |
|
* |
| 81 |
|
* If you just want to fill the ray queue without checking for |
| 86 |
|
* ray_psend(&myRay); |
| 87 |
|
* } |
| 88 |
|
* |
| 89 |
< |
* The ray_presult() and/or ray_pqueue() functions may then be |
| 90 |
< |
* called to read back the results. |
| 89 |
> |
* Note that it is a fatal error to call ra_psend() when |
| 90 |
> |
* ray_pnidle is zero. The ray_presult() and/or ray_pqueue() |
| 91 |
> |
* functions may be called subsequently to read back the results. |
| 92 |
|
* |
| 93 |
|
* When you are done, you may call ray_pdone(1) to close |
| 94 |
|
* all child processes and clean up memory used by Radiance. |
| 95 |
|
* Any queued ray calculations will be awaited and discarded. |
| 96 |
|
* As with ray_done(), ray_pdone(0) hangs onto data files |
| 97 |
|
* and fonts that are likely to be used in subsequent renderings. |
| 98 |
< |
* Whether you want to bother cleaning up memory or not, you |
| 99 |
< |
* should at least call ray_pclose(0) to clean the child processes. |
| 98 |
> |
* Whether you need to clean up memory or not, you should |
| 99 |
> |
* at least call ray_pclose(0) to await the child processes. |
| 100 |
|
* |
| 101 |
|
* Warning: You cannot affect any of the rendering processes |
| 102 |
|
* by changing global parameter values onece ray_pinit() has |
| 105 |
|
* If you just want to reap children so that you can alter the |
| 106 |
|
* rendering parameters without reloading the scene, use the |
| 107 |
|
* ray_pclose(0) and ray_popen(nproc) calls to close |
| 108 |
< |
* then restart the child processes. |
| 108 |
> |
* then restart the child processes after the changes are made. |
| 109 |
|
* |
| 110 |
|
* Note: These routines are written to coordinate with the |
| 111 |
|
* definitions in raycalls.c, and in fact depend on them. |
| 112 |
|
* If you want to trace a ray and get a result synchronously, |
| 113 |
|
* use the ray_trace() call to compute it in the parent process. |
| 114 |
+ |
* This will not interfere with any subprocess calculations, |
| 115 |
+ |
* but beware that a fatal error may end with a call to quit(). |
| 116 |
|
* |
| 117 |
|
* Note: One of the advantages of using separate processes |
| 118 |
|
* is that it gives the calling program some immunity from |
| 119 |
|
* fatal rendering errors. As discussed in raycalls.c, |
| 120 |
|
* Radiance tends to throw up its hands and exit at the |
| 121 |
|
* first sign of trouble, calling quit() to return control |
| 122 |
< |
* to the system. Although you can avoid exit() with |
| 122 |
> |
* to the top level. Although you can avoid exit() with |
| 123 |
|
* your own longjmp() in quit(), the cleanup afterwards |
| 124 |
|
* is always suspect. Through the use of subprocesses, |
| 125 |
|
* we avoid this pitfall by closing the processes and |
| 126 |
|
* returning a negative value from ray_pqueue() or |
| 127 |
|
* ray_presult(). If you get a negative value from either |
| 128 |
|
* of these calls, you can assume that the processes have |
| 129 |
< |
* been cleaned up with a call to ray_close(), though you |
| 129 |
> |
* been cleaned up with a call to ray_pclose(), though you |
| 130 |
|
* will have to call ray_pdone() yourself if you want to |
| 131 |
< |
* free memory. Obviously, you cannot continue rendering, |
| 132 |
< |
* but otherwise your process should not be compromised. |
| 131 |
> |
* free memory. Obviously, you cannot continue rendering |
| 132 |
> |
* without risking further errors, but otherwise your |
| 133 |
> |
* process should not be compromised. |
| 134 |
|
*/ |
| 135 |
|
|
| 136 |
+ |
#include "rtprocess.h" |
| 137 |
|
#include "ray.h" |
| 138 |
< |
|
| 138 |
> |
#include "ambient.h" |
| 139 |
> |
#include <sys/types.h> |
| 140 |
> |
#include <sys/wait.h> |
| 141 |
|
#include "selcall.h" |
| 142 |
|
|
| 143 |
|
#ifndef RAYQLEN |
| 144 |
< |
#define RAYQLEN 16 /* # rays to send at once */ |
| 144 |
> |
#define RAYQLEN 12 /* # rays to send at once */ |
| 145 |
|
#endif |
| 146 |
|
|
| 147 |
|
#ifndef MAX_RPROCS |
| 154 |
|
|
| 155 |
|
extern char *shm_boundary; /* boundary of shared memory */ |
| 156 |
|
|
| 157 |
+ |
int ray_pfifo = 0; /* maintain ray call order? */ |
| 158 |
|
int ray_pnprocs = 0; /* number of child processes */ |
| 159 |
|
int ray_pnidle = 0; /* number of idle children */ |
| 160 |
|
|
| 163 |
|
int fd_send; /* write to child here */ |
| 164 |
|
int fd_recv; /* read from child here */ |
| 165 |
|
int npending; /* # rays in process */ |
| 166 |
< |
unsigned long rno[RAYQLEN]; /* working on these rays */ |
| 166 |
> |
RNUMBER rno[RAYQLEN]; /* working on these rays */ |
| 167 |
|
} r_proc[MAX_NPROCS]; /* our child processes */ |
| 168 |
|
|
| 169 |
|
static RAY r_queue[2*RAYQLEN]; /* ray i/o buffer */ |
| 173 |
|
|
| 174 |
|
#define sendq_full() (r_send_next >= RAYQLEN) |
| 175 |
|
|
| 176 |
+ |
static int ray_pflush(void); |
| 177 |
+ |
static void ray_pchild(int fd_in, int fd_out); |
| 178 |
|
|
| 179 |
< |
void |
| 180 |
< |
ray_pinit(otnm, nproc) /* initialize ray-tracing processes */ |
| 181 |
< |
char *otnm; |
| 182 |
< |
int nproc; |
| 179 |
> |
|
| 180 |
> |
extern void |
| 181 |
> |
ray_pinit( /* initialize ray-tracing processes */ |
| 182 |
> |
char *otnm, |
| 183 |
> |
int nproc |
| 184 |
> |
) |
| 185 |
|
{ |
| 186 |
|
if (nobjects > 0) /* close old calculation */ |
| 187 |
|
ray_pdone(0); |
| 188 |
|
|
| 189 |
|
ray_init(otnm); /* load the shared scene */ |
| 190 |
|
|
| 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 |
– |
|
| 191 |
|
r_send_next = 0; /* set up queue */ |
| 192 |
|
r_recv_first = r_recv_next = RAYQLEN; |
| 193 |
|
|
| 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 |
|
|
| 230 |
|
} |
| 231 |
|
|
| 232 |
|
|
| 233 |
< |
void |
| 234 |
< |
ray_psend(r) /* add a ray to our send queue */ |
| 235 |
< |
RAY *r; |
| 233 |
> |
extern void |
| 234 |
> |
ray_psend( /* add a ray to our send queue */ |
| 235 |
> |
RAY *r |
| 236 |
> |
) |
| 237 |
|
{ |
| 238 |
|
if (r == NULL) |
| 239 |
|
return; |
| 241 |
|
if (sendq_full() && ray_pflush() <= 0) |
| 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; |
| 248 |
> |
extern int |
| 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; |
| 277 |
> |
extern int |
| 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; |
| 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 |
|
n = ray_pnprocs - ray_pnidle; /* pending before flush? */ |
| 303 |
|
n = ray_pnprocs - ray_pnidle; |
| 304 |
|
if (n <= 0) /* return if nothing to await */ |
| 305 |
|
return(0); |
| 306 |
+ |
if (!poll && ray_pnprocs == 1) /* one process -> skip select() */ |
| 307 |
+ |
FD_SET(r_proc[0].fd_recv, &readset); |
| 308 |
+ |
|
| 309 |
|
getready: /* any children waiting for us? */ |
| 310 |
|
for (pn = ray_pnprocs; pn--; ) |
| 311 |
|
if (FD_ISSET(r_proc[pn].fd_recv, &readset) || |
| 366 |
|
rp->slights = NULL; |
| 367 |
|
} |
| 368 |
|
/* return first ray received */ |
| 369 |
< |
*r = r_queue[r_recv_first]; |
| 358 |
< |
r_recv_first++; |
| 369 |
> |
*r = r_queue[r_recv_first++]; |
| 370 |
|
return(1); |
| 371 |
|
} |
| 372 |
|
|
| 373 |
|
|
| 374 |
< |
void |
| 375 |
< |
ray_pdone(freall) /* reap children and free data */ |
| 376 |
< |
int freall; |
| 374 |
> |
extern void |
| 375 |
> |
ray_pdone( /* reap children and free data */ |
| 376 |
> |
int freall |
| 377 |
> |
) |
| 378 |
|
{ |
| 379 |
|
ray_pclose(0); /* close child processes */ |
| 380 |
|
|
| 387 |
|
|
| 388 |
|
|
| 389 |
|
static void |
| 390 |
< |
ray_pchild(fd_in, fd_out) /* process rays (never returns) */ |
| 391 |
< |
int fd_in; |
| 392 |
< |
int fd_out; |
| 390 |
> |
ray_pchild( /* process rays (never returns) */ |
| 391 |
> |
int fd_in, |
| 392 |
> |
int fd_out |
| 393 |
> |
) |
| 394 |
|
{ |
| 395 |
|
int n; |
| 396 |
|
register int i; |
| 397 |
+ |
/* flag child process for quit() */ |
| 398 |
+ |
ray_pnprocs = -1; |
| 399 |
|
/* read each ray request set */ |
| 400 |
|
while ((n = read(fd_in, (char *)r_queue, sizeof(r_queue))) > 0) { |
| 401 |
|
int n2; |
| 402 |
< |
if (n % sizeof(RAY)) |
| 402 |
> |
if (n < sizeof(RAY)) |
| 403 |
|
break; |
| 389 |
– |
n /= sizeof(RAY); |
| 404 |
|
/* get smuggled set length */ |
| 405 |
< |
n2 = r_queue[0].crtype - n; |
| 405 |
> |
n2 = sizeof(RAY)*r_queue[0].crtype - n; |
| 406 |
|
if (n2 < 0) |
| 407 |
|
error(INTERNAL, "buffer over-read in ray_pchild"); |
| 408 |
|
if (n2 > 0) { /* read the rest of the set */ |
| 409 |
< |
i = readbuf(fd_in, (char *)(r_queue+n), |
| 410 |
< |
sizeof(RAY)*n2); |
| 397 |
< |
if (i != sizeof(RAY)*n2) |
| 409 |
> |
i = readbuf(fd_in, (char *)r_queue + n, n2); |
| 410 |
> |
if (i != n2) |
| 411 |
|
break; |
| 412 |
|
n += n2; |
| 413 |
|
} |
| 414 |
+ |
n /= sizeof(RAY); |
| 415 |
|
/* evaluate rays */ |
| 416 |
|
for (i = 0; i < n; i++) { |
| 417 |
|
r_queue[i].crtype = r_queue[i].rtype; |
| 418 |
|
r_queue[i].parent = NULL; |
| 419 |
|
r_queue[i].clipset = NULL; |
| 420 |
|
r_queue[i].slights = NULL; |
| 421 |
< |
r_queue[i].revf = raytrace; |
| 421 |
> |
r_queue[i].rlvl = 0; |
| 422 |
|
samplendx++; |
| 423 |
|
rayclear(&r_queue[i]); |
| 424 |
|
rayvalue(&r_queue[i]); |
| 435 |
|
} |
| 436 |
|
|
| 437 |
|
|
| 438 |
< |
void |
| 439 |
< |
ray_popen(nadd) /* open the specified # processes */ |
| 440 |
< |
int nadd; |
| 438 |
> |
extern void |
| 439 |
> |
ray_popen( /* open the specified # processes */ |
| 440 |
> |
int nadd |
| 441 |
> |
) |
| 442 |
|
{ |
| 443 |
|
/* check if our table has room */ |
| 444 |
|
if (ray_pnprocs + nadd > MAX_NPROCS) |
| 445 |
|
nadd = MAX_NPROCS - ray_pnprocs; |
| 446 |
|
if (nadd <= 0) |
| 447 |
|
return; |
| 448 |
< |
fflush(stderr); /* clear pending output */ |
| 449 |
< |
fflush(stdout); |
| 448 |
> |
ambsync(); /* load any new ambient values */ |
| 449 |
> |
if (shm_boundary == NULL) { /* first child process? */ |
| 450 |
> |
preload_objs(); /* preload auxiliary data */ |
| 451 |
> |
/* set shared memory boundary */ |
| 452 |
> |
shm_boundary = (char *)malloc(16); |
| 453 |
> |
strcpy(shm_boundary, "SHM_BOUNDARY"); |
| 454 |
> |
} |
| 455 |
> |
fflush(NULL); /* clear pending output */ |
| 456 |
|
while (nadd--) { /* fork each new process */ |
| 457 |
|
int p0[2], p1[2]; |
| 458 |
|
if (pipe(p0) < 0 || pipe(p1) < 0) |
| 470 |
|
if (r_proc[ray_pnprocs].pid < 0) |
| 471 |
|
error(SYSTEM, "cannot fork child process"); |
| 472 |
|
close(p1[0]); close(p0[1]); |
| 473 |
+ |
/* |
| 474 |
+ |
* Close write stream on exec to avoid multiprocessing deadlock. |
| 475 |
+ |
* No use in read stream without it, so set flag there as well. |
| 476 |
+ |
*/ |
| 477 |
+ |
fcntl(p1[1], F_SETFD, FD_CLOEXEC); |
| 478 |
+ |
fcntl(p0[0], F_SETFD, FD_CLOEXEC); |
| 479 |
|
r_proc[ray_pnprocs].fd_send = p1[1]; |
| 480 |
|
r_proc[ray_pnprocs].fd_recv = p0[0]; |
| 481 |
|
r_proc[ray_pnprocs].npending = 0; |
| 485 |
|
} |
| 486 |
|
|
| 487 |
|
|
| 488 |
< |
void |
| 489 |
< |
ray_pclose(nsub) /* close one or more child processes */ |
| 490 |
< |
int nsub; |
| 488 |
> |
extern void |
| 489 |
> |
ray_pclose( /* close one or more child processes */ |
| 490 |
> |
int nsub |
| 491 |
> |
) |
| 492 |
|
{ |
| 493 |
|
static int inclose = 0; |
| 494 |
|
RAY res; |
| 508 |
|
ray_pnprocs--; |
| 509 |
|
close(r_proc[ray_pnprocs].fd_recv); |
| 510 |
|
close(r_proc[ray_pnprocs].fd_send); |
| 511 |
< |
while (wait(&status) != r_proc[ray_pnprocs].pid) |
| 512 |
< |
; |
| 511 |
> |
if (waitpid(r_proc[ray_pnprocs].pid, &status, 0) < 0) |
| 512 |
> |
status = 127<<8; |
| 513 |
|
if (status) { |
| 514 |
|
sprintf(errmsg, |
| 515 |
|
"rendering process %d exited with code %d", |
| 526 |
|
quit(ec) /* make sure exit is called */ |
| 527 |
|
int ec; |
| 528 |
|
{ |
| 529 |
+ |
if (ray_pnprocs > 0) /* close children if any */ |
| 530 |
+ |
ray_pclose(0); |
| 531 |
|
exit(ec); |
| 532 |
|
} |
