src/rt/raypcalls.c

#ifndef lint
static const char       RCSid[] = "$Id: raypcalls.c,v 2.19 2008/02/20 05:21:29 greg Exp $";
#endif
/*
 *  raypcalls.c - interface for parallel rendering using Radiance
 *
 *  External symbols declared in ray.h
 */

#include "copyright.h"

/*
 *  These calls are designed similarly to the ones in raycalls.c,
 *  but allow for multiple rendering processes on the same host
 *  machine.  There is no sense in specifying more child processes
 *  than you have processors, but one child may help by allowing
 *  asynchronous ray computation in an interactive program, and
 *  will protect the caller from fatal rendering errors.
 *
 *  You should first read and undrstand the header in raycalls.c,
 *  as some things are explained there that are not repated here.
 *
 *  The first step is opening one or more rendering processes
 *  with a call to ray_pinit(oct, nproc).  Before calling fork(),
 *  ray_pinit() loads the octree and data structures into the
 *  caller's memory, and ray_popen() synchronizes the ambient
 *  file, if any.  Shared memory permits all sorts of queries
 *  that wouldn't be possible otherwise, without causing any real
 *  memory overhead, since all the static data are shared
 *  between processes.  Rays are then traced using a simple
 *  queuing mechanism, explained below.
 *
 *  The ray queue buffers RAYQLEN rays before sending to
 *  children, each of which may internally buffer RAYQLEN rays.
 *
 *  Rays are queued and returned by a single
 *  ray_pqueue() call.  A ray_pqueue() return
 *  value of 0 indicates that no rays are ready
 *  and the queue is not yet full.  A return value of 1
 *  indicates that a ray was returned, though it is probably
 *  not the one you just requested.  Rays may be identified by
 *  the rno member of the RAY struct, which is incremented
 *  by the rayorigin() call, or may be set explicitly by
 *  the caller.  Below is an example call sequence:
 *
 *      myRay.rorg = ( ray origin point )
 *      myRay.rdir = ( normalized ray direction )
 *      myRay.rmax = ( maximum length, or zero for no limit )
 *      rayorigin(&myRay, PRIMARY, NULL, NULL);
 *      myRay.rno = ( my personal ray identifier )
 *      if (ray_pqueue(&myRay) == 1)
 *              { do something with results }
 *
 *  Note the differences between this and the simpler ray_trace()
 *  call.  In particular, the call may or may not return a value
 *  in the passed ray structure.  Also, you need to call rayorigin()
 *  yourself, which is normally called for you by ray_trace().  The
 *  benefit is that ray_pqueue() will trace rays faster in
 *  proportion to the number of CPUs you have available on your
 *  system.  If the ray queue is full before the call, ray_pqueue()
 *  will block until a result is ready so it can queue this one.
 *  The global int ray_pnidle indicates the number of currently idle
 *  children.  If you want to check for completed rays without blocking,
 *  or get the results from rays that have been queued without
 *  queuing any new ones, the ray_presult() call is for you:
 *
 *      if (ray_presult(&myRay, 1) == 1)
 *              { do something with results }
 *
 *  If the second argument is 1, the call won't block when
 *  results aren't ready, but will immediately return 0.
 *  If the second argument is 0, the call will block
 *  until a value is available, returning 0 only if the
 *  queue is completely empty.  A negative return value
 *  indicates that a rendering process died.  If this
 *  happens, ray_close(0) is automatically called to close
 *  all child processes, and ray_pnprocs is set to zero.
 *
 *  If you just want to fill the ray queue without checking for
 *  results, check ray_pnidle and call ray_psend():
 *
 *      while (ray_pnidle) {
 *              ( set up ray )
 *              ray_psend(&myRay);
 *      }
 *
 *  Note that it is a fatal error to call ra_psend() when
 *  ray_pnidle is zero.  The ray_presult() and/or ray_pqueue()
 *  functions may be called subsequently to read back the results.
 *
 *  When you are done, you may call ray_pdone(1) to close
 *  all child processes and clean up memory used by Radiance.
 *  Any queued ray calculations will be awaited and discarded.
 *  As with ray_done(), ray_pdone(0) hangs onto data files
 *  and fonts that are likely to be used in subsequent renderings.
 *  Whether you want to bother cleaning up memory or not, you
 *  should at least call ray_pclose(0) to clean the child processes.
 *
 *  Warning:  You cannot affect any of the rendering processes
 *  by changing global parameter values onece ray_pinit() has
 *  been called.  Changing global parameters will have no effect
 *  until the next call to ray_pinit(), which restarts everything.
 *  If you just want to reap children so that you can alter the
 *  rendering parameters without reloading the scene, use the
 *  ray_pclose(0) and ray_popen(nproc) calls to close
 *  then restart the child processes after the changes are made.
 *
 *  Note:  These routines are written to coordinate with the
 *  definitions in raycalls.c, and in fact depend on them.
 *  If you want to trace a ray and get a result synchronously,
 *  use the ray_trace() call to compute it in the parent process.
 *  This will not interfere with any subprocess calculations,
 *  but beware that a fatal error may end with a call to quit().
 *
 *  Note:  One of the advantages of using separate processes
 *  is that it gives the calling program some immunity from
 *  fatal rendering errors.  As discussed in raycalls.c,
 *  Radiance tends to throw up its hands and exit at the
 *  first sign of trouble, calling quit() to return control
 *  to the top level.  Although you can avoid exit() with
 *  your own longjmp() in quit(), the cleanup afterwards
 *  is always suspect.  Through the use of subprocesses,
 *  we avoid this pitfall by closing the processes and
 *  returning a negative value from ray_pqueue() or
 *  ray_presult().  If you get a negative value from either
 *  of these calls, you can assume that the processes have
 *  been cleaned up with a call to ray_close(), though you
 *  will have to call ray_pdone() yourself if you want to
 *  free memory.  Obviously, you cannot continue rendering
 *  without risking further errors, but otherwise your
 *  process should not be compromised.
 */

#include  "rtprocess.h"
#include  "ray.h"
#include  "ambient.h"
#include  <sys/types.h>
#include  <sys/wait.h>
#include  "selcall.h"

#ifndef RAYQLEN
#define RAYQLEN         12              /* # rays to send at once */
#endif

#ifndef MAX_RPROCS
#if (FD_SETSIZE/2-4 < 64) 
#define MAX_NPROCS      (FD_SETSIZE/2-4)
#else
#define MAX_NPROCS      64              /* max. # rendering processes */
#endif
#endif

extern char     *shm_boundary;          /* boundary of shared memory */

int             ray_pnprocs = 0;        /* number of child processes */
int             ray_pnidle = 0;         /* number of idle children */

static struct child_proc {
        int     pid;                            /* child process id */
        int     fd_send;                        /* write to child here */
        int     fd_recv;                        /* read from child here */
        int     npending;                       /* # rays in process */
        unsigned long  rno[RAYQLEN];            /* working on these rays */
} r_proc[MAX_NPROCS];                   /* our child processes */

static RAY      r_queue[2*RAYQLEN];     /* ray i/o buffer */
static int      r_send_next;            /* next send ray placement */
static int      r_recv_first;           /* position of first unreported ray */
static int      r_recv_next;            /* next receive ray placement */

#define sendq_full()    (r_send_next >= RAYQLEN)

static int ray_pflush(void);
static void ray_pchild(int fd_in, int fd_out);


extern void
ray_pinit(              /* initialize ray-tracing processes */
        char    *otnm,
        int     nproc
)
{
        if (nobjects > 0)               /* close old calculation */
                ray_pdone(0);

        ray_init(otnm);                 /* load the shared scene */

        r_send_next = 0;                /* set up queue */
        r_recv_first = r_recv_next = RAYQLEN;

        ray_popen(nproc);               /* fork children */
}


static int
ray_pflush(void)                        /* send queued rays to idle children */
{
        int     nc, n, nw, i, sfirst;

        if ((ray_pnidle <= 0) | (r_send_next <= 0))
                return(0);              /* nothing we can send */
        
        sfirst = 0;                     /* divvy up labor */
        nc = ray_pnidle;
        for (i = ray_pnprocs; nc && i--; ) {
                if (r_proc[i].npending > 0)
                        continue;       /* child looks busy */
                n = (r_send_next - sfirst)/nc--;
                if (!n)
                        continue;
                                        /* smuggle set size in crtype */
                r_queue[sfirst].crtype = n;
                nw = writebuf(r_proc[i].fd_send, (char *)&r_queue[sfirst],
                                sizeof(RAY)*n);
                if (nw != sizeof(RAY)*n)
                        return(-1);     /* write error */
                r_proc[i].npending = n;
                while (n--)             /* record ray IDs */
                        r_proc[i].rno[n] = r_queue[sfirst+n].rno;
                sfirst += r_proc[i].npending;
                ray_pnidle--;           /* now she's busy */
        }
        if (sfirst != r_send_next)
                error(CONSISTENCY, "code screwup in ray_pflush");
        r_send_next = 0;
        return(sfirst);                 /* return total # sent */
}


extern void
ray_psend(                      /* add a ray to our send queue */
        RAY     *r
)
{
        if (r == NULL)
                return;
                                        /* flush output if necessary */
        if (sendq_full() && ray_pflush() <= 0)
                error(INTERNAL, "ray_pflush failed in ray_psend");

        r_queue[r_send_next++] = *r;
}


extern int
ray_pqueue(                     /* queue a ray for computation */
        RAY     *r
)
{
        if (r == NULL)
                return(0);
                                        /* check for full send queue */
        if (sendq_full()) {
                RAY     mySend = *r;
                                        /* wait for a result */
                if (ray_presult(r, 0) <= 0)
                        return(-1);
                                        /* put new ray in queue */
                r_queue[r_send_next++] = mySend;
                                /* XXX r_send_next may now be > RAYQLEN */
                return(1);
        }
                                        /* else add ray to send queue */
        r_queue[r_send_next++] = *r;
                                        /* check for returned ray... */
        if (r_recv_first >= r_recv_next)
                return(0);
                                        /* ...one is sitting in queue */
        *r = r_queue[r_recv_first++];
        return(1);
}


extern int
ray_presult(            /* check for a completed ray */
        RAY     *r,
        int     poll
)
{
        static struct timeval   tpoll;  /* zero timeval struct */
        static fd_set   readset, errset;
        int     n, ok;
        register int    pn;

        if (r == NULL)
                return(0);
                                        /* check queued results first */
        if (r_recv_first < r_recv_next) {
                *r = r_queue[r_recv_first++];
                return(1);
        }
        n = ray_pnprocs - ray_pnidle;   /* pending before flush? */

        if (ray_pflush() < 0)           /* send new rays to process */
                return(-1);
                                        /* reset receive queue */
        r_recv_first = r_recv_next = RAYQLEN;

        if (!poll)                      /* count newly sent unless polling */
                n = ray_pnprocs - ray_pnidle;
        if (n <= 0)                     /* return if nothing to await */
                return(0);
        if (!poll && ray_pnprocs == 1)  /* one process -> skip select() */
                FD_SET(r_proc[0].fd_recv, &readset);

getready:                               /* any children waiting for us? */
        for (pn = ray_pnprocs; pn--; )
                if (FD_ISSET(r_proc[pn].fd_recv, &readset) ||
                                FD_ISSET(r_proc[pn].fd_recv, &errset))
                        break;
                                        /* call select if we must */
        if (pn < 0) {
                FD_ZERO(&readset); FD_ZERO(&errset); n = 0;
                for (pn = ray_pnprocs; pn--; ) {
                        if (r_proc[pn].npending > 0)
                                FD_SET(r_proc[pn].fd_recv, &readset);
                        FD_SET(r_proc[pn].fd_recv, &errset);
                        if (r_proc[pn].fd_recv >= n)
                                n = r_proc[pn].fd_recv + 1;
                }
                                        /* find out who is ready */
                while ((n = select(n, &readset, (fd_set *)NULL, &errset,
                                poll ? &tpoll : (struct timeval *)NULL)) < 0)
                        if (errno != EINTR) {
                                error(WARNING,
                                        "select call failed in ray_presult");
                                ray_pclose(0);
                                return(-1);
                        }
                if (n > 0)              /* go back and get it */
                        goto getready;
                return(0);              /* else poll came up empty */
        }
        if (r_recv_next + r_proc[pn].npending > sizeof(r_queue)/sizeof(RAY))
                error(CONSISTENCY, "buffer shortage in ray_presult()");

                                        /* read rendered ray data */
        n = readbuf(r_proc[pn].fd_recv, (char *)&r_queue[r_recv_next],
                        sizeof(RAY)*r_proc[pn].npending);
        if (n > 0) {
                r_recv_next += n/sizeof(RAY);
                ok = (n == sizeof(RAY)*r_proc[pn].npending);
        } else
                ok = 0;
                                        /* reset child's status */
        FD_CLR(r_proc[pn].fd_recv, &readset);
        if (n <= 0)
                FD_CLR(r_proc[pn].fd_recv, &errset);
        r_proc[pn].npending = 0;
        ray_pnidle++;
                                        /* check for rendering errors */
        if (!ok) {
                ray_pclose(0);          /* process died -- clean up */
                return(-1);
        }
                                        /* preen returned rays */
        for (n = r_recv_next - r_recv_first; n--; ) {
                register RAY    *rp = &r_queue[r_recv_first + n];
                rp->rno = r_proc[pn].rno[n];
                rp->parent = NULL;
                rp->newcset = rp->clipset = NULL;
                rp->rox = NULL;
                rp->slights = NULL;
        }
                                        /* return first ray received */
        *r = r_queue[r_recv_first++];
        return(1);
}


extern void
ray_pdone(              /* reap children and free data */
        int     freall
)
{
        ray_pclose(0);                  /* close child processes */

        if (shm_boundary != NULL) {     /* clear shared memory boundary */
                free((void *)shm_boundary);
                shm_boundary = NULL;
        }
        ray_done(freall);               /* free rendering data */
}


static void
ray_pchild(     /* process rays (never returns) */
        int     fd_in,
        int     fd_out
)
{
        int     n;
        register int    i;
                                        /* flag child process for quit() */
        ray_pnprocs = -1;
                                        /* read each ray request set */
        while ((n = read(fd_in, (char *)r_queue, sizeof(r_queue))) > 0) {
                int     n2;
                if (n < sizeof(RAY))
                        break;
                                        /* get smuggled set length */
                n2 = sizeof(RAY)*r_queue[0].crtype - n;
                if (n2 < 0)
                        error(INTERNAL, "buffer over-read in ray_pchild");
                if (n2 > 0) {           /* read the rest of the set */
                        i = readbuf(fd_in, (char *)r_queue + n, n2);
                        if (i != n2)
                                break;
                        n += n2;
                }
                n /= sizeof(RAY);
                                        /* evaluate rays */
                for (i = 0; i < n; i++) {
                        r_queue[i].crtype = r_queue[i].rtype;
                        r_queue[i].parent = NULL;
                        r_queue[i].clipset = NULL;
                        r_queue[i].slights = NULL;
                        samplendx++;
                        rayclear(&r_queue[i]);
                        rayvalue(&r_queue[i]);
                }
                                        /* write back our results */
                i = writebuf(fd_out, (char *)r_queue, sizeof(RAY)*n);
                if (i != sizeof(RAY)*n)
                        error(SYSTEM, "write error in ray_pchild");
        }
        if (n)
                error(SYSTEM, "read error in ray_pchild");
        ambsync();
        quit(0);                        /* normal exit */
}


extern void
ray_popen(                      /* open the specified # processes */
        int     nadd
)
{
                                        /* check if our table has room */
        if (ray_pnprocs + nadd > MAX_NPROCS)
                nadd = MAX_NPROCS - ray_pnprocs;
        if (nadd <= 0)
                return;
        ambsync();                      /* load any new ambient values */
        if (shm_boundary == NULL) {     /* first child process? */
                preload_objs();         /* preload auxiliary data */
                                        /* set shared memory boundary */
                shm_boundary = (char *)malloc(16);
                strcpy(shm_boundary, "SHM_BOUNDARY");
        }
        fflush(NULL);                   /* clear pending output */
        while (nadd--) {                /* fork each new process */
                int     p0[2], p1[2];
                if (pipe(p0) < 0 || pipe(p1) < 0)
                        error(SYSTEM, "cannot create pipe");
                if ((r_proc[ray_pnprocs].pid = fork()) == 0) {
                        int     pn;     /* close others' descriptors */
                        for (pn = ray_pnprocs; pn--; ) {
                                close(r_proc[pn].fd_send);
                                close(r_proc[pn].fd_recv);
                        }
                        close(p0[0]); close(p1[1]);
                                        /* following call never returns */
                        ray_pchild(p1[0], p0[1]);
                }
                if (r_proc[ray_pnprocs].pid < 0)
                        error(SYSTEM, "cannot fork child process");
                close(p1[0]); close(p0[1]);
                /*
                 * Close write stream on exec to avoid multiprocessing deadlock.
                 * No use in read stream without it, so set flag there as well.
                 */
                fcntl(p1[1], F_SETFD, FD_CLOEXEC);
                fcntl(p0[0], F_SETFD, FD_CLOEXEC);
                r_proc[ray_pnprocs].fd_send = p1[1];
                r_proc[ray_pnprocs].fd_recv = p0[0];
                r_proc[ray_pnprocs].npending = 0;
                ray_pnprocs++;
                ray_pnidle++;
        }
}


extern void
ray_pclose(             /* close one or more child processes */
        int     nsub
)
{
        static int      inclose = 0;
        RAY     res;
                                        /* check recursion */
        if (inclose)
                return;
        inclose++;
                                        /* check argument */
        if ((nsub <= 0) | (nsub > ray_pnprocs))
                nsub = ray_pnprocs;
                                        /* clear our ray queue */
        while (ray_presult(&res,0) > 0)
                ;
                                        /* clean up children */
        while (nsub--) {
                int     status;
                ray_pnprocs--;
                close(r_proc[ray_pnprocs].fd_recv);
                close(r_proc[ray_pnprocs].fd_send);
                if (waitpid(r_proc[ray_pnprocs].pid, &status, 0) < 0)
                        status = 127<<8;
                if (status) {
                        sprintf(errmsg,
                                "rendering process %d exited with code %d",
                                        r_proc[ray_pnprocs].pid, status>>8);
                        error(WARNING, errmsg);
                }
                ray_pnidle--;
        }
        inclose--;
}


void
quit(ec)                        /* make sure exit is called */
int     ec;
{
        if (ray_pnprocs > 0)    /* close children if any */
                ray_pclose(0);          
        exit(ec);
}
Revision:	2.20
Committed:	Tue Dec 2 23:28:34 2008 UTC (15 years, 11 months ago) by greg
Content type:	text/plain
Branch:	MAIN
Changes since 2.19:	+7 -7 lines
Log Message:	Delayed loading of object data until it is needed
#	User	Rev	Content
1	greg	2.1	#ifndef lint
2	greg	2.20	static const char RCSid[] = "$Id: raypcalls.c,v 2.19 2008/02/20 05:21:29 greg Exp $";
3	greg	2.1	#endif
4			/*
5			* raypcalls.c - interface for parallel rendering using Radiance
6			*
7			* External symbols declared in ray.h
8			*/
9
10	greg	2.2	#include "copyright.h"
11	greg	2.1
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	greg	2.13	* 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	greg	2.1	* 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	greg	2.14	* The ray queue buffers RAYQLEN rays before sending to
34			* children, each of which may internally buffer RAYQLEN rays.
35			*
36	greg	2.13	* Rays are queued and returned by a single
37	greg	2.1	* 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	greg	2.11	* rayorigin(&myRay, PRIMARY, NULL, NULL);
50	greg	2.1	* 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	greg	2.7	* yourself, which is normally called for you by ray_trace(). The
58			* benefit is that ray_pqueue() will trace rays faster in
59	greg	2.1	* 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	greg	2.3	* The global int ray_pnidle indicates the number of currently idle
63	greg	2.1	* 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	greg	2.3	* all child processes, and ray_pnprocs is set to zero.
78	greg	2.1	*
79			* If you just want to fill the ray queue without checking for
80	greg	2.3	* results, check ray_pnidle and call ray_psend():
81	greg	2.1	*
82	greg	2.3	* while (ray_pnidle) {
83	greg	2.1	* ( set up ray )
84			* ray_psend(&myRay);
85			* }
86			*
87	greg	2.7	* 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	greg	2.1	*
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	greg	2.7	* then restart the child processes after the changes are made.
107	greg	2.1	*
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	greg	2.13	* use the ray_trace() call to compute it in the parent process.
112	greg	2.7	* This will not interfere with any subprocess calculations,
113			* but beware that a fatal error may end with a call to quit().
114	greg	2.1	*
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	greg	2.7	* to the top level. Although you can avoid exit() with
121	greg	2.1	* 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	greg	2.7	* free memory. Obviously, you cannot continue rendering
130			* without risking further errors, but otherwise your
131			* process should not be compromised.
132	greg	2.1	*/
133
134	schorsch	2.6	#include "rtprocess.h"
135	greg	2.1	#include "ray.h"
136	schorsch	2.6	#include "ambient.h"
137	greg	2.18	#include <sys/types.h>
138			#include <sys/wait.h>
139	greg	2.1	#include "selcall.h"
140
141			#ifndef RAYQLEN
142	greg	2.13	#define RAYQLEN 12 /* # rays to send at once */
143	greg	2.1	#endif
144
145			#ifndef MAX_RPROCS
146			#if (FD_SETSIZE/2-4 < 64)
147			#define MAX_NPROCS (FD_SETSIZE/2-4)
148			#else
149			#define MAX_NPROCS 64 /* max. # rendering processes */
150			#endif
151			#endif
152
153			extern char shm_boundary; / boundary of shared memory */
154
155	greg	2.3	int ray_pnprocs = 0; /* number of child processes */
156			int ray_pnidle = 0; /* number of idle children */
157	greg	2.1
158			static struct child_proc {
159			int pid; /* child process id */
160			int fd_send; /* write to child here */
161			int fd_recv; /* read from child here */
162			int npending; /* # rays in process */
163			unsigned long rno[RAYQLEN]; /* working on these rays */
164			} r_proc[MAX_NPROCS]; /* our child processes */
165
166			static RAY r_queue[2RAYQLEN]; / ray i/o buffer */
167			static int r_send_next; /* next send ray placement */
168			static int r_recv_first; /* position of first unreported ray */
169			static int r_recv_next; /* next receive ray placement */
170
171			#define sendq_full() (r_send_next >= RAYQLEN)
172
173	schorsch	2.6	static int ray_pflush(void);
174	greg	2.13	static void ray_pchild(int fd_in, int fd_out);
175	greg	2.1
176	schorsch	2.6
177			extern void
178			ray_pinit( /* initialize ray-tracing processes */
179			char *otnm,
180			int nproc
181			)
182	greg	2.1	{
183			if (nobjects > 0) /* close old calculation */
184			ray_pdone(0);
185
186			ray_init(otnm); /* load the shared scene */
187
188			r_send_next = 0; /* set up queue */
189			r_recv_first = r_recv_next = RAYQLEN;
190
191			ray_popen(nproc); /* fork children */
192			}
193
194
195			static int
196	schorsch	2.6	ray_pflush(void) /* send queued rays to idle children */
197	greg	2.1	{
198			int nc, n, nw, i, sfirst;
199
200	schorsch	2.5	if ((ray_pnidle <= 0) \| (r_send_next <= 0))
201	greg	2.1	return(0); /* nothing we can send */
202
203			sfirst = 0; /* divvy up labor */
204	greg	2.3	nc = ray_pnidle;
205			for (i = ray_pnprocs; nc && i--; ) {
206	greg	2.1	if (r_proc[i].npending > 0)
207			continue; /* child looks busy */
208			n = (r_send_next - sfirst)/nc--;
209			if (!n)
210			continue;
211			/* smuggle set size in crtype */
212			r_queue[sfirst].crtype = n;
213			nw = writebuf(r_proc[i].fd_send, (char *)&r_queue[sfirst],
214			sizeof(RAY)*n);
215			if (nw != sizeof(RAY)*n)
216			return(-1); /* write error */
217			r_proc[i].npending = n;
218			while (n--) /* record ray IDs */
219			r_proc[i].rno[n] = r_queue[sfirst+n].rno;
220			sfirst += r_proc[i].npending;
221	greg	2.3	ray_pnidle--; /* now she's busy */
222	greg	2.1	}
223			if (sfirst != r_send_next)
224			error(CONSISTENCY, "code screwup in ray_pflush");
225			r_send_next = 0;
226			return(sfirst); /* return total # sent */
227			}
228
229
230	schorsch	2.6	extern void
231			ray_psend( /* add a ray to our send queue */
232			RAY *r
233			)
234	greg	2.1	{
235			if (r == NULL)
236			return;
237			/* flush output if necessary */
238			if (sendq_full() && ray_pflush() <= 0)
239			error(INTERNAL, "ray_pflush failed in ray_psend");
240
241	greg	2.14	r_queue[r_send_next++] = *r;
242	greg	2.1	}
243
244
245	schorsch	2.6	extern int
246			ray_pqueue( /* queue a ray for computation */
247			RAY *r
248			)
249	greg	2.1	{
250			if (r == NULL)
251			return(0);
252			/* check for full send queue */
253			if (sendq_full()) {
254	greg	2.19	RAY mySend = *r;
255	greg	2.1	/* wait for a result */
256	greg	2.19	if (ray_presult(r, 0) <= 0)
257			return(-1);
258	greg	2.1	/* put new ray in queue */
259	greg	2.14	r_queue[r_send_next++] = mySend;
260	greg	2.19	/* XXX r_send_next may now be > RAYQLEN */
261			return(1);
262	greg	2.1	}
263	greg	2.13	/* else add ray to send queue */
264	greg	2.14	r_queue[r_send_next++] = *r;
265	greg	2.1	/* check for returned ray... */
266			if (r_recv_first >= r_recv_next)
267			return(0);
268			/* ...one is sitting in queue */
269	greg	2.14	*r = r_queue[r_recv_first++];
270	greg	2.1	return(1);
271			}
272
273
274	schorsch	2.6	extern int
275			ray_presult( /* check for a completed ray */
276			RAY *r,
277			int poll
278			)
279	greg	2.1	{
280			static struct timeval tpoll; /* zero timeval struct */
281			static fd_set readset, errset;
282			int n, ok;
283			register int pn;
284
285			if (r == NULL)
286			return(0);
287			/* check queued results first */
288			if (r_recv_first < r_recv_next) {
289	greg	2.14	*r = r_queue[r_recv_first++];
290	greg	2.1	return(1);
291			}
292	greg	2.3	n = ray_pnprocs - ray_pnidle; /* pending before flush? */
293	greg	2.1
294			if (ray_pflush() < 0) /* send new rays to process */
295			return(-1);
296			/* reset receive queue */
297			r_recv_first = r_recv_next = RAYQLEN;
298
299			if (!poll) /* count newly sent unless polling */
300	greg	2.3	n = ray_pnprocs - ray_pnidle;
301	greg	2.1	if (n <= 0) /* return if nothing to await */
302			return(0);
303	greg	2.16	if (!poll && ray_pnprocs == 1) /* one process -> skip select() */
304			FD_SET(r_proc[0].fd_recv, &readset);
305
306	greg	2.1	getready: /* any children waiting for us? */
307	greg	2.3	for (pn = ray_pnprocs; pn--; )
308	greg	2.1	if (FD_ISSET(r_proc[pn].fd_recv, &readset) \|\|
309			FD_ISSET(r_proc[pn].fd_recv, &errset))
310			break;
311			/* call select if we must */
312			if (pn < 0) {
313			FD_ZERO(&readset); FD_ZERO(&errset); n = 0;
314	greg	2.3	for (pn = ray_pnprocs; pn--; ) {
315	greg	2.1	if (r_proc[pn].npending > 0)
316			FD_SET(r_proc[pn].fd_recv, &readset);
317			FD_SET(r_proc[pn].fd_recv, &errset);
318			if (r_proc[pn].fd_recv >= n)
319			n = r_proc[pn].fd_recv + 1;
320			}
321			/* find out who is ready */
322			while ((n = select(n, &readset, (fd_set *)NULL, &errset,
323			poll ? &tpoll : (struct timeval *)NULL)) < 0)
324			if (errno != EINTR) {
325			error(WARNING,
326			"select call failed in ray_presult");
327			ray_pclose(0);
328			return(-1);
329			}
330			if (n > 0) /* go back and get it */
331			goto getready;
332			return(0); /* else poll came up empty */
333			}
334			if (r_recv_next + r_proc[pn].npending > sizeof(r_queue)/sizeof(RAY))
335			error(CONSISTENCY, "buffer shortage in ray_presult()");
336
337			/* read rendered ray data */
338			n = readbuf(r_proc[pn].fd_recv, (char *)&r_queue[r_recv_next],
339			sizeof(RAY)*r_proc[pn].npending);
340			if (n > 0) {
341			r_recv_next += n/sizeof(RAY);
342			ok = (n == sizeof(RAY)*r_proc[pn].npending);
343			} else
344			ok = 0;
345			/* reset child's status */
346			FD_CLR(r_proc[pn].fd_recv, &readset);
347			if (n <= 0)
348			FD_CLR(r_proc[pn].fd_recv, &errset);
349			r_proc[pn].npending = 0;
350	greg	2.3	ray_pnidle++;
351	greg	2.1	/* check for rendering errors */
352			if (!ok) {
353			ray_pclose(0); /* process died -- clean up */
354			return(-1);
355			}
356			/* preen returned rays */
357			for (n = r_recv_next - r_recv_first; n--; ) {
358			register RAY *rp = &r_queue[r_recv_first + n];
359			rp->rno = r_proc[pn].rno[n];
360			rp->parent = NULL;
361			rp->newcset = rp->clipset = NULL;
362			rp->rox = NULL;
363			rp->slights = NULL;
364			}
365			/* return first ray received */
366	greg	2.13	*r = r_queue[r_recv_first++];
367	greg	2.1	return(1);
368			}
369
370
371	schorsch	2.6	extern void
372			ray_pdone( /* reap children and free data */
373			int freall
374			)
375	greg	2.1	{
376			ray_pclose(0); /* close child processes */
377
378			if (shm_boundary != NULL) { /* clear shared memory boundary */
379			free((void *)shm_boundary);
380			shm_boundary = NULL;
381			}
382			ray_done(freall); /* free rendering data */
383			}
384
385
386			static void
387	schorsch	2.6	ray_pchild( /* process rays (never returns) */
388			int fd_in,
389			int fd_out
390			)
391	greg	2.1	{
392			int n;
393			register int i;
394	greg	2.15	/* flag child process for quit() */
395			ray_pnprocs = -1;
396	greg	2.1	/* read each ray request set */
397			while ((n = read(fd_in, (char *)r_queue, sizeof(r_queue))) > 0) {
398			int n2;
399	greg	2.12	if (n < sizeof(RAY))
400	greg	2.1	break;
401			/* get smuggled set length */
402	greg	2.12	n2 = sizeof(RAY)*r_queue[0].crtype - n;
403	greg	2.1	if (n2 < 0)
404			error(INTERNAL, "buffer over-read in ray_pchild");
405			if (n2 > 0) { /* read the rest of the set */
406	greg	2.12	i = readbuf(fd_in, (char *)r_queue + n, n2);
407			if (i != n2)
408	greg	2.1	break;
409			n += n2;
410			}
411	greg	2.12	n /= sizeof(RAY);
412	greg	2.1	/* evaluate rays */
413			for (i = 0; i < n; i++) {
414			r_queue[i].crtype = r_queue[i].rtype;
415			r_queue[i].parent = NULL;
416			r_queue[i].clipset = NULL;
417			r_queue[i].slights = NULL;
418			samplendx++;
419			rayclear(&r_queue[i]);
420			rayvalue(&r_queue[i]);
421			}
422			/* write back our results */
423			i = writebuf(fd_out, (char )r_queue, sizeof(RAY)n);
424			if (i != sizeof(RAY)*n)
425			error(SYSTEM, "write error in ray_pchild");
426			}
427			if (n)
428			error(SYSTEM, "read error in ray_pchild");
429			ambsync();
430			quit(0); /* normal exit */
431			}
432
433
434	schorsch	2.6	extern void
435			ray_popen( /* open the specified # processes */
436			int nadd
437			)
438	greg	2.1	{
439			/* check if our table has room */
440	greg	2.3	if (ray_pnprocs + nadd > MAX_NPROCS)
441			nadd = MAX_NPROCS - ray_pnprocs;
442	greg	2.1	if (nadd <= 0)
443			return;
444	greg	2.13	ambsync(); /* load any new ambient values */
445	greg	2.20	if (shm_boundary == NULL) { /* first child process? */
446			preload_objs(); /* preload auxiliary data */
447			/* set shared memory boundary */
448			shm_boundary = (char *)malloc(16);
449			strcpy(shm_boundary, "SHM_BOUNDARY");
450			}
451	greg	2.13	fflush(NULL); /* clear pending output */
452	greg	2.1	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	greg	2.3	if ((r_proc[ray_pnprocs].pid = fork()) == 0) {
457	greg	2.1	int pn; /* close others' descriptors */
458	greg	2.3	for (pn = ray_pnprocs; pn--; ) {
459	greg	2.1	close(r_proc[pn].fd_send);
460			close(r_proc[pn].fd_recv);
461			}
462			close(p0[0]); close(p1[1]);
463			/* following call never returns */
464			ray_pchild(p1[0], p0[1]);
465			}
466	greg	2.3	if (r_proc[ray_pnprocs].pid < 0)
467	greg	2.1	error(SYSTEM, "cannot fork child process");
468			close(p1[0]); close(p0[1]);
469	greg	2.9	/*
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	greg	2.3	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	greg	2.1	}
481			}
482
483
484	schorsch	2.6	extern void
485			ray_pclose( /* close one or more child processes */
486			int nsub
487			)
488	greg	2.1	{
489			static int inclose = 0;
490			RAY res;
491			/* check recursion */
492			if (inclose)
493			return;
494			inclose++;
495			/* check argument */
496	schorsch	2.5	if ((nsub <= 0) \| (nsub > ray_pnprocs))
497	greg	2.3	nsub = ray_pnprocs;
498	greg	2.1	/* clear our ray queue */
499			while (ray_presult(&res,0) > 0)
500			;
501			/* clean up children */
502			while (nsub--) {
503			int status;
504	greg	2.3	ray_pnprocs--;
505			close(r_proc[ray_pnprocs].fd_recv);
506			close(r_proc[ray_pnprocs].fd_send);
507	greg	2.8	if (waitpid(r_proc[ray_pnprocs].pid, &status, 0) < 0)
508			status = 127<<8;
509	greg	2.1	if (status) {
510			sprintf(errmsg,
511			"rendering process %d exited with code %d",
512	greg	2.3	r_proc[ray_pnprocs].pid, status>>8);
513	greg	2.1	error(WARNING, errmsg);
514			}
515	greg	2.3	ray_pnidle--;
516	greg	2.1	}
517			inclose--;
518			}
519
520
521			void
522			quit(ec) /* make sure exit is called */
523			int ec;
524			{
525	greg	2.15	if (ray_pnprocs > 0) /* close children if any */
526			ray_pclose(0);
527	greg	2.1	exit(ec);
528			}