src/rt/raypcalls.c

#ifndef lint
static const char       RCSid[] = "$Id: raypcalls.c,v 2.39 2024/07/02 23:54:16 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 processor cores, 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 understand 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 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
 *  during evaluation.  Rays are not returned in the order
 *  they are sent when multiple processes are open.
 *
 *  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.  Setting the second argument
 *  to -1 returns 0 unless a ray is ready in the queue and
 *  no system calls are needed.  A negative return value
 *  indicates that a rendering process died.  If this
 *  happens, ray_pclose(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 mistake to call ra_psend() when
 *  ray_pnidle is zero, and nothing will be sent in
 *  this case.  Otherwise, the ray_presult() and/or ray_pqueue()
 *  functions may be called subsequently to read back the results
 *  of rays queued by ray_psend().
 *
 *  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 need to clean up memory or not, you should
 *  at least call ray_pclose(0) to await the child processes.
 *  The caller should define a quit() function that calls
 *  ray_pclose(0) if ray_pnprocs > 0.
 *
 *  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_pclose(), 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         24              /* # 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

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

static struct child_proc {
        RT_PID  pid;                            /* child process id */
        int     fd_send;                        /* write to child here */
        int     fd_recv;                        /* read from child here */
        int     npending;                       /* # rays in process */
        RNUMBER 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 = 0;        /* next send ray placement */
static int      r_recv_first = RAYQLEN; /* position of first unreported ray */
static int      r_recv_next = RAYQLEN;  /* next received ray placement */

static int      samplestep = 1;         /* sample step size */

#define sendq_full()    (r_send_next >= RAYQLEN)

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


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 */

        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, &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 */
}


int
ray_psend(                      /* add a ray to our send queue */
        RAY     *r
)
{
        int     rv;

        if ((r == NULL) | (ray_pnidle <= 0))
                return(0);
                                        /* flush output if necessary */
        if (sendq_full() && (rv = ray_pflush()) <= 0)
                return(rv);

        r_queue[r_send_next++] = *r;
        return(1);
}


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;

                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);
}


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;
        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);
        }
        if (poll < 0)                   /* immediate polling mode? */
                return(0);

        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, &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--; ) {
                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);
}


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

        cow_doneshare();                /* clear shared memory boundary */

        ray_done(freall);               /* free rendering data */
}


static void
ray_pchild(     /* process rays (never returns) */
        int     fd_in,
        int     fd_out
)
{
        int     n;
        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;
                        r_queue[i].rlvl = 0;
                        samplendx += samplestep;
                        rayclear(&r_queue[i]);
                        rayvalue(&r_queue[i]);
                }
                                        /* write back our results */
                i = writebuf(fd_out, 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 */
}


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;
        if (nobjects <= 0)
                error(CONSISTENCY, "ray_popen() called before scene loaded");
        ambsync();                      /* load any new ambient values */
        cow_memshare();                 /* copy-on-write shared memory */
        fflush(NULL);                   /* clear pending output */
        samplestep = ray_pnprocs + nadd;
        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]);
                        close(0);       /* don't share stdin */
                                        /* 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]);
                if (rand_samp)          /* decorrelate random sequence */
                        srandom(random());
                else
                        samplendx++;
                /*
                 * 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++;
        }
}


void
ray_pclose(             /* close one or more child processes */
        int     nsub
)
{
        static int      inclose = 0;
        RAY             res;
        int             i, status = 0;
                                        /* check no child / in child */
        if (ray_pnprocs <= 0)
                return;
                                        /* check recursion */
        if (inclose)
                return;
        inclose++;
                                        /* check argument */
        if ((nsub <= 0) | (nsub > ray_pnprocs))
                nsub = ray_pnprocs;
                                        /* clear our ray queue */
        i = r_send_next;
        r_send_next = 0;
        while (ray_presult(&res,0) > 0)
                ++i;
        if (i) {
                sprintf(errmsg, "dropped %d rays in ray_pclose()", i);
                error(WARNING, errmsg);
        }
        r_recv_first = r_recv_next = RAYQLEN;
                                        /* close send pipes */
        for (i = ray_pnprocs-nsub; i < ray_pnprocs; i++)
                close(r_proc[i].fd_send);

        if (nsub == 1) {                /* awaiting single process? */
                if (waitpid(r_proc[ray_pnprocs-1].pid, &status, 0) < 0)
                        status = 127<<8;
                close(r_proc[ray_pnprocs-1].fd_recv);
        } else                          /* else unordered wait */
                for (i = 0; i < nsub; ) {
                        int     j, mystatus;
                        RT_PID  pid = wait(&mystatus);
                        if (pid < 0) {
                                status = 127<<8;
                                break;
                        }
                        for (j = ray_pnprocs-nsub; j < ray_pnprocs; j++)
                                if (r_proc[j].pid == pid) {
                                        if (mystatus)
                                                status = mystatus;
                                        close(r_proc[j].fd_recv);
                                        ++i;
                                }
                }
        ray_pnprocs -= nsub;
        ray_pnidle -= nsub;
        if (status) {
                sprintf(errmsg, "rendering process exited with code %d", status>>8);
                error(WARNING, errmsg);
        }
        inclose--;
}
Revision:	2.40
Committed:	Wed Aug 21 20:42:20 2024 UTC (8 months, 1 week ago) by greg
Content type:	text/plain
Branch:	MAIN
CVS Tags:	HEAD
Changes since 2.39:	+3 -13 lines
Log Message:	refactor: moved memory sharing routines to preload.c with common call interface
#	User	Rev	Content
1	greg	2.1	#ifndef lint
2	greg	2.40	static const char RCSid[] = "$Id: raypcalls.c,v 2.39 2024/07/02 23:54:16 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	greg	2.21	* than you have processor cores, but one child may help by allowing
17	greg	2.1	* asynchronous ray computation in an interactive program, and
18			* will protect the caller from fatal rendering errors.
19			*
20	greg	2.21	* You should first read and understand the header in raycalls.c,
21	greg	2.1	* 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	greg	2.21	* that wouldn't be possible otherwise without causing any real
29	greg	2.1	* memory overhead, since all the static data are shared
30	greg	2.21	* between processes. Rays are traced using a simple
31	greg	2.1	* queuing mechanism, explained below.
32			*
33	greg	2.14	* The ray queue buffers RAYQLEN rays before sending to
34	greg	2.21	* 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	greg	2.14	*
38	greg	2.13	* Rays are queued and returned by a single
39	greg	2.1	* 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
42			* indicates that a ray was returned, though it is probably
43			* not the one you just requested. Rays may be identified by
44			* the rno member of the RAY struct, which is incremented
45			* by the rayorigin() call, or may be set explicitly by
46			* the caller. Below is an example call sequence:
47			*
48			* myRay.rorg = ( ray origin point )
49			* myRay.rdir = ( normalized ray direction )
50			* myRay.rmax = ( maximum length, or zero for no limit )
51	greg	2.11	* rayorigin(&myRay, PRIMARY, NULL, NULL);
52	greg	2.1	* myRay.rno = ( my personal ray identifier )
53			* if (ray_pqueue(&myRay) == 1)
54			* { do something with results }
55			*
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	greg	2.7	* yourself, which is normally called for you by ray_trace(). The
60			* benefit is that ray_pqueue() will trace rays faster in
61	greg	2.1	* 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.
64	greg	2.3	* The global int ray_pnidle indicates the number of currently idle
65	greg	2.1	* children. If you want to check for completed rays without blocking,
66			* or get the results from rays that have been queued without
67			* queuing any new ones, the ray_presult() call is for you:
68			*
69			* if (ray_presult(&myRay, 1) == 1)
70			* { do something with results }
71			*
72			* If the second argument is 1, the call won't block when
73			* results aren't ready, but will immediately return 0.
74			* If the second argument is 0, the call will block
75			* until a value is available, returning 0 only if the
76	greg	2.26	* queue is completely empty. Setting the second argument
77			* to -1 returns 0 unless a ray is ready in the queue and
78			* no system calls are needed. A negative return value
79	greg	2.1	* indicates that a rendering process died. If this
80	greg	2.21	* happens, ray_pclose(0) is automatically called to close
81	greg	2.3	* all child processes, and ray_pnprocs is set to zero.
82	greg	2.1	*
83			* If you just want to fill the ray queue without checking for
84	greg	2.3	* results, check ray_pnidle and call ray_psend():
85	greg	2.1	*
86	greg	2.3	* while (ray_pnidle) {
87	greg	2.1	* ( set up ray )
88			* ray_psend(&myRay);
89			* }
90			*
91	greg	2.25	* Note that it is a mistake to call ra_psend() when
92			* ray_pnidle is zero, and nothing will be sent in
93			* this case. Otherwise, the ray_presult() and/or ray_pqueue()
94			* functions may be called subsequently to read back the results
95			* of rays queued by ray_psend().
96	greg	2.1	*
97			* When you are done, you may call ray_pdone(1) to close
98			* all child processes and clean up memory used by Radiance.
99			* Any queued ray calculations will be awaited and discarded.
100			* As with ray_done(), ray_pdone(0) hangs onto data files
101			* and fonts that are likely to be used in subsequent renderings.
102	greg	2.21	* Whether you need to clean up memory or not, you should
103			* at least call ray_pclose(0) to await the child processes.
104	greg	2.23	* The caller should define a quit() function that calls
105			* ray_pclose(0) if ray_pnprocs > 0.
106	greg	2.1	*
107			* Warning: You cannot affect any of the rendering processes
108			* by changing global parameter values onece ray_pinit() has
109			* been called. Changing global parameters will have no effect
110			* until the next call to ray_pinit(), which restarts everything.
111			* If you just want to reap children so that you can alter the
112			* rendering parameters without reloading the scene, use the
113			* ray_pclose(0) and ray_popen(nproc) calls to close
114	greg	2.7	* then restart the child processes after the changes are made.
115	greg	2.1	*
116			* Note: These routines are written to coordinate with the
117			* definitions in raycalls.c, and in fact depend on them.
118			* If you want to trace a ray and get a result synchronously,
119	greg	2.13	* use the ray_trace() call to compute it in the parent process.
120	greg	2.7	* This will not interfere with any subprocess calculations,
121			* but beware that a fatal error may end with a call to quit().
122	greg	2.1	*
123			* Note: One of the advantages of using separate processes
124			* is that it gives the calling program some immunity from
125			* fatal rendering errors. As discussed in raycalls.c,
126			* Radiance tends to throw up its hands and exit at the
127			* first sign of trouble, calling quit() to return control
128	greg	2.7	* to the top level. Although you can avoid exit() with
129	greg	2.1	* your own longjmp() in quit(), the cleanup afterwards
130			* is always suspect. Through the use of subprocesses,
131			* we avoid this pitfall by closing the processes and
132			* returning a negative value from ray_pqueue() or
133			* ray_presult(). If you get a negative value from either
134			* of these calls, you can assume that the processes have
135	greg	2.21	* been cleaned up with a call to ray_pclose(), though you
136	greg	2.1	* will have to call ray_pdone() yourself if you want to
137	greg	2.7	* free memory. Obviously, you cannot continue rendering
138			* without risking further errors, but otherwise your
139			* process should not be compromised.
140	greg	2.1	*/
141
142	schorsch	2.6	#include "rtprocess.h"
143	greg	2.1	#include "ray.h"
144	schorsch	2.6	#include "ambient.h"
145	greg	2.18	#include <sys/types.h>
146			#include <sys/wait.h>
147	greg	2.1	#include "selcall.h"
148
149			#ifndef RAYQLEN
150	greg	2.32	#define RAYQLEN 24 /* # rays to send at once */
151	greg	2.1	#endif
152
153			#ifndef MAX_RPROCS
154			#if (FD_SETSIZE/2-4 < 64)
155			#define MAX_NPROCS (FD_SETSIZE/2-4)
156			#else
157			#define MAX_NPROCS 64 /* max. # rendering processes */
158			#endif
159			#endif
160
161	greg	2.3	int ray_pnprocs = 0; /* number of child processes */
162			int ray_pnidle = 0; /* number of idle children */
163	greg	2.1
164			static struct child_proc {
165	greg	2.29	RT_PID pid; /* child process id */
166	greg	2.1	int fd_send; /* write to child here */
167			int fd_recv; /* read from child here */
168			int npending; /* # rays in process */
169	greg	2.21	RNUMBER rno[RAYQLEN]; /* working on these rays */
170	greg	2.1	} r_proc[MAX_NPROCS]; /* our child processes */
171
172			static RAY r_queue[2RAYQLEN]; / ray i/o buffer */
173	greg	2.23	static int r_send_next = 0; /* next send ray placement */
174			static int r_recv_first = RAYQLEN; /* position of first unreported ray */
175			static int r_recv_next = RAYQLEN; /* next received ray placement */
176	greg	2.1
177	greg	2.28	static int samplestep = 1; /* sample step size */
178
179	greg	2.1	#define sendq_full() (r_send_next >= RAYQLEN)
180
181	schorsch	2.6	static int ray_pflush(void);
182	greg	2.13	static void ray_pchild(int fd_in, int fd_out);
183	greg	2.1
184	schorsch	2.6
185	greg	2.22	void
186	schorsch	2.6	ray_pinit( /* initialize ray-tracing processes */
187			char *otnm,
188			int nproc
189			)
190	greg	2.1	{
191			if (nobjects > 0) /* close old calculation */
192			ray_pdone(0);
193
194			ray_init(otnm); /* load the shared scene */
195
196			ray_popen(nproc); /* fork children */
197			}
198
199
200			static int
201	schorsch	2.6	ray_pflush(void) /* send queued rays to idle children */
202	greg	2.1	{
203			int nc, n, nw, i, sfirst;
204
205	schorsch	2.5	if ((ray_pnidle <= 0) \| (r_send_next <= 0))
206	greg	2.1	return(0); /* nothing we can send */
207
208			sfirst = 0; /* divvy up labor */
209	greg	2.3	nc = ray_pnidle;
210			for (i = ray_pnprocs; nc && i--; ) {
211	greg	2.1	if (r_proc[i].npending > 0)
212			continue; /* child looks busy */
213	greg	2.34	n = (r_send_next - sfirst) / nc--;
214	greg	2.1	if (!n)
215			continue;
216			/* smuggle set size in crtype */
217			r_queue[sfirst].crtype = n;
218	greg	2.39	nw = writebuf(r_proc[i].fd_send, &r_queue[sfirst],
219	greg	2.1	sizeof(RAY)*n);
220			if (nw != sizeof(RAY)*n)
221			return(-1); /* write error */
222			r_proc[i].npending = n;
223			while (n--) /* record ray IDs */
224			r_proc[i].rno[n] = r_queue[sfirst+n].rno;
225			sfirst += r_proc[i].npending;
226	greg	2.3	ray_pnidle--; /* now she's busy */
227	greg	2.1	}
228			if (sfirst != r_send_next)
229	greg	2.24	error(CONSISTENCY, "code screwup in ray_pflush()");
230	greg	2.1	r_send_next = 0;
231			return(sfirst); /* return total # sent */
232			}
233
234
235	greg	2.25	int
236	schorsch	2.6	ray_psend( /* add a ray to our send queue */
237			RAY *r
238			)
239	greg	2.1	{
240	greg	2.25	int rv;
241
242			if ((r == NULL) \| (ray_pnidle <= 0))
243			return(0);
244	greg	2.1	/* flush output if necessary */
245	greg	2.25	if (sendq_full() && (rv = ray_pflush()) <= 0)
246			return(rv);
247	greg	2.1
248	greg	2.14	r_queue[r_send_next++] = *r;
249	greg	2.25	return(1);
250	greg	2.1	}
251
252
253	greg	2.22	int
254	schorsch	2.6	ray_pqueue( /* queue a ray for computation */
255			RAY *r
256			)
257	greg	2.1	{
258			if (r == NULL)
259			return(0);
260			/* check for full send queue */
261			if (sendq_full()) {
262	greg	2.19	RAY mySend = *r;
263	greg	2.1	/* wait for a result */
264	greg	2.19	if (ray_presult(r, 0) <= 0)
265			return(-1);
266	greg	2.1	/* put new ray in queue */
267	greg	2.14	r_queue[r_send_next++] = mySend;
268	greg	2.25
269	greg	2.19	return(1);
270	greg	2.1	}
271	greg	2.13	/* else add ray to send queue */
272	greg	2.14	r_queue[r_send_next++] = *r;
273	greg	2.1	/* check for returned ray... */
274			if (r_recv_first >= r_recv_next)
275			return(0);
276			/* ...one is sitting in queue */
277	greg	2.14	*r = r_queue[r_recv_first++];
278	greg	2.1	return(1);
279			}
280
281
282	greg	2.22	int
283	schorsch	2.6	ray_presult( /* check for a completed ray */
284			RAY *r,
285			int poll
286			)
287	greg	2.1	{
288			static struct timeval tpoll; /* zero timeval struct */
289			static fd_set readset, errset;
290			int n, ok;
291	greg	2.30	int pn;
292	greg	2.1
293			if (r == NULL)
294			return(0);
295			/* check queued results first */
296			if (r_recv_first < r_recv_next) {
297	greg	2.14	*r = r_queue[r_recv_first++];
298	greg	2.1	return(1);
299			}
300	greg	2.23	if (poll < 0) /* immediate polling mode? */
301			return(0);
302
303	greg	2.3	n = ray_pnprocs - ray_pnidle; /* pending before flush? */
304	greg	2.1
305			if (ray_pflush() < 0) /* send new rays to process */
306			return(-1);
307			/* reset receive queue */
308			r_recv_first = r_recv_next = RAYQLEN;
309
310			if (!poll) /* count newly sent unless polling */
311	greg	2.3	n = ray_pnprocs - ray_pnidle;
312	greg	2.1	if (n <= 0) /* return if nothing to await */
313			return(0);
314	greg	2.16	if (!poll && ray_pnprocs == 1) /* one process -> skip select() */
315			FD_SET(r_proc[0].fd_recv, &readset);
316
317	greg	2.1	getready: /* any children waiting for us? */
318	greg	2.3	for (pn = ray_pnprocs; pn--; )
319	greg	2.1	if (FD_ISSET(r_proc[pn].fd_recv, &readset) \|\|
320			FD_ISSET(r_proc[pn].fd_recv, &errset))
321			break;
322	greg	2.22	/* call select() if we must */
323	greg	2.1	if (pn < 0) {
324			FD_ZERO(&readset); FD_ZERO(&errset); n = 0;
325	greg	2.3	for (pn = ray_pnprocs; pn--; ) {
326	greg	2.1	if (r_proc[pn].npending > 0)
327			FD_SET(r_proc[pn].fd_recv, &readset);
328			FD_SET(r_proc[pn].fd_recv, &errset);
329			if (r_proc[pn].fd_recv >= n)
330			n = r_proc[pn].fd_recv + 1;
331			}
332			/* find out who is ready */
333			while ((n = select(n, &readset, (fd_set *)NULL, &errset,
334			poll ? &tpoll : (struct timeval *)NULL)) < 0)
335			if (errno != EINTR) {
336			error(WARNING,
337	greg	2.24	"select call failed in ray_presult()");
338	greg	2.1	ray_pclose(0);
339			return(-1);
340			}
341			if (n > 0) /* go back and get it */
342			goto getready;
343			return(0); /* else poll came up empty */
344			}
345			if (r_recv_next + r_proc[pn].npending > sizeof(r_queue)/sizeof(RAY))
346			error(CONSISTENCY, "buffer shortage in ray_presult()");
347
348			/* read rendered ray data */
349	greg	2.39	n = readbuf(r_proc[pn].fd_recv, &r_queue[r_recv_next],
350	greg	2.1	sizeof(RAY)*r_proc[pn].npending);
351			if (n > 0) {
352			r_recv_next += n/sizeof(RAY);
353			ok = (n == sizeof(RAY)*r_proc[pn].npending);
354			} else
355			ok = 0;
356			/* reset child's status */
357			FD_CLR(r_proc[pn].fd_recv, &readset);
358			if (n <= 0)
359			FD_CLR(r_proc[pn].fd_recv, &errset);
360			r_proc[pn].npending = 0;
361	greg	2.3	ray_pnidle++;
362	greg	2.1	/* check for rendering errors */
363			if (!ok) {
364			ray_pclose(0); /* process died -- clean up */
365			return(-1);
366			}
367			/* preen returned rays */
368			for (n = r_recv_next - r_recv_first; n--; ) {
369	greg	2.30	RAY *rp = &r_queue[r_recv_first + n];
370	greg	2.1	rp->rno = r_proc[pn].rno[n];
371			rp->parent = NULL;
372			rp->newcset = rp->clipset = NULL;
373			rp->rox = NULL;
374			rp->slights = NULL;
375			}
376			/* return first ray received */
377	greg	2.13	*r = r_queue[r_recv_first++];
378	greg	2.1	return(1);
379			}
380
381
382	greg	2.22	void
383	schorsch	2.6	ray_pdone( /* reap children and free data */
384			int freall
385			)
386	greg	2.1	{
387			ray_pclose(0); /* close child processes */
388
389	greg	2.40	cow_doneshare(); /* clear shared memory boundary */
390	greg	2.23
391	greg	2.1	ray_done(freall); /* free rendering data */
392			}
393
394
395			static void
396	schorsch	2.6	ray_pchild( /* process rays (never returns) */
397			int fd_in,
398			int fd_out
399			)
400	greg	2.1	{
401			int n;
402	greg	2.30	int i;
403	greg	2.15	/* flag child process for quit() */
404			ray_pnprocs = -1;
405	greg	2.1	/* read each ray request set */
406			while ((n = read(fd_in, (char *)r_queue, sizeof(r_queue))) > 0) {
407			int n2;
408	greg	2.12	if (n < sizeof(RAY))
409	greg	2.1	break;
410			/* get smuggled set length */
411	greg	2.12	n2 = sizeof(RAY)*r_queue[0].crtype - n;
412	greg	2.1	if (n2 < 0)
413	greg	2.24	error(INTERNAL, "buffer over-read in ray_pchild()");
414	greg	2.1	if (n2 > 0) { /* read the rest of the set */
415	greg	2.12	i = readbuf(fd_in, (char *)r_queue + n, n2);
416			if (i != n2)
417	greg	2.1	break;
418			n += n2;
419			}
420	greg	2.12	n /= sizeof(RAY);
421	greg	2.1	/* evaluate rays */
422			for (i = 0; i < n; i++) {
423			r_queue[i].crtype = r_queue[i].rtype;
424			r_queue[i].parent = NULL;
425			r_queue[i].clipset = NULL;
426			r_queue[i].slights = NULL;
427	greg	2.21	r_queue[i].rlvl = 0;
428	greg	2.28	samplendx += samplestep;
429	greg	2.1	rayclear(&r_queue[i]);
430			rayvalue(&r_queue[i]);
431			}
432			/* write back our results */
433	greg	2.39	i = writebuf(fd_out, r_queue, sizeof(RAY)*n);
434	greg	2.1	if (i != sizeof(RAY)*n)
435	greg	2.24	error(SYSTEM, "write error in ray_pchild()");
436	greg	2.1	}
437			if (n)
438	greg	2.24	error(SYSTEM, "read error in ray_pchild()");
439	greg	2.1	ambsync();
440			quit(0); /* normal exit */
441			}
442
443
444	greg	2.22	void
445	schorsch	2.6	ray_popen( /* open the specified # processes */
446			int nadd
447			)
448	greg	2.1	{
449			/* check if our table has room */
450	greg	2.3	if (ray_pnprocs + nadd > MAX_NPROCS)
451			nadd = MAX_NPROCS - ray_pnprocs;
452	greg	2.1	if (nadd <= 0)
453			return;
454	greg	2.38	if (nobjects <= 0)
455			error(CONSISTENCY, "ray_popen() called before scene loaded");
456	greg	2.13	ambsync(); /* load any new ambient values */
457	greg	2.40	cow_memshare(); /* copy-on-write shared memory */
458	greg	2.13	fflush(NULL); /* clear pending output */
459	greg	2.28	samplestep = ray_pnprocs + nadd;
460	greg	2.1	while (nadd--) { /* fork each new process */
461			int p0[2], p1[2];
462			if (pipe(p0) < 0 \|\| pipe(p1) < 0)
463			error(SYSTEM, "cannot create pipe");
464	greg	2.3	if ((r_proc[ray_pnprocs].pid = fork()) == 0) {
465	greg	2.1	int pn; /* close others' descriptors */
466	greg	2.3	for (pn = ray_pnprocs; pn--; ) {
467	greg	2.1	close(r_proc[pn].fd_send);
468			close(r_proc[pn].fd_recv);
469			}
470			close(p0[0]); close(p1[1]);
471	greg	2.24	close(0); /* don't share stdin */
472	greg	2.1	/* following call never returns */
473			ray_pchild(p1[0], p0[1]);
474			}
475	greg	2.3	if (r_proc[ray_pnprocs].pid < 0)
476	greg	2.1	error(SYSTEM, "cannot fork child process");
477			close(p1[0]); close(p0[1]);
478	greg	2.28	if (rand_samp) /* decorrelate random sequence */
479			srandom(random());
480			else
481			samplendx++;
482	greg	2.9	/*
483			* Close write stream on exec to avoid multiprocessing deadlock.
484			* No use in read stream without it, so set flag there as well.
485			*/
486			fcntl(p1[1], F_SETFD, FD_CLOEXEC);
487			fcntl(p0[0], F_SETFD, FD_CLOEXEC);
488	greg	2.3	r_proc[ray_pnprocs].fd_send = p1[1];
489			r_proc[ray_pnprocs].fd_recv = p0[0];
490			r_proc[ray_pnprocs].npending = 0;
491			ray_pnprocs++;
492			ray_pnidle++;
493	greg	2.1	}
494			}
495
496
497	greg	2.22	void
498	schorsch	2.6	ray_pclose( /* close one or more child processes */
499			int nsub
500			)
501	greg	2.1	{
502			static int inclose = 0;
503	greg	2.29	RAY res;
504			int i, status = 0;
505	greg	2.35	/* check no child / in child */
506			if (ray_pnprocs <= 0)
507			return;
508	greg	2.1	/* check recursion */
509			if (inclose)
510			return;
511			inclose++;
512			/* check argument */
513	schorsch	2.5	if ((nsub <= 0) \| (nsub > ray_pnprocs))
514	greg	2.3	nsub = ray_pnprocs;
515	greg	2.37	/* clear our ray queue */
516			i = r_send_next;
517			r_send_next = 0;
518	greg	2.1	while (ray_presult(&res,0) > 0)
519	greg	2.36	++i;
520			if (i) {
521	greg	2.37	sprintf(errmsg, "dropped %d rays in ray_pclose()", i);
522	greg	2.36	error(WARNING, errmsg);
523			}
524	greg	2.23	r_recv_first = r_recv_next = RAYQLEN;
525	greg	2.29	/* close send pipes */
526			for (i = ray_pnprocs-nsub; i < ray_pnprocs; i++)
527			close(r_proc[i].fd_send);
528
529			if (nsub == 1) { /* awaiting single process? */
530			if (waitpid(r_proc[ray_pnprocs-1].pid, &status, 0) < 0)
531	greg	2.8	status = 127<<8;
532	greg	2.29	close(r_proc[ray_pnprocs-1].fd_recv);
533			} else /* else unordered wait */
534			for (i = 0; i < nsub; ) {
535			int j, mystatus;
536			RT_PID pid = wait(&mystatus);
537	greg	2.37	if (pid < 0) {
538			status = 127<<8;
539			break;
540			}
541	greg	2.29	for (j = ray_pnprocs-nsub; j < ray_pnprocs; j++)
542			if (r_proc[j].pid == pid) {
543			if (mystatus)
544			status = mystatus;
545			close(r_proc[j].fd_recv);
546			++i;
547			}
548	greg	2.1	}
549	greg	2.29	ray_pnprocs -= nsub;
550			ray_pnidle -= nsub;
551			if (status) {
552			sprintf(errmsg, "rendering process exited with code %d", status>>8);
553			error(WARNING, errmsg);
554	greg	2.1	}
555			inclose--;
556			}