src/rt/raypcalls.c

#ifndef lint
static const char       RCSid[] = "$Id: raypcalls.c,v 2.15 2007/09/12 03:57:00 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 <stdio.h>
#include <sys/types.h>
#include <sys/wait.h> /* XXX platform */

#include  "rtprocess.h"
#include  "ray.h"
#include  "ambient.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 */

        preload_objs();                 /* preload auxiliary data */

                                        /* set shared memory boundary */
        shm_boundary = (char *)malloc(16);
        strcpy(shm_boundary, "SHM_BOUNDARY");

        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;
                int     rval;
                mySend = *r;
                                        /* wait for a result */
                rval = ray_presult(r, 0);
                                        /* put new ray in queue */
                r_queue[r_send_next++] = mySend;
                return(rval);           /* done */
        }
                                        /* 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 */
        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:	2.16
Committed:	Sat Sep 15 02:47:56 2007 UTC (16 years, 7 months ago) by greg
Content type:	text/plain
Branch:	MAIN
Changes since 2.15:	+4 -1 lines
Log Message:	Eliminated unnecessary call to select(2) for single process
#	User	Rev	Content
1	greg	2.1	#ifndef lint
2	greg	2.16	static const char RCSid[] = "$Id: raypcalls.c,v 2.15 2007/09/12 03:57:00 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 <stdio.h>
135			#include <sys/types.h>
136			#include <sys/wait.h> /* XXX platform */
137
138			#include "rtprocess.h"
139	greg	2.1	#include "ray.h"
140	schorsch	2.6	#include "ambient.h"
141	greg	2.1	#include "selcall.h"
142
143			#ifndef RAYQLEN
144	greg	2.13	#define RAYQLEN 12 /* # rays to send at once */
145	greg	2.1	#endif
146
147			#ifndef MAX_RPROCS
148			#if (FD_SETSIZE/2-4 < 64)
149			#define MAX_NPROCS (FD_SETSIZE/2-4)
150			#else
151			#define MAX_NPROCS 64 /* max. # rendering processes */
152			#endif
153			#endif
154
155			extern char shm_boundary; / boundary of shared memory */
156
157	greg	2.3	int ray_pnprocs = 0; /* number of child processes */
158			int ray_pnidle = 0; /* number of idle children */
159	greg	2.1
160			static struct child_proc {
161			int pid; /* child process id */
162			int fd_send; /* write to child here */
163			int fd_recv; /* read from child here */
164			int npending; /* # rays in process */
165			unsigned long rno[RAYQLEN]; /* working on these rays */
166			} r_proc[MAX_NPROCS]; /* our child processes */
167
168			static RAY r_queue[2RAYQLEN]; / ray i/o buffer */
169			static int r_send_next; /* next send ray placement */
170			static int r_recv_first; /* position of first unreported ray */
171			static int r_recv_next; /* next receive ray placement */
172
173			#define sendq_full() (r_send_next >= RAYQLEN)
174
175	schorsch	2.6	static int ray_pflush(void);
176	greg	2.13	static void ray_pchild(int fd_in, int fd_out);
177	greg	2.1
178	schorsch	2.6
179			extern void
180			ray_pinit( /* initialize ray-tracing processes */
181			char *otnm,
182			int nproc
183			)
184	greg	2.1	{
185			if (nobjects > 0) /* close old calculation */
186			ray_pdone(0);
187
188			ray_init(otnm); /* load the shared scene */
189
190			preload_objs(); /* preload auxiliary data */
191
192			/* set shared memory boundary */
193			shm_boundary = (char *)malloc(16);
194			strcpy(shm_boundary, "SHM_BOUNDARY");
195
196			r_send_next = 0; /* set up queue */
197			r_recv_first = r_recv_next = RAYQLEN;
198
199			ray_popen(nproc); /* fork children */
200			}
201
202
203			static int
204	schorsch	2.6	ray_pflush(void) /* send queued rays to idle children */
205	greg	2.1	{
206			int nc, n, nw, i, sfirst;
207
208	schorsch	2.5	if ((ray_pnidle <= 0) \| (r_send_next <= 0))
209	greg	2.1	return(0); /* nothing we can send */
210
211			sfirst = 0; /* divvy up labor */
212	greg	2.3	nc = ray_pnidle;
213			for (i = ray_pnprocs; nc && i--; ) {
214	greg	2.1	if (r_proc[i].npending > 0)
215			continue; /* child looks busy */
216			n = (r_send_next - sfirst)/nc--;
217			if (!n)
218			continue;
219			/* smuggle set size in crtype */
220			r_queue[sfirst].crtype = n;
221			nw = writebuf(r_proc[i].fd_send, (char *)&r_queue[sfirst],
222			sizeof(RAY)*n);
223			if (nw != sizeof(RAY)*n)
224			return(-1); /* write error */
225			r_proc[i].npending = n;
226			while (n--) /* record ray IDs */
227			r_proc[i].rno[n] = r_queue[sfirst+n].rno;
228			sfirst += r_proc[i].npending;
229	greg	2.3	ray_pnidle--; /* now she's busy */
230	greg	2.1	}
231			if (sfirst != r_send_next)
232			error(CONSISTENCY, "code screwup in ray_pflush");
233			r_send_next = 0;
234			return(sfirst); /* return total # sent */
235			}
236
237
238	schorsch	2.6	extern void
239			ray_psend( /* add a ray to our send queue */
240			RAY *r
241			)
242	greg	2.1	{
243			if (r == NULL)
244			return;
245			/* flush output if necessary */
246			if (sendq_full() && ray_pflush() <= 0)
247			error(INTERNAL, "ray_pflush failed in ray_psend");
248
249	greg	2.14	r_queue[r_send_next++] = *r;
250	greg	2.1	}
251
252
253	schorsch	2.6	extern int
254			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			RAY mySend;
263			int rval;
264	schorsch	2.4	mySend = *r;
265	greg	2.1	/* wait for a result */
266			rval = ray_presult(r, 0);
267			/* put new ray in queue */
268	greg	2.14	r_queue[r_send_next++] = mySend;
269	greg	2.1	return(rval); /* done */
270			}
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	schorsch	2.6	extern int
283			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			register int pn;
292
293			if (r == NULL)
294			return(0);
295			/* check queued results first */
296			if (r_recv_first < r_recv_next) {
297	greg	2.14	*r = r_queue[r_recv_first++];
298	greg	2.1	return(1);
299			}
300	greg	2.3	n = ray_pnprocs - ray_pnidle; /* pending before flush? */
301	greg	2.1
302			if (ray_pflush() < 0) /* send new rays to process */
303			return(-1);
304			/* reset receive queue */
305			r_recv_first = r_recv_next = RAYQLEN;
306
307			if (!poll) /* count newly sent unless polling */
308	greg	2.3	n = ray_pnprocs - ray_pnidle;
309	greg	2.1	if (n <= 0) /* return if nothing to await */
310			return(0);
311	greg	2.16	if (!poll && ray_pnprocs == 1) /* one process -> skip select() */
312			FD_SET(r_proc[0].fd_recv, &readset);
313
314	greg	2.1	getready: /* any children waiting for us? */
315	greg	2.3	for (pn = ray_pnprocs; pn--; )
316	greg	2.1	if (FD_ISSET(r_proc[pn].fd_recv, &readset) \|\|
317			FD_ISSET(r_proc[pn].fd_recv, &errset))
318			break;
319			/* call select if we must */
320			if (pn < 0) {
321			FD_ZERO(&readset); FD_ZERO(&errset); n = 0;
322	greg	2.3	for (pn = ray_pnprocs; pn--; ) {
323	greg	2.1	if (r_proc[pn].npending > 0)
324			FD_SET(r_proc[pn].fd_recv, &readset);
325			FD_SET(r_proc[pn].fd_recv, &errset);
326			if (r_proc[pn].fd_recv >= n)
327			n = r_proc[pn].fd_recv + 1;
328			}
329			/* find out who is ready */
330			while ((n = select(n, &readset, (fd_set *)NULL, &errset,
331			poll ? &tpoll : (struct timeval *)NULL)) < 0)
332			if (errno != EINTR) {
333			error(WARNING,
334			"select call failed in ray_presult");
335			ray_pclose(0);
336			return(-1);
337			}
338			if (n > 0) /* go back and get it */
339			goto getready;
340			return(0); /* else poll came up empty */
341			}
342			if (r_recv_next + r_proc[pn].npending > sizeof(r_queue)/sizeof(RAY))
343			error(CONSISTENCY, "buffer shortage in ray_presult()");
344
345			/* read rendered ray data */
346			n = readbuf(r_proc[pn].fd_recv, (char *)&r_queue[r_recv_next],
347			sizeof(RAY)*r_proc[pn].npending);
348			if (n > 0) {
349			r_recv_next += n/sizeof(RAY);
350			ok = (n == sizeof(RAY)*r_proc[pn].npending);
351			} else
352			ok = 0;
353			/* reset child's status */
354			FD_CLR(r_proc[pn].fd_recv, &readset);
355			if (n <= 0)
356			FD_CLR(r_proc[pn].fd_recv, &errset);
357			r_proc[pn].npending = 0;
358	greg	2.3	ray_pnidle++;
359	greg	2.1	/* check for rendering errors */
360			if (!ok) {
361			ray_pclose(0); /* process died -- clean up */
362			return(-1);
363			}
364			/* preen returned rays */
365			for (n = r_recv_next - r_recv_first; n--; ) {
366			register RAY *rp = &r_queue[r_recv_first + n];
367			rp->rno = r_proc[pn].rno[n];
368			rp->parent = NULL;
369			rp->newcset = rp->clipset = NULL;
370			rp->rox = NULL;
371			rp->slights = NULL;
372			}
373			/* return first ray received */
374	greg	2.13	*r = r_queue[r_recv_first++];
375	greg	2.1	return(1);
376			}
377
378
379	schorsch	2.6	extern void
380			ray_pdone( /* reap children and free data */
381			int freall
382			)
383	greg	2.1	{
384			ray_pclose(0); /* close child processes */
385
386			if (shm_boundary != NULL) { /* clear shared memory boundary */
387			free((void *)shm_boundary);
388			shm_boundary = NULL;
389			}
390			ray_done(freall); /* free rendering data */
391			}
392
393
394			static void
395	schorsch	2.6	ray_pchild( /* process rays (never returns) */
396			int fd_in,
397			int fd_out
398			)
399	greg	2.1	{
400			int n;
401			register int i;
402	greg	2.15	/* flag child process for quit() */
403			ray_pnprocs = -1;
404	greg	2.1	/* read each ray request set */
405			while ((n = read(fd_in, (char *)r_queue, sizeof(r_queue))) > 0) {
406			int n2;
407	greg	2.12	if (n < sizeof(RAY))
408	greg	2.1	break;
409			/* get smuggled set length */
410	greg	2.12	n2 = sizeof(RAY)*r_queue[0].crtype - n;
411	greg	2.1	if (n2 < 0)
412			error(INTERNAL, "buffer over-read in ray_pchild");
413			if (n2 > 0) { /* read the rest of the set */
414	greg	2.12	i = readbuf(fd_in, (char *)r_queue + n, n2);
415			if (i != n2)
416	greg	2.1	break;
417			n += n2;
418			}
419	greg	2.12	n /= sizeof(RAY);
420	greg	2.1	/* evaluate rays */
421			for (i = 0; i < n; i++) {
422			r_queue[i].crtype = r_queue[i].rtype;
423			r_queue[i].parent = NULL;
424			r_queue[i].clipset = NULL;
425			r_queue[i].slights = NULL;
426			samplendx++;
427			rayclear(&r_queue[i]);
428			rayvalue(&r_queue[i]);
429			}
430			/* write back our results */
431			i = writebuf(fd_out, (char )r_queue, sizeof(RAY)n);
432			if (i != sizeof(RAY)*n)
433			error(SYSTEM, "write error in ray_pchild");
434			}
435			if (n)
436			error(SYSTEM, "read error in ray_pchild");
437			ambsync();
438			quit(0); /* normal exit */
439			}
440
441
442	schorsch	2.6	extern void
443			ray_popen( /* open the specified # processes */
444			int nadd
445			)
446	greg	2.1	{
447			/* check if our table has room */
448	greg	2.3	if (ray_pnprocs + nadd > MAX_NPROCS)
449			nadd = MAX_NPROCS - ray_pnprocs;
450	greg	2.1	if (nadd <= 0)
451			return;
452	greg	2.13	ambsync(); /* load any new ambient values */
453			fflush(NULL); /* clear pending output */
454	greg	2.1	while (nadd--) { /* fork each new process */
455			int p0[2], p1[2];
456			if (pipe(p0) < 0 \|\| pipe(p1) < 0)
457			error(SYSTEM, "cannot create pipe");
458	greg	2.3	if ((r_proc[ray_pnprocs].pid = fork()) == 0) {
459	greg	2.1	int pn; /* close others' descriptors */
460	greg	2.3	for (pn = ray_pnprocs; pn--; ) {
461	greg	2.1	close(r_proc[pn].fd_send);
462			close(r_proc[pn].fd_recv);
463			}
464			close(p0[0]); close(p1[1]);
465			/* following call never returns */
466			ray_pchild(p1[0], p0[1]);
467			}
468	greg	2.3	if (r_proc[ray_pnprocs].pid < 0)
469	greg	2.1	error(SYSTEM, "cannot fork child process");
470			close(p1[0]); close(p0[1]);
471	greg	2.9	/*
472			* Close write stream on exec to avoid multiprocessing deadlock.
473			* No use in read stream without it, so set flag there as well.
474			*/
475			fcntl(p1[1], F_SETFD, FD_CLOEXEC);
476			fcntl(p0[0], F_SETFD, FD_CLOEXEC);
477	greg	2.3	r_proc[ray_pnprocs].fd_send = p1[1];
478			r_proc[ray_pnprocs].fd_recv = p0[0];
479			r_proc[ray_pnprocs].npending = 0;
480			ray_pnprocs++;
481			ray_pnidle++;
482	greg	2.1	}
483			}
484
485
486	schorsch	2.6	extern void
487			ray_pclose( /* close one or more child processes */
488			int nsub
489			)
490	greg	2.1	{
491			static int inclose = 0;
492			RAY res;
493			/* check recursion */
494			if (inclose)
495			return;
496			inclose++;
497			/* check argument */
498	schorsch	2.5	if ((nsub <= 0) \| (nsub > ray_pnprocs))
499	greg	2.3	nsub = ray_pnprocs;
500	greg	2.1	/* clear our ray queue */
501			while (ray_presult(&res,0) > 0)
502			;
503			/* clean up children */
504			while (nsub--) {
505			int status;
506	greg	2.3	ray_pnprocs--;
507			close(r_proc[ray_pnprocs].fd_recv);
508			close(r_proc[ray_pnprocs].fd_send);
509	greg	2.8	if (waitpid(r_proc[ray_pnprocs].pid, &status, 0) < 0)
510			status = 127<<8;
511	greg	2.1	if (status) {
512			sprintf(errmsg,
513			"rendering process %d exited with code %d",
514	greg	2.3	r_proc[ray_pnprocs].pid, status>>8);
515	greg	2.1	error(WARNING, errmsg);
516			}
517	greg	2.3	ray_pnidle--;
518	greg	2.1	}
519			inclose--;
520			}
521
522
523			void
524			quit(ec) /* make sure exit is called */
525			int ec;
526			{
527	greg	2.15	if (ray_pnprocs > 0) /* close children if any */
528			ray_pclose(0);
529	greg	2.1	exit(ec);
530			}