src/rt/raypcalls.c

#ifndef lint
static const char       RCSid[] = "$Id: raypcalls.c,v 2.23 2009/12/12 19:01:00 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 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.
 *  (A special value of -1 returns 0 unless a ray is
 *  ready in the queue and no system calls are needed.)
 *  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_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 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 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         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 */
        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 */

#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, (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 */
}


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;
}


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);
}


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);
        }
        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, (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);
}


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;
                        r_queue[i].rlvl = 0;
                        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 */
}


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]);
                        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]);
                /*
                 * 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;
                                        /* 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)
                ;
        r_send_next = 0;                /* hard reset in case of error */
        r_recv_first = r_recv_next = RAYQLEN;
                                        /* clean up children */
        while (nsub--) {
                int     status;
                ray_pnprocs--;
                close(r_proc[ray_pnprocs].fd_send);
                if (waitpid(r_proc[ray_pnprocs].pid, &status, 0) < 0)
                        status = 127<<8;
                close(r_proc[ray_pnprocs].fd_recv);
                if (status) {
                        sprintf(errmsg,
                                "rendering process %d exited with code %d",
                                        r_proc[ray_pnprocs].pid, status>>8);
                        error(WARNING, errmsg);
                }
                ray_pnidle--;
        }
        inclose--;
}
Revision:	2.24
Committed:	Sat Dec 12 23:08:13 2009 UTC (14 years, 4 months ago) by greg
Content type:	text/plain
Branch:	MAIN
Changes since 2.23:	+9 -8 lines
Log Message:	Bug fixes and performance improvements to rtrace -n option
#	User	Rev	Content
1	greg	2.1	#ifndef lint
2	greg	2.24	static const char RCSid[] = "$Id: raypcalls.c,v 2.23 2009/12/12 19:01:00 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	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	greg	2.23	* (A special value of -1 returns 0 unless a ray is
75			* ready in the queue and no system calls are needed.)
76	greg	2.1	* If the second argument is 0, the call will block
77			* until a value is available, returning 0 only if the
78			* queue is completely empty. A negative return value
79			* 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	2.7	* Note that it is a fatal error to call ra_psend() when
92			* ray_pnidle is zero. The ray_presult() and/or ray_pqueue()
93			* functions may be called subsequently to read back the results.
94	greg	2.1	*
95			* When you are done, you may call ray_pdone(1) to close
96			* all child processes and clean up memory used by Radiance.
97			* Any queued ray calculations will be awaited and discarded.
98			* As with ray_done(), ray_pdone(0) hangs onto data files
99			* and fonts that are likely to be used in subsequent renderings.
100	greg	2.21	* Whether you need to clean up memory or not, you should
101			* at least call ray_pclose(0) to await the child processes.
102	greg	2.23	* The caller should define a quit() function that calls
103			* ray_pclose(0) if ray_pnprocs > 0.
104	greg	2.1	*
105			* Warning: You cannot affect any of the rendering processes
106			* by changing global parameter values onece ray_pinit() has
107			* been called. Changing global parameters will have no effect
108			* until the next call to ray_pinit(), which restarts everything.
109			* If you just want to reap children so that you can alter the
110			* rendering parameters without reloading the scene, use the
111			* ray_pclose(0) and ray_popen(nproc) calls to close
112	greg	2.7	* then restart the child processes after the changes are made.
113	greg	2.1	*
114			* Note: These routines are written to coordinate with the
115			* definitions in raycalls.c, and in fact depend on them.
116			* If you want to trace a ray and get a result synchronously,
117	greg	2.13	* use the ray_trace() call to compute it in the parent process.
118	greg	2.7	* This will not interfere with any subprocess calculations,
119			* but beware that a fatal error may end with a call to quit().
120	greg	2.1	*
121			* Note: One of the advantages of using separate processes
122			* is that it gives the calling program some immunity from
123			* fatal rendering errors. As discussed in raycalls.c,
124			* Radiance tends to throw up its hands and exit at the
125			* first sign of trouble, calling quit() to return control
126	greg	2.7	* to the top level. Although you can avoid exit() with
127	greg	2.1	* your own longjmp() in quit(), the cleanup afterwards
128			* is always suspect. Through the use of subprocesses,
129			* we avoid this pitfall by closing the processes and
130			* returning a negative value from ray_pqueue() or
131			* ray_presult(). If you get a negative value from either
132			* of these calls, you can assume that the processes have
133	greg	2.21	* been cleaned up with a call to ray_pclose(), though you
134	greg	2.1	* will have to call ray_pdone() yourself if you want to
135	greg	2.7	* free memory. Obviously, you cannot continue rendering
136			* without risking further errors, but otherwise your
137			* process should not be compromised.
138	greg	2.1	*/
139
140	schorsch	2.6	#include "rtprocess.h"
141	greg	2.1	#include "ray.h"
142	schorsch	2.6	#include "ambient.h"
143	greg	2.18	#include <sys/types.h>
144			#include <sys/wait.h>
145	greg	2.1	#include "selcall.h"
146
147			#ifndef RAYQLEN
148	greg	2.13	#define RAYQLEN 12 /* # rays to send at once */
149	greg	2.1	#endif
150
151			#ifndef MAX_RPROCS
152			#if (FD_SETSIZE/2-4 < 64)
153			#define MAX_NPROCS (FD_SETSIZE/2-4)
154			#else
155			#define MAX_NPROCS 64 /* max. # rendering processes */
156			#endif
157			#endif
158
159			extern char shm_boundary; / boundary of shared memory */
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			int pid; /* child process id */
166			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			#define sendq_full() (r_send_next >= RAYQLEN)
178
179	schorsch	2.6	static int ray_pflush(void);
180	greg	2.13	static void ray_pchild(int fd_in, int fd_out);
181	greg	2.1
182	schorsch	2.6
183	greg	2.22	void
184	schorsch	2.6	ray_pinit( /* initialize ray-tracing processes */
185			char *otnm,
186			int nproc
187			)
188	greg	2.1	{
189			if (nobjects > 0) /* close old calculation */
190			ray_pdone(0);
191
192			ray_init(otnm); /* load the shared scene */
193
194			ray_popen(nproc); /* fork children */
195			}
196
197
198			static int
199	schorsch	2.6	ray_pflush(void) /* send queued rays to idle children */
200	greg	2.1	{
201			int nc, n, nw, i, sfirst;
202
203	schorsch	2.5	if ((ray_pnidle <= 0) \| (r_send_next <= 0))
204	greg	2.1	return(0); /* nothing we can send */
205
206			sfirst = 0; /* divvy up labor */
207	greg	2.3	nc = ray_pnidle;
208			for (i = ray_pnprocs; nc && i--; ) {
209	greg	2.1	if (r_proc[i].npending > 0)
210			continue; /* child looks busy */
211			n = (r_send_next - sfirst)/nc--;
212			if (!n)
213			continue;
214			/* smuggle set size in crtype */
215			r_queue[sfirst].crtype = n;
216			nw = writebuf(r_proc[i].fd_send, (char *)&r_queue[sfirst],
217			sizeof(RAY)*n);
218			if (nw != sizeof(RAY)*n)
219			return(-1); /* write error */
220			r_proc[i].npending = n;
221			while (n--) /* record ray IDs */
222			r_proc[i].rno[n] = r_queue[sfirst+n].rno;
223			sfirst += r_proc[i].npending;
224	greg	2.3	ray_pnidle--; /* now she's busy */
225	greg	2.1	}
226			if (sfirst != r_send_next)
227	greg	2.24	error(CONSISTENCY, "code screwup in ray_pflush()");
228	greg	2.1	r_send_next = 0;
229			return(sfirst); /* return total # sent */
230			}
231
232
233	greg	2.22	void
234	schorsch	2.6	ray_psend( /* add a ray to our send queue */
235			RAY *r
236			)
237	greg	2.1	{
238			if (r == NULL)
239			return;
240			/* flush output if necessary */
241			if (sendq_full() && ray_pflush() <= 0)
242	greg	2.24	error(INTERNAL, "ray_pflush failed in ray_psend()");
243	greg	2.1
244	greg	2.14	r_queue[r_send_next++] = *r;
245	greg	2.1	}
246
247
248	greg	2.22	int
249	schorsch	2.6	ray_pqueue( /* queue a ray for computation */
250			RAY *r
251			)
252	greg	2.1	{
253			if (r == NULL)
254			return(0);
255			/* check for full send queue */
256			if (sendq_full()) {
257	greg	2.19	RAY mySend = *r;
258	greg	2.1	/* wait for a result */
259	greg	2.19	if (ray_presult(r, 0) <= 0)
260			return(-1);
261	greg	2.1	/* put new ray in queue */
262	greg	2.14	r_queue[r_send_next++] = mySend;
263	greg	2.19	/* XXX r_send_next may now be > RAYQLEN */
264			return(1);
265	greg	2.1	}
266	greg	2.13	/* else add ray to send queue */
267	greg	2.14	r_queue[r_send_next++] = *r;
268	greg	2.1	/* check for returned ray... */
269			if (r_recv_first >= r_recv_next)
270			return(0);
271			/* ...one is sitting in queue */
272	greg	2.14	*r = r_queue[r_recv_first++];
273	greg	2.1	return(1);
274			}
275
276
277	greg	2.22	int
278	schorsch	2.6	ray_presult( /* check for a completed ray */
279			RAY *r,
280			int poll
281			)
282	greg	2.1	{
283			static struct timeval tpoll; /* zero timeval struct */
284			static fd_set readset, errset;
285			int n, ok;
286			register int pn;
287
288			if (r == NULL)
289			return(0);
290			/* check queued results first */
291			if (r_recv_first < r_recv_next) {
292	greg	2.14	*r = r_queue[r_recv_first++];
293	greg	2.1	return(1);
294			}
295	greg	2.23	if (poll < 0) /* immediate polling mode? */
296			return(0);
297
298	greg	2.3	n = ray_pnprocs - ray_pnidle; /* pending before flush? */
299	greg	2.1
300			if (ray_pflush() < 0) /* send new rays to process */
301			return(-1);
302			/* reset receive queue */
303			r_recv_first = r_recv_next = RAYQLEN;
304
305			if (!poll) /* count newly sent unless polling */
306	greg	2.3	n = ray_pnprocs - ray_pnidle;
307	greg	2.1	if (n <= 0) /* return if nothing to await */
308			return(0);
309	greg	2.16	if (!poll && ray_pnprocs == 1) /* one process -> skip select() */
310			FD_SET(r_proc[0].fd_recv, &readset);
311
312	greg	2.1	getready: /* any children waiting for us? */
313	greg	2.3	for (pn = ray_pnprocs; pn--; )
314	greg	2.1	if (FD_ISSET(r_proc[pn].fd_recv, &readset) \|\|
315			FD_ISSET(r_proc[pn].fd_recv, &errset))
316			break;
317	greg	2.22	/* call select() if we must */
318	greg	2.1	if (pn < 0) {
319			FD_ZERO(&readset); FD_ZERO(&errset); n = 0;
320	greg	2.3	for (pn = ray_pnprocs; pn--; ) {
321	greg	2.1	if (r_proc[pn].npending > 0)
322			FD_SET(r_proc[pn].fd_recv, &readset);
323			FD_SET(r_proc[pn].fd_recv, &errset);
324			if (r_proc[pn].fd_recv >= n)
325			n = r_proc[pn].fd_recv + 1;
326			}
327			/* find out who is ready */
328			while ((n = select(n, &readset, (fd_set *)NULL, &errset,
329			poll ? &tpoll : (struct timeval *)NULL)) < 0)
330			if (errno != EINTR) {
331			error(WARNING,
332	greg	2.24	"select call failed in ray_presult()");
333	greg	2.1	ray_pclose(0);
334			return(-1);
335			}
336			if (n > 0) /* go back and get it */
337			goto getready;
338			return(0); /* else poll came up empty */
339			}
340			if (r_recv_next + r_proc[pn].npending > sizeof(r_queue)/sizeof(RAY))
341			error(CONSISTENCY, "buffer shortage in ray_presult()");
342
343			/* read rendered ray data */
344			n = readbuf(r_proc[pn].fd_recv, (char *)&r_queue[r_recv_next],
345			sizeof(RAY)*r_proc[pn].npending);
346			if (n > 0) {
347			r_recv_next += n/sizeof(RAY);
348			ok = (n == sizeof(RAY)*r_proc[pn].npending);
349			} else
350			ok = 0;
351			/* reset child's status */
352			FD_CLR(r_proc[pn].fd_recv, &readset);
353			if (n <= 0)
354			FD_CLR(r_proc[pn].fd_recv, &errset);
355			r_proc[pn].npending = 0;
356	greg	2.3	ray_pnidle++;
357	greg	2.1	/* check for rendering errors */
358			if (!ok) {
359			ray_pclose(0); /* process died -- clean up */
360			return(-1);
361			}
362			/* preen returned rays */
363			for (n = r_recv_next - r_recv_first; n--; ) {
364			register RAY *rp = &r_queue[r_recv_first + n];
365			rp->rno = r_proc[pn].rno[n];
366			rp->parent = NULL;
367			rp->newcset = rp->clipset = NULL;
368			rp->rox = NULL;
369			rp->slights = NULL;
370			}
371			/* return first ray received */
372	greg	2.13	*r = r_queue[r_recv_first++];
373	greg	2.1	return(1);
374			}
375
376
377	greg	2.22	void
378	schorsch	2.6	ray_pdone( /* reap children and free data */
379			int freall
380			)
381	greg	2.1	{
382			ray_pclose(0); /* close child processes */
383
384			if (shm_boundary != NULL) { /* clear shared memory boundary */
385			free((void *)shm_boundary);
386			shm_boundary = NULL;
387			}
388	greg	2.23
389	greg	2.1	ray_done(freall); /* free rendering data */
390			}
391
392
393			static void
394	schorsch	2.6	ray_pchild( /* process rays (never returns) */
395			int fd_in,
396			int fd_out
397			)
398	greg	2.1	{
399			int n;
400			register int i;
401	greg	2.15	/* flag child process for quit() */
402			ray_pnprocs = -1;
403	greg	2.1	/* read each ray request set */
404			while ((n = read(fd_in, (char *)r_queue, sizeof(r_queue))) > 0) {
405			int n2;
406	greg	2.12	if (n < sizeof(RAY))
407	greg	2.1	break;
408			/* get smuggled set length */
409	greg	2.12	n2 = sizeof(RAY)*r_queue[0].crtype - n;
410	greg	2.1	if (n2 < 0)
411	greg	2.24	error(INTERNAL, "buffer over-read in ray_pchild()");
412	greg	2.1	if (n2 > 0) { /* read the rest of the set */
413	greg	2.12	i = readbuf(fd_in, (char *)r_queue + n, n2);
414			if (i != n2)
415	greg	2.1	break;
416			n += n2;
417			}
418	greg	2.12	n /= sizeof(RAY);
419	greg	2.1	/* evaluate rays */
420			for (i = 0; i < n; i++) {
421			r_queue[i].crtype = r_queue[i].rtype;
422			r_queue[i].parent = NULL;
423			r_queue[i].clipset = NULL;
424			r_queue[i].slights = NULL;
425	greg	2.21	r_queue[i].rlvl = 0;
426	greg	2.1	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	greg	2.24	error(SYSTEM, "write error in ray_pchild()");
434	greg	2.1	}
435			if (n)
436	greg	2.24	error(SYSTEM, "read error in ray_pchild()");
437	greg	2.1	ambsync();
438			quit(0); /* normal exit */
439			}
440
441
442	greg	2.22	void
443	schorsch	2.6	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	greg	2.20	if (shm_boundary == NULL) { /* first child process? */
454			preload_objs(); /* preload auxiliary data */
455			/* set shared memory boundary */
456			shm_boundary = (char *)malloc(16);
457			strcpy(shm_boundary, "SHM_BOUNDARY");
458			}
459	greg	2.13	fflush(NULL); /* clear pending output */
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	2.9	/*
479			* Close write stream on exec to avoid multiprocessing deadlock.
480			* No use in read stream without it, so set flag there as well.
481			*/
482			fcntl(p1[1], F_SETFD, FD_CLOEXEC);
483			fcntl(p0[0], F_SETFD, FD_CLOEXEC);
484	greg	2.3	r_proc[ray_pnprocs].fd_send = p1[1];
485			r_proc[ray_pnprocs].fd_recv = p0[0];
486			r_proc[ray_pnprocs].npending = 0;
487			ray_pnprocs++;
488			ray_pnidle++;
489	greg	2.1	}
490			}
491
492
493	greg	2.22	void
494	schorsch	2.6	ray_pclose( /* close one or more child processes */
495			int nsub
496			)
497	greg	2.1	{
498			static int inclose = 0;
499			RAY res;
500			/* check recursion */
501			if (inclose)
502			return;
503			inclose++;
504			/* check argument */
505	schorsch	2.5	if ((nsub <= 0) \| (nsub > ray_pnprocs))
506	greg	2.3	nsub = ray_pnprocs;
507	greg	2.1	/* clear our ray queue */
508			while (ray_presult(&res,0) > 0)
509			;
510	greg	2.23	r_send_next = 0; /* hard reset in case of error */
511			r_recv_first = r_recv_next = RAYQLEN;
512	greg	2.1	/* clean up children */
513			while (nsub--) {
514			int status;
515	greg	2.3	ray_pnprocs--;
516			close(r_proc[ray_pnprocs].fd_send);
517	greg	2.8	if (waitpid(r_proc[ray_pnprocs].pid, &status, 0) < 0)
518			status = 127<<8;
519	greg	2.24	close(r_proc[ray_pnprocs].fd_recv);
520	greg	2.1	if (status) {
521			sprintf(errmsg,
522			"rendering process %d exited with code %d",
523	greg	2.3	r_proc[ray_pnprocs].pid, status>>8);
524	greg	2.1	error(WARNING, errmsg);
525			}
526	greg	2.3	ray_pnidle--;
527	greg	2.1	}
528			inclose--;
529			}