/* Copyright (C) 2001-2012 Artifex Software, Inc.
   All Rights Reserved.

   This software is provided AS-IS with no warranty, either express or
   implied.

   This software is distributed under license and may not be copied,
   modified or distributed except as expressly authorized under the terms
   of the license contained in the file LICENSE in this distribution.

   Refer to licensing information at http://www.artifex.com or contact
   Artifex Software, Inc.,  7 Mt. Lassen Drive - Suite A-134, San Rafael,
   CA  94903, U.S.A., +1(415)492-9861, for further information.
*/


/* Display PostScript context operators */
#include "memory_.h"
#include "ghost.h"
#include "gp.h"			/* for usertime */
#include "oper.h"
#include "gsexit.h"
#include "gsgc.h"
#include "gsstruct.h"
#include "gsutil.h"
#include "gxalloc.h"
#include "gxstate.h"		/* for copying gstate stack */
#include "stream.h"		/* for files.h */
#include "files.h"
#include "idict.h"
#include "igstate.h"
#include "icontext.h"
#include "interp.h"
#include "isave.h"
#include "istruct.h"
#include "dstack.h"
#include "estack.h"
#include "ostack.h"
#include "store.h"

/*
 * Define the rescheduling interval.  A value of max_int effectively
 * disables scheduling.  The only reason not to make this const is to
 * allow it to be changed during testing.
 */
static int reschedule_interval = 100;

/* Context structure */
typedef enum {
    cs_active,
    cs_done
} ctx_status_t;
typedef long ctx_index_t;	/* >= 0 */
typedef struct gs_context_s gs_context_t;
typedef struct gs_scheduler_s gs_scheduler_t;

/*
 * If several contexts share local VM, then if any one of them has done an
 * unmatched save, the others are not allowed to run.  We handle this by
 * maintaining the following invariant:
 *      When control reaches the point in the scheduler that decides
 *      what context to run next, then for each group of contexts
 *      sharing local VM, if the save level for that VM is non-zero,
 *      saved_local_vm is only set in the context that has unmatched
 *      saves.
 * We maintain this invariant as follows: when control enters the
 * scheduler, if a context was running, we set its saved_local_vm flag
 * to (save_level > 0).  When selecting a context to run, we ignore
 * contexts where saved_local_vm is false and the local VM save_level > 0.
 */
struct gs_context_s {
    gs_context_state_t state;	/* (must be first for subclassing) */
    /* Private state */
    gs_scheduler_t *scheduler;
    ctx_status_t status;
    ctx_index_t index;		/* > 0 */
    bool detach;		/* true if a detach has been */
                                /* executed for this context */
    bool saved_local_vm;	/* (see above) */
    bool visible;		/* during GC, true if visible; */
                                /* otherwise, always true */
    ctx_index_t next_index;	/* next context with same status */
                                /* (active, waiting on same lock, */
                                /* waiting on same condition, */
                                /* waiting to be destroyed) */
    ctx_index_t joiner_index;	/* context waiting on a join */
                                /* for this one */
    gs_context_t *table_next;	/* hash table chain -- this must be a real */
                                /* pointer, for looking up indices */
};
static inline bool
context_is_visible(const gs_context_t *pctx)
{
    return (pctx && pctx->visible);
}
static inline gs_context_t *
visible_context(gs_context_t *pctx)
{
    return (pctx && pctx->visible ? pctx : (gs_context_t *)0);
}

/* GC descriptor */
static
CLEAR_MARKS_PROC(context_clear_marks)
{
    gs_context_t *const pctx = vptr;

    (*st_context_state.clear_marks)
        (cmem, &pctx->state, sizeof(pctx->state), &st_context_state);
}
static
ENUM_PTRS_WITH(context_enum_ptrs, gs_context_t *pctx)
ENUM_PREFIX(st_context_state, 2);
case 0: return ENUM_OBJ(pctx->scheduler);
case 1: {
    /* Return the next *visible* context. */
    const gs_context_t *next = pctx->table_next;

    while (next && !next->visible)
        next = next->table_next;
    return ENUM_OBJ(next);
}
ENUM_PTRS_END
static RELOC_PTRS_WITH(context_reloc_ptrs, gs_context_t *pctx)
    RELOC_PREFIX(st_context_state);
    RELOC_VAR(pctx->scheduler);
    /* Don't relocate table_next -- the scheduler object handles that. */
RELOC_PTRS_END
gs_private_st_complex_only(st_context, gs_context_t, "gs_context_t",
             context_clear_marks, context_enum_ptrs, context_reloc_ptrs, 0);

/*
 * Context list structure.  Note that this uses context indices, not
 * pointers, to avoid having to worry about pointers between local VMs.
 */
typedef struct ctx_list_s {
    ctx_index_t head_index;
    ctx_index_t tail_index;
} ctx_list_t;

/* Condition structure */
typedef struct gs_condition_s {
    ctx_list_t waiting;	/* contexts waiting on this condition */
} gs_condition_t;
gs_private_st_simple(st_condition, gs_condition_t, "conditiontype");

/* Lock structure */
typedef struct gs_lock_s {
    ctx_list_t waiting;		/* contexts waiting for this lock, */
                                /* must be first for subclassing */
    ctx_index_t holder_index;	/* context holding the lock, if any */
    gs_scheduler_t *scheduler;
} gs_lock_t;
gs_private_st_ptrs1(st_lock, gs_lock_t, "locktype",
                    lock_enum_ptrs, lock_reloc_ptrs, scheduler);

/* Global state */
/*typedef struct gs_scheduler_s gs_scheduler_t; *//* (above) */
struct gs_scheduler_s {
    gs_context_t *current;
    long usertime_initial;	/* usertime when current started running */
    ctx_list_t active;
    vm_reclaim_proc((*save_vm_reclaim));
    ctx_index_t dead_index;
#define CTX_TABLE_SIZE 19
    gs_context_t *table[CTX_TABLE_SIZE];
};

/* Convert a context index to a context pointer. */
static gs_context_t *
index_context(const gs_scheduler_t *psched, long index)
{
    gs_context_t *pctx;

    if (index == 0)
        return 0;
    pctx = psched->table[index % CTX_TABLE_SIZE];
    while (pctx != 0 && pctx->index != index)
        pctx = pctx->table_next;
    return pctx;
}

/* Structure definition */
gs_private_st_composite(st_scheduler, gs_scheduler_t, "gs_scheduler",
                        scheduler_enum_ptrs, scheduler_reloc_ptrs);
/*
 * The only cross-local-VM pointers in the context machinery are the
 * table_next pointers in contexts, and the current and table[] pointers
 * in the scheduler.  We need to handle all of these specially.
 */
static ENUM_PTRS_WITH(scheduler_enum_ptrs, gs_scheduler_t *psched)
{
    index -= 1;
    if (index < CTX_TABLE_SIZE) {
        gs_context_t *pctx = psched->table[index];

        while (pctx && !pctx->visible)
            pctx = pctx->table_next;
        return ENUM_OBJ(pctx);
    }
    return 0;
}
case 0: return ENUM_OBJ(visible_context(psched->current));
ENUM_PTRS_END
static RELOC_PTRS_WITH(scheduler_reloc_ptrs, gs_scheduler_t *psched)
{
    if (psched->current->visible)
        RELOC_VAR(psched->current);
    {
        int i;

        for (i = 0; i < CTX_TABLE_SIZE; ++i) {
            gs_context_t **ppctx = &psched->table[i];
            gs_context_t **pnext;

            for (; *ppctx; ppctx = pnext) {
                pnext = &(*ppctx)->table_next;
                if ((*ppctx)->visible)
                    RELOC_VAR(*ppctx);
            }
        }
    }
}
RELOC_PTRS_END

/*
 * The context scheduler requires special handling during garbage
 * collection, since it is the only structure that can legitimately
 * reference objects in multiple local VMs.  To deal with this, we wrap the
 * interpreter's garbage collector with code that prevents it from seeing
 * contexts in other than the current local VM.  ****** WORKS FOR LOCAL GC,
 * NOT FOR GLOBAL ******
 */
static void
context_reclaim(vm_spaces * pspaces, bool global)
{
    /*
     * Search through the registered roots to find the current context.
     * (This is a hack so we can find the scheduler.)
     */
    int i;
    gs_context_t *pctx = 0;	/* = 0 is bogus to pacify compilers */
    gs_scheduler_t *psched = 0;
    gs_ref_memory_t *lmem = 0;	/* = 0 is bogus to pacify compilers */
    chunk_locator_t loc;

    for (i = countof(pspaces->memories.indexed) - 1; psched == 0 && i > 0; --i) {
        gs_ref_memory_t *mem = pspaces->memories.indexed[i];
        const gs_gc_root_t *root = mem->roots;

        for (; root; root = root->next) {
            if (gs_object_type((gs_memory_t *)mem, *root->p) == &st_context) {
                pctx = *root->p;
                psched = pctx->scheduler;
                lmem = mem;
                break;
            }
        }
    }

    /* Hide all contexts in other (local) VMs. */
    /*
     * See context_create below for why we look for the context
     * in stable memory.
     */
    loc.memory = (gs_ref_memory_t *)gs_memory_stable((gs_memory_t *)lmem);
    loc.cp = 0;
    for (i = 0; i < CTX_TABLE_SIZE; ++i)
        for (pctx = psched->table[i]; pctx; pctx = pctx->table_next)
            pctx->visible = chunk_locate_ptr(pctx, &loc);

#ifdef DEBUG
    if (!psched->current->visible) {
        lprintf("Current context is invisible!\n");
        gs_abort((gs_memory_t *)lmem);
    }
#endif

    /* Do the actual garbage collection. */
    psched->save_vm_reclaim(pspaces, global);

    /* Make all contexts visible again. */
    for (i = 0; i < CTX_TABLE_SIZE; ++i)
        for (pctx = psched->table[i]; pctx; pctx = pctx->table_next)
            pctx->visible = true;
}

/* Forward references */
static int context_create(gs_scheduler_t *, gs_context_t **,
                           const gs_dual_memory_t *,
                           const gs_context_state_t *, bool);
static long context_usertime(void);
static int context_param(const gs_scheduler_t *, os_ptr, gs_context_t **);
static void context_destroy(gs_context_t *);
static void stack_copy(ref_stack_t *, const ref_stack_t *, uint, uint);
static int lock_acquire(os_ptr, gs_context_t *);
static int lock_release(ref *);

/* Internal procedures */
static void
context_load(gs_scheduler_t *psched, gs_context_t *pctx)
{
    if_debug1('"', "[\"]loading %ld\n", pctx->index);
    if ( pctx->state.keep_usertime )
      psched->usertime_initial = context_usertime();
    context_state_load(&pctx->state);
}
static void
context_store(gs_scheduler_t *psched, gs_context_t *pctx)
{
    if_debug1('"', "[\"]storing %ld\n", pctx->index);
    context_state_store(&pctx->state);
    if ( pctx->state.keep_usertime )
      pctx->state.usertime_total +=
        context_usertime() - psched->usertime_initial;
}

/* List manipulation */
static void
add_last(const gs_scheduler_t *psched, ctx_list_t *pl, gs_context_t *pc)
{
    pc->next_index = 0;
    if (pl->head_index == 0)
        pl->head_index = pc->index;
    else
        index_context(psched, pl->tail_index)->next_index = pc->index;
    pl->tail_index = pc->index;
}

/* ------ Initialization ------ */

static int ctx_initialize(i_ctx_t **);
static int ctx_reschedule(i_ctx_t **);
static int ctx_time_slice(i_ctx_t **);
static int
zcontext_init(i_ctx_t *i_ctx_p)
{
    /* Complete initialization after the interpreter is entered. */
    i_ctx_p->reschedule_proc = ctx_initialize;
    i_ctx_p->time_slice_proc = ctx_initialize;
    i_ctx_p->time_slice_ticks = 0;
    return 0;
}
/*
 * The interpreter calls this procedure at the first reschedule point.
 * It completes context initialization.
 */
static int
ctx_initialize(i_ctx_t **pi_ctx_p)
{
    i_ctx_t *i_ctx_p = *pi_ctx_p; /* for gs_imemory */
    gs_ref_memory_t *imem = iimemory_system;
    gs_scheduler_t *psched =
        gs_alloc_struct_immovable((gs_memory_t *) imem, gs_scheduler_t,
                                  &st_scheduler, "gs_scheduler");

    psched->current = 0;
    psched->active.head_index = psched->active.tail_index = 0;
    psched->save_vm_reclaim = i_ctx_p->memory.spaces.vm_reclaim;
    i_ctx_p->memory.spaces.vm_reclaim = context_reclaim;
    psched->dead_index = 0;
    memset(psched->table, 0, sizeof(psched->table));
    /* Create an initial context. */
    if (context_create(psched, &psched->current, &gs_imemory, *pi_ctx_p, true) < 0) {
        lprintf("Can't create initial context!");
        gs_abort(imemory);
    }
    psched->current->scheduler = psched;
    /* Hook into the interpreter. */
    *pi_ctx_p = &psched->current->state;
    psched->current->state.reschedule_proc = ctx_reschedule;
    psched->current->state.time_slice_proc = ctx_time_slice;
    psched->current->state.time_slice_ticks = reschedule_interval;
    return 0;
}

/* ------ Interpreter interface to scheduler ------ */

/* When an operator decides it is time to run a new context, */
/* it returns o_reschedule.  The interpreter saves all its state in */
/* memory, calls ctx_reschedule, and then loads the state from memory. */
static int
ctx_reschedule(i_ctx_t **pi_ctx_p)
{
    gs_context_t *current = (gs_context_t *)*pi_ctx_p;
    gs_scheduler_t *psched = current->scheduler;

#ifdef DEBUG
    if (*pi_ctx_p != &current->state) {
        lprintf2("current->state = 0x%lx, != i_ctx_p = 0x%lx!\n",
                 (ulong)&current->state, (ulong)*pi_ctx_p);
    }
#endif
    /* If there are any dead contexts waiting to be released, */
    /* take care of that now. */
    while (psched->dead_index != 0) {
        gs_context_t *dead = index_context(psched, psched->dead_index);
        long next_index = dead->next_index;

        if (current == dead) {
            if_debug1('"', "[\"]storing dead %ld\n", current->index);
            context_state_store(&current->state);
            current = 0;
        }
        context_destroy(dead);
        psched->dead_index = next_index;
    }
    /* Update saved_local_vm.  See above for the invariant. */
    if (current != 0)
        current->saved_local_vm =
            current->state.memory.space_local->saved != 0;
    /* Run the first ready context, taking the 'save' lock into account. */
    {
        gs_context_t *prev = 0;
        gs_context_t *ready;

        for (ready = index_context(psched, psched->active.head_index);;
             prev = ready, ready = index_context(psched, ready->next_index)
            ) {
            if (ready == 0) {
                if (current != 0)
                    context_store(psched, current);
                lprintf("No context to run!");
                return_error(e_Fatal);
            }
            /* See above for an explanation of the following test. */
            if (ready->state.memory.space_local->saved != 0 &&
                !ready->saved_local_vm
                )
                continue;
            /* Found a context to run. */
            {
                ctx_index_t next_index = ready->next_index;

                if (prev)
                    prev->next_index = next_index;
                else
                    psched->active.head_index = next_index;
                if (!next_index)
                    psched->active.tail_index = (prev ? prev->index : 0);
            }
            break;
        }
        if (ready == current)
            return 0;		/* no switch */
        /*
         * Save the state of the current context in psched->current,
         * if any context is current.
         */
        if (current != 0)
            context_store(psched, current);
        psched->current = ready;
        /* Load the state of the new current context. */
        context_load(psched, ready);
        /* Switch the interpreter's context state pointer. */
        *pi_ctx_p = &ready->state;
    }
    return 0;
}

/* If the interpreter wants to time-slice, it saves its state, */
/* calls ctx_time_slice, and reloads its state. */
static int
ctx_time_slice(i_ctx_t **pi_ctx_p)
{
    gs_scheduler_t *psched = ((gs_context_t *)*pi_ctx_p)->scheduler;

    if (psched->active.head_index == 0)
        return 0;
    if_debug0('"', "[\"]time-slice\n");
    add_last(psched, &psched->active, psched->current);
    return ctx_reschedule(pi_ctx_p);
}

/* ------ Context operators ------ */

/* - currentcontext <context> */
static int
zcurrentcontext(i_ctx_t *i_ctx_p)
{
    os_ptr op = osp;
    const gs_context_t *current = (const gs_context_t *)i_ctx_p;

    push(1);
    make_int(op, current->index);
    return 0;
}

/* <context> detach - */
static int
zdetach(i_ctx_t *i_ctx_p)
{
    os_ptr op = osp;
    const gs_scheduler_t *psched = ((gs_context_t *)i_ctx_p)->scheduler;
    gs_context_t *pctx;
    int code;

    if ((code = context_param(psched, op, &pctx)) < 0)
        return code;
    if_debug2('\'', "[']detach %ld, status = %d\n",
              pctx->index, pctx->status);
    if (pctx->joiner_index != 0 || pctx->detach)
        return_error(e_invalidcontext);
    switch (pctx->status) {
        case cs_active:
            pctx->detach = true;
            break;
        case cs_done:
            context_destroy(pctx);
    }
    pop(1);
    return 0;
}

static int
    do_fork(i_ctx_t *i_ctx_p, os_ptr op, const ref * pstdin,
            const ref * pstdout, uint mcount, bool local),
    values_older_than(const ref_stack_t * pstack, uint first, uint last,
                      int max_space);
static int
    fork_done(i_ctx_t *),
    fork_done_with_error(i_ctx_t *),
    finish_join(i_ctx_t *),
    reschedule_now(i_ctx_t *);

/* <mark> <obj1> ... <objN> <proc> .fork <context> */
/* <mark> <obj1> ... <objN> <proc> <stdin|null> <stdout|null> */
/*   .localfork <context> */
static int
zfork(i_ctx_t *i_ctx_p)
{
    os_ptr op = osp;
    uint mcount = ref_stack_counttomark(&o_stack);
    ref rnull;

    if (mcount == 0)
        return_error(e_unmatchedmark);
    make_null(&rnull);
    return do_fork(i_ctx_p, op, &rnull, &rnull, mcount, false);
}
static int
zlocalfork(i_ctx_t *i_ctx_p)
{
    os_ptr op = osp;
    uint mcount = ref_stack_counttomark(&o_stack);
    int code;

    if (mcount == 0)
        return_error(e_unmatchedmark);
    code = values_older_than(&o_stack, 1, mcount - 1, avm_local);
    if (code < 0)
        return code;
    code = do_fork(i_ctx_p, op - 2, op - 1, op, mcount - 2, true);
    if (code < 0)
        return code;
    op = osp;
    op[-2] = *op;
    pop(2);
    return code;
}

/* Internal procedure to actually do the fork operation. */
static int
do_fork(i_ctx_t *i_ctx_p, os_ptr op, const ref * pstdin, const ref * pstdout,
        uint mcount, bool local)
{
    gs_context_t *pcur = (gs_context_t *)i_ctx_p;
    gs_scheduler_t *psched = pcur->scheduler;
    stream *s;
    gs_dual_memory_t dmem;
    gs_context_t *pctx;
    ref old_userdict, new_userdict;
    int code;

    check_proc(*op);
    if (iimemory_local->save_level)
        return_error(e_invalidcontext);
    if (r_has_type(pstdout, t_null)) {
        code = zget_stdout(i_ctx_p, &s);
        if (code < 0)
            return code;
        pstdout = &ref_stdio[1];
    } else
        check_read_file(i_ctx_p, s, pstdout);
    if (r_has_type(pstdin, t_null)) {
        code = zget_stdin(i_ctx_p, &s);
        if (code < 0)
            return code;
        pstdin = &ref_stdio[0];
    } else
        check_read_file(i_ctx_p, s, pstdin);
    dmem = gs_imemory;
    if (local) {
        /* Share global VM, private local VM. */
        ref *puserdict;
        uint userdict_size;
        gs_memory_t *parent = iimemory_local->non_gc_memory;
        gs_ref_memory_t *lmem;
        gs_ref_memory_t *lmem_stable;

        if (dict_find_string(systemdict, "userdict", &puserdict) <= 0 ||
            !r_has_type(puserdict, t_dictionary)
            )
            return_error(e_Fatal);
        old_userdict = *puserdict;
        userdict_size = dict_maxlength(&old_userdict);
        lmem = ialloc_alloc_state(parent, iimemory_local->chunk_size);
        lmem_stable = ialloc_alloc_state(parent, iimemory_local->chunk_size);
        if (lmem == 0 || lmem_stable == 0) {
            gs_free_object(parent, lmem_stable, "do_fork");
            gs_free_object(parent, lmem, "do_fork");
            return_error(e_VMerror);
        }
        lmem->space = avm_local;
        lmem_stable->space = avm_local;
        lmem->stable_memory = (gs_memory_t *)lmem_stable;
        dmem.space_local = lmem;
        code = context_create(psched, &pctx, &dmem, &pcur->state, false);
        if (code < 0) {
            /****** FREE lmem ******/
            return code;
        }
        /*
         * Create a new userdict.  PostScript code will take care of
         * the rest of the initialization of the new context.
         */
        code = dict_alloc(lmem, userdict_size, &new_userdict);
        if (code < 0) {
            context_destroy(pctx);
            /****** FREE lmem ******/
            return code;
        }
    } else {
        /* Share global and local VM. */
        code = context_create(psched, &pctx, &dmem, &pcur->state, false);
        if (code < 0) {
            /****** FREE lmem ******/
            return code;
        }
        /*
         * Copy the gstate stack.  The current method is not elegant;
         * in fact, I'm not entirely sure it works.
         */
        {
            int n;
            const gs_state *old;
            gs_state *new;

            for (n = 0, old = igs; old != 0; old = gs_state_saved(old))
                ++n;
            for (old = pctx->state.pgs; old != 0; old = gs_state_saved(old))
                --n;
            for (; n > 0 && code >= 0; --n)
                code = gs_gsave(pctx->state.pgs);
            if (code < 0) {
/****** FREE lmem & GSTATES ******/
                return code;
            }
            for (old = igs, new = pctx->state.pgs;
                 old != 0 /* (== new != 0) */  && code >= 0;
                 old = gs_state_saved(old), new = gs_state_saved(new)
                )
                code = gs_setgstate(new, old);
            if (code < 0) {
/****** FREE lmem & GSTATES ******/
                return code;
            }
        }
    }
    pctx->state.op_array_table_global = i_ctx_p->op_array_table_global;
    pctx->state.op_array_table_local  = i_ctx_p->op_array_table_local;
    pctx->state.language_level = i_ctx_p->language_level;
    pctx->state.dict_stack.min_size = idict_stack.min_size;
    pctx->state.dict_stack.userdict_index = idict_stack.userdict_index;
    pctx->state.stdio[0] = *pstdin;
    pctx->state.stdio[1] = *pstdout;
    pctx->state.stdio[2] = pcur->state.stdio[2];
    /* Initialize the interpreter stacks. */
    {
        ref_stack_t *dstack = (ref_stack_t *)&pctx->state.dict_stack;
        uint count = ref_stack_count(&d_stack);
        uint copy = (local ? min_dstack_size : count);

        ref_stack_push(dstack, copy);
        stack_copy(dstack, &d_stack, copy, count - copy);
        if (local) {
            /* Substitute the new userdict for the old one. */
            long i;

            for (i = 0; i < copy; ++i) {
                ref *pdref = ref_stack_index(dstack, i);

                if (obj_eq(imemory, pdref, &old_userdict))
                    *pdref = new_userdict;
            }
        }
    }
    {
        ref_stack_t *estack = (ref_stack_t *)&pctx->state.exec_stack;

        ref_stack_push(estack, 3);
        /* fork_done must be executed in both normal and error cases. */
        make_mark_estack(estack->p - 2, es_other, fork_done_with_error);
        make_oper(estack->p - 1, 0, fork_done);
        *estack->p = *op;
    }
    {
        ref_stack_t *ostack = (ref_stack_t *)&pctx->state.op_stack;
        uint count = mcount - 2;

        ref_stack_push(ostack, count);
        stack_copy(ostack, &o_stack, count, osp - op + 1);
    }
    pctx->state.binary_object_format = pcur->state.binary_object_format;
    add_last(psched, &psched->active, pctx);
    pop(mcount - 1);
    op = osp;
    make_int(op, pctx->index);
    return 0;
}

/*
 * Check that all values being passed by fork or join are old enough
 * to be valid in the environment to which they are being transferred.
 */
static int
values_older_than(const ref_stack_t * pstack, uint first, uint last,
                  int next_space)
{
    uint i;

    for (i = first; i <= last; ++i)
        if (r_space(ref_stack_index(pstack, (long)i)) >= next_space)
            return_error(e_invalidaccess);
    return 0;
}

/* This gets executed when a context terminates normally. */
/****** MUST DO ALL RESTORES ******/
/****** WHAT IF invalidrestore? ******/
static int
fork_done(i_ctx_t *i_ctx_p)
{
    os_ptr op = osp;
    gs_context_t *pcur = (gs_context_t *)i_ctx_p;
    gs_scheduler_t *psched = pcur->scheduler;

    if_debug2('\'', "[']done %ld%s\n", pcur->index,
              (pcur->detach ? ", detached" : ""));
    /*
     * Clear the context's dictionary, execution and graphics stacks
     * now, to retain as little as possible in case of a garbage
     * collection or restore.  We know that fork_done is the
     * next-to-bottom entry on the execution stack.
     */
    ref_stack_pop_to(&d_stack, min_dstack_size);
    pop_estack(&pcur->state, ref_stack_count(&e_stack) - 1);
    gs_grestoreall(igs);
    /*
     * If there are any unmatched saves, we need to execute restores
     * until there aren't.  An invalidrestore is possible and will
     * result in an error termination.
     */
    if (iimemory_local->save_level) {
        ref *prestore;

        if (dict_find_string(systemdict, "restore", &prestore) <= 0) {
            lprintf("restore not found in systemdict!");
            return_error(e_Fatal);
        }
        if (pcur->detach) {
            ref_stack_clear(&o_stack);	/* help avoid invalidrestore */
            op = osp;
        }
        push(1);
        make_tv(op, t_save, saveid, alloc_save_current_id(&gs_imemory));
        push_op_estack(fork_done);
        ++esp;
        ref_assign(esp, prestore);
        return o_push_estack;
    }
    if (pcur->detach) {
        /*
         * We would like to free the context's memory, but we can't do
         * it yet, because the interpreter still has references to it.
         * Instead, queue the context to be freed the next time we
         * reschedule.  We can, however, clear its operand stack now.
         */
        ref_stack_clear(&o_stack);
        context_store(psched, pcur);
        pcur->next_index = psched->dead_index;
        psched->dead_index = pcur->index;
        psched->current = 0;
    } else {
        gs_context_t *pctx = index_context(psched, pcur->joiner_index);

        pcur->status = cs_done;
        /* Schedule the context waiting to join this one, if any. */
        if (pctx != 0)
            add_last(psched, &psched->active, pctx);
    }
    return o_reschedule;
}
/*
 * This gets executed when the stack is being unwound for an error
 * termination.
 */
static int
fork_done_with_error(i_ctx_t *i_ctx_p)
{
/****** WHAT TO DO? ******/
    return fork_done(i_ctx_p);
}

/* <context> join <mark> <obj1> ... <objN> */
static int
zjoin(i_ctx_t *i_ctx_p)
{
    os_ptr op = osp;
    gs_context_t *current = (gs_context_t *)i_ctx_p;
    gs_scheduler_t *psched = current->scheduler;
    gs_context_t *pctx;
    int code;

    if ((code = context_param(psched, op, &pctx)) < 0)
        return code;
    if_debug2('\'', "[']join %ld, status = %d\n",
              pctx->index, pctx->status);
    /*
     * It doesn't seem logically necessary, but the Red Book says that
     * the context being joined must share both global and local VM with
     * the current context.
     */
    if (pctx->joiner_index != 0 || pctx->detach || pctx == current ||
        pctx->state.memory.space_global !=
          current->state.memory.space_global ||
        pctx->state.memory.space_local !=
          current->state.memory.space_local ||
        iimemory_local->save_level != 0
        )
        return_error(e_invalidcontext);
    switch (pctx->status) {
        case cs_active:
            /*
             * We need to re-execute the join after the joined
             * context is done.  Since we can't return both
             * o_push_estack and o_reschedule, we push a call on
             * reschedule_now, which accomplishes the latter.
             */
            check_estack(2);
            push_op_estack(finish_join);
            push_op_estack(reschedule_now);
            pctx->joiner_index = current->index;
            return o_push_estack;
        case cs_done:
            {
                const ref_stack_t *ostack =
                    (ref_stack_t *)&pctx->state.op_stack;
                uint count = ref_stack_count(ostack);

                push(count);
                {
                    ref *rp = ref_stack_index(&o_stack, count);

                    make_mark(rp);
                }
                stack_copy(&o_stack, ostack, count, 0);
                context_destroy(pctx);
            }
    }
    return 0;
}

/* Finish a deferred join. */
static int
finish_join(i_ctx_t *i_ctx_p)
{
    os_ptr op = osp;
    gs_context_t *current = (gs_context_t *)i_ctx_p;
    gs_scheduler_t *psched = current->scheduler;
    gs_context_t *pctx;
    int code;

    if ((code = context_param(psched, op, &pctx)) < 0)
        return code;
    if_debug2('\'', "[']finish_join %ld, status = %d\n",
              pctx->index, pctx->status);
    if (pctx->joiner_index != current->index)
        return_error(e_invalidcontext);
    pctx->joiner_index = 0;
    return zjoin(i_ctx_p);
}

/* Reschedule now. */
static int
reschedule_now(i_ctx_t *i_ctx_p)
{
    return o_reschedule;
}

/* - yield - */
static int
zyield(i_ctx_t *i_ctx_p)
{
    gs_context_t *current = (gs_context_t *)i_ctx_p;
    gs_scheduler_t *psched = current->scheduler;

    if (psched->active.head_index == 0)
        return 0;
    if_debug0('"', "[\"]yield\n");
    add_last(psched, &psched->active, current);
    return o_reschedule;
}

/* ------ Condition and lock operators ------ */

static int
    monitor_cleanup(i_ctx_t *),
    monitor_release(i_ctx_t *),
    await_lock(i_ctx_t *);
static void
     activate_waiting(gs_scheduler_t *, ctx_list_t * pcl);

/* - condition <condition> */
static int
zcondition(i_ctx_t *i_ctx_p)
{
    os_ptr op = osp;
    gs_condition_t *pcond =
        ialloc_struct(gs_condition_t, &st_condition, "zcondition");

    if (pcond == 0)
        return_error(e_VMerror);
    pcond->waiting.head_index = pcond->waiting.tail_index = 0;
    push(1);
    make_istruct(op, a_all, pcond);
    return 0;
}

/* - lock <lock> */
static int
zlock(i_ctx_t *i_ctx_p)
{
    os_ptr op = osp;
    gs_lock_t *plock = ialloc_struct(gs_lock_t, &st_lock, "zlock");

    if (plock == 0)
        return_error(e_VMerror);
    plock->holder_index = 0;
    plock->waiting.head_index = plock->waiting.tail_index = 0;
    push(1);
    make_istruct(op, a_all, plock);
    return 0;
}

/* <lock> <proc> monitor - */
static int
zmonitor(i_ctx_t *i_ctx_p)
{
    gs_context_t *current = (gs_context_t *)i_ctx_p;
    os_ptr op = osp;
    gs_lock_t *plock;
    gs_context_t *pctx;
    int code;

    check_stype(op[-1], st_lock);
    check_proc(*op);
    plock = r_ptr(op - 1, gs_lock_t);
    pctx = index_context(current->scheduler, plock->holder_index);
    if_debug1('\'', "[']monitor 0x%lx\n", (ulong) plock);
    if (pctx != 0) {
        if (pctx == current ||
            (iimemory_local->save_level != 0 &&
             pctx->state.memory.space_local ==
             current->state.memory.space_local)
            )
            return_error(e_invalidcontext);
    }
    /*
     * We push on the e-stack:
     *      The lock object
     *      An e-stack mark with monitor_cleanup, to release the lock
     *        in case of an error
     *      monitor_release, to release the lock in the normal case
     *      The procedure to execute
     */
    check_estack(4);
    code = lock_acquire(op - 1, current);
    if (code != 0) {		/* We didn't acquire the lock.  Re-execute this later. */
        push_op_estack(zmonitor);
        return code;		/* o_reschedule */
    }
    *++esp = op[-1];
    push_mark_estack(es_other, monitor_cleanup);
    push_op_estack(monitor_release);
    *++esp = *op;
    pop(2);
    return o_push_estack;
}
/* Release the monitor lock when unwinding for an error or exit. */
static int
monitor_cleanup(i_ctx_t *i_ctx_p)
{
    int code = lock_release(esp);

    if (code < 0)
        return code;
    --esp;
    return o_pop_estack;
}
/* Release the monitor lock when the procedure completes. */
static int
monitor_release(i_ctx_t *i_ctx_p)
{
    int code = lock_release(esp - 1);

    if (code < 0)
        return code;
    esp -= 2;
    return o_pop_estack;
}

/* <condition> notify - */
static int
znotify(i_ctx_t *i_ctx_p)
{
    os_ptr op = osp;
    gs_context_t *current = (gs_context_t *)i_ctx_p;
    gs_condition_t *pcond;

    check_stype(*op, st_condition);
    pcond = r_ptr(op, gs_condition_t);
    if_debug1('"', "[\"]notify 0x%lx\n", (ulong) pcond);
    pop(1);
    op--;
    if (pcond->waiting.head_index == 0)	/* nothing to do */
        return 0;
    activate_waiting(current->scheduler, &pcond->waiting);
    return zyield(i_ctx_p);
}

/* <lock> <condition> wait - */
static int
zwait(i_ctx_t *i_ctx_p)
{
    os_ptr op = osp;
    gs_context_t *current = (gs_context_t *)i_ctx_p;
    gs_scheduler_t *psched = current->scheduler;
    gs_lock_t *plock;
    gs_context_t *pctx;
    gs_condition_t *pcond;

    check_stype(op[-1], st_lock);
    plock = r_ptr(op - 1, gs_lock_t);
    check_stype(*op, st_condition);
    pcond = r_ptr(op, gs_condition_t);
    if_debug2('"', "[\"]wait lock 0x%lx, condition 0x%lx\n",
              (ulong) plock, (ulong) pcond);
    pctx = index_context(psched, plock->holder_index);
    if (pctx == 0 || pctx != psched->current ||
        (iimemory_local->save_level != 0 &&
         (r_space(op - 1) == avm_local || r_space(op) == avm_local))
        )
        return_error(e_invalidcontext);
    check_estack(1);
    lock_release(op - 1);
    add_last(psched, &pcond->waiting, pctx);
    push_op_estack(await_lock);
    return o_reschedule;
}
/* When the condition is signaled, wait for acquiring the lock. */
static int
await_lock(i_ctx_t *i_ctx_p)
{
    gs_context_t *current = (gs_context_t *)i_ctx_p;
    os_ptr op = osp;
    int code = lock_acquire(op - 1, current);

    if (code == 0) {
        pop(2);
        return 0;
    }
    /* We didn't acquire the lock.  Re-execute the wait. */
    push_op_estack(await_lock);
    return code;		/* o_reschedule */
}

/* Activate a list of waiting contexts, and reset the list. */
static void
activate_waiting(gs_scheduler_t *psched, ctx_list_t * pcl)
{
    gs_context_t *pctx = index_context(psched, pcl->head_index);
    gs_context_t *next;

    for (; pctx != 0; pctx = next) {
        next = index_context(psched, pctx->next_index);
        add_last(psched, &psched->active, pctx);
    }
    pcl->head_index = pcl->tail_index = 0;
}

/* ------ Miscellaneous operators ------ */

/* - usertime <int> */
static int
zusertime_context(i_ctx_t *i_ctx_p)
{
    gs_context_t *current = (gs_context_t *)i_ctx_p;
    gs_scheduler_t *psched = current->scheduler;
    os_ptr op = osp;
    long utime = context_usertime();

    push(1);
    if (!current->state.keep_usertime) {
        /*
         * This is the first time this context has executed usertime:
         * we must track its execution time from now on.
         */
        psched->usertime_initial = utime;
        current->state.keep_usertime = true;
    }
    make_int(op, current->state.usertime_total + utime -
             psched->usertime_initial);
    return 0;
}

/* ------ Internal procedures ------ */

/* Create a context. */
static int
context_create(gs_scheduler_t * psched, gs_context_t ** ppctx,
               const gs_dual_memory_t * dmem,
               const gs_context_state_t *i_ctx_p, bool copy_state)
{
    /*
     * Contexts are always created at the outermost save level, so they do
     * not need to be allocated in stable memory for the sake of
     * save/restore.  However, context_reclaim needs to be able to test
     * whether a given context belongs to a given local VM, and allocating
     * contexts in stable local VM avoids the need to scan multiple save
     * levels when making this test.
     */
    gs_memory_t *mem = gs_memory_stable((gs_memory_t *)dmem->space_local);
    gs_context_t *pctx;
    int code;
    long ctx_index;
    gs_context_t **pte;

    pctx = gs_alloc_struct(mem, gs_context_t, &st_context, "context_create");
    if (pctx == 0)
        return_error(e_VMerror);
    if (copy_state) {
        pctx->state = *i_ctx_p;
    } else {
        gs_context_state_t *pctx_st = &pctx->state;

        code = context_state_alloc(&pctx_st, systemdict, dmem);
        if (code < 0) {
            gs_free_object(mem, pctx, "context_create");
            return code;
        }
    }
    ctx_index = gs_next_ids(mem, 1);
    pctx->scheduler = psched;
    pctx->status = cs_active;
    pctx->index = ctx_index;
    pctx->detach = false;
    pctx->saved_local_vm = false;
    pctx->visible = true;
    pctx->next_index = 0;
    pctx->joiner_index = 0;
    pte = &psched->table[ctx_index % CTX_TABLE_SIZE];
    pctx->table_next = *pte;
    *pte = pctx;
    *ppctx = pctx;
    if (gs_debug_c('\'') | gs_debug_c('"'))
        dlprintf2("[']create %ld at 0x%lx\n", ctx_index, (ulong) pctx);
    return 0;
}

/* Check a context ID.  Note that we do not check for context validity. */
static int
context_param(const gs_scheduler_t * psched, os_ptr op, gs_context_t ** ppctx)
{
    gs_context_t *pctx;

    check_type(*op, t_integer);
    pctx = index_context(psched, op->value.intval);
    if (pctx == 0)
        return_error(e_invalidcontext);
    *ppctx = pctx;
    return 0;
}

/* Read the usertime as a single value. */
static long
context_usertime(void)
{
    long secs_ns[2];

    gp_get_usertime(secs_ns);
    return secs_ns[0] * 1000 + secs_ns[1] / 1000000;
}

/* Destroy a context. */
static void
context_destroy(gs_context_t * pctx)
{
    gs_ref_memory_t *mem = pctx->state.memory.space_local;
    gs_scheduler_t *psched = pctx->scheduler;
    gs_context_t **ppctx = &psched->table[pctx->index % CTX_TABLE_SIZE];

    while (*ppctx != pctx)
        ppctx = &(*ppctx)->table_next;
    *ppctx = (*ppctx)->table_next;
    if (gs_debug_c('\'') | gs_debug_c('"'))
        dlprintf3("[']destroy %ld at 0x%lx, status = %d\n",
                  pctx->index, (ulong) pctx, pctx->status);
    if (!context_state_free(&pctx->state))
        gs_free_object((gs_memory_t *) mem, pctx, "context_destroy");
}

/* Copy the top elements of one stack to another. */
/* Note that this does not push the elements: */
/* the destination stack must have enough space preallocated. */
static void
stack_copy(ref_stack_t * to, const ref_stack_t * from, uint count,
           uint from_index)
{
    long i;

    for (i = (long)count - 1; i >= 0; --i)
        *ref_stack_index(to, i) = *ref_stack_index(from, i + from_index);
}

/* Acquire a lock.  Return 0 if acquired, o_reschedule if not. */
static int
lock_acquire(os_ptr op, gs_context_t * pctx)
{
    gs_lock_t *plock = r_ptr(op, gs_lock_t);

    if (plock->holder_index == 0) {
        plock->holder_index = pctx->index;
        plock->scheduler = pctx->scheduler;
        return 0;
    }
    add_last(pctx->scheduler, &plock->waiting, pctx);
    return o_reschedule;
}

/* Release a lock.  Return 0 if OK, e_invalidcontext if not. */
static int
lock_release(ref * op)
{
    gs_lock_t *plock = r_ptr(op, gs_lock_t);
    gs_scheduler_t *psched = plock->scheduler;
    gs_context_t *pctx = index_context(psched, plock->holder_index);

    if (pctx != 0 && pctx == psched->current) {
        plock->holder_index = 0;
        activate_waiting(psched, &plock->waiting);
        return 0;
    }
    return_error(e_invalidcontext);
}

/* ------ Initialization procedure ------ */

/* We need to split the table because of the 16-element limit. */
const op_def zcontext1_op_defs[] = {
    {"0condition", zcondition},
    {"0currentcontext", zcurrentcontext},
    {"1detach", zdetach},
    {"2.fork", zfork},
    {"1join", zjoin},
    {"4.localfork", zlocalfork},
    {"0lock", zlock},
    {"2monitor", zmonitor},
    {"1notify", znotify},
    {"2wait", zwait},
    {"0yield", zyield},
                /* Note that the following replace prior definitions */
                /* in the indicated files: */
    {"0usertime", zusertime_context},	/* zmisc.c */
    op_def_end(0)
};
const op_def zcontext2_op_defs[] = {
                /* Internal operators */
    {"0%fork_done", fork_done},
    {"1%finish_join", finish_join},
    {"0%monitor_cleanup", monitor_cleanup},
    {"0%monitor_release", monitor_release},
    {"2%await_lock", await_lock},
    op_def_end(zcontext_init)
};