Home Research Papers Posters Manuals Talks Download People Help Internal

conv-core/threads.C

Go to the documentation of this file.

/**************************************************************************
 *
 * typedef CthThread
 *
 *   - a first-class thread object.
 *
 * CthThread CthSelf()
 *
 *   - returns the current thread.
 *
 * void CthResume(CthThread t)
 *
 *   - Immediately transfers control to thread t.  Note: normally, the user
 *     of a thread package wouldn't explicitly choose which thread to transfer
 *     to.  Instead, the user would rely upon a "scheduler" to choose the
 *     next thread.  Therefore, this routine is primarily intended for people
 *     who are implementing schedulers, not for end-users.  End-users should
 *     probably call CthSuspend or CthAwaken (see below).
 *
 * CthThread CthCreate(CthVoidFn fn, void *arg, int size)
 *
 *   - Creates a new thread object.  The thread is not given control yet.
 *     The thread is not passed to the scheduler.  When (and if) the thread
 *     eventually receives control, it will begin executing the specified 
 *     function 'fn' with the specified argument.  The 'size' parameter
 *     specifies the stack size, 0 means use the default size.
 *
 * void CthFree(CthThread t)
 *
 *   - Frees thread t.  You may free the currently-executing thread, although
 *     the free will actually be postponed until the thread suspends.
 *
 *
 * In addition to the routines above, the threads package assumes that there
 * will be a "scheduler" of some sort, whose job is to select which threads
 * to execute.  The threads package does not provide a scheduler (although
 * converse may provide one or more schedulers separately).  However, for
 * standardization reasons, it does define an interface to which all schedulers
 * can comply.  A scheduler consists of a pair of functions:
 *
 *   - An awaken-function.  The awaken-function is called to
 *     notify the scheduler that a particular thread needs the CPU.  The
 *     scheduler is expected to respond to this by inserting the thread
 *     into a ready-pool of some sort.
 *
 *   - A choose-next function.  The choose-next function is called to
 *     to ask the scheduler which thread to execute next.
 *
 * The interface to the scheduler is formalized in the following functions:
 *
 * void CthSchedInit()
 *
 *   - you must call this before any of the following functions will work.
 *
 * void CthSuspend()
 *
 *   - The thread calls this function, which in turn calls the scheduler's
 *     choose-next function.  It then resumes whatever thread is returned
 *     by the choose-next function.
 *
 * void CthAwaken(CthThread t)
 *
 *   - The thread-package user calls this function, which in turn calls the
 *     scheduler's awaken-function to awaken thread t.  This probably causes
 *     the thread t to be inserted in the ready-pool.
 *
 * void CthSetStrategy(CthThread t, CthAwkFn awakenfn, CthThFn choosefn)
 *
 *   - This specifies the scheduling functions to be used for thread 't'.
 *     The scheduling functions must have the following prototypes:
 *
 *          void awakenfn(CthThread t);
 *          CthThread choosefn();
 *
 *     These functions must be provided on a per-thread basis.  (Eg, if you
 *     CthAwaken a thread X, then X's awakefn will be called.  If a thread Y
 *     calls CthSuspend, then Y's choosefn will be called to pick the next
 *     thread.)  Of course, you may use the same functions for all threads
 *     (the common case), but the specification on a per-thread basis gives
 *     you maximum flexibility in controlling scheduling.
 *
 *     See also: common code, CthSetStrategyDefault.
 *
 * void CthYield()
 *
 *   - simply executes { CthAwaken(CthSelf()); CthSuspend(); }.  This
 *     combination gives up control temporarily, but ensures that control
 *     will eventually return.
 *
 *
 * Note: there are several possible ways to implement threads.
 *****************************************************************************/

#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#include "converse.h"

#if CMK_MEMORY_PROTECTABLE
#if CMK_HAS_MALLOC_H
#include <malloc.h> /*<- for memalign*/
#else
CLINKAGE void *memalign(size_t align, size_t size) CMK_THROW;
#endif
#endif

#include "qt.h"
#include "memory-isomalloc.h"

#include "conv-trace.h"
#include <sys/types.h>

#ifndef _WIN32
#include "valgrind.h"
#endif

#ifndef CMK_STACKSIZE_DEFAULT
#define CMK_STACKSIZE_DEFAULT 32768
#endif

#if ! CMK_THREADS_BUILD_DEFAULT
#undef CMK_THREADS_USE_JCONTEXT
#undef CMK_THREADS_USE_FCONTEXT
#undef CMK_THREADS_USE_CONTEXT
#undef CMK_THREADS_ARE_WIN32_FIBERS
#undef CMK_THREADS_USE_PTHREADS

#if CMK_THREADS_BUILD_CONTEXT
#define CMK_THREADS_USE_CONTEXT       1
#elif CMK_THREADS_BUILD_FCONTEXT
#define CMK_THREADS_USE_FCONTEXT      1
#elif CMK_THREADS_BUILD_JCONTEXT
#define CMK_THREADS_USE_JCONTEXT       1
#elif  CMK_THREADS_BUILD_FIBERS
#define CMK_THREADS_ARE_WIN32_FIBERS  1
#elif  CMK_THREADS_BUILD_PTHREADS
#define CMK_THREADS_USE_PTHREADS      1
#elif  CMK_THREADS_BUILD_STACKCOPY
#define CMK_THREADS_USE_STACKCOPY      1
#endif

#endif

#if CMK_THREADS_BUILD_TLS
#include "cmitls.h"
#endif
#define CthDebug(...)  //CmiPrintf(__VA_ARGS__)

  /**************************** Shared Base Thread Class ***********************/
  /*
     CMK_FAULT_EVAC
     Moved the cmicore converse header from CthThreadBase to CthThreadToken.
     The CthThreadToken gets enqueued in the converse queue instead of the
     CthThread. This allows the thread to be moved out of a processor even
     if there is an awaken call for it enqueued in the scheduler

*/

#define THD_MAGIC_NUM 0x12345678

  typedef struct CthThreadBase
{
  CthThreadToken *token; /* token that shall be enqueued into the ready queue*/
  int scheduled;         /* has this thread been added to the ready queue ? */

  CmiObjId   tid;        /* globally unique tid */
  CthAwkFn   awakenfn;   /* Insert this thread into the ready queue */
  CthThFn    choosefn;   /* Return the next ready thread */
  CthThread  next; /* Next active thread */
#if CMK_OMP
  CthThread  prev; /* Previous active thread */
#endif
  int        suspendable; /* Can this thread be blocked */
  int        exiting;    /* Is this thread finished */

  char      *data;       /* thread private data */
  size_t     datasize;   /* size of thread-private data, in bytes */

  int isMigratable; /* thread is migratable (isomalloc or aliased stack) */
#if CMK_THREADS_ALIAS_STACK
  int aliasStackHandle; /* handle for aliased stack */
#endif
  CmiIsomallocBlockList *isomallocBlockList;

  void      *stack; /*Pointer to thread stack*/
  int        stacksize; /*Size of thread stack (bytes)*/
  struct CthThreadListener *listener; /* pointer to the first of the listeners */

#ifndef _WIN32
  unsigned int valgrindStackID;
#endif

#if CMK_THREADS_BUILD_TLS
  tlsseg_t tlsseg;
#endif

#if CMK_TRACE_ENABLED
  int eventID;
  int srcPE;
#endif
  int magic; /* magic number for checking corruption */

} CthThreadBase;

/* By default, there are no flags */
static int CmiThreadIs_flag=0;

int CmiThreadIs(int flag)
{
  return (CmiThreadIs_flag&flag)==flag;
}

/*Macros to convert between base and specific thread types*/
#define B(t) ((CthThreadBase *)(t))
#define S(t) ((CthThread)(t))


CthThreadToken *CthGetToken(CthThread t){
  return B(t)->token;
}

CpvStaticDeclare(int, Cth_serialNo);

/*********************** Stack Aliasing *********************
  Stack aliasing: instead of consuming virtual address space
  with isomalloc, map all stacks to the same virtual addresses
  (like stack copying), but use the VM hardware to make thread
  swaps fast.  This implementation uses *files* to store the stacks,
  and mmap to drop the stack data into the right address.  Despite
  the presence of files, at least under Linux with local disks context
  switches are still fast; the mmap overhead is less than 5us per
  switch, even for multi-megabyte stacks.

WARNING: Does NOT work in SMP mode, because all the processors
on a node will try to map their stacks to the same place.
WARNING: Does NOT work if switching directly from one migrateable
thread to another, because it blows away the stack you're currently
running on.  The usual Charm approach of switching to the 
(non-migratable) scheduler thread on context switch works fine.
*/

#ifndef CMK_THREADS_ALIAS_STACK
#define CMK_THREADS_ALIAS_STACK 0
#endif

#if CMK_OSF1
#define CMK_THREADS_ALIAS_LOCATION   ((void *)0xe00000000)
#else
#define CMK_THREADS_ALIAS_LOCATION   ((void *)0xb0000000)
#endif

#if CMK_THREADS_ALIAS_STACK
#include <stdlib.h> /* for mkstemp */
#include <sys/mman.h> /* for mmap */
#include <errno.h> /* for perror */
#include <unistd.h> /* for unlink, lseek, close */

int CthAliasCreate(int stackSize)
{
  /* Make a file to contain thread stack */
  char tmpName[128];
  char lastByte=0;
  int fd;
  sprintf(tmpName,"/tmp/charmThreadStackXXXXXX");
  fd=mkstemp(tmpName);
  if (fd==-1) CmiAbort("threads.C> Cannot create /tmp file to contain thread stack");
  unlink(tmpName); /* delete file when it gets closed */

  /* Make file big enough for stack, by writing one byte at end */
  lseek(fd,stackSize-sizeof(lastByte),SEEK_SET);
  if (write(fd,&lastByte,sizeof(lastByte)) != sizeof(lastByte)) {
     CmiAbort("CthThread> writing thread stack to file failed!");
  }

  return fd;
}

void CthAliasFree(int fd) {
  close(fd);
}

#endif

#if CMK_THREADS_ALIAS_STACK
CthThreadBase *_curMappedStack=0;
void CthAliasEnable(CthThreadBase *t) {
  void *s;
  int flags=MAP_FIXED|MAP_SHARED; /* Posix flags */
  if (!t->isMigratable) return;
  if (t==_curMappedStack) return; /* don't re-map */
  _curMappedStack=t;
  CthDebug("Mmapping in thread %p from runtime stack %p\n", (void*)t, (void*)&s);

  /* Linux mmap flag MAP_POPULATE, to pre-fault in all the pages,
     only seems to slow down overall performance. */
  /* Linux mmap flag MAP_GROWSDOWN is rejected at runtime under 2.4.25 */
#if CMK_AIX
  if (_curMappedStack) munmap(_curMappedStack->stack,_curMappedStack->stacksize);
#endif
  s=mmap(t->stack,t->stacksize,
      PROT_READ|PROT_WRITE|PROT_EXEC, /* exec for gcc nested function thunks */
      flags, t->aliasStackHandle,0);
  if (s!=t->stack) {
    perror("threads.C CthAliasEnable mmap");
    CmiAbort("threads.C CthAliasEnable mmap failed");
  }
}
#else
#define CthAliasEnable(t) /* empty */
#endif





/*********** Thread-local storage *********/

CthCpvStatic(CthThread,  CthCurrent); /*Current thread*/
CthCpvDeclare(char *,    CthData); /*Current thread's private data (externally visible)*/
CthCpvStatic(size_t,     CthDatasize);

void CthSetThreadID(CthThread th, int a, int b, int c)
{
  B(th)->tid.id[0] = a;
  B(th)->tid.id[1] = b;
  B(th)->tid.id[2] = c;
}

/* possible hack? CW */
CmiObjId *CthGetThreadID(CthThread th)
{
  return &(B(th)->tid);
}

char *CthGetData(CthThread t) { return B(t)->data; }

/* Ensure this thread has at least enough 
   room for all the thread-local variables 
   initialized so far on this processor.
   */
#if CMK_C_INLINE
inline
#endif
static void CthFixData(CthThread t)
{
  size_t newsize = CthCpvAccess(CthDatasize);
  size_t oldsize = B(t)->datasize;
  if (oldsize < newsize) {
    newsize = 2*newsize;
    B(t)->datasize = newsize;
    /* Note: realloc(NULL,size) is equivalent to malloc(size) */
    B(t)->data = (char *)realloc(B(t)->data, newsize);
    memset(B(t)->data+oldsize, 0, newsize-oldsize);
  }
}

size_t CthRegister(size_t size)
{
  size_t datasize=CthCpvAccess(CthDatasize);
  CthThreadBase *th=(CthThreadBase *)CthCpvAccess(CthCurrent);
  size_t result, align = 1;
  while (size>align) align<<=1;
  datasize = (datasize+align-1) & ~(align-1);
  result = datasize;
  datasize += size;
  CthCpvAccess(CthDatasize) = datasize;
  CthFixData(S(th)); /*Make the current thread have this much storage*/
  CthCpvAccess(CthData) = th->data;
  return result;
}

void CthRegistered(size_t maxOffset) {
  if (CthCpvAccess(CthDatasize)<maxOffset) {
    CthThreadBase *th=(CthThreadBase *)CthCpvAccess(CthCurrent);
    CthCpvAccess(CthDatasize) = maxOffset;
    CthFixData(S(th)); /*Make the current thread have this much storage*/
    CthCpvAccess(CthData) = th->data;
  }
}

/*********** Creation and Deletion **********/
CthCpvStatic(int, _defaultStackSize);

void CthSetSerialNo(CthThread t, int no)
{
  B(t)->token->serialNo = no;
}

static void CthThreadBaseInit(CthThreadBase *th)
{
  static int serialno = 1;
  th->token = (CthThreadToken *)malloc(sizeof(CthThreadToken));
  th->token->thread = S(th);
#if CMK_BIGSIM_CHARM
  th->token->serialNo = -1;
#else
  th->token->serialNo = CpvAccess(Cth_serialNo)++;
#endif
  th->scheduled = 0;

  th->awakenfn = 0;
  th->choosefn = 0;
  th->next=0;
  th->suspendable = 1;
  th->exiting = 0;

  th->data=0;
  th->datasize=0;
  CthFixData(S(th));

  CthSetStrategyDefault(S(th));

  th->isMigratable=0;
#if CMK_THREADS_ALIAS_STACK
  th->aliasStackHandle=0;
#endif
  th->isomallocBlockList = NULL;

  th->stack=NULL;
  th->stacksize=0;

  th->tid.id[0] = CmiMyPe();
  CmiMemoryAtomicFetchAndInc(serialno, th->tid.id[1]);
#if CMK_OMP
  th->tid.id[2] = -1;
#else
  th->tid.id[2] = 0;
#endif

  th->listener = NULL;

  th->magic = THD_MAGIC_NUM;
}

static void *CthAllocateStack(CthThreadBase *th,int *stackSize,int useMigratable)
{
  void *ret=NULL;
  if (*stackSize==0) *stackSize=CthCpvAccess(_defaultStackSize);
  th->stacksize=*stackSize;
  if (!useMigratable || !CmiIsomallocEnabled()) {
    ret=malloc(*stackSize); 
  } else {
    th->isMigratable=1;
#if CMK_THREADS_ALIAS_STACK
    th->aliasStackHandle=CthAliasCreate(*stackSize);
    ret=CMK_THREADS_ALIAS_LOCATION;
#else /* isomalloc */
    if (th->isomallocBlockList == NULL)
      th->isomallocBlockList = CmiIsomallocBlockListNew();
#if defined(__APPLE__) && CMK_64BIT
      /* MAC OS needs 16-byte aligned stack */
    ret=CmiIsomallocBlockListMallocAlign(th->isomallocBlockList, 16, *stackSize);
#else
    ret=CmiIsomallocBlockListMalloc(th->isomallocBlockList, *stackSize);
#endif
#endif
  }
  _MEMCHECK(ret);
  th->stack=ret;

#ifndef _WIN32
  th->valgrindStackID = VALGRIND_STACK_REGISTER(ret, (char *)ret + *stackSize);
#endif

  return ret;
}
static void CthThreadBaseFree(CthThreadBase *th)
{
  struct CthThreadListener *l,*lnext;
  /*
   * remove the token if it is not queued in the converse scheduler     
   */
  if(th->scheduled == 0){
    free(th->token);
  }else{
    th->token->thread = NULL;
  }
  /* Call the free function pointer on all the listeners on
     this thread and also delete the thread listener objects
     */
  for(l=th->listener;l!=NULL;l=lnext){
    lnext=l->next;
    l->next=0;
    if (l->free) l->free(l);
  }
  th->listener = NULL;
  free(th->data);
  if (th->isMigratable) {
#if CMK_THREADS_ALIAS_STACK
    CthAliasFree(th->aliasStackHandle);
#endif
  } 
  else if (th->stack!=NULL) {
    free(th->stack);
  }
#ifndef _WIN32
  VALGRIND_STACK_DEREGISTER(th->valgrindStackID);
#endif
  th->stack=NULL;

  if (th->isomallocBlockList)
  {
    CmiIsomallocBlockListDelete(th->isomallocBlockList);
    th->isomallocBlockList = NULL;
  }
}

#if CMK_THREADS_BUILD_TLS
CthCpvStatic(tlsseg_t, _ctgTLS);

// Generic TLS interface for callers that don't know which thread
// they are switching to (main thread or CthThread).
void CtgInstallTLS(void *cur, void *next)
{
  tlsseg_t *newtls = (tlsseg_t *)next;
  if (newtls == NULL)   newtls = &CthCpvAccess(_ctgTLS);
  switchTLS((tlsseg_t *)cur, newtls);
}

// TLS interface for callers who know if they are context switching
// to the main thread or a CthThread (this avoids branching).
void CtgInstallMainThreadTLS(void *cur)
{
  switchTLS((tlsseg_t *)cur, &CthCpvAccess(_ctgTLS));
}

void CtgInstallCthTLS(void *cur, void *next)
{
  CmiAssert(next != NULL);
  switchTLS((tlsseg_t *)cur, (tlsseg_t *)next);
}

static tlsseg_t  _oldtlsseg[128] = {0};
static int       tlsseg_ptr = -1;
void CmiDisableTLS(void)
{
  tlsseg_ptr++;
  CmiAssert(tlsseg_ptr < 128);
  switchTLS(&_oldtlsseg[tlsseg_ptr], &CthCpvAccess(_ctgTLS));
}

void CmiEnableTLS(void)
{
  tlsseg_t  dummy;
  switchTLS(&dummy, &_oldtlsseg[tlsseg_ptr]);
  tlsseg_ptr --;
}
#else
void CtgInstallTLS(void *cur, void *next) {}
void CtgInstallMainThreadTLS(void *cur) {}
void CtgInstallCthTLS(void *cur, void *next) {}
void CmiDisableTLS(void) {}
void CmiEnableTLS(void) {}
#endif

static void CthBaseInit(char **argv)
{
  char *str;

  CthCpvInitialize(int,  _defaultStackSize);
  CthCpvAccess(_defaultStackSize)=CMK_STACKSIZE_DEFAULT;
/*
  CmiGetArgIntDesc(argv,"+stacksize",&CthCpvAccess(_defaultStackSize),
      "Default user-level thread stack size");  
*/
  if (CmiGetArgStringDesc(argv,"+stacksize",&str,"Default user-level thread stack size"))  {
      CthCpvAccess(_defaultStackSize) = CmiReadSize(str);
  }

  CthCpvInitialize(CthThread,  CthCurrent);
  CthCpvInitialize(char *, CthData);
  CthCpvInitialize(size_t, CthDatasize);

  CthCpvAccess(CthData)=0;
  CthCpvAccess(CthDatasize)=0;

  CpvInitialize(int, Cth_serialNo);
  CpvAccess(Cth_serialNo) = 1;

#if CMK_THREADS_BUILD_TLS
  CmiTLSInit();
  CthCpvInitialize(tlsseg_t, _ctgTLS);
  CmiThreadIs_flag |= CMI_THREAD_IS_TLS;
  currentTLS(&CthCpvAccess(_ctgTLS));
#endif
}

int CthImplemented(void) { return 1; }

CthThread CthSelf(void)
{
  return CthCpvAccess(CthCurrent);
}

void CthPupBase(pup_er p,CthThreadBase *t,int useMigratable)
{
#if CMK_ERROR_CHECKING
  if (!pup_isSizing(p) && (CthThread)t==CthCpvAccess(CthCurrent))
    CmiAbort("CthPupBase: Cannot pack running thread!");
#endif
  /*
   * Token will never be freed, so its pointer should be pupped.
   * When packing, set the thread pointer in this token to be NULL.
   * When unpacking, reset the thread pointer in token to this thread.
   */

  if(_BgOutOfCoreFlag!=0){
    pup_bytes(p, &t->token, sizeof(void *));
    if(!pup_isUnpacking(p)){
      t->token->thread = NULL;
    }
    pup_int(p, &t->scheduled);
  }
  if(pup_isUnpacking(p)){
    if(_BgOutOfCoreFlag==0){
      t->token = (CthThreadToken *)malloc(sizeof(CthThreadToken));
      t->token->thread = S(t);
      t->token->serialNo = CpvAccess(Cth_serialNo)++;
      /*For normal runs where this pup is needed,
        set scheduled to 0 in the unpacking period since the thread has
        not been scheduled */
      t->scheduled = 0;
    }else{
      /* During out-of-core emulation */
      /* 
       * When t->scheduled is set, the thread is in the queue so the token
       * should be kept. Otherwise, allocate a new space for the token
       */
      if(t->scheduled==0){
        /*CmiPrintf("Creating a new token for %p!\n", t->token);*/
        t->token = (CthThreadToken *)malloc(sizeof(CthThreadToken));
      }
      t->token->thread = S(t);
      t->token->serialNo = CpvAccess(Cth_serialNo)++;
    }
  }

  /*BIGSIM_OOC DEBUGGING */
  /*if(_BgOutOfCoreFlag!=0){
    if(pup_isUnpacking(p)){
    CmiPrintf("Unpacking: ");
    }else{
    CmiPrintf("Packing: ");
    }   
    CmiPrintf("thd=%p, its token=%p, token's thd=%p\n", t, t->token, t->token->thread); 
    } */

  /*Really need a pup_functionPtr here:*/
  pup_bytes(p,&t->awakenfn,sizeof(t->awakenfn));
  pup_bytes(p,&t->choosefn,sizeof(t->choosefn));
  pup_bytes(p,&t->next,sizeof(t->next));
  pup_int(p,&t->suspendable);
  pup_size_t(p,&t->datasize);
  if (pup_isUnpacking(p)) { 
    t->data = (char *) malloc(t->datasize);_MEMCHECK(t->data);
  }
  pup_bytes(p,(void *)t->data,t->datasize);
  pup_int(p,&t->isMigratable);
  pup_int(p,&t->stacksize);
  CmiIsomallocBlockListPup(p, &t->isomallocBlockList);
  if (t->isMigratable) {
#if CMK_THREADS_ALIAS_STACK
    if (pup_isUnpacking(p)) { 
      CthAllocateStack(t,&t->stacksize,1);
    }
    CthAliasEnable(t);
    pup_bytes(p,t->stack,t->stacksize);
#elif CMK_THREADS_USE_STACKCOPY
    /* do nothing */
#else /* isomalloc */
    pup_bytes(p,&t->stack,sizeof(char*));
#endif
  } 
  else {
    if (useMigratable)
      CmiAbort("You must use CthCreateMigratable to use CthPup!\n");
    /*Pup the stack pointer as raw bytes*/
    pup_bytes(p,&t->stack,sizeof(t->stack));
  }
  if (pup_isUnpacking(p)) { 
    /* FIXME:  restore thread listener */
    t->listener = NULL;
  }

  pup_int(p, &t->magic);

#if CMK_THREADS_BUILD_TLS && CMK_HAS_TLS_VARIABLES
  {
    void* aux;
    pup_bytes(p, &t->tlsseg, sizeof(tlsseg_t));
    aux = ((char*)(t->tlsseg.memseg)) - t->tlsseg.size;
    /* fixme: tls global variables handling needs isomalloc */
    pup_bytes(p,&aux,sizeof(char*));
    /* printf("[%d] %s %p\n", CmiMyPe(), pup_typeString(p), t->tlsseg.memseg); */
  }
#endif
}

static void CthThreadFinished(CthThread t)
{
  B(t)->exiting=1;
  CthSuspend();
}


/************ Scheduler Interface **********/

void CthSetSuspendable(CthThread t, int val) { B(t)->suspendable = val; }
int CthIsSuspendable(CthThread t) { return B(t)->suspendable; }

void CthSetNext(CthThread t, CthThread v) { B(t)->next = v; }
CthThread CthGetNext(CthThread t) { return B(t)->next; }
#if CMK_OMP
void CthSetPrev(CthThread t, CthThread v) { B(t)->prev = v;}
#endif
#if CMK_TRACE_ENABLED
void CthSetEventInfo(CthThread t, int event, int srcPE ) {
  B(t)->eventID = event;
  B(t)->srcPE = srcPE;
}
#endif
static void CthNoStrategy(void)
{
  CmiAbort("Called CthAwaken or CthSuspend before calling CthSetStrategy.\n");
}

void CthSetStrategy(CthThread t, CthAwkFn awkfn, CthThFn chsfn)
{
  B(t)->awakenfn = awkfn;
  B(t)->choosefn = chsfn;
}

#if CMK_OMP
inline int CthScheduled(CthThread t) {
  return B(t)->scheduled;
}

/* The next scheduled thread decrements 'scheduled' of the previous thread.*/
void CthScheduledDecrement() {
    CthThread prevCurrent = B(CthSelf())->prev;
    if (!B(prevCurrent))
      return;
    CthDebug("[%f][%d] scheduled before decremented: %d\n", CmiWallTimer(), CmiMyRank(), B(prevCurrent)->scheduled);
    /* CthMainThread should have empty stack(stack == NULL) and never scheduled(scheduled == 0) */
    if (B(prevCurrent)->stack && B(prevCurrent)->scheduled > 0) {
        CmiMemoryAtomicDecrement(B(prevCurrent)->scheduled, memory_order_release);
        CthDebug("[%f][%d] scheduled decremented: %d\n", CmiWallTimer(), CmiMyRank(), B(prevCurrent)->scheduled);
    }
#if CMK_ERROR_CHECKING
    if(B(prevCurrent)->scheduled < 0)
      CmiAbort("A thread's scheduler should not be less than 0!\n");
#endif
}
#endif

#if CMK_C_INLINE
inline
#endif
static void CthBaseResume(CthThread t)
{
  struct CthThreadListener *l;
  for(l=B(t)->listener;l!=NULL;l=l->next){
    if (l->resume) l->resume(l);
  }
  CthFixData(t); /*Thread-local storage may have changed in other thread.*/
  CthCpvAccess(CthCurrent) = t;
  CthCpvAccess(CthData) = B(t)->data;
  CthAliasEnable(B(t));
}

void CthSwitchThread(CthThread t)
{
  CthBaseResume(t);
}

#if CMK_ERROR_CHECKING

/* check for sanity of the thread:
   1- stack might have grown too much
   2- memory corruptions might have occured
 */
void CthCheckThreadSanity(void)
{
#if !CMK_THREADS_USE_FCONTEXT
  /* use the address of a dummy variable on stack to see how large the stack is currently */
  int tmp;
  char* curr_stack;
  char* base_stack;
  CthThreadBase *base_thread=B(CthCpvAccess(CthCurrent));
  
  curr_stack = (char*)(&tmp);
  base_stack = (char*)(base_thread->stack);

  /* stack pointer should be between start and end addresses of stack, regardless of direction */ 
  /* check to see if we actually allocated a stack (it is not main thread) */
  if ( base_thread->magic != THD_MAGIC_NUM ||
      (base_stack != 0 && (curr_stack < base_stack || curr_stack > base_stack + base_thread->stacksize)))
    CmiAbort("Thread meta data is not sane! Check for memory corruption and stack overallocation. Use +stacksize to"
        "increase stack size or allocate in heap instead of stack.");
#endif
}
#endif


/*
Suspend: finds the next thread to execute, and resumes it
*/
void CthSuspend(void)
{

  CthThread next;
  struct CthThreadListener *l;
  CthThreadBase *cur=B(CthCpvAccess(CthCurrent));

#if CMK_ERROR_CHECKING
  CthCheckThreadSanity();
#endif

  if (cur->suspendable == 0)
    CmiAbort("Fatal Error> trying to suspend a non-suspendable thread!\n");

  /*
     Call the suspend function on listeners
     */
  for(l=cur->listener;l!=NULL;l=l->next){
    if (l->suspend) l->suspend(l);
  }
  if (cur->choosefn == 0) CthNoStrategy();
  next = cur->choosefn();
#if CMK_OMP
  cur->tid.id[2] = CmiMyRank();
#else
  /*cur->scheduled=0;*/
  /*changed due to out-of-core emulation in BigSim*/
  if(cur->scheduled > 0)
    cur->scheduled--;

#if CMK_ERROR_CHECKING
  if(cur->scheduled<0)
    CmiAbort("A thread's scheduler should not be less than 0!\n");
#endif
#endif
#if CMK_TRACE_ENABLED
#if !CMK_TRACE_IN_CHARM
  if(CpvAccess(traceOn))
    traceSuspend();
#endif
#endif
  CthDebug("[%f] next(%p) resumed\n",CmiWallTimer(), next);
#if CMK_OMP
  /* If this thread is supposed to terminate after CthResume, then the next thread cannot get access to this thread to decrement 'scheduled' */
  if (cur->exiting)
    CthSetPrev(next, 0);
  else
    CthSetPrev(next, CthCpvAccess(CthCurrent));
#endif
  CthResume(next);
#if CMK_OMP
  CthScheduledDecrement();
  CthSetPrev(CthSelf(), 0);
#endif
}

void CthAwaken(CthThread th)
{
  if (B(th)->awakenfn == 0) CthNoStrategy();

  /*BIGSIM_OOC DEBUGGING
    if(B(th)->scheduled==1){
    CmiPrintf("====Thread %p is already scheduled!!!!\n", th);
    return;
    } */

#if CMK_TRACE_ENABLED
#if ! CMK_TRACE_IN_CHARM
  if(CpvAccess(traceOn))
    traceAwaken(th);
#endif
#endif

  B(th)->scheduled++;
  B(th)->awakenfn(B(th)->token, CQS_QUEUEING_FIFO, 0, 0);
  /*B(th)->scheduled = 1; */
  /*changed due to out-of-core emulation in BigSim */

}

void CthYield(void)
{
#if CMK_OMP
  B(CthCpvAccess(CthCurrent))->scheduled--;
#endif
  CthAwaken(CthCpvAccess(CthCurrent));
  CthSuspend();
}

void CthAwakenPrio(CthThread th, int s, int pb, unsigned int *prio)
{
  if (B(th)->awakenfn == 0) CthNoStrategy();
#if CMK_TRACE_ENABLED
#if ! CMK_TRACE_IN_CHARM
  if(CpvAccess(traceOn))
    traceAwaken(th);
#endif
#endif
  B(th)->awakenfn(B(th)->token, s, pb, prio);
  /*B(th)->scheduled = 1; */
  /*changed due to out-of-core emulation in BigSim */
  B(th)->scheduled++;
}

void CthYieldPrio(int s, int pb, unsigned int *prio)
{
  CthAwakenPrio(CthCpvAccess(CthCurrent), s, pb, prio);
  CthSuspend();
}

/*
   Add a new thread listener to a thread 
   */
void CthAddListener(CthThread t,struct CthThreadListener *l){
  struct CthThreadListener *p=B(t)->listener;
  if(p== NULL){ /* first listener */
    B(t)->listener=l;
    l->thread = t;
    l->next=NULL;
    return; 
  }
  /* Add l at end of current chain of listeners: */
  while(p->next != NULL){
    p = p->next;
  }
  p->next = l;
  l->next = NULL;
  l->thread = t;
}

/*************************** Stack-Copying Threads (obsolete) *******************
  Basic idea: switch from thread A (currently running) to thread B by copying
  A's stack from the system stack area into A's buffer in the heap, then
  copy B's stack from its heap buffer onto the system stack.

  This allows thread migration, because the system stack is in the same
  location on every processor; but the context-switching overhead (especially
  for threads with deep stacks) is extremely high.

  Written by Josh Yelon around May 1999

  stack grows down like:
  lo   <- savedptr    <- savedstack

  ...

  high <- stackbase

NOTE: this only works when system stack base is same on all processors.
Which is not the case on my FC4 laptop ?!

extended it to work for tcharm and registered user data migration
tested platforms: opteron, Cygwin.

For Fedora and Ubuntu, run the following command as root
echo 0 > /proc/sys/kernel/randomize_va_space
will disable the randomization of the stack pointer

Gengbin Zheng March, 2006
*/

#if CMK_THREADS_USE_STACKCOPY

#define SWITCHBUF_SIZE 32768

typedef struct CthProcInfo_s *CthProcInfo;

typedef struct CthThreadStruct
{
  CthThreadBase base;
  CthVoidFn  startfn;    /* function that thread will execute */
  void      *startarg;   /* argument that start function will be passed */
  qt_t      *savedstack; /* pointer to saved stack */
  int        savedsize;  /* length of saved stack (zero when running) */    
  int        stacklen;   /* length of the allocated savedstack >= savedsize */
  qt_t      *savedptr;   /* stack pointer */    
} CthThreadStruct;

int CthMigratable(void)
{
  return 1;
}

CthThread CthPup(pup_er p, CthThread t)
{
  if (pup_isUnpacking(p))
  { t = (CthThread) malloc(sizeof(CthThreadStruct));_MEMCHECK(t);}
  pup_bytes(p, (void*) t, sizeof(CthThreadStruct)); 
  CthPupBase(p,&t->base,0);
  pup_int(p,&t->savedsize);
  if (pup_isUnpacking(p)) {
    t->savedstack = (qt_t*) malloc(t->savedsize);_MEMCHECK(t->savedstack);
    t->stacklen = t->savedsize;       /* reflect actual size */
  }
  pup_bytes(p, (void*) t->savedstack, t->savedsize);

  /* assume system stacks are same on all processors !! */
  pup_bytes(p,&t->savedptr,sizeof(t->savedptr));  

  if (pup_isDeleting(p))
  {CthFree(t);t=0;}
  return t;
}



struct CthProcInfo_s
{
  qt_t      *stackbase;
  qt_t      *switchbuf_sp;
  qt_t      *switchbuf;
};

CthCpvDeclare(CthProcInfo, CthProc);

static void CthThreadInit(CthThread t, CthVoidFn fn, void *arg)
{
  CthThreadBaseInit(&t->base);
  t->startfn = fn;
  t->startarg = arg;
  t->savedstack = 0;
  t->savedsize = 0;
  t->stacklen = 0;
  t->savedptr = 0;
}

static void CthThreadFree(CthThread t)
{
  CthThreadBaseFree(&t->base);
  if (t->savedstack) free(t->savedstack);
  free(t);
}

void CthFree(CthThread t)
{
  CthProcInfo proc;
  if (t==NULL) return;
  proc = CthCpvAccess(CthProc);

  if (t != CthSelf()) {
    CthThreadFree(t);
  } else
    t->base.exiting = 1;
}

void CthDummy(void) { }

void CthInit(char **argv)
{
  CthThread t; CthProcInfo p; qt_t *switchbuf, *sp;

  CthCpvInitialize(CthProcInfo, CthProc);

  CthBaseInit(argv);
  t = (CthThread)malloc(sizeof(struct CthThreadStruct));
  _MEMCHECK(t);
  CthCpvAccess(CthCurrent)=t;
  CthThreadInit(t,0,0);

  p = (CthProcInfo)malloc(sizeof(struct CthProcInfo_s));
  _MEMCHECK(p);
  CthCpvAccess(CthProc)=p;

  /* leave some space for current stack frame < 256 bytes */
  /* sp must be same on all processors for migration to work ! */
  sp = (qt_t*)(((size_t)&t) & ~((size_t)0xFF));
  /*printf("[%d] System stack base: %p\n", CmiMyPe(), sp);*/
  p->stackbase = QT_SP(sp, 0x100);

  /* printf("sp: %p\n", sp); */

  switchbuf = (qt_t*)malloc(QT_STKALIGN + SWITCHBUF_SIZE);
  _MEMCHECK(switchbuf);
  switchbuf = (qt_t*)((((size_t)switchbuf)+QT_STKALIGN) & ~(QT_STKALIGN-1));
  p->switchbuf = switchbuf;
  sp = QT_SP(switchbuf, SWITCHBUF_SIZE);
  sp = QT_ARGS(sp,0,0,0,(qt_only_t*)CthDummy);
  p->switchbuf_sp = sp;

  CmiThreadIs_flag |= CMI_THREAD_IS_STACKCOPY;
}

static void CthOnly(CthThread t, void *dum1, void *dum2)
{
  t->startfn(t->startarg);
  CthThreadFinished(t);
}

#define USE_SPECIAL_STACKPOINTER    1

/* p is a pointer on stack */
size_t CthStackOffset(CthThread t, char *p)
{
  CthProcInfo proc = CthCpvAccess(CthProc);
  return p - (char *)proc->stackbase;
}

char * CthPointer(CthThread t, size_t pos)
{
  char *stackbase, *p;
  CthProcInfo proc;
  CmiAssert(t);
  proc = CthCpvAccess(CthProc);
  if (CthCpvAccess(CthCurrent) == t)    /* current thread uses current stack */
    stackbase = (char *)proc->stackbase;
  else                                  /* sleep thread uses its saved stack */
    stackbase = (char *)t->savedstack;
#ifdef QT_GROW_DOWN
  p = stackbase + t->savedsize + pos;
#else
  p = stackbase + pos;
#endif
  return p;
}

static void CthResume1(qt_t *sp, CthProcInfo proc, CthThread t)
{
  int bytes; qt_t *lo, *hi;
  CthThread old = CthCpvAccess(CthCurrent);
  CthBaseResume(t);
  if (old->base.exiting) {
    CthThreadFree(old);
  } else {
#ifdef QT_GROW_DOWN
    lo = sp; hi = proc->stackbase;
#else
    hi = sp; lo = proc->stackbase;
#endif
    bytes = ((size_t)hi)-((size_t)lo);
    if(bytes > old->stacklen) {
      if(old->savedstack) free((void *)old->savedstack);
      old->savedstack = (qt_t*)malloc(bytes);
      _MEMCHECK(old->savedstack);
      old->stacklen = bytes;
    }
    old->savedsize = bytes;
    old->savedptr = sp;
    memcpy(old->savedstack, lo, bytes);
  }
  if (t->savedstack) {
#ifdef QT_GROW_DOWN
    lo = t->savedptr;
#else
    lo = proc->stackbase;
#endif
    memcpy(lo, t->savedstack, t->savedsize);
    t->savedsize=0;
    sp = t->savedptr;
  } else {
    sp = proc->stackbase;
    sp = QT_ARGS(sp,t,0,0,(qt_only_t*)CthOnly);
  }
  QT_ABORT((qt_helper_t*)CthDummy,0,0,sp);
}

void CthResume(CthThread t)
{
  CthProcInfo proc = CthCpvAccess(CthProc);
  QT_BLOCK((qt_helper_t*)CthResume1, proc, t, proc->switchbuf_sp);
}

CthThread CthCreate(CthVoidFn fn,void *arg,int size)
{
  CthThread result = (CthThread)malloc(sizeof(struct CthThreadStruct));
  _MEMCHECK(result);
  CthThreadInit(result, fn, arg);
  return result;
}
CthThread CthCreateMigratable(CthVoidFn fn,void *arg,int size)
{
  /*All threads are migratable under stack copying*/
  return CthCreate(fn,arg,size);
}

/**************************************************************************
  QuickThreads does not work on Win32-- our stack-shifting large allocas
  fail a stack depth check.  Windows NT and 98 provide a user-level thread 
  interface called "Fibers", used here.

  Written by Sameer Paranjpye around October 2000
  */
#elif  CMK_THREADS_ARE_WIN32_FIBERS
#if defined _WIN32
#include <windows.h>
#include <winbase.h>

struct CthThreadStruct
{
  CthThreadBase base;
  LPVOID     fiber;
};

CthCpvStatic(CthThread,  CthPrevious);

typedef CthThread *threadTable;
CthCpvStatic(int,     tablesize);
CthCpvStatic(threadTable, exitThreads);
CthCpvStatic(int,     nExit);

static void CthThreadInit(CthThread t)
{
  CthThreadBaseInit(&t->base);
}

void CthInit(char **argv)
{
  CthThread t;
  int i;

  CthCpvInitialize(CthThread,  CthPrevious);
  CthCpvInitialize(int,        nExit);
  CthCpvInitialize(threadTable,        exitThreads);
  CthCpvInitialize(int, tablesize);

#define INITIALSIZE 128
  CthCpvAccess(tablesize) = INITIALSIZE;   /* initial size */
  CthCpvAccess(exitThreads) = (threadTable)malloc(sizeof(CthThread)*INITIALSIZE);
  for (i=0; i<INITIALSIZE; i++) CthCpvAccess(exitThreads)[i] = NULL;

  CthCpvAccess(CthPrevious)=0;
  CthCpvAccess(nExit)=0;

  CthBaseInit(argv);  
  t = (CthThread)malloc(sizeof(struct CthThreadStruct));
  _MEMCHECK(t);
  CthCpvAccess(CthCurrent)=t;
  CthThreadInit(t);
  t->fiber = ConvertThreadToFiber(t);
  _MEMCHECK(t->fiber);

  CmiThreadIs_flag |= CMI_THREAD_IS_FIBERS;
}

void CthThreadFree(CthThread old)
{
  CthThreadBaseFree(&old->base);
  if (old->fiber) DeleteFiber((PVOID)old->fiber);
  free(old);
}

static void CthClearThreads(void)
{
  int i,p,m;
  int n = CthCpvAccess(nExit);
  CthThread tc = CthCpvAccess(CthCurrent);
  CthThread tp = CthCpvAccess(CthPrevious);
  m = n;
  p=0;
  for (i=0; i<m; i++) {
    CthThread t = CthCpvAccess(exitThreads)[i];
    if (t && t != tc && t != tp) {
      CthThreadFree(t);
      CthCpvAccess(nExit) --;
    }
    else {
      if (p != i) CthCpvAccess(exitThreads)[p] = t;
      p++;
    }
  }
  if (m!=p)
    for (i=m; i<n; i++,p++) {
      CthCpvAccess(exitThreads)[p] = CthCpvAccess(exitThreads)[i];
    }
}

void CthFree(CthThread t)
{
  int i;
  if (t==NULL) return;

  if(CthCpvAccess(nExit) >= CthCpvAccess(tablesize)) {   /* expand */
    threadTable newtable;
    int oldsize = CthCpvAccess(tablesize);
    CthCpvAccess(tablesize) *= 2;
    newtable = (threadTable)malloc(sizeof(CthThread)*CthCpvAccess(tablesize));
    for (i=0; i<CthCpvAccess(tablesize); i++) newtable[i] = NULL;
    for (i=0; i<oldsize; i++) newtable[i] = CthCpvAccess(exitThreads)[i];
    free(CthCpvAccess(exitThreads));
    CthCpvAccess(exitThreads) = newtable;
  }

  /* store into exiting threads table to avoid delete thread itself */
  CthCpvAccess(exitThreads)[CthCpvAccess(nExit)++] = t;
  if (t==CthCpvAccess(CthCurrent)) 
  {
    t->base.exiting = 1;
  } 
  else 
  {
    CthClearThreads();
    /*  was
        if (t->data) free(t->data);
        DeleteFiber(t->fiber);
        free(t);
        */
  }
}

#if 0
void CthFiberBlock(CthThread t)
{
  CthThread tp;

  SwitchToFiber(t->fiber);
  tp = CthCpvAccess(CthPrevious);
  if (tp != 0 && tp->killed == 1)
    CthThreadFree(tp);
}
#endif

void CthResume(CthThread t)
{
  CthThread tc;

  tc = CthCpvAccess(CthCurrent);
  if (t == tc) return;
  CthBaseResume(t);
  CthCpvAccess(CthPrevious)=tc;
#if 0
  if (tc->base.exiting) 
  {
    SwitchToFiber(t->fiber);
  } 
  else 
    CthFiberBlock(t);
#endif
  SwitchToFiber(t->fiber);
}

VOID CALLBACK FiberSetUp(PVOID fiberData)
{
  void **ptr = (void **) fiberData;
  qt_userf_t* fn = (qt_userf_t *)ptr[0];
  void *arg = ptr[1];
  CthThread  t = CthSelf();

  CthClearThreads();

  fn(arg);

  CthCpvAccess(exitThreads)[CthCpvAccess(nExit)++] = t;
  CthThreadFinished(t);
}

CthThread CthCreate(CthVoidFn fn, void *arg, int size)
{
  CthThread result; 
  void**    fiberData;
  fiberData = (void**)malloc(2*sizeof(void *));
  fiberData[0] = (void *)fn;
  fiberData[1] = arg;

  result = (CthThread)malloc(sizeof(struct CthThreadStruct));
  _MEMCHECK(result);
  CthThreadInit(result);
  /* result->fiber = CreateFiber(size, FiberSetUp, (PVOID) fiberData); */
  result->fiber = CreateFiberEx(size, size, 0, FiberSetUp, (PVOID) fiberData);
  if (!result->fiber)
    CmiAbort("CthCreate failed to create fiber!\n");

  return result;
}

int CthMigratable(void)
{
  return 0;
}
CthThread CthPup(pup_er p, CthThread t)
{
  CmiAbort("CthPup not implemented.\n");
  return 0;
}
CthThread CthCreateMigratable(CthVoidFn fn,void *arg,int size)
{
  /*Fibers are never migratable, unless we can figure out how to set their stacks*/
  return CthCreate(fn,arg,size);
}
#else /* defined _WIN32 */
struct CthThreadStruct
{
  CthThreadBase base;
};
#endif /* defined _WIN32 */

/***************************************************
  Use Posix Threads to simulate cooperative user-level
  threads.  This version is very portable but inefficient.

  Written by Milind Bhandarkar around November 2000

IMPORTANT:  for SUN, must link with -mt compiler flag
Rewritten by Gengbin Zheng
*/
#elif CMK_THREADS_USE_PTHREADS

#include <pthread.h>
#include <errno.h> /* for perror */

struct CthThreadStruct
{
  CthThreadBase base;
  pthread_t  self;
  pthread_cond_t cond;
  pthread_cond_t *creator;
  CthVoidFn  fn;
  void      *arg;
  char       inited;
};

CthCpvStatic(pthread_mutex_t, sched_mutex);

static void CthThreadInit(CthThread t)
{
  CthThreadBaseInit(&t->base);
  t->inited = 0;
  pthread_cond_init(&(t->cond) , (pthread_condattr_t *) 0);
}

void CthInit(char **argv)
{
  CthThread t;

  CthCpvInitialize(pthread_mutex_t, sched_mutex);

  pthread_mutex_init(&CthCpvAccess(sched_mutex), (pthread_mutexattr_t *) 0);
  pthread_mutex_lock(&CthCpvAccess(sched_mutex));
  CthBaseInit(argv); 
  t = (CthThread)malloc(sizeof(struct CthThreadStruct));
  _MEMCHECK(t);
  CthCpvAccess(CthCurrent)=t;
  CthThreadInit(t);
  t->self = pthread_self();

  CmiThreadIs_flag |= CMI_THREAD_IS_PTHREADS;
}

void CthFree(CthThread t)
{
  if (t==NULL) return;
  if (t==CthCpvAccess(CthCurrent)) {
    t->base.exiting = 1;
  } else {
    CthThreadBaseFree(&t->base);
    free(t);
  }
}

void CthResume(CthThread t)
{
  CthThread tc = CthCpvAccess(CthCurrent);
  if (t == tc) return;
  CthBaseResume(t);
  pthread_cond_signal(&(t->cond)); /* wake up the next thread */
  if (tc->base.exiting) {
    pthread_mutex_unlock(&CthCpvAccess(sched_mutex));
    pthread_exit(0);
  } else {
    /* pthread_cond_wait might (with low probability) return when the 
       condition variable has not been signaled, guarded with 
       predicate checks */
    do {
      pthread_cond_wait(&(tc->cond), &CthCpvAccess(sched_mutex));
    } while (tc!=CthCpvAccess(CthCurrent)) ;
  }
}

static void *CthOnly(void * arg)
{
  CthThread th = (CthThread)arg;
  th->inited = 1;
  pthread_detach(pthread_self());
  pthread_mutex_lock(&CthCpvAccess(sched_mutex));
  pthread_cond_signal(th->creator);
  do {
    pthread_cond_wait(&(th->cond), &CthCpvAccess(sched_mutex));
  } while (arg!=CthCpvAccess(CthCurrent)) ;
  th->fn(th->arg);
  CthThreadFinished(th);
  return 0;
}

CthThread CthCreate(CthVoidFn fn, void *arg, int size)
{
  static int reported = 0;
  pthread_attr_t attr;
  int r;
  CthThread result;
  CthThread self = CthSelf();
  /* size is ignored in this version */
  result = (CthThread)malloc(sizeof(struct CthThreadStruct));
  _MEMCHECK(result);
  CthThreadInit(result);
  result->fn = fn;
  result->arg = arg;
  result->creator = &(self->cond);

  /* try set pthread stack, not necessarily supported on all platforms */
  pthread_attr_init(&attr);
  /* pthread_attr_setscope(&attr, PTHREAD_SCOPE_SYSTEM); */
  if (size<1024) size = CthCpvAccess(_defaultStackSize);
  if (0!=(r=pthread_attr_setstacksize(&attr,size))) {
    if (!reported) {
      CmiPrintf("Warning: pthread_attr_setstacksize failed\n");
      errno = r;
      perror("pthread_attr_setstacksize");
      reported = 1;
    }
  }

  /* **CWL** Am assuming Gengbin left this unchanged because the macro
     re-definition of pthread_create would not happen before this part of
     the code. If the assumption is not true, then we can simply replace
     this hash-if with the else portion.
     */
#if CMK_WITH_TAU
  r = tau_pthread_create(&(result->self), &attr, CthOnly, (void*) result);
#else
  r = pthread_create(&(result->self), &attr, CthOnly, (void*) result);
#endif
  if (0 != r) {
    CmiPrintf("pthread_create failed with %d\n", r);
    CmiAbort("CthCreate failed to created a new pthread\n");
  }
  do {
    pthread_cond_wait(&(self->cond), &CthCpvAccess(sched_mutex));
  } while (result->inited==0);
  return result;
}

int CthMigratable(void)
{
  return 0;
}
CthThread CthPup(pup_er p, CthThread t)
{
  CmiAbort("CthPup not implemented.\n");
  return 0;
}
CthThread CthCreateMigratable(CthVoidFn fn,void *arg,int size)
{
  /*Pthreads are never migratable, unless we can figure out how to set their stacks*/
  return CthCreate(fn,arg,size);
}

/***************************************************************
  Use SysV r3 setcontext/getcontext calls instead of
  quickthreads.  This works on lots of architectures (such as
  SUN, IBM SP, O2K, DEC Alpha, IA64, Cray X1, Linux with newer version of 
  glibc such as the one with RH9) 
  On some machine such as IA64 and Cray X1, the context version is the 
  only thread package that is working. 

  Porting should be easy. To port context threads, one need to set the 
  direction of the thread stack properly in conv-mach.h.

Note: on some machine like Sun and IBM SP, one needs to link with memory gnuold
to have this context thread working.

Written by Gengbin Zheng around April 2001

For thread checkpoint/restart, isomalloc requires that iso region from all
machines are same.

For Fedora and Ubuntu, run the following command as root
echo 0 > /proc/sys/kernel/randomize_va_space
will disable the randomization of the stack pointer

Gengbin Zheng October, 2007

*/
#elif (CMK_THREADS_USE_CONTEXT || CMK_THREADS_USE_JCONTEXT || CMK_THREADS_USE_FCONTEXT)

#include <signal.h>
#include <errno.h>

#if CMK_THREADS_USE_CONTEXT
/* system builtin context routines: */

#if defined(__APPLE__)
#define _XOPEN_SOURCE
#endif
#include <ucontext.h>
#if defined(__APPLE__)
#undef _XOPEN_SOURCE
#endif

typedef ucontext_t uJcontext_t;
typedef void (*uJcontext_fn_t)(void);
#ifdef __GNUC__
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wdeprecated-declarations"
#endif
/* function 'getJcontext' can never be inlined because it uses setjmp */
#define getJcontext(ucp) getcontext(ucp)
static CMI_FORCE_INLINE int setJcontext(const uJcontext_t *ucp)
{
  return setcontext(ucp);
}
static CMI_FORCE_INLINE int swapJcontext(uJcontext_t *oucp, const uJcontext_t *ucp)
{
  return swapcontext(oucp, ucp);
}
static CMI_FORCE_INLINE void makeJcontext(uJcontext_t *ucp, uJcontext_fn_t func, int argc, void *a1, void *a2)
{
  makecontext(ucp, func, argc, a1, a2);
}
#ifdef __GNUC__
#pragma GCC diagnostic pop
#endif
#elif CMK_THREADS_USE_FCONTEXT
#include "uFcontext.h"
#define uJcontext_t uFcontext_t
#else /* CMK_THREADS_USE_JCONTEXT */
/* Orion's setjmp-based context routines: */
#include "uJcontext.h"
#include "uJcontext.C"

#endif


struct CthThreadStruct
{
  CthThreadBase base;
  double * dummy;
  uJcontext_t context;
};


static void CthThreadInit(CthThread t)
{
  CthThreadBaseInit(&t->base);
}

/* Threads waiting to be destroyed */
CpvStaticDeclare(CthThread , doomedThreadPool);

void CthInit(char **argv)
{
  CthThread t;

  CthBaseInit(argv);
  t = (CthThread)malloc(sizeof(struct CthThreadStruct));
  _MEMCHECK(t);
  CthCpvAccess(CthCurrent)=t;
#ifdef __GNUC__
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wdeprecated-declarations"
#endif
  if (0 != getJcontext(&t->context))
    CmiAbort("CthInit: getcontext failed.\n");
#ifdef __GNUC__
#pragma GCC diagnostic pop
#endif
  CthThreadInit(t);
  CpvInitialize(CthThread, doomedThreadPool);
  CpvAccess(doomedThreadPool) = (CthThread)NULL;

  /* don't trust the _defaultStackSize */
 if (CmiMyRank() == 0) {
#ifdef MINSIGSTKSZ
    if (CthCpvAccess(_defaultStackSize) < MINSIGSTKSZ) 
      CthCpvAccess(_defaultStackSize) = MINSIGSTKSZ;
#endif
#if CMK_THREADS_USE_CONTEXT
    CmiThreadIs_flag |= CMI_THREAD_IS_CONTEXT;
#else
    CmiThreadIs_flag |= CMI_THREAD_IS_UJCONTEXT;
#endif
#if CMK_THREADS_ALIAS_STACK
    CmiThreadIs_flag |= CMI_THREAD_IS_ALIAS;
#endif
 }
}

static void CthThreadFree(CthThread t)
{
  /* avoid freeing thread while it is being used, store in pool and 
     free it next time. Note the last thread in pool won't be free'd! */
  CthThread doomed=CpvAccess(doomedThreadPool);
  CpvAccess(doomedThreadPool) = t;
  if (doomed != NULL) {
    CthThreadBaseFree(&doomed->base);
    free(doomed);
  }
}

void CthFree(CthThread t)
{
  if (t==NULL) return;
  if (t==CthCpvAccess(CthCurrent)) {
    t->base.exiting = 1; /* thread is already running-- free on next swap out */
  } else {
    CthThreadFree(t);
  }
}

#if CMK_THREADS_BUILD_TLS
void CthResume(CthThread) CMI_NOOPTIMIZE;
#endif

void CthResume(CthThread t)
{
  CthThread tc;
  tc = CthCpvAccess(CthCurrent);

#if CMK_THREADS_BUILD_TLS
  switchTLS(&B(tc)->tlsseg, &B(t)->tlsseg);
#endif

  if (t != tc) { /* Actually switch threads */
    CthBaseResume(t);
    if (!tc->base.exiting)
    {
      CthDebug("[%d][%f] swap starts from %p to %p\n",CmiMyRank(), CmiWallTimer() ,tc, t);
      if (0 != swapJcontext(&tc->context, &t->context)) {
        CmiAbort("CthResume: swapcontext failed.\n");
      }
    }
    else /* tc->base.exiting, so jump directly to next context */ 
    {
      CthThreadFree(tc);
      setJcontext(&t->context);
    }
  }

  /*This check will mistakenly fail if the thread migrates (changing tc)
    if (tc!=CthCpvAccess(CthCurrent)) { CmiAbort("Stack corrupted?\n"); }
    */
}

#if CMK_THREADS_USE_CONTEXT && CMK_64BIT /* makecontext only pass integer arguments */
void CthStartThread(CmiUInt4 fn1, CmiUInt4 fn2, CmiUInt4 arg1, CmiUInt4 arg2)
{
  CmiUInt8 fn0 =  (((CmiUInt8)fn1) << 32) | fn2;
  CmiUInt8 arg0 = (((CmiUInt8)arg1) << 32) | arg2;
  void *arg = (void *)arg0;
  qt_userf_t *fn = (qt_userf_t*)fn0;
  CthDebug("[%f] thread: %p resumed, arg: %p, fn: %p\n", CmiWallTimer(), CthSelf(), arg, fn);
  (*fn)(arg);
  CthThreadFinished(CthSelf());
}
#elif CMK_THREADS_USE_FCONTEXT
void CthStartThread(transfer_t arg)
{
  data_t *data = (data_t *)arg.data;
  uFcontext_t *old_ucp  = (uFcontext_t *)data->from;
  old_ucp->fctx = arg.fctx;
  uFcontext_t *cur_ucp = (uFcontext_t *)data->data;
  cur_ucp->func(cur_ucp->arg);
  CthThreadFinished(CthSelf());
}
#else
void CthStartThread(qt_userf_t fn,void *arg)
{
  fn(arg);
  CthThreadFinished(CthSelf());
}
#endif

#define STP_STKALIGN(sp, alignment) \
  ((void *)((((qt_word_t)(sp)) + (alignment) - 1) & ~((alignment)-1)))

int ptrDiffLen(const void *a,const void *b) {
  char *ac=(char *)a, *bc=(char *)b;
  int ret=ac-bc;
  if (ret<0) ret=-ret;
  return ret;
}

static CthThread CthCreateInner(CthVoidFn fn,void *arg,int size,int migratable)
{
  CthThread result;
  char *stack, *ss_sp, *ss_end;

  result = (CthThread)malloc(sizeof(struct CthThreadStruct));
  _MEMCHECK(result);
  CthThreadInit(result);
#ifdef MINSIGSTKSZ
  /* if (size<MINSIGSTKSZ) size = CthCpvAccess(_defaultStackSize); */
  if (size && size<MINSIGSTKSZ) size = MINSIGSTKSZ;
#endif
  CthAllocateStack(&result->base,&size,migratable);
  stack = (char *)result->base.stack;
#if !CMK_THREADS_USE_FCONTEXT
#ifdef __GNUC__
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wdeprecated-declarations"
#endif
  if (0 != getJcontext(&result->context))
    CmiAbort("CthCreateInner: getcontext failed.\n");
#ifdef __GNUC__
#pragma GCC diagnostic pop
#endif
#endif
  ss_end = stack + size;

#if CMK_THREADS_USE_JCONTEXT /* Jcontext is always STACKBEGIN */
  ss_sp = stack;
#elif CMK_THREADS_USE_FCONTEXT
  ss_sp = (char *)stack+size;
  ss_end = stack;
#elif CMK_CONTEXT_STACKEND /* ss_sp should point to *end* of buffer */
  ss_sp = stack+size-MINSIGSTKSZ; /* the MINSIGSTKSZ seems like a hack */
  ss_end = stack;
#elif CMK_CONTEXT_STACKMIDDLE /* ss_sp should point to *middle* of buffer */
  ss_sp = stack+size/2;
#else /* CMK_CONTEXT_STACKBEGIN, the usual case  */
  ss_sp = stack;
#endif

#if CMK_THREADS_USE_FCONTEXT
  result->context.uc_stack.ss_sp = ss_sp;
  result->context.uc_stack.ss_size = size;
#else
  result->context.uc_stack.ss_sp = STP_STKALIGN(ss_sp,sizeof(char *)*8);
  result->context.uc_stack.ss_size = ptrDiffLen(result->context.uc_stack.ss_sp,ss_end);
#endif
  result->context.uc_stack.ss_flags = 0;
  result->context.uc_link = 0;

  CthAliasEnable(B(result)); /* Change to new thread's stack while building context */
  errno = 0;
#if CMK_THREADS_USE_FCONTEXT
  makeJcontext(&result->context, (uFcontext_fn_t)CthStartThread, fn, arg);
#else
#if CMK_THREADS_USE_CONTEXT
  if (sizeof(void *) == 8) {
    CmiUInt4 fn1 = ((CmiUInt8)(uintptr_t)fn) >> 32;
    CmiUInt4 fn2 = (CmiUInt8)(uintptr_t)fn & 0xFFFFFFFF;
    CmiUInt4 arg1 = ((CmiUInt8)(uintptr_t)arg) >> 32;
    CmiUInt4 arg2 = (CmiUInt8)(uintptr_t)arg & 0xFFFFFFFF;
#ifdef __GNUC__
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wdeprecated-declarations"
#endif
    makecontext(&result->context, (uJcontext_fn_t)CthStartThread, 4, fn1, fn2, arg1, arg2);
#ifdef __GNUC__
#pragma GCC diagnostic pop
#endif
  }
  else
#endif
    makeJcontext(&result->context, (uJcontext_fn_t)CthStartThread, 2, (void *)fn,(void *)arg);
#endif
  if(errno !=0) { 
    perror("makecontext"); 
    CmiAbort("CthCreateInner: makecontext failed.\n");
  }
  CthAliasEnable(B(CthCpvAccess(CthCurrent)));

#if CMK_THREADS_BUILD_TLS
  allocNewTLSSeg(&B(result)->tlsseg, result);
#endif

  return result;  
}

CthThread CthCreate(CthVoidFn fn,void *arg,int size)
{
  return CthCreateInner(fn,arg,size,0);
}
CthThread CthCreateMigratable(CthVoidFn fn,void *arg,int size)
{
  return CthCreateInner(fn,arg,size,1);
}

int CthMigratable(void)
{
  return CmiIsomallocEnabled();
}

CthThread CthPup(pup_er p, CthThread t)
{
  int flag;
  if (pup_isUnpacking(p)) {
    t=(CthThread)malloc(sizeof(struct CthThreadStruct));
    _MEMCHECK(t);
    CthThreadInit(t);
  }
  CthPupBase(p,&t->base,1);

  /*Pup the processor context as bytes-- this is not guarenteed to work!*/
  /* so far, context and context-memoryalias works for IA64, not ia32 */
  /* so far, uJcontext and context-memoryalias works for IA32, not ia64 */
  pup_bytes(p,&t->context,sizeof(t->context));
#if !CMK_THREADS_USE_FCONTEXT && !CMK_THREADS_USE_JCONTEXT && CMK_CONTEXT_FPU_POINTER
#if ! CMK_CONTEXT_FPU_POINTER_UCREGS
  /* context is not portable for ia32 due to pointer in uc_mcontext.fpregs,
     pup it separately */
  if (!pup_isUnpacking(p)) flag = t->context.uc_mcontext.fpregs != NULL;
  pup_int(p,&flag);
  if (flag) {
    if (pup_isUnpacking(p)) {
      t->context.uc_mcontext.fpregs = (struct _libc_fpstate *)malloc(sizeof(struct _libc_fpstate));
    }
    pup_bytes(p,t->context.uc_mcontext.fpregs,sizeof(struct _libc_fpstate));
  }
#else             /* net-linux-ppc 32 bit */
  if (!pup_isUnpacking(p)) flag = t->context.uc_mcontext.uc_regs != NULL;
  pup_int(p,&flag);
  if (flag) {
    if (pup_isUnpacking(p)) {
      t->context.uc_mcontext.uc_regs = (mcontext_t *)malloc(sizeof(mcontext_t));
    }
    pup_bytes(p,t->context.uc_mcontext.uc_regs,sizeof(mcontext_t));
  }
#endif
#endif
#if !CMK_THREADS_USE_FCONTEXT && !CMK_THREADS_USE_JCONTEXT && CMK_CONTEXT_V_REGS
  /* linux-ppc  64 bit */
  if (pup_isUnpacking(p)) {
    t->context.uc_mcontext.v_regs = (vrregset_t *)malloc(sizeof(vrregset_t));
  }
  pup_bytes(p,t->context.uc_mcontext.v_regs,sizeof(vrregset_t));
#endif
  if (pup_isUnpacking(p)) {
    t->context.uc_link = 0;
  }
  if (pup_isDeleting(p)) {
    CthFree(t);
    return 0;
  }
  return t;
}

#else 
/***************************************************************
  Basic qthreads implementation. 

  These threads can also add a "protection block" of
  inaccessible memory to detect stack overflows, which
  would otherwise just trash the heap.

  (7/13/2001 creation times on 300MHz AMD K6-3 x86, Linux 2.2.18:
  Qt setjmp, without stackprotect: 18.5 us
  Qt i386, without stackprotect: 17.9 us
  Qt setjmp, with stackprotect: 68.6 us
  )

  Written by Josh Yelon around 1995
  */

#if !CMK_ERROR_CHECKING || (!CMK_MEMORY_PROTECTABLE)
#  define CMK_STACKPROTECT 0

#  define CthMemAlign(x,n) 0
#  define CthMemoryProtect(m,p,l) CmiAbort("Shouldn't call CthMemoryProtect!\n")
#  define CthMemoryUnprotect(m,p,l) CmiAbort("Shouldn't call CthMemoryUnprotect!\n")
#else
#  define CMK_STACKPROTECT 1

extern void setProtection(char*, char*, int, int);
#  include "sys/mman.h"
#  define CthMemAlign(x,n) memalign((x),(n))
#  define CthMemoryProtect(m,p,l) mprotect(p,l,PROT_NONE);setProtection((char*)m,p,l,1);
#  define CthMemoryUnprotect(m,p,l) mprotect(p,l,PROT_READ | PROT_WRITE);setProtection((char*)m,p,l,0);
#endif

struct CthThreadStruct
{
  CthThreadBase base;

  char      *protect;
  int        protlen;

  qt_t      *stack;
  qt_t      *stackp;
};

static CthThread CthThreadInit(void)
{
  CthThread ret=(CthThread)malloc(sizeof(struct CthThreadStruct));
  _MEMCHECK(ret);
  CthThreadBaseInit(&ret->base);
  ret->protect = 0;
  ret->protlen = 0;

  return ret;
}

static void CthThreadFree(CthThread t)
{
  if (t->protlen!=0) {
    CthMemoryUnprotect(t->stack, t->protect, t->protlen);
  }
  CthThreadBaseFree(&t->base);
  free(t);
}

void CthInit(char **argv)
{
  CthThread mainThread;

  CthBaseInit(argv);  
  mainThread=CthThreadInit();
  CthCpvAccess(CthCurrent)=mainThread;
  /* mainThread->base.suspendable=0;*/ /*Can't suspend main thread (trashes Quickthreads jump buffer)*/

  CmiThreadIs_flag |= CMI_THREAD_IS_QT;
#if CMK_THREADS_ALIAS_STACK
  CmiThreadIs_flag |= CMI_THREAD_IS_ALIAS;
#endif
}

void CthFree(CthThread t)
{
  if (t==NULL) return;
  if (t==CthCpvAccess(CthCurrent)) {
    t->base.exiting = 1;
  } else {
    CthThreadFree(t);
  }
}

static void *CthAbortHelp(qt_t *sp, CthThread old, void *null)
{
  CthThreadFree(old);
  return (void *) 0;
}

static void *CthBlockHelp(qt_t *sp, CthThread old, void *null)
{
  old->stackp = sp;
  return (void *) 0;
}

#if CMK_THREADS_BUILD_TLS
void CthResume(CthThread) CMI_NOOPTIMIZE;
#endif

void CthResume(CthThread t)
{
  CthThread tc = CthCpvAccess(CthCurrent);
  if (t == tc) return;

#if CMK_THREADS_BUILD_TLS
  switchTLS(&B(tc)->tlsseg, &B(t)->tlsseg);
#endif

  CthBaseResume(t);
  if (tc->base.exiting) {
    QT_ABORT((qt_helper_t*)CthAbortHelp, tc, 0, t->stackp);
  } else {
    QT_BLOCK((qt_helper_t*)CthBlockHelp, tc, 0, t->stackp);
  }
  /*This check will mistakenly fail if the thread migrates (changing tc)
    if (tc!=CthCpvAccess(CthCurrent)) { CmiAbort("Stack corrupted?\n"); }
    */
}

static void CthOnly(void *arg, void *vt, qt_userf_t fn)
{
  fn(arg);
  CthThreadFinished(CthSelf());
}

static CthThread CthCreateInner(CthVoidFn fn, void *arg, int size,int Migratable)
{
  CthThread result; qt_t *stack, *stackbase, *stackp;
  const size_t pagesize = CmiGetPageSize();
  int doProtect=(!Migratable) && CMK_STACKPROTECT;
  result=CthThreadInit();
  if (doProtect) 
  { /*Can only protect on a page boundary-- allocate an extra page and align stack*/
    if (size==0) size=CthCpvAccess(_defaultStackSize);
    size = (size+(pagesize*2)-1) & ~(pagesize-1);
    stack = (qt_t*)CthMemAlign(pagesize, size);
    B(result)->stack = stack;
    B(result)->stacksize = size;
  } else
    stack = (qt_t *)CthAllocateStack(&result->base,&size,Migratable);
  CthAliasEnable(B(result)); /* Change to new thread's stack while setting args */
  stackbase = QT_SP(stack, size);
  stackp = QT_ARGS(stackbase, arg, result, (qt_userf_t *)fn, CthOnly);
  CthAliasEnable(B(CthCpvAccess(CthCurrent)));
  result->stack = stack;
  B(result)->stacksize = size;
  result->stackp = stackp;
  if (doProtect) {
#ifdef QT_GROW_UP
    /*Stack grows up-- protect highest page of stack*/
    result->protect = ((char*)stack) + size - pagesize;
#else
    /*Stack grows down-- protect lowest page in stack*/
    result->protect = ((char*)stack);
#endif
    result->protlen = pagesize;
    CthMemoryProtect(stack, result->protect, result->protlen);
  }

#if CMK_THREADS_BUILD_TLS
  allocNewTLSSeg(&B(result)->tlsseg, result);
#endif

  return result;
}

CthThread CthCreate(CthVoidFn fn, void *arg, int size)
{ return CthCreateInner(fn,arg,size,0);}

CthThread CthCreateMigratable(CthVoidFn fn, void *arg, int size)
{ return CthCreateInner(fn,arg,size,1);}

int CthMigratable(void)
{
#if CMK_THREADS_ALIAS_STACK
  return 1;
#else
  return CmiIsomallocEnabled();
#endif
}

CthThread CthPup(pup_er p, CthThread t)
{
  if (pup_isUnpacking(p)) {
    t=CthThreadInit();
  }
#if CMK_THREADS_ALIAS_STACK
  CthPupBase(p,&t->base,0);
#else
  CthPupBase(p,&t->base,1);
#endif

  /*Pup the stack pointer as bytes-- this works because stack is migratable*/
  pup_bytes(p,&t->stackp,sizeof(t->stackp));

  /*Don't worry about stack protection on migration*/  

  if (pup_isDeleting(p)) {
    CthFree(t);
    return 0;
  }
  return t;
}

/* Functions that help debugging of out-of-core emulation in BigSim */
void CthPrintThdStack(CthThread t){
  CmiPrintf("thread=%p, base stack=%p, stack pointer=%p\n", t, t->base.stack, t->stackp);
}
#endif

#if ! USE_SPECIAL_STACKPOINTER
size_t CthStackOffset(CthThread t, char *p)
{
  size_t s;
  CmiAssert(t);
  if (B(t)->stack == NULL)      /* fiber, pthread */
    s = p - (char *)t;
  else
    s = p - (char *)B(t)->stack;
  /* size_t s = (size_t)p; */
  return s;
}

char * CthPointer(CthThread t, size_t pos)
{
  char *p;
  CmiAssert(t);
  if (B(t)->stack == NULL)      /* fiber, pthread */
    p = (char*)t + pos;
  else
    p = (char *)B(t)->stack + pos;
  /* char *p = (char *)size; */
  return p;
}
#endif

#if CMK_TRACE_ENABLED
void CthTraceResume(CthThread t)
{
  traceResume(B(t)->eventID, B(t)->srcPE,&t->base.tid);
}
#endif
/* Functions that help debugging of out-of-core emulation in BigSim */
void CthPrintThdMagic(CthThread t){
  CmiPrintf("CthThread[%p]'s magic: %x\n", t, t->base.magic);
}

void *CmiIsomallocMallocForThread(CthThread th, size_t nBytes)
{
  if (B(th)->isomallocBlockList == NULL && CmiIsomallocEnabled())
    B(th)->isomallocBlockList = CmiIsomallocBlockListNew();
  return CmiIsomallocBlockListMalloc(B(th)->isomallocBlockList, nBytes);
}

void *CmiIsomallocMallocAlignForThread(CthThread th, size_t align, size_t nBytes)
{
  if (B(th)->isomallocBlockList == NULL && CmiIsomallocEnabled())
    B(th)->isomallocBlockList = CmiIsomallocBlockListNew();
  return CmiIsomallocBlockListMallocAlign(B(th)->isomallocBlockList, align, nBytes);
}

---