/* initnode node table reply format:
+-------------------------------------------------------
| 4 bytes | Number of nodes n ^
| | (big-endian binary integer) 4+12*n bytes
+------------------------------------------------- |
^ | (one entry for each node) ^ |
| | 4 bytes | Number of PEs for this node | |
n | 4 bytes | IP address of this node 12*n bytes |
| | 4 bytes | Data (UDP) port of this node | |
v | | (big-endian binary integers) v v
---+----------------------------------------------------
*/
static void node_addresses_store(ChMessage *msg)
{
ChMessageInt_t *n32=(ChMessageInt_t *)msg->data;
ChNodeinfo *d=(ChNodeinfo *)(n32+1);
int nodestart;
int i,j,n;
MACHSTATE(1,"node_addresses_store {");
_Cmi_numnodes=ChMessageInt(n32[0]);
if ((sizeof(ChMessageInt_t)+sizeof(ChNodeinfo)*_Cmi_numnodes)
!=(unsigned int)msg->len)
{printf("Node table has inconsistent length!");machine_exit(1);}
nodes = (OtherNode)malloc(_Cmi_numnodes * sizeof(struct OtherNodeStruct));
nodestart=0;
for (i=0; i<_Cmi_numnodes; i++) {
nodes[i].nodestart = nodestart;
nodes[i].nodesize = ChMessageInt(d[i].nPE);
MACHSTATE2(3,"node %d nodesize %d",i,nodes[i].nodesize);
nodes[i].mach_id = ChMessageInt(d[i].mach_id);
nodes[i].IP=d[i].IP;
if (i==_Cmi_mynode) {
Cmi_nodestart=nodes[i].nodestart;
_Cmi_mynodesize=nodes[i].nodesize;
Cmi_self_IP=nodes[i].IP;
}
nodes[i].dataport = ChMessageInt(d[i].dataport);
nodes[i].addr = skt_build_addr(nodes[i].IP,nodes[i].dataport);
#if CMK_USE_TCP
nodes[i].sock = INVALID_SOCKET;
#endif
nodestart+=nodes[i].nodesize;
}
_Cmi_numpes=nodestart;
n = _Cmi_numpes;
#ifdef CMK_CPV_IS_SMP
n += _Cmi_numnodes;
#endif
nodes_by_pe = (OtherNode*)malloc(n * sizeof(OtherNode));
_MEMCHECK(nodes_by_pe);
for (i=0; i<_Cmi_numnodes; i++) {
OtherNode node = nodes + i;
OtherNode_init(node);
for (j=0; j<node->nodesize; j++) {
nodes_by_pe[j + node->nodestart] = node;
}
}
#ifdef CMK_CPV_IS_SMP
/* index for communication threads */
for (i=_Cmi_numpes; i<_Cmi_numpes+_Cmi_numnodes; i++) {
OtherNode node = nodes + i-_Cmi_numpes;
nodes_by_pe[i] = node;
}
#endif
MACHSTATE(1,"} node_addresses_store");
}
static void CmiIdleLock_sleep(CmiIdleLock *l,int msTimeout) {
if (l->hasMessages) return;
l->isSleeping=1;
MACHSTATE(4,"Processor going to sleep {")
WaitForSingleObject(l->sem,msTimeout);
MACHSTATE(4,"} Processor awake again")
l->isSleeping=0;
}
/**
* lapi completion handler on the recv side. It's responsible to push messages
* to the destination proc or relay broadcast messages. --Chao Mei
*
* Note: The completion handler could be executed on any cores within a node ???
* So in SMP mode when there's a comm thread, the completion handler should be carefully
* dealt with.
*
* Given lapi also provides an internal lapi thread to deal with network progress which
* will call this function (???), we should be careful with the following situations:
* 1) non SMP mode, with interrupt (lapi internal completion thread)
* 2) non SMP mode, with polling (machine layer is responsible for network progress)
* 3) SMP mode, no comm thread, with polling
* 4) SMP mode, no comm thread, with interrupt
* 5) SMP mode, with comm thread, with polling (not yet implemented, comm server is empty right now)
* 6) SMP mode, with comm thread, with interrupt??
*
* Currently, SMP mode without comm thread is undergoing implementation.
*
* This function is executed by LAPI internal threads. It seems that the number of internal
* completion handler threads could vary during the program. LAPI adaptively creates more
* threads if there are more outstanding messages!!!! This means pcqueue needs protection
* even in the nonsmp case!!!!
*
* --Chao Mei
*/
static void PumpMsgsComplete(lapi_handle_t *myLapiContext, void *am_info) {
int i;
char *msg = am_info;
int broot, destrank;
MACHSTATE3(2,"[%d] PumpMsgsComplete with msg %p (isImm=%d) begin {",CmiMyNode(), msg, CmiIsImmediate(msg));
#if ENSURE_MSG_PAIRORDER
MACHSTATE3(2,"msg %p info: srcpe=%d, seqno=%d", msg, CMI_MSG_SRCPE(msg), CMI_MSG_SEQNO(msg));
#endif
/**
* First, we check if the msg is a broadcast msg via spanning
* tree. If it is, it needs to call SendSpanningChildren to
* relay the broadcast, and then send the msg to every cores on
* this node.
*
* After the first check, we deal with normal messages.
* --Chao Mei
*/
/* It's the right place to relay the broadcast message */
/**
* 1. For in-order delivery, because this is the handler for
* receiving a message, and we assume the cross-network msgs are
* always delivered to the first proc (rank 0) of this node, we
* select the srcpe of the bcast msgs and the next msg seq no
* correspondingly.
*
* --Chao Mei
*/
#if ENSURE_MSG_PAIRORDER
broot = CMI_BROADCAST_ROOT(msg);
destrank = CMI_DEST_RANK(msg);
/* Only check proc-level msgs */
if (broot>=0
#if CMK_NODE_QUEUE_AVAILABLE
&& destrank != DGRAM_NODEMESSAGE
#endif
) {
MsgOrderInfo *info;
info = &CpvAccessOther(p2pMsgSeqInfo, destrank);
MACHSTATE1(2, "Check msg in-order for p2p msg %p", msg);
if (checkMsgInOrder(msg,info)) {
MACHSTATE(2,"} PumpMsgsComplete end ");
return;
}
}
#endif
handleOneRecvedMsg(CMI_MSG_SIZE(msg), msg);
MACHSTATE(2,"} PumpMsgsComplete end ");
return;
}
static void CmiIdleLock_sleep(CmiIdleLock *l,int msTimeout) {
struct timespec wakeup;
if (l->hasMessages) return;
l->isSleeping=1;
MACHSTATE(4,"Processor going to sleep {")
pthread_mutex_lock(&l->mutex);
getTimespec(msTimeout,&wakeup);
while (!l->hasMessages)
if (ETIMEDOUT==pthread_cond_timedwait(&l->cond,&l->mutex,&wakeup))
break;
pthread_mutex_unlock(&l->mutex);
MACHSTATE(4,"} Processor awake again")
l->isSleeping=0;
}
static INLINE_KEYWORD void lapiSendFn(int destNode, int size, char *msg, scompl_hndlr_t *shdlr, void *sinfo) {
lapi_xfer_t xfer_cmd;
MACHSTATE3(2,"lapiSendFn to destNode=%d with msg %p (isImm=%d) begin {",destNode,msg, CmiIsImmediate(msg));
MACHSTATE3(2, "inside lapiSendFn 1: size=%d, sinfo=%p, deliverable=%d", size, sinfo, deliverable);
MACHSTATE2(2, "Ready to call LAPI_Xfer with destNode=%d, destRank=%d",destNode,CMI_DEST_RANK(msg));
xfer_cmd.Am.Xfer_type = LAPI_AM_XFER;
xfer_cmd.Am.flags = 0;
xfer_cmd.Am.tgt = destNode;
xfer_cmd.Am.hdr_hdl = lapiHeaderHandler;
xfer_cmd.Am.uhdr_len = 0;
xfer_cmd.Am.uhdr = NULL;
xfer_cmd.Am.udata = msg;
xfer_cmd.Am.udata_len = size;
xfer_cmd.Am.shdlr = shdlr;
xfer_cmd.Am.sinfo = sinfo;
xfer_cmd.Am.tgt_cntr = NULL;
xfer_cmd.Am.org_cntr = NULL;
xfer_cmd.Am.cmpl_cntr = NULL;
check_lapi(LAPI_Xfer,(lapiContext, &xfer_cmd));
MACHSTATE(2,"} lapiSendFn end");
}
/******************
* Initialization routine
* currently just testing start up
* ****************/
void CmiInitSysvshm(char **argv){
MACHSTATE(3,"CminitSysvshm start");
sysvshmContext = (SysvshmContext *)malloc(sizeof(SysvshmContext));
if(Cmi_charmrun_pid <= 0){
CmiAbort("sysvshm must be run with charmrun");
}
calculateNodeSizeAndRank(argv);
if(sysvshmContext->nodesize == 1){
return;
}
MACHSTATE1(3,"CminitSysvshm %d calculateNodeSizeAndRank",sysvshmContext->nodesize);
setupSharedBuffers();
MACHSTATE2(3,"CminitSysvshm %d %d setupSharedBuffers",Cmi_charmrun_pid,sysvshmContext->nodesize);
initAllSendQs();
MACHSTATE2(3,"CminitSysvshm %d %d initAllSendQs",Cmi_charmrun_pid,sysvshmContext->nodesize);
MACHSTATE2(3,"CminitSysvshm %d %d done",Cmi_charmrun_pid,sysvshmContext->nodesize);
#if SYSVSHM_STATS
sysvshmContext->sendCount=0;
sysvshmContext->sendTime=0.0;
sysvshmContext->validCheckCount=0;
sysvshmContext->validCheckTime=0.0;
sysvshmContext->commServerTime = 0;
sysvshmContext->lockRecvCount = 0;
#endif
};
static void CmiIdleLock_addMessage(CmiIdleLock *l) {
l->hasMessages=1;
if (l->isSleeping) { /*The PE is sleeping on this lock-- wake him*/
MACHSTATE(4,"Waking sleeping processor")
ReleaseSemaphore(l->sem,1,NULL);
}
}
CmiCommHandle CmiAsyncBroadcastFn(int size, char *msg) {
#if ENSURE_MSG_PAIRORDER
/* Not sure how to add the msg seq no for async broadcast messages --Chao Mei */
/* so abort here ! */
CmiAssert(0);
return 0;
#else
int i, rank;
int mype = CmiMyPe();
#if ENABLE_CONVERSE_QD
CQdCreate(CpvAccess(cQdState), CmiNumPes()-1);
#endif
MACHSTATE1(3,"[%d] Sending async broadcast message from {",CmiMyNode());
CMI_BROADCAST_ROOT(msg) = 0;
void *handle = malloc(sizeof(int));
*((int *)handle) = CmiNumPes()-1;
for (i=mype+1; i<CmiNumPes(); i++) {
CMI_DEST_RANK(msg) = CmiRankOf(i);
lapiSendFn(CmiNodeOf(i), size, msg, DeliveredMsg, handle);
}
for (i=0; i<mype; i++) {
CMI_DEST_RANK(msg) = CmiRankOf(i);
lapiSendFn(CmiNodeOf(i), size, msg, DeliveredMsg, handle);
}
MACHSTATE(3,"} Sending async broadcast message end");
return handle;
#endif
}
static void CmiNotifyStillIdle(CmiIdleState *s)
{
#if CMK_SHARED_VARS_UNAVAILABLE
/*No comm. thread-- listen on sockets for incoming messages*/
MACHSTATE(1,"idle commserver {")
CommunicationServer(Cmi_idlepoll?0:10, COMM_SERVER_FROM_SMP);
MACHSTATE(1,"} idle commserver")
#else
#if CMK_SHARED_VARS_POSIX_THREADS_SMP
if(_Cmi_sleepOnIdle ){
#endif
int nSpins=20; /*Number of times to spin before sleeping*/
s->nIdles++;
if (s->nIdles>nSpins) { /*Start giving some time back to the OS*/
s->sleepMs+=2;
if (s->sleepMs>10) s->sleepMs=10;
}
/*Comm. thread will listen on sockets-- just sleep*/
if (s->sleepMs>0) {
MACHSTATE1(3,"idle lock(%d) {",CmiMyPe())
CmiIdleLock_sleep(&s->cs->idle,s->sleepMs);
CsdResetPeriodic(); /* check ccd callbacks when I am awakened */
MACHSTATE1(3,"} idle lock(%d)",CmiMyPe())
}
#if CMK_SHARED_VARS_POSIX_THREADS_SMP
}
static void CmiIdleLock_wakeup(CmiIdleLock *l) {
l->hasMessages=1;
MACHSTATE(4,"Waking sleeping processor")
/*The PE is sleeping on this condition variable-- wake him*/
pthread_mutex_lock(&l->mutex);
pthread_cond_signal(&l->cond);
pthread_mutex_unlock(&l->mutex);
}
static void CommunicationsClock(void)
{
MACHSTATE(1,"CommunicationsClock");
Cmi_clock = GetClock();
if (Cmi_clock > Cmi_ack_last + 0.5*Cmi_ack_delay) {
MACHSTATE(2,"CommunicationsClock timing out acks");
Cmi_ack_last=Cmi_clock;
writeableAcks=1;
writeableDgrams=1;
}
if (Cmi_clock > Cmi_check_last + Cmi_check_delay) {
MACHSTATE(4,"CommunicationsClock pinging charmrun");
Cmi_check_last = Cmi_clock;
ctrl_sendone_nolock("ping",NULL,0,NULL,0); /*Charmrun may have died*/
}
}
void CmiSendMessageSysvshm(OutgoingMsg ogm,OtherNode node,int rank,unsigned int broot){
struct sembuf sb;
#if SYSVSHM_STATS
double _startSendTime = CmiWallTimer();
#endif
int dstRank = SysvshmRank(ogm->dst);
MEMDEBUG(CmiMemoryCheck());
DgramHeaderMake(ogm->data,rank,ogm->src,Cmi_charmrun_pid,1, broot);
MACHSTATE4(3,"Send Msg Sysvshm ogm %p size %d dst %d dstRank %d",ogm,ogm->size,ogm->dst,dstRank);
CmiAssert(dstRank >=0 && dstRank != sysvshmContext->noderank);
sharedBufData *dstBuf = &(sysvshmContext->sendBufs[dstRank]);
ACQUIRENW(sysvshmContext->noderank);
if(semop(dstBuf->semid, &sb, 1)<0) {
/**failed to get the lock
insert into q and retain the message*/
pushSendQ(sysvshmContext->sendQs[dstRank],ogm);
ogm->refcount++;
MEMDEBUG(CmiMemoryCheck());
return;
}else{
/***
* We got the lock for this buffer
* first write all the messages in the sendQ and then write this guy
* */
if(sysvshmContext->sendQs[dstRank]->numEntries == 0){
/* send message user event */
int ret = sendMessage(ogm,dstBuf,sysvshmContext->sendQs[dstRank]);
MACHSTATE(3,"Sysvshm Send succeeded immediately");
}else{
ogm->refcount+=2;/*this message should not get deleted when the queue is flushed*/
pushSendQ(sysvshmContext->sendQs[dstRank],ogm);
MACHSTATE3(3,"Sysvshm ogm %p pushed to sendQ length %d refcount %d",ogm,sysvshmContext->sendQs[dstRank]->numEntries,ogm->refcount);
int sent = flushSendQ(dstRank);
ogm->refcount--; /*if it has been sent, can be deleted by caller, if not will be deleted when queue is flushed*/
MACHSTATE1(3,"Sysvshm flushSendQ sent %d messages",sent);
}
/* unlock the recvbuffer*/
RELEASE(sysvshmContext->noderank);
CmiAssert(semop(dstBuf->semid, &sb, 1)>=0);
}
#if SYSVSHM_STATS
sysvshmContext->sendCount ++;
sysvshmContext->sendTime += (CmiWallTimer()-_startSendTime);
#endif
MEMDEBUG(CmiMemoryCheck());
};
/***************
* calculate the name of the shared objects and semaphores
*
* name scheme
* shared memory: charm_pxshm_<recvernoderank>_<sendernoderank>
* semaphore : charm_pxshm_<recvernoderank>_<sendernoderank>.sem for semaphore for that shared object
* the semaphore name used by us is the same as the shared memory object name
* the posix library adds the semaphore tag // in linux at least . other machines might need more portable code
*
* open these shared objects and semaphores
* *********/
void setupSharedBuffers(){
int i=0;
allocBufNameStrings(&(pxshmContext->recvBufNames));
MACHSTATE(3,"allocBufNameStrings for recvBufNames done");
MEMDEBUG(CmiMemoryCheck());
allocBufNameStrings((&pxshmContext->sendBufNames));
MACHSTATE(3,"allocBufNameStrings for sendBufNames done");
for(i=0;i<pxshmContext->nodesize;i++){
if(i != pxshmContext->noderank){
snprintf(pxshmContext->recvBufNames[i],NAMESTRLEN-1,"%s_%d_%d",pxshmContext->prefixStr,pxshmContext->noderank+pxshmContext->nodestart,i+pxshmContext->nodestart);
MACHSTATE2(3,"recvBufName %s with rank %d",pxshmContext->recvBufNames[i],i)
snprintf(pxshmContext->sendBufNames[i],NAMESTRLEN-1,"%s_%d_%d",pxshmContext->prefixStr,i+pxshmContext->nodestart,pxshmContext->noderank+pxshmContext->nodestart);
MACHSTATE2(3,"sendBufName %s with rank %d",pxshmContext->sendBufNames[i],i);
}
}
createShmObjectsAndSems(&(pxshmContext->recvBufs),pxshmContext->recvBufNames);
createShmObjectsAndSems(&(pxshmContext->sendBufs),pxshmContext->sendBufNames);
for(i=0;i<pxshmContext->nodesize;i++){
if(i != pxshmContext->noderank){
//CmiAssert(pxshmContext->sendBufs[i].header->count == 0);
pxshmContext->sendBufs[i].header->count = 0;
pxshmContext->sendBufs[i].header->bytes = 0;
}
}
#if CMK_SMP && ( CMK_CRAYXE || CMK_CRAYXC )
if (PMI_Barrier() != GNI_RC_SUCCESS) return;
#else
if (CmiBarrier() != 0) return;
#endif
freeSharedBuffers();
pxshm_freed = 1;
}
void calculateNodeSizeAndRank(char **argv){
sysvshmContext->nodesize=1;
MACHSTATE(3,"calculateNodeSizeAndRank start");
CmiGetArgIntDesc(argv, "+nodesize", &(sysvshmContext->nodesize),"Number of cores in this node");
MACHSTATE1(3,"calculateNodeSizeAndRank argintdesc %d",sysvshmContext->nodesize);
sysvshmContext->noderank = _Cmi_mynode % (sysvshmContext->nodesize);
MACHSTATE1(3,"calculateNodeSizeAndRank noderank %d",sysvshmContext->noderank);
sysvshmContext->nodestart = _Cmi_mynode -sysvshmContext->noderank;
MACHSTATE(3,"calculateNodeSizeAndRank nodestart ");
sysvshmContext->nodeend = sysvshmContext->nodestart + sysvshmContext->nodesize -1;
if(sysvshmContext->nodeend >= _Cmi_numnodes){
sysvshmContext->nodeend = _Cmi_numnodes-1;
sysvshmContext->nodesize = (sysvshmContext->nodeend - sysvshmContext->nodestart) +1;
}
MACHSTATE3(3,"calculateNodeSizeAndRank nodestart %d nodesize %d noderank %d",sysvshmContext->nodestart,sysvshmContext->nodesize,sysvshmContext->noderank);
}
void calculateNodeSizeAndRank(char **argv){
pxshmContext->nodesize=1;
MACHSTATE(3,"calculateNodeSizeAndRank start");
//CmiGetArgIntDesc(argv, "+nodesize", &(pxshmContext->nodesize),"Number of cores in this node (for non-smp case).Used by the shared memory communication layer");
CmiGetArgIntDesc(argv, "+nodesize", &(pxshmContext->nodesize),"Number of cores in this node");
MACHSTATE1(3,"calculateNodeSizeAndRank argintdesc %d",pxshmContext->nodesize);
pxshmContext->noderank = _Cmi_mynode % (pxshmContext->nodesize);
MACHSTATE1(3,"calculateNodeSizeAndRank noderank %d",pxshmContext->noderank);
pxshmContext->nodestart = _Cmi_mynode -pxshmContext->noderank;
MACHSTATE(3,"calculateNodeSizeAndRank nodestart ");
pxshmContext->nodeend = pxshmContext->nodestart + pxshmContext->nodesize -1;
if(pxshmContext->nodeend >= _Cmi_numnodes){
pxshmContext->nodeend = _Cmi_numnodes-1;
pxshmContext->nodesize = (pxshmContext->nodeend - pxshmContext->nodestart) +1;
}
MACHSTATE3(3,"calculateNodeSizeAndRank nodestart %d nodesize %d noderank %d",pxshmContext->nodestart,pxshmContext->nodesize,pxshmContext->noderank);
}
/** lapi header handler: executed on the recv side, when the
* first packet of the recving msg arrives, it is called to
* prepare the memory buffer in the user space for recving the
* data --Chao Mei
*/
static void* PumpMsgsBegin(lapi_handle_t *myLapiContext,
void *hdr, uint *uhdr_len,
lapi_return_info_t *msg_info,
compl_hndlr_t **comp_h, void **comp_am_info) {
void *msg_buf;
MACHSTATE1(2,"[%d] PumpMsgsBegin begin {",CmiMyNode());
/* prepare the space for receiving the data, set the completion handler to
be executed inline */
msg_buf = (void *)CmiAlloc(msg_info->msg_len);
msg_info->ret_flags = LAPI_SEND_REPLY;
*comp_h = PumpMsgsComplete;
*comp_am_info = msg_buf;
MACHSTATE(2,"} PumpMsgsBegin end");
return msg_buf;
}
void CmiStateInit(int pe, int rank, CmiState state)
{
#if CMK_SMP_MULTIQ
int i;
#endif
MACHSTATE(4,"StateInit")
state->pe = pe;
state->rank = rank;
if (rank==CmiMyNodeSize()) return; /* Communications thread */
#if !CMK_SMP_MULTIQ
state->recv = CMIQueueCreate();
#else
for(i=0; i<MULTIQ_GRPSIZE; i++) state->recv[i]=CMIQueueCreate();
state->myGrpIdx = rank % MULTIQ_GRPSIZE;
state->curPolledIdx = 0;
#endif
state->localqueue = CdsFifo_Create();
CmiIdleLock_init(&state->idle);
}
CmiCommHandle CmiAsyncNodeBroadcastFn(int size, char *msg) {
int i;
#if ENABLE_CONVERSE_QD
CQdCreate(CpvAccess(cQdState), CmiNumNodes()-1);
#endif
MACHSTATE1(3,"[%d] Sending async node broadcast message from {",CmiMyNode());
CMI_BROADCAST_ROOT(msg) = 0;
CMI_DEST_RANK(msg) =DGRAM_NODEMESSAGE;
void *handle = malloc(sizeof(int));
*((int *)handle) = CmiNumNodes()-1;
for (i=CmiMyNode()+1; i<CmiNumNodes(); i++) {
lapiSendFn(i, size, msg, DeliveredMsg, handle);
}
for (i=0; i<CmiMyNode(); i++) {
lapiSendFn(i, size, msg, DeliveredMsg, handle);
}
MACHSTATE(3,"} Sending async broadcast message end");
return handle;
}
/* The following two are callbacks for sync and async send respectively */
static void ReleaseMsg(lapi_handle_t *myLapiContext, void *msg, lapi_sh_info_t *info) {
MACHSTATE2(2,"[%d] ReleaseMsg begin %p {",CmiMyNode(),msg);
check_lapi_err(info->reason, "ReleaseMsg", __LINE__);
CmiFree(msg);
MACHSTATE(2,"} ReleaseMsg end");
}
void CmiSendMessagePxshm(char *msg, int size, int dstnode, int *refcount)
{
#if PXSHM_STATS
double _startSendTime = CmiWallTimer();
#endif
LrtsPrepareEnvelope(msg, size);
int dstRank = PxshmRank(dstnode);
MEMDEBUG(CmiMemoryCheck());
/*
MACHSTATE4(3,"Send Msg Pxshm ogm %p size %d dst %d dstRank %d",ogm,ogm->size,ogm->dst,dstRank);
MACHSTATE4(3,"Send Msg Pxshm ogm %p size %d dst %d dstRank %d",ogm,ogm->size,ogm->dst,dstRank);
*/
CmiAssert(dstRank >=0 && dstRank != pxshmContext->noderank);
sharedBufData *dstBuf = &(pxshmContext->sendBufs[dstRank]);
PxshmSendQ *sendQ = pxshmContext->sendQs[dstRank];
#if PXSHM_OSSPINLOCK
if(! OSSpinLockTry(&dstBuf->header->lock)){
#elif PXSHM_LOCK
if(sem_trywait(dstBuf->mutex) < 0){
#elif PXSHM_FENCE
dstBuf->header->flagSender = 1;
dstBuf->header->turn = RECEIVER;
CmiMemoryReadFence(0,0);
CmiMemoryWriteFence(0,0);
//if(dstBuf->header->flagReceiver && dstBuf->header->turn == RECEIVER){
if(dstBuf->header->flagReceiver){
dstBuf->header->flagSender = 0;
#endif
/**failed to get the lock
insert into q and retain the message*/
#if SENDQ_LIST
if (sendQ->numEntries == 0 && sendQ->next == -2) {
sendQ->next = sendQ_head_index;
sendQ_head_index = dstRank;
}
#endif
pushSendQ(pxshmContext->sendQs[dstRank], msg, size, refcount);
(*refcount)++;
MEMDEBUG(CmiMemoryCheck());
return;
}else{
/***
* We got the lock for this buffer
* first write all the messages in the sendQ and then write this guy
* */
if(pxshmContext->sendQs[dstRank]->numEntries == 0){
// send message user event
int ret = sendMessage(msg,size,refcount,dstBuf,pxshmContext->sendQs[dstRank]);
#if SENDQ_LIST
if (sendQ->numEntries > 0 && sendQ->next == -2)
{
sendQ->next = sendQ_head_index;
sendQ_head_index = dstRank;
}
#endif
MACHSTATE(3,"Pxshm Send succeeded immediately");
}else{
(*refcount)+=2;/*this message should not get deleted when the queue is flushed*/
pushSendQ(pxshmContext->sendQs[dstRank],msg,size,refcount);
// MACHSTATE3(3,"Pxshm ogm %p pushed to sendQ length %d refcount %d",ogm,pxshmContext->sendQs[dstRank]->numEntries,ogm->refcount);
int sent = flushSendQ(sendQ);
(*refcount)--; /*if it has been sent, can be deleted by caller, if not will be deleted when queue is flushed*/
MACHSTATE1(3,"Pxshm flushSendQ sent %d messages",sent);
}
/* unlock the recvbuffer*/
#if PXSHM_OSSPINLOCK
OSSpinLockUnlock(&dstBuf->header->lock);
#elif PXSHM_LOCK
sem_post(dstBuf->mutex);
#elif PXSHM_FENCE
CmiMemoryReadFence(0,0);
CmiMemoryWriteFence(0,0);
dstBuf->header->flagSender = 0;
#endif
}
#if PXSHM_STATS
pxshmContext->sendCount ++;
pxshmContext->sendTime += (CmiWallTimer()-_startSendTime);
#endif
MEMDEBUG(CmiMemoryCheck());
};
inline void emptyAllRecvBufs();
inline void flushAllSendQs();
/**********
* Extract all the messages from the recvBuffers you can
* Flush all sendQs
* ***/
inline void CommunicationServerPxshm(){
#if PXSHM_STATS
double _startCommServerTime =CmiWallTimer();
#endif
MEMDEBUG(CmiMemoryCheck());
emptyAllRecvBufs();
flushAllSendQs();
#if PXSHM_STATS
pxshmContext->commServerTime += (CmiWallTimer()-_startCommServerTime);
#endif
MEMDEBUG(CmiMemoryCheck());
};
static void CmiNotifyStillIdlePxshm(CmiIdleState *s){
CommunicationServerPxshm();
}
static void CmiNotifyBeginIdlePxshm(CmiIdleState *s)
{
CmiNotifyStillIdle(s);
}
/******************
* Initialization routine
* currently just testing start up
* ****************/
void CmiInitPxshm(char **argv){
char *env;
MACHSTATE(3,"CminitPxshm start");
pxshmContext = (PxshmContext *)calloc(1,sizeof(PxshmContext));
calculateNodeSizeAndRank(argv);
if(pxshmContext->nodesize == 1) return;
MACHSTATE1(3,"CminitPxshm %d calculateNodeSizeAndRank",pxshmContext->nodesize);
env = getenv("CHARM_PXSHM_POOL_SIZE");
if (env) {
SHMBUFLEN = CmiReadSize(env);
}
env = getenv("CHARM_PXSHM_MESSAGE_MAX_SIZE");
if (env) {
SHMMAXSIZE = CmiReadSize(env);
}
if (SHMMAXSIZE > SHMBUFLEN)
CmiAbort("Error> Pxshm pool size is set too small in env variable CHARM_PXSHM_POOL_SIZE");
SENDQSTARTSIZE = 32 * pxshmContext->nodesize;
if (_Cmi_mynode == 0)
printf("Charm++> pxshm enabled: %d cores per node, buffer size: %.1fMB\n", pxshmContext->nodesize, SHMBUFLEN/1024.0/1024.0);
#if CMK_CRAYXE || CMK_CRAYXC
srand(getpid());
int Cmi_charmrun_pid = rand();
PMI_Bcast(&Cmi_charmrun_pid, sizeof(int));
snprintf(&(pxshmContext->prefixStr[0]),PREFIXSTRLEN-1,"charm_pxshm_%d",Cmi_charmrun_pid);
#endif
MACHSTATE2(3,"CminitPxshm %s %d pre setupSharedBuffers",pxshmContext->prefixStr,pxshmContext->nodesize);
setupSharedBuffers();
MACHSTATE2(3,"CminitPxshm %s %d setupSharedBuffers",pxshmContext->prefixStr,pxshmContext->nodesize);
initAllSendQs();
MACHSTATE2(3,"CminitPxshm %s %d initAllSendQs",pxshmContext->prefixStr,pxshmContext->nodesize);
MACHSTATE2(3,"CminitPxshm %s %d done",pxshmContext->prefixStr,pxshmContext->nodesize);
#if PXSHM_STATS
pxshmContext->sendCount=0;
pxshmContext->sendTime=0.0;
pxshmContext->validCheckCount=0;
pxshmContext->validCheckTime=0.0;
pxshmContext->commServerTime = 0;
pxshmContext->lockRecvCount = 0;
#endif
signal(SIGSEGV, cleanupOnAllSigs);
signal(SIGFPE, cleanupOnAllSigs);
signal(SIGILL, cleanupOnAllSigs);
signal(SIGTERM, cleanupOnAllSigs);
signal(SIGABRT, cleanupOnAllSigs);
signal(SIGQUIT, cleanupOnAllSigs);
signal(SIGBUS, cleanupOnAllSigs);
signal(SIGINT, cleanupOnAllSigs);
signal(SIGTRAP, cleanupOnAllSigs);
#if 0
char name[64];
gethostname(name,64);
printf("[%d] name: %s\n", myrank, name);
#endif
};
static void DeliveredMsg(lapi_handle_t *myLapiContext, void *msg, lapi_sh_info_t *info) {
MACHSTATE1(2,"[%d] DeliveredMsg begin {",CmiMyNode());
check_lapi_err(info->reason, "DeliveredMsg", __LINE__);
*((int *)msg) = *((int *)msg) - 1;
MACHSTATE(2,"} DeliveredMsg end");
}
static void CmiStartThreads(char **argv)
{
pthread_t pid;
size_t i;
int ok, tocreate;
pthread_attr_t attr;
int start, end;
MACHSTATE(4,"CmiStartThreads")
CmiMemLock_lock=CmiCreateLock();
comm_mutex=CmiCreateLock();
_smp_mutex = CmiCreateLock();
#if defined(CMK_NO_ASM_AVAILABLE) && CMK_PCQUEUE_LOCK
cmiMemoryLock = CmiCreateLock();
if (CmiMyNode()==0) printf("Charm++ warning> fences and atomic operations not available in native assembly\n");
#endif
#if ! (CMK_HAS_TLS_VARIABLES && !CMK_NOT_USE_TLS_THREAD)
pthread_key_create(&Cmi_state_key, 0);
Cmi_state_vector =
(CmiState)calloc(_Cmi_mynodesize+1, sizeof(struct CmiStateStruct));
for (i=0; i<_Cmi_mynodesize; i++)
CmiStateInit(i+Cmi_nodestart, i, CmiGetStateN(i));
/*Create a fake state structure for the comm. thread*/
/* CmiStateInit(-1,_Cmi_mynodesize,CmiGetStateN(_Cmi_mynodesize)); */
CmiStateInit(_Cmi_mynode+CmiNumPes(),_Cmi_mynodesize,CmiGetStateN(_Cmi_mynodesize));
#else
/* for main thread */
Cmi_state_vector = (CmiState *)calloc(_Cmi_mynodesize+1, sizeof(CmiState));
#if CMK_CONVERSE_MPI
/* main thread is communication thread */
if(!CharmLibInterOperate) {
CmiStateInit(_Cmi_mynode+CmiNumPes(), _Cmi_mynodesize, &Cmi_mystate);
Cmi_state_vector[_Cmi_mynodesize] = &Cmi_mystate;
} else
#endif
{
/* main thread is of rank 0 */
CmiStateInit(Cmi_nodestart, 0, &Cmi_mystate);
Cmi_state_vector[0] = &Cmi_mystate;
}
#endif
#if CMK_MULTICORE || CMK_SMP_NO_COMMTHD
if (!Cmi_commthread)
tocreate = _Cmi_mynodesize-1;
else
#endif
tocreate = _Cmi_mynodesize;
#if CMK_CONVERSE_MPI
if(!CharmLibInterOperate) {
start = 0;
end = tocreate - 1; /* skip comm thread */
} else
#endif
{
start = 1;
end = tocreate; /* skip rank 0 main thread */
}
for (i=start; i<=end; i++) {
pthread_attr_init(&attr);
pthread_attr_setscope(&attr, PTHREAD_SCOPE_SYSTEM);
ok = pthread_create(&pid, &attr, call_startfn, (void *)i);
if (ok<0) PerrorExit("pthread_create");
pthread_attr_destroy(&attr);
}
#if ! (CMK_HAS_TLS_VARIABLES && !CMK_NOT_USE_TLS_THREAD)
#if CMK_CONVERSE_MPI
if(!CharmLibInterOperate)
pthread_setspecific(Cmi_state_key, Cmi_state_vector+_Cmi_mynodesize);
else
#endif
pthread_setspecific(Cmi_state_key, Cmi_state_vector);
#endif
MACHSTATE(4,"CmiStartThreads done")
}
