Robust/src/Runtime/DSTM/interface/trans.c

   1 #include "dstm.h"
   2 #include "ip.h"
   3 #include "clookup.h"
   4 #include "machinepile.h"
   5 #include "mlookup.h"
   6 #include "llookup.h"
   7 #include "plookup.h"
   8 #include "prelookup.h"
   9 #include "queue.h"
  10 #include <pthread.h>
  11 #include <sys/types.h>
  12 #include <sys/socket.h>
  13 #include <netdb.h>
  14 #include <netinet/in.h>
  15 #include <sys/types.h>
  16 #include <unistd.h>
  17 #include <errno.h>
  18 #include <time.h>
  19 #include <string.h>
  20 #include <pthread.h>
  21
  22 #define LISTEN_PORT 2156
  23 #define RECEIVE_BUFFER_SIZE 2048
  24 #define NUM_THREADS 10
  25 #define PREFETCH_CACHE_SIZE 1048576 //1MB
  26
  27 /* Global Variables */
  28 extern int classsize[];
  29 extern primarypfq_t pqueue; // shared prefetch queue
  30 extern mcpileq_t mcqueue;  //Shared queue containing prefetch requests sorted by remote machineids
  31 objstr_t *prefetchcache; //Global Prefetch cache
  32 extern prehashtable_t pflookup; //Global Prefetch cache's lookup table
  33 pthread_t wthreads[NUM_THREADS]; //Worker threads for working on the prefetch queue
  34 pthread_t tPrefetch;
  35 extern objstr_t *mainobjstore;
  36
  37 plistnode_t *createPiles(transrecord_t *);
  38 inline int arrayLength(int *array) {
  39         int i;
  40         for(i=0 ;array[i] != -1; i++)
  41                 ;
  42         return i;
  43 }
  44 inline int findmax(int *array, int arraylength) {
  45         int max, i;
  46         max = array[0];
  47         for(i = 0; i < arraylength; i++){
  48                 if(array[i] > max) {
  49                         max = array[i];
  50                 }
  51         }
  52         return max;
  53 }
  54 /* This function is a prefetch call generated by the compiler that
  55  * populates the shared primary prefetch queue*/
  56 void prefetch(int ntuples, unsigned int *oids, short *endoffsets, short *arrayfields) {
  57         int qnodesize;
  58         int len = 0;
  59
  60         /* Allocate for the queue node*/
  61         char *node;
  62         qnodesize = sizeof(prefetchqelem_t) + sizeof(int) + ntuples * (sizeof(short) + sizeof(unsigned int)) + endoffsets[ntuples - 1] * sizeof(short);
  63         if((node = calloc(1, qnodesize)) == NULL) {
  64                 printf("Calloc Error %s, %d\n", __FILE__, __LINE__);
  65                 return;
  66         }
  67         /* Set queue node values */
  68         len = sizeof(prefetchqelem_t);
  69         memcpy(node + len, &ntuples, sizeof(int));
  70         len += sizeof(int);
  71         memcpy(node + len, oids, ntuples*sizeof(unsigned int));
  72         len += ntuples * sizeof(unsigned int);
  73         memcpy(node + len, endoffsets, ntuples*sizeof(short));
  74         len += ntuples * sizeof(short);
  75         memcpy(node + len, arrayfields, endoffsets[ntuples-1]*sizeof(short));
  76         /* Lock and insert into primary prefetch queue */
  77         pthread_mutex_lock(&pqueue.qlock);
  78         pre_enqueue((prefetchqelem_t *)node);
  79         pthread_cond_signal(&pqueue.qcond);
  80         pthread_mutex_unlock(&pqueue.qlock);
  81 }
  82
  83 /* This function allocates an object on the local machine */
  84 //FIXME
  85
  86 void * dstmalloc(transrecord_t *trans, int size) {
  87   objheader_t * newobj=(objheader_t *)objstrAlloc(trans->cache, size+sizeof(objheader_t));
  88   //Need to assign OID
  89
  90   return newobj;
  91 }
  92
  93 /* This function starts up the transaction runtime. */
  94 int dstmStartup(const char * option) {
  95   pthread_t thread_Listen;
  96   pthread_attr_t attr;
  97   int master=strcmp(option, "master")==0;
  98
  99   dstmInit();
 100   transInit();
 101
 102   if (master) {
 103     pthread_attr_init(&attr);
 104     pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE);
 105     pthread_create(&thread_Listen, &attr, dstmListen, NULL);
 106     return 1;
 107   } else {
 108     dstmListen();
 109     return 0;
 110   }
 111
 112 }
 113
 114
 115 /* This function initiates the prefetch thread
 116  * A queue is shared between the main thread of execution
 117  * and the prefetch thread to process the prefetch call
 118  * Call from compiler populates the shared queue with prefetch requests while prefetch thread
 119  * processes the prefetch requests */
 120 void transInit() {
 121         int t, rc;
 122         //Create and initialize prefetch cache structure
 123         prefetchcache = objstrCreate(PREFETCH_CACHE_SIZE);
 124         //Create prefetch cache lookup table
 125         if(prehashCreate(HASH_SIZE, LOADFACTOR))
 126                 return; //Failure
 127         //Initialize primary shared queue
 128         queueInit();
 129         //Initialize machine pile w/prefetch oids and offsets shared queue
 130         mcpileqInit();
 131         //Create the primary prefetch thread
 132         pthread_create(&tPrefetch, NULL, transPrefetch, NULL);
 133         //Create and Initialize a pool of threads
 134         /* Threads are active for the entire period runtime is running */
 135         for(t = 0; t< NUM_THREADS; t++) {
 136                 rc = pthread_create(&wthreads[t], NULL, mcqProcess, (void *)t);
 137                 if (rc) {
 138                         printf("Thread create error %s, %d\n", __FILE__, __LINE__);
 139                         return;
 140                 }
 141         }
 142 }
 143
 144 /* This function stops the threads spawned */
 145 void transExit() {
 146         int t;
 147         pthread_cancel(tPrefetch);
 148         for(t = 0; t < NUM_THREADS; t++)
 149                 pthread_cancel(wthreads[t]);
 150
 151         return;
 152 }
 153
 154 /* This functions inserts randowm wait delays in the order of msec
 155  * Mostly used when transaction commits retry*/
 156 void randomdelay(void)
 157 {
 158         struct timespec req, rem;
 159         time_t t;
 160
 161         t = time(NULL);
 162         req.tv_sec = 0;
 163         req.tv_nsec = (long)(1000000 + (t%10000000)); //1-11 msec
 164         nanosleep(&req, &rem);
 165         return;
 166 }
 167
 168 /* This function initializes things required in the transaction start*/
 169 transrecord_t *transStart()
 170 {
 171         transrecord_t *tmp = malloc(sizeof(transrecord_t));
 172         tmp->cache = objstrCreate(1048576);
 173         tmp->lookupTable = chashCreate(HASH_SIZE, LOADFACTOR);
 174 #ifdef COMPILER
 175         tmp->revertlist=NULL;
 176 #endif
 177         return tmp;
 178 }
 179
 180 /* This function finds the location of the objects involved in a transaction
 181  * and returns the pointer to the object if found in a remote location */
 182 objheader_t *transRead(transrecord_t *record, unsigned int oid) {
 183         printf("Inside transaction read call\n");
 184         unsigned int machinenumber;
 185         objheader_t *tmp, *objheader;
 186         void *objcopy;
 187         int size, rc, found = 0;
 188         void *buf;
 189         struct timespec ts;
 190         struct timeval tp;
 191
 192         rc = gettimeofday(&tp, NULL);
 193
 194         /* Convert from timeval to timespec */
 195         ts.tv_nsec = tp.tv_usec * 1000;
 196
 197         /* Search local transaction cache */
 198         if((objheader = (objheader_t *)chashSearch(record->lookupTable, oid)) != NULL){
 199                 printf("Inside transaction cache \n");
 200                 return(objheader);
 201         } else if ((objheader = (objheader_t *) mhashSearch(oid)) != NULL) {
 202                 /* Look up in machine lookup table  and copy  into cache*/
 203                 printf("Inside mainobject store \n");
 204                 tmp = mhashSearch(oid);
 205                 size = sizeof(objheader_t)+classsize[TYPE(tmp)];
 206                 objcopy = objstrAlloc(record->cache, size);
 207                 memcpy(objcopy, (void *)objheader, size);
 208                 /* Insert into cache's lookup table */
 209                 chashInsert(record->lookupTable, OID(objheader), objcopy);
 210                 return(objcopy);
 211         } else if((tmp = (objheader_t *) prehashSearch(oid)) != NULL) { /* Look up in prefetch cache */
 212                 printf("Inside prefetch cache \n");
 213                 found = 1;
 214                 size = sizeof(objheader_t)+classsize[TYPE(tmp)];
 215                 objcopy = objstrAlloc(record->cache, size);
 216                 memcpy(objcopy, (void *)tmp, size);
 217                 /* Insert into cache's lookup table */
 218                 chashInsert(record->lookupTable, OID(tmp), objcopy);
 219                 return(objcopy);
 220         } else { /* If not found anywhere, then block until object appears in prefetch cache */
 221                 printf("Inside remote machine\n");
 222                 pthread_mutex_lock(&pflookup.lock);
 223                 while(!found) {
 224                         rc = pthread_cond_timedwait(&pflookup.cond, &pflookup.lock, &ts);
 225                         if(rc == ETIMEDOUT) {
 226                                 printf("Wait timed out\n");
 227                                 /* Check Prefetch cache again */
 228                                 if((tmp = (objheader_t *) prehashSearch(oid)) != NULL) { /* Look up in prefetch cache */
 229                                         found = 1;
 230                                         size = sizeof(objheader_t)+classsize[TYPE(tmp)];
 231                                         objcopy = objstrAlloc(record->cache, size);
 232                                         memcpy(objcopy, (void *)tmp, size);
 233                                         /* Insert into cache's lookup table */
 234                                         chashInsert(record->lookupTable, OID(tmp), objcopy);
 235                                         return(objcopy);
 236                                 } else {
 237                                         pthread_mutex_unlock(&pflookup.lock);
 238                                         break;
 239                                 }
 240                                 pthread_mutex_unlock(&pflookup.lock);
 241                         }
 242                 }
 243                 /* Get the object from the remote location */
 244                 machinenumber = lhashSearch(oid);
 245                 objcopy = getRemoteObj(record, machinenumber, oid);
 246                 if(objcopy == NULL) {
 247                         //If object is not found in Remote location
 248                         //printf("Object oid = %d not found in Machine %d\n", oid, machinenumber);
 249                         return NULL;
 250                 }
 251                 else {
 252                         //printf("Object oid = %d found in Machine %d\n", oid, machinenumber);
 253                         return(objcopy);
 254                 }
 255         }
 256 }
 257
 258 /* This function creates objects in the transaction record */
 259 objheader_t *transCreateObj(transrecord_t *record, unsigned short type)
 260 {
 261         objheader_t *tmp = (objheader_t *) objstrAlloc(record->cache, (sizeof(objheader_t) + classsize[type]));
 262         OID(tmp) = getNewOID();
 263         TYPE(tmp) = type;
 264         tmp->version = 1;
 265         tmp->rcount = 0; //? not sure how to handle this yet
 266         STATUS(tmp) = NEW;
 267         chashInsert(record->lookupTable, OID(tmp), tmp);
 268         return tmp;
 269 }
 270
 271 /* This function creates machine piles based on all machines involved in a
 272  * transaction commit request */
 273 plistnode_t *createPiles(transrecord_t *record) {
 274         int i = 0;
 275         unsigned int size;/* Represents number of bins in the chash table */
 276         chashlistnode_t *curr, *ptr, *next;
 277         plistnode_t *pile = NULL;
 278         unsigned int machinenum;
 279         void *localmachinenum;
 280         objheader_t *headeraddr;
 281
 282         ptr = record->lookupTable->table;
 283         size = record->lookupTable->size;
 284
 285         for(i = 0; i < size ; i++) {
 286                 curr = &ptr[i];
 287                 /* Inner loop to traverse the linked list of the cache lookupTable */
 288                 while(curr != NULL) {
 289                         //if the first bin in hash table is empty
 290                         if(curr->key == 0) {
 291                                 break;
 292                         }
 293                         next = curr->next;
 294                         //Get machine location for object id
 295
 296                         if ((machinenum = lhashSearch(curr->key)) == 0) {
 297                                 printf("Error: No such machine %s, %d\n", __FILE__, __LINE__);
 298                                 return NULL;
 299                         }
 300
 301                         if ((headeraddr = chashSearch(record->lookupTable, curr->key)) == NULL) {
 302                                 printf("Error: No such oid %s, %d\n", __FILE__, __LINE__);
 303                                 return NULL;
 304                         }
 305                         //Make machine groups
 306                         if ((pile = pInsert(pile, headeraddr, machinenum, record->lookupTable->numelements)) == NULL) {
 307                                 printf("pInsert error %s, %d\n", __FILE__, __LINE__);
 308                                 return NULL;
 309                         }
 310
 311                         /* Check if local or not */
 312                         if((localmachinenum = mhashSearch(curr->key)) != NULL) {
 313                                 /* Set the pile->local flag*/
 314                                 pile->local = 1; //True i.e. local
 315                         }
 316
 317                         curr = next;
 318                 }
 319         }
 320         return pile;
 321 }
 322
 323 /* This function initiates the transaction commit process
 324  * Spawns threads for each of the new connections with Participants
 325  * and creates new piles by calling the createPiles(),
 326  * Fills the piles with necesaary information and
 327  * Sends a transrequest() to each pile*/
 328 int transCommit(transrecord_t *record) {
 329         unsigned int tot_bytes_mod, *listmid;
 330         plistnode_t *pile, *pile_ptr;
 331         int i, rc, val;
 332         int pilecount = 0, offset, threadnum = 0, trecvcount = 0, tmachcount = 0;
 333         char buffer[RECEIVE_BUFFER_SIZE],control;
 334         char transid[TID_LEN];
 335         trans_req_data_t *tosend;
 336         trans_commit_data_t transinfo;
 337         static int newtid = 0;
 338         char treplyctrl = 0, treplyretry = 0; /* keeps track of the common response that needs to be sent */
 339         char localstat = 0;
 340
 341
 342         /* Look through all the objects in the transaction record and make piles
 343          * for each machine involved in the transaction*/
 344         pile_ptr = pile = createPiles(record);
 345
 346         /* Create the packet to be sent in TRANS_REQUEST */
 347
 348         /* Count the number of participants */
 349         pilecount = pCount(pile);
 350
 351         /* Create a list of machine ids(Participants) involved in transaction   */
 352         if((listmid = calloc(pilecount, sizeof(unsigned int))) == NULL) {
 353                 printf("Calloc error %s, %d\n", __FILE__, __LINE__);
 354                 return 1;
 355         }
 356         pListMid(pile, listmid);
 357
 358
 359         /* Initialize thread variables,
 360          * Spawn a thread for each Participant involved in a transaction */
 361         pthread_t thread[pilecount];
 362         pthread_attr_t attr;
 363         pthread_cond_t tcond;
 364         pthread_mutex_t tlock;
 365         pthread_mutex_t tlshrd;
 366
 367         thread_data_array_t *thread_data_array;
 368         thread_data_array = (thread_data_array_t *) malloc(sizeof(thread_data_array_t)*pilecount);
 369         local_thread_data_array_t *ltdata;
 370         if((ltdata = calloc(1, sizeof(local_thread_data_array_t))) == NULL) {
 371                 printf("Calloc error %s, %d\n", __FILE__, __LINE__);
 372                 return 1;
 373         }
 374
 375         thread_response_t rcvd_control_msg[pilecount];  /* Shared thread array that keeps track of responses of participants */
 376
 377         /* Initialize and set thread detach attribute */
 378         pthread_attr_init(&attr);
 379         pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE);
 380         pthread_mutex_init(&tlock, NULL);
 381         pthread_cond_init(&tcond, NULL);
 382
 383         /* Process each machine pile */
 384         while(pile != NULL) {
 385                 //Create transaction id
 386                 newtid++;
 387                 if ((tosend = calloc(1, sizeof(trans_req_data_t))) == NULL) {
 388                         printf("Calloc error %s, %d\n", __FILE__, __LINE__);
 389                         return 1;
 390                 }
 391                 tosend->f.control = TRANS_REQUEST;
 392                 sprintf(tosend->f.trans_id, "%x_%d", pile->mid, newtid);
 393                 tosend->f.mcount = pilecount;
 394                 tosend->f.numread = pile->numread;
 395                 tosend->f.nummod = pile->nummod;
 396                 tosend->f.sum_bytes = pile->sum_bytes;
 397                 tosend->listmid = listmid;
 398                 tosend->objread = pile->objread;
 399                 tosend->oidmod = pile->oidmod;
 400                 thread_data_array[threadnum].thread_id = threadnum;
 401                 thread_data_array[threadnum].mid = pile->mid;
 402                 thread_data_array[threadnum].pilecount = pilecount;
 403                 thread_data_array[threadnum].buffer = tosend;
 404                 thread_data_array[threadnum].recvmsg = rcvd_control_msg;
 405                 thread_data_array[threadnum].threshold = &tcond;
 406                 thread_data_array[threadnum].lock = &tlock;
 407                 thread_data_array[threadnum].count = &trecvcount;
 408                 thread_data_array[threadnum].replyctrl = &treplyctrl;
 409                 thread_data_array[threadnum].replyretry = &treplyretry;
 410                 thread_data_array[threadnum].rec = record;
 411                 /* If local do not create any extra connection */
 412                 if(pile->local != 1) { /* Not local */
 413                         rc = pthread_create(&thread[threadnum], NULL, transRequest, (void *) &thread_data_array[threadnum]);
 414                         if (rc) {
 415                                 perror("Error in pthread create\n");
 416                                 return 1;
 417                         }
 418                 } else { /*Local*/
 419                         /*Unset the pile->local flag*/
 420                         pile->local = 0;
 421                         /*Set flag to identify that Local machine is involved*/
 422                         ltdata->tdata = &thread_data_array[threadnum];
 423                         ltdata->transinfo = &transinfo;
 424                         val = pthread_create(&thread[threadnum], NULL, handleLocalReq, (void *) ltdata);
 425                         if (val) {
 426                                 perror("Error in pthread create\n");
 427                                 return 1;
 428                         }
 429                 }
 430                 threadnum++;
 431                 pile = pile->next;
 432         }
 433
 434         /* Free attribute and wait for the other threads */
 435         pthread_attr_destroy(&attr);
 436         for (i = 0 ;i < pilecount ; i++) {
 437                 rc = pthread_join(thread[i], NULL);
 438                 if (rc)
 439                 {
 440                         printf("ERROR return code from pthread_join() is %d\n", rc);
 441                         return 1;
 442                 }
 443                 free(thread_data_array[i].buffer);
 444         }
 445
 446         /* Free resources */
 447         pthread_cond_destroy(&tcond);
 448         pthread_mutex_destroy(&tlock);
 449         free(listmid);
 450         pDelete(pile_ptr);
 451         free(thread_data_array);
 452         free(ltdata);
 453
 454         /* Retry trans commit procedure if not sucessful in the first try */
 455         if(treplyretry == 1) {
 456                 /* wait a random amount of time */
 457                 randomdelay();
 458                 /* Retry the commiting transaction again */
 459                 transCommit(record);
 460         }
 461
 462         return 0;
 463 }
 464
 465 /* This function sends information involved in the transaction request and
 466  * accepts a response from particpants.
 467  * It calls decideresponse() to decide on what control message
 468  * to send next and sends the message using sendResponse()*/
 469 void *transRequest(void *threadarg) {
 470         int sd, i, n;
 471         struct sockaddr_in serv_addr;
 472         struct hostent *server;
 473         thread_data_array_t *tdata;
 474         objheader_t *headeraddr;
 475         char buffer[RECEIVE_BUFFER_SIZE], control, recvcontrol;
 476         char machineip[16], retval;
 477
 478         tdata = (thread_data_array_t *) threadarg;
 479
 480         /* Send Trans Request */
 481         if ((sd = socket(AF_INET, SOCK_STREAM, 0)) < 0) {
 482                 perror("Error in socket for TRANS_REQUEST\n");
 483                 return NULL;
 484         }
 485         bzero((char*) &serv_addr, sizeof(serv_addr));
 486         serv_addr.sin_family = AF_INET;
 487         serv_addr.sin_port = htons(LISTEN_PORT);
 488         midtoIP(tdata->mid,machineip);
 489         machineip[15] = '\0';
 490         serv_addr.sin_addr.s_addr = inet_addr(machineip);
 491         /* Open Connection */
 492         if (connect(sd, (struct sockaddr *) &serv_addr, sizeof(struct sockaddr)) < 0) {
 493                 perror("Error in connect for TRANS_REQUEST\n");
 494                 return NULL;
 495         }
 496
 497         printf("DEBUG-> trans.c Sending TRANS_REQUEST to mid %s\n", machineip);
 498         /* Send bytes of data with TRANS_REQUEST control message */
 499         if (send(sd, &(tdata->buffer->f), sizeof(fixed_data_t),MSG_NOSIGNAL) < sizeof(fixed_data_t)) {
 500                 perror("Error sending fixed bytes for thread\n");
 501                 return NULL;
 502         }
 503         /* Send list of machines involved in the transaction */
 504         {
 505                 int size=sizeof(unsigned int)*tdata->pilecount;
 506                 if (send(sd, tdata->buffer->listmid, size, MSG_NOSIGNAL) < size) {
 507                         perror("Error sending list of machines for thread\n");
 508                         return NULL;
 509                 }
 510         }
 511         /* Send oids and version number tuples for objects that are read */
 512         {
 513                 int size=(sizeof(unsigned int)+sizeof(short))*tdata->buffer->f.numread;
 514                 if (send(sd, tdata->buffer->objread, size, MSG_NOSIGNAL) < size) {
 515                         perror("Error sending tuples for thread\n");
 516                         return NULL;
 517                 }
 518         }
 519         /* Send objects that are modified */
 520         for(i = 0; i < tdata->buffer->f.nummod ; i++) {
 521                 int size;
 522                 headeraddr = chashSearch(tdata->rec->lookupTable, tdata->buffer->oidmod[i]);
 523                 size=sizeof(objheader_t)+classsize[TYPE(headeraddr)];
 524                 if (send(sd, headeraddr, size, MSG_NOSIGNAL)  < size) {
 525                         perror("Error sending obj modified for thread\n");
 526                         return NULL;
 527                 }
 528         }
 529
 530         /* Read control message from Participant */
 531         if((n = read(sd, &control, sizeof(char))) <= 0) {
 532                 perror("Error in reading control message from Participant\n");
 533                 return NULL;
 534         }
 535         recvcontrol = control;
 536
 537         /* Update common data structure and increment count */
 538         tdata->recvmsg[tdata->thread_id].rcv_status = recvcontrol;
 539
 540         /* Lock and update count */
 541         //Thread sleeps until all messages from pariticipants are received by coordinator
 542         pthread_mutex_lock(tdata->lock);
 543
 544         (*(tdata->count))++; /* keeps track of no of messages received by the coordinator */
 545
 546         /* Wake up the threads and invoke decideResponse (once) */
 547         if(*(tdata->count) == tdata->pilecount) {
 548                 if (decideResponse(tdata) != 0) {
 549                         printf("decideResponse returned error %s,%d\n", __FILE__, __LINE__);
 550                         pthread_mutex_unlock(tdata->lock);
 551                         close(sd);
 552                         return NULL;
 553                 }
 554                 pthread_cond_broadcast(tdata->threshold);
 555         } else {
 556                 pthread_cond_wait(tdata->threshold, tdata->lock);
 557         }
 558         pthread_mutex_unlock(tdata->lock);
 559
 560         /* Send the final response such as TRANS_COMMIT or TRANS_ABORT t
 561          * to all participants in their respective socket */
 562         if (sendResponse(tdata, sd) == 0) {
 563                 printf("sendResponse returned error %s,%d\n", __FILE__, __LINE__);
 564                 pthread_mutex_unlock(tdata->lock);
 565                 close(sd);
 566                 return NULL;
 567         }
 568
 569         /* Close connection */
 570         close(sd);
 571         pthread_exit(NULL);
 572 }
 573
 574 /* This function decides the reponse that needs to be sent to
 575  * all Participant machines involved in the transaction commit */
 576 int decideResponse(thread_data_array_t *tdata) {
 577         char control;
 578         int i, transagree = 0, transdisagree = 0, transsoftabort = 0; /* Counters to formulate decision of what
 579                                                                          message to send */
 580
 581         //Check common data structure
 582         for (i = 0 ; i < tdata->pilecount ; i++) {
 583                 /*Switch on response from Participant */
 584                 control = tdata->recvmsg[i].rcv_status; /* tdata: keeps track of all participant responses
 585                                                            written onto the shared array */
 586                 switch(control) {
 587                         case TRANS_DISAGREE:
 588                                 printf("DEBUG-> trans.c Recv TRANS_DISAGREE\n");
 589                                 transdisagree++;
 590                                 break;
 591
 592                         case TRANS_AGREE:
 593                                 printf("DEBUG-> trans.c Recv TRANS_AGREE\n");
 594                                 transagree++;
 595                                 break;
 596
 597                         case TRANS_SOFT_ABORT:
 598                                 printf("DEBUG-> trans.c Recv TRANS_SOFT_ABORT\n");
 599                                 transsoftabort++;
 600                                 break;
 601                         default:
 602                                 printf("Participant sent unknown message in %s, %d\n", __FILE__, __LINE__);
 603                                 return -1;
 604                 }
 605         }
 606
 607         /* Decide what control message to send to Participant */
 608         if(transdisagree > 0) {
 609                 /* Send Abort */
 610                 *(tdata->replyctrl) = TRANS_ABORT;
 611                 printf("DEBUG-> trans.c Sending TRANS_ABORT\n");
 612                 /* Free resources */
 613                 objstrDelete(tdata->rec->cache);
 614                 chashDelete(tdata->rec->lookupTable);
 615                 free(tdata->rec);
 616         } else if(transagree == tdata->pilecount){
 617                 /* Send Commit */
 618                 *(tdata->replyctrl) = TRANS_COMMIT;
 619                 printf("DEBUG-> trans.c Sending TRANS_COMMIT\n");
 620                 /* Free resources */
 621                 objstrDelete(tdata->rec->cache);
 622                 chashDelete(tdata->rec->lookupTable);
 623                 free(tdata->rec);
 624         } else if(transsoftabort > 0 && transdisagree == 0) {
 625                 /* Send Abort in soft abort case followed by retry commiting transaction again*/
 626                 *(tdata->replyctrl) = TRANS_ABORT;
 627                 *(tdata->replyretry) = 1;
 628                 printf("DEBUG-> trans.c Sending TRANS_ABORT\n");
 629         } else {
 630                 printf("DEBUG -> %s, %d: Error: undecided response\n", __FILE__, __LINE__);
 631                 return -1;
 632         }
 633
 634         return 0;
 635 }
 636 /* This function sends the final response to remote machines per thread in their respective socket id */
 637 char sendResponse(thread_data_array_t *tdata, int sd) {
 638         int n, N, sum, oidcount = 0;
 639         char *ptr, retval = 0;
 640         unsigned int *oidnotfound;
 641
 642         /* If the decided response is due to a soft abort and missing objects at the Participant's side */
 643         if(tdata->recvmsg[tdata->thread_id].rcv_status == TRANS_SOFT_ABORT) {
 644                 /* Read list of objects missing */
 645                 if((read(sd, &oidcount, sizeof(int)) != 0) && (oidcount != 0)) {
 646                         N = oidcount * sizeof(unsigned int);
 647                         if((oidnotfound = calloc(oidcount, sizeof(unsigned int))) == NULL) {
 648                                 printf("Calloc error %s, %d\n", __FILE__, __LINE__);
 649                         }
 650                         ptr = (char *) oidnotfound;
 651                         do {
 652                                 n = read(sd, ptr+sum, N-sum);
 653                                 sum += n;
 654                         } while(sum < N && n !=0);
 655                 }
 656                 retval =  TRANS_SOFT_ABORT;
 657         }
 658         /* If the decided response is TRANS_ABORT */
 659         if(*(tdata->replyctrl) == TRANS_ABORT) {
 660                 retval = TRANS_ABORT;
 661         } else if(*(tdata->replyctrl) == TRANS_COMMIT) { /* If the decided response is TRANS_COMMIT */
 662                 retval = TRANS_COMMIT;
 663         }
 664         /* Send response to the Participant */
 665         if (send(sd, tdata->replyctrl, sizeof(char),MSG_NOSIGNAL) < sizeof(char)) {
 666                 perror("Error sending ctrl message for participant\n");
 667         }
 668
 669         return retval;
 670 }
 671
 672 /* This function opens a connection, places an object read request to the
 673  * remote machine, reads the control message and object if available  and
 674  * copies the object and its header to the local cache.
 675  * TODO replace mnum and midtoIP() with MACHINE_IP address later */
 676
 677 void *getRemoteObj(transrecord_t *record, unsigned int mnum, unsigned int oid) {
 678         int sd, size, val;
 679         struct sockaddr_in serv_addr;
 680         struct hostent *server;
 681         char control;
 682         char machineip[16];
 683         objheader_t *h;
 684         void *objcopy;
 685
 686         if ((sd = socket(AF_INET, SOCK_STREAM, 0)) < 0) {
 687                 perror("Error in socket\n");
 688                 return NULL;
 689         }
 690         bzero((char*) &serv_addr, sizeof(serv_addr));
 691         serv_addr.sin_family = AF_INET;
 692         serv_addr.sin_port = htons(LISTEN_PORT);
 693         //serv_addr.sin_addr.s_addr = inet_addr(MACHINE_IP);
 694         midtoIP(mnum,machineip);
 695         machineip[15] = '\0';
 696         serv_addr.sin_addr.s_addr = inet_addr(machineip);
 697         /* Open connection */
 698         if (connect(sd, (struct sockaddr *) &serv_addr, sizeof(struct sockaddr)) < 0) {
 699                 perror("Error in connect\n");
 700                 return NULL;
 701         }
 702         char readrequest[sizeof(char)+sizeof(unsigned int)];
 703         readrequest[0] = READ_REQUEST;
 704         *((unsigned int *)(&readrequest[1])) = oid;
 705         if (send(sd, &readrequest, sizeof(readrequest), MSG_NOSIGNAL) < sizeof(readrequest)) {
 706                 perror("Error sending message\n");
 707                 return NULL;
 708         }
 709
 710 #ifdef DEBUG1
 711         printf("DEBUG -> ready to rcv ...\n");
 712 #endif
 713         /* Read response from the Participant */
 714         if((val = read(sd, &control, sizeof(char))) <= 0) {
 715                 perror("No control response for getRemoteObj sent\n");
 716                 return NULL;
 717         }
 718         switch(control) {
 719                 case OBJECT_NOT_FOUND:
 720                         printf("DEBUG -> Control OBJECT_NOT_FOUND received\n");
 721                         return NULL;
 722                 case OBJECT_FOUND:
 723                         /* Read object if found into local cache */
 724                         if((val = read(sd, &size, sizeof(int))) <= 0) {
 725                                 perror("No size is read from the participant\n");
 726                                 return NULL;
 727                         }
 728                         objcopy = objstrAlloc(record->cache, size);
 729                         if((val = read(sd, objcopy, size)) <= 0) {
 730                                 perror("No objects are read from the remote participant\n");
 731                                 return NULL;
 732                         }
 733                         /* Insert into cache's lookup table */
 734                         chashInsert(record->lookupTable, oid, objcopy);
 735                         break;
 736                 default:
 737                         printf("Error in recv request from participant on a READ_REQUEST %s, %d\n",__FILE__, __LINE__);
 738                         return NULL;
 739         }
 740         /* Close connection */
 741         close(sd);
 742         return objcopy;
 743 }
 744
 745 /*This function handles the local trans requests involved in a transaction commiting process
 746  * makes a decision if the local machine sends AGREE or DISAGREE or SOFT_ABORT
 747  * Activates the other nonlocal threads that are waiting for the decision and the
 748  * based on common decision by all groups involved in the transaction it
 749  * either commits or aborts the transaction.
 750  * It also frees the calloced memory resources
 751  */
 752
 753 void *handleLocalReq(void *threadarg) {
 754         int val, i = 0;
 755         short version;
 756         char control = 0, *ptr;
 757         unsigned int oid;
 758         unsigned int *oidnotfound = NULL, *oidlocked = NULL, *oidmod = NULL;
 759         void *mobj, *modptr;
 760         objheader_t *headptr;
 761         local_thread_data_array_t *localtdata;
 762
 763         localtdata = (local_thread_data_array_t *) threadarg;
 764
 765         /* Counters and arrays to formulate decision on control message to be sent */
 766         oidnotfound = (unsigned int *) calloc((localtdata->tdata->buffer->f.numread + localtdata->tdata->buffer->f.nummod), sizeof(unsigned int));
 767         oidlocked = (unsigned int *) calloc((localtdata->tdata->buffer->f.numread + localtdata->tdata->buffer->f.nummod), sizeof(unsigned int));
 768         oidmod = (unsigned int *) calloc(localtdata->tdata->buffer->f.nummod, sizeof(unsigned int));
 769         int objnotfound = 0, objlocked = 0, objmod =0, v_nomatch = 0, v_matchlock = 0, v_matchnolock = 0;
 770         int objmodnotfound = 0, nummodfound = 0;
 771
 772         /* modptr points to the beginning of the object store
 773          * created at the Pariticipant */
 774         if ((modptr = objstrAlloc(mainobjstore, localtdata->tdata->buffer->f.sum_bytes)) == NULL) {
 775                 printf("objstrAlloc error for modified objects %s, %d\n", __FILE__, __LINE__);
 776                 return NULL;
 777         }
 778
 779         ptr = modptr;
 780
 781         /* Process each oid in the machine pile/ group per thread */
 782         for (i = 0; i < localtdata->tdata->buffer->f.numread + localtdata->tdata->buffer->f.nummod; i++) {
 783                 if (i < localtdata->tdata->buffer->f.numread) {//Objs only read and not modified
 784                         int incr = sizeof(unsigned int) + sizeof(short);// Offset that points to next position in the objread array
 785                         incr *= i;
 786                         oid = *((unsigned int *)(localtdata->tdata->buffer->objread + incr));
 787                         incr += sizeof(unsigned int);
 788                         version = *((short *)(localtdata->tdata->buffer->objread + incr));
 789                 } else {//Objs modified
 790                         headptr = (objheader_t *) ptr;
 791                         oid = OID(headptr);
 792                         oidmod[objmod] = oid;//Array containing modified oids
 793                         objmod++;
 794                         version = headptr->version;
 795                         ptr += sizeof(objheader_t) + classsize[TYPE(headptr)];
 796                 }
 797
 798                 /* Check if object is still present in the machine since the beginning of TRANS_REQUEST */
 799
 800                 /* Save the oids not found and number of oids not found for later use */
 801                 if ((mobj = mhashSearch(oid)) == NULL) {/* Obj not found */
 802                         /* Save the oids not found and number of oids not found for later use */
 803
 804                         oidnotfound[objnotfound] = OID(((objheader_t *)mobj));
 805                         objnotfound++;
 806                 } else { /* If Obj found in machine (i.e. has not moved) */
 807                         /* Check if Obj is locked by any previous transaction */
 808                         if (STATUS(((objheader_t *)mobj)) & LOCK) {
 809                                 if (version == ((objheader_t *)mobj)->version) {      /* If not locked then match versions */
 810                                         v_matchlock++;
 811                                 } else {/* If versions don't match ...HARD ABORT */
 812                                         v_nomatch++;
 813                                         /* Send TRANS_DISAGREE to Coordinator */
 814                                         localtdata->tdata->recvmsg[localtdata->tdata->thread_id].rcv_status = TRANS_DISAGREE;
 815                                         printf("DEBUG -> Sending TRANS_DISAGREE\n");
 816                                         //return tdata->recvmsg[tdata->thread_id].rcv_status;
 817                                 }
 818                         } else {/* If Obj is not locked then lock object */
 819                                 STATUS(((objheader_t *)mobj)) |= LOCK;
 820                                 //TODO Remove this for Testing
 821                                 randomdelay();
 822
 823                                 /* Save all object oids that are locked on this machine during this transaction request call */
 824                                 oidlocked[objlocked] = OID(((objheader_t *)mobj));
 825                                 objlocked++;
 826                                 if (version == ((objheader_t *)mobj)->version) { /* Check if versions match */
 827                                         v_matchnolock++;
 828                                 } else { /* If versions don't match ...HARD ABORT */
 829                                         v_nomatch++;
 830                                         /* Send TRANS_DISAGREE to Coordinator */
 831                                         localtdata->tdata->recvmsg[localtdata->tdata->thread_id].rcv_status = TRANS_DISAGREE;
 832                                         printf("DEBUG -> Sending TRANS_DISAGREE\n");
 833                                         //      return tdata->recvmsg[tdata->thread_id].rcv_status;
 834                                 }
 835                         }
 836                 }
 837         }
 838
 839         /*Decide the response to be sent to the Coordinator( the local machine in this case)*/
 840
 841         /* Condition to send TRANS_AGREE */
 842         if(v_matchnolock == localtdata->tdata->buffer->f.numread + localtdata->tdata->buffer->f.nummod) {
 843                 localtdata->tdata->recvmsg[localtdata->tdata->thread_id].rcv_status = TRANS_AGREE;
 844                 printf("DEBUG -> Sending TRANS_AGREE\n");
 845         }
 846         /* Condition to send TRANS_SOFT_ABORT */
 847         if((v_matchlock > 0 && v_nomatch == 0) || (objnotfound > 0 && v_nomatch == 0)) {
 848                 localtdata->tdata->recvmsg[localtdata->tdata->thread_id].rcv_status = TRANS_SOFT_ABORT;
 849                 printf("DEBUG -> Sending TRANS_SOFT_ABORT\n");
 850                 //TODO  currently the only soft abort case that is supported is when object locked by previous
 851                 //transaction => v_matchlock > 0
 852                 //The other case for SOFT ABORT i.e. when object is not found but versions match is not supported
 853                 /* Send number of oids not found and the missing oids if objects are missing in the machine */
 854                 /* TODO Remember to store the oidnotfound for later use
 855                    if(objnotfound != 0) {
 856                    int size = sizeof(unsigned int)* objnotfound;
 857                    }
 858                    */
 859         }
 860
 861         /* Fill out the trans_commit_data_t data structure. This is required for a trans commit process
 862          * if Participant receives a TRANS_COMMIT */
 863         localtdata->transinfo->objmod = oidmod;
 864         localtdata->transinfo->objlocked = oidlocked;
 865         localtdata->transinfo->objnotfound = oidnotfound;
 866         localtdata->transinfo->modptr = modptr;
 867         localtdata->transinfo->nummod = localtdata->tdata->buffer->f.nummod;
 868         localtdata->transinfo->numlocked = objlocked;
 869         localtdata->transinfo->numnotfound = objnotfound;
 870
 871         /*Set flag to show that common data structure for this individual thread has been written to */
 872         //*(tdata->localstatus) |= LM_UPDATED;
 873
 874         /* Lock and update count */
 875         //Thread sleeps until all messages from pariticipants are received by coordinator
 876         pthread_mutex_lock(localtdata->tdata->lock);
 877         (*(localtdata->tdata->count))++; /* keeps track of no of messages received by the coordinator */
 878
 879         /* Wake up the threads and invoke decideResponse (once) */
 880         if(*(localtdata->tdata->count) == localtdata->tdata->pilecount) {
 881                 if (decideResponse(localtdata->tdata) != 0) {
 882                         printf("decideResponse returned error %s,%d\n", __FILE__, __LINE__);
 883                         pthread_mutex_unlock(localtdata->tdata->lock);
 884                         return NULL;
 885                 }
 886                 pthread_cond_broadcast(localtdata->tdata->threshold);
 887         } else {
 888                 pthread_cond_wait(localtdata->tdata->threshold, localtdata->tdata->lock);
 889         }
 890         pthread_mutex_unlock(localtdata->tdata->lock);
 891
 892         /*Based on DecideResponse(), Either COMMIT or ABORT the operation*/
 893         if(*(localtdata->tdata->replyctrl) == TRANS_ABORT){
 894                 if(transAbortProcess(modptr,oidlocked, localtdata->transinfo->numlocked, localtdata->transinfo->nummod, localtdata->tdata->buffer->f.numread) != 0) {
 895                         printf("Error in transAbortProcess() %s,%d\n", __FILE__, __LINE__);
 896                         return NULL;
 897                 }
 898         }else if(*(localtdata->tdata->replyctrl) == TRANS_COMMIT){
 899                 if(transComProcess(localtdata->transinfo) != 0) {
 900                         printf("Error in transComProcess() %s,%d\n", __FILE__, __LINE__);
 901                         return NULL;
 902                 }
 903         }
 904
 905         /* Free memory */
 906         printf("DEBUG -> Freeing...\n");
 907         fflush(stdout);
 908         if (localtdata->transinfo->objmod != NULL) {
 909                 free(localtdata->transinfo->objmod);
 910                 localtdata->transinfo->objmod = NULL;
 911         }
 912         if (localtdata->transinfo->objlocked != NULL) {
 913                 free(localtdata->transinfo->objlocked);
 914                 localtdata->transinfo->objlocked = NULL;
 915         }
 916         if (localtdata->transinfo->objnotfound != NULL) {
 917                 free(localtdata->transinfo->objnotfound);
 918                 localtdata->transinfo->objnotfound = NULL;
 919         }
 920
 921         pthread_exit(NULL);
 922 }
 923 /* This function completes the ABORT process if the transaction is aborting
 924 */
 925 int transAbortProcess(void *modptr, unsigned int *objlocked, int numlocked, int nummod, int numread) {
 926         char *ptr;
 927         int i;
 928         objheader_t *tmp_header;
 929         void *header;
 930
 931         printf("DEBUG -> Recv TRANS_ABORT\n");
 932         /* Set all ref counts as 1 and do garbage collection */
 933         ptr = modptr;
 934         for(i = 0; i< nummod; i++) {
 935                 tmp_header = (objheader_t *)ptr;
 936                 tmp_header->rcount = 1;
 937                 ptr += sizeof(objheader_t) + classsize[TYPE(tmp_header)];
 938         }
 939         /* Unlock objects that was locked due to this transaction */
 940         for(i = 0; i< numlocked; i++) {
 941                 header = mhashSearch(objlocked[i]);// find the header address
 942                 STATUS(((objheader_t *)header)) &= ~(LOCK);
 943         }
 944
 945         /* Send ack to Coordinator */
 946         printf("DEBUG-> TRANS_SUCCESSFUL\n");
 947
 948         /*Free the pointer */
 949         ptr = NULL;
 950         return 0;
 951 }
 952
 953 /*This function completes the COMMIT process is the transaction is commiting
 954 */
 955 int transComProcess(trans_commit_data_t *transinfo) {
 956         objheader_t *header;
 957         int i = 0, offset = 0;
 958         char control;
 959
 960         printf("DEBUG -> Recv TRANS_COMMIT\n");
 961         /* Process each modified object saved in the mainobject store */
 962         for(i=0; i<transinfo->nummod; i++) {
 963                 if((header = (objheader_t *) mhashSearch(transinfo->objmod[i])) == NULL) {
 964                         printf("mhashsearch returns NULL at %s, %d\n", __FILE__, __LINE__);
 965                 }
 966                 /* Change reference count of older address and free space in objstr ?? */
 967                 header->rcount = 1; //TODO Not sure what would be the val
 968
 969                 /* Change ptr address in mhash table */
 970                 mhashRemove(transinfo->objmod[i]);
 971                 mhashInsert(transinfo->objmod[i], (transinfo->modptr + offset));
 972                 offset += sizeof(objheader_t) + classsize[TYPE(header)];
 973
 974                 /* Update object version number */
 975                 header = (objheader_t *) mhashSearch(transinfo->objmod[i]);
 976                 header->version += 1;
 977         }
 978
 979         /* Unlock locked objects */
 980         for(i=0; i<transinfo->numlocked; i++) {
 981                 header = (objheader_t *) mhashSearch(transinfo->objlocked[i]);
 982                 STATUS(header) &= ~(LOCK);
 983         }
 984
 985         //TODO Update location lookup table
 986
 987         /* Send ack to Coordinator */
 988         printf("DEBUG-> TRANS_SUCESSFUL\n");
 989         return 0;
 990 }
 991
 992 /* This function checks if the prefetch oids are same and have same offsets
 993  * for case x.a.b and y.a.b where x and y have same oid's
 994  * or if a.b.c is a subset of x.b.c.d*/
 995 /* check for case where the generated request a.y.z or x.y.z.g then
 996  * prefetch needs to be generated for x.y.z.g  if oid of a and x are same*/
 997 void checkPrefetchTuples(prefetchqelem_t *node) {
 998         int i,j, count,k, sindex, index;
 999         char *ptr, *tmp;
1000         int ntuples, slength;
1001         unsigned int *oid;
1002         short *endoffsets, *arryfields;
1003
1004         /* Check for the case x.y.z and a.b.c are same oids */
1005         ptr = (char *) node;
1006         ntuples = *(GET_NTUPLES(ptr));
1007         oid = GET_PTR_OID(ptr);
1008         endoffsets = GET_PTR_EOFF(ptr, ntuples);
1009         arryfields = GET_PTR_ARRYFLD(ptr, ntuples);
1010         /* Find offset length for each tuple */
1011         int numoffset[ntuples];
1012         numoffset[0] = endoffsets[0];
1013         for(i = 1; i<ntuples; i++) {
1014                 numoffset[i] = endoffsets[i] - endoffsets[i-1];
1015         }
1016         /* Check for redundant tuples by comparing oids of each tuple */
1017         for(i = 0; i < ntuples; i++) {
1018                 if(oid[i] == -1)
1019                         continue;
1020                 for(j = i+1 ; j < ntuples; j++) {
1021                         if(oid[j] == -1)
1022                                 continue;
1023                         /*If oids of tuples match */
1024                         if (oid[i] == oid[j]) {
1025                                 /* Find the smallest offset length of two tuples*/
1026                                 if(numoffset[i] >  numoffset[j]){
1027                                         slength = numoffset[j];
1028                                         sindex = j;
1029                                 }
1030                                 else {
1031                                         slength = numoffset[i];
1032                                         sindex = i;
1033                                 }
1034
1035                                 /* Compare the offset values based on the current indices
1036                                  * break if they do not match
1037                                  * if all offset values match then pick the largest tuple*/
1038
1039                                 if(i == 0) {
1040                                         k = 0;
1041                                         index = endoffsets[j -1];
1042                                         for(count = 0; count < slength; count ++) {
1043                                                 if (arryfields[k] != arryfields[index]) {
1044                                                         break;
1045                                                 }
1046                                                 index++;
1047                                                 k++;
1048                                         }
1049                                 } else {
1050                                         k = endoffsets[i-1];
1051                                         index = endoffsets[j-1];
1052                                         printf("Value of slength = %d\n", slength);
1053                                         for(count = 0; count < slength; count++) {
1054                                                 if(arryfields[k] != arryfields[index]) {
1055                                                         break;
1056                                                 }
1057                                                 index++;
1058                                                 k++;
1059                                         }
1060                                 }
1061
1062                                 if(slength == count) {
1063                                         oid[sindex] = -1;
1064                                 }
1065                         }
1066                 }
1067         }
1068 }
1069
1070 void checkPreCache(prefetchqelem_t *node, int *numoffset, int counter, int loopcount, unsigned int objoid, int index, int iter, int oidnfound) {
1071         char *ptr, *tmp;
1072         int ntuples, i, k, flag;
1073         unsigned int * oid;
1074         short *endoffsets, *arryfields;
1075         objheader_t *header;
1076
1077         ptr = (char *) node;
1078         ntuples = *(GET_NTUPLES(ptr));
1079         oid = GET_PTR_OID(ptr);
1080         endoffsets = GET_PTR_EOFF(ptr, ntuples);
1081         arryfields = GET_PTR_ARRYFLD(ptr, ntuples);
1082
1083         if(oidnfound == 1) {
1084                 if((header = (objheader_t *) prehashSearch(objoid)) == NULL) {
1085                         return;
1086                 } else { //Found in Prefetch Cache
1087                         //TODO Decide if object is too old, if old remove from cache
1088                         tmp = (char *) header;
1089                         /* Check if any of the offset oid is available in the Prefetch cache */
1090                         for(i = counter; i < loopcount; i++) {
1091                                 objoid = *(tmp + sizeof(objheader_t) + arryfields[counter]);
1092                                 if((header = (objheader_t *)prehashSearch(objoid)) != NULL) {
1093                                         flag = 0;
1094                                 } else {
1095                                         flag = 1;
1096                                         break;
1097                                 }
1098                         }
1099                 }
1100         } else {
1101                 for(i = counter; i<loopcount; i++) {
1102                         if((header = (objheader_t *)prehashSearch(objoid)) != NULL) {
1103                                 tmp = (char *) header;
1104                                 objoid = *(tmp + sizeof(objheader_t) + arryfields[index]);
1105                                 flag = 0;
1106                                 index++;
1107                         } else {
1108                                 flag = 1;
1109                                 break;
1110                         }
1111                 }
1112         }
1113
1114         /* If oid not found locally or in prefetch cache then
1115          * assign the latest oid found as the new oid
1116          * and copy left over offsets into the arrayoffsetfieldarray*/
1117         oid[iter] = objoid;
1118         numoffset[iter] = numoffset[iter] - (i+1);
1119         for(k = 0; k < numoffset[iter] ; k++) {
1120                 arryfields[endoffsets[counter]+k] = arryfields[endoffsets[counter]+k+1];
1121         }
1122
1123         if(flag == 0) {
1124                 oid[iter] = -1;
1125                 numoffset[iter] = 0;
1126         }
1127 }
1128
1129 /* This function makes machine piles to be added into the machine pile queue for each prefetch call */
1130 prefetchpile_t *makePreGroups(prefetchqelem_t *node, int *numoffset) {
1131         char *ptr, *tmp;
1132         int ntuples, slength, i, machinenum;
1133         int maxoffset;
1134         unsigned int *oid;
1135         short *endoffsets, *arryfields, *offset;
1136         prefetchpile_t *head = NULL;
1137
1138         /* Check for the case x.y.z and a.b.c are same oids */
1139         ptr = (char *) node;
1140         ntuples = *(GET_NTUPLES(ptr));
1141         oid = GET_PTR_OID(ptr);
1142         endoffsets = GET_PTR_EOFF(ptr, ntuples);
1143         arryfields = GET_PTR_ARRYFLD(ptr, ntuples);
1144
1145         /* Check for redundant tuples by comparing oids of each tuple */
1146         for(i = 0; i < ntuples; i++) {
1147                 if(oid[i] == -1)
1148                         continue;
1149                 /* For each tuple make piles */
1150                 if ((machinenum = lhashSearch(oid[i])) == 0) {
1151                         printf("Error: No such Machine %s, %d\n", __FILE__, __LINE__);
1152                         return NULL;
1153                 }
1154                 /* Insert into machine pile */
1155                 offset = &arryfields[endoffsets[i-1]];
1156                 insertPile(machinenum, oid[i], numoffset[i], offset, head);
1157         }
1158
1159         return head;
1160 }
1161
1162
1163 /* This function checks if the oids within the prefetch tuples are available locally.
1164  * If yes then makes the tuple invalid. If no then rearranges oid and offset values in
1165  * the prefetchqelem_t node to represent a new prefetch tuple */
1166 prefetchpile_t *foundLocal(prefetchqelem_t *node) {
1167         int ntuples,i, j, k, oidnfound = 0, index, flag;
1168         unsigned int *oid;
1169         unsigned int  objoid;
1170         char *ptr, *tmp;
1171         objheader_t *objheader;
1172         short *endoffsets, *arryfields;
1173         prefetchpile_t *head = NULL;
1174
1175         ptr = (char *) node;
1176         ntuples = *(GET_NTUPLES(ptr));
1177         oid = GET_PTR_OID(ptr);
1178         endoffsets = GET_PTR_EOFF(ptr, ntuples);
1179         arryfields = GET_PTR_ARRYFLD(ptr, ntuples);
1180         /* Find offset length for each tuple */
1181         int numoffset[ntuples];//Number of offsets for each tuple
1182         numoffset[0] = endoffsets[0];
1183         for(i = 1; i<ntuples; i++) {
1184                 numoffset[i] = endoffsets[i] - endoffsets[i-1];
1185         }
1186         for(i = 0; i < ntuples; i++) {
1187                 if(oid[i] == -1)
1188                         continue;
1189                 /* If object found locally */
1190                 if((objheader = (objheader_t*) mhashSearch(oid[i])) != NULL) {
1191                         oidnfound = 0;
1192                         tmp = (char *) objheader;
1193                         /* Find the oid of its offset value */
1194                         if(i == 0)
1195                                 index = 0;
1196                         else
1197                                 index = endoffsets[i - 1];
1198                         for(j = 0 ; j < numoffset[i] ; j++) {
1199                                 objoid = *(tmp + sizeof(objheader_t) + arryfields[index]);
1200                                 /*If oid found locally then
1201                                  *assign the latest oid found as the new oid
1202                                  *and copy left over offsets into the arrayoffsetfieldarray*/
1203                                 oid[i] = objoid;
1204                                 numoffset[i] = numoffset[i] - (j+1);
1205                                 for(k = 0; k < numoffset[i]; k++)
1206                                         arryfields[endoffsets[j]+ k] = arryfields[endoffsets[j]+k+1];
1207                                 index++;
1208                                 /*New offset oid not found */
1209                                 if((objheader = (objheader_t*) mhashSearch(objoid)) == NULL) {
1210                                         flag = 1;
1211                                         checkPreCache(node, numoffset, j, numoffset[i], objoid, index, i, oidnfound);
1212                                         break;
1213                                 } else
1214                                         flag = 0;
1215                         }
1216
1217                         /*If all offset oids are found locally,make the prefetch tuple invalid */
1218                         if(flag == 0) {
1219                                 oid[i] = -1;
1220                                 numoffset[i] = 0;
1221                         }
1222                 } else {
1223                         oidnfound = 1;
1224                         /* Look in Prefetch cache */
1225                         checkPreCache(node, numoffset, 0, numoffset[i], oid[i], 0, i, oidnfound);
1226                 }
1227
1228         }
1229         /* Make machine groups */
1230         head = makePreGroups(node, numoffset);
1231         return head;
1232 }
1233
1234 /* This function is called by the thread calling transPrefetch */
1235 void *transPrefetch(void *t) {
1236         //int *offstarray = NULL;
1237         prefetchqelem_t *qnode;
1238         prefetchpile_t *pilehead = NULL;
1239
1240         while(1) {
1241                 /* lock mutex of primary prefetch queue */
1242                 pthread_mutex_lock(&pqueue.qlock);
1243                 /* while primary queue is empty, then wait */
1244                 while((pqueue.front == NULL) && (pqueue.rear == NULL)) {
1245                         pthread_cond_wait(&pqueue.qcond, &pqueue.qlock);
1246                 }
1247
1248                 /* dequeue node to create a machine piles and  finally unlock mutex */
1249                 if((qnode = pre_dequeue()) == NULL) {
1250                         printf("Error: No node returned %s, %d\n", __FILE__, __LINE__);
1251                         return NULL;
1252                 }
1253                 pthread_mutex_unlock(&pqueue.qlock);
1254                 /* Reduce redundant prefetch requests */
1255                 checkPrefetchTuples(qnode);
1256                 /* Check if the tuples are found locally, if yes then reduce them further*/
1257                 /* and group requests by remote machine ids by calling the makePreGroups() */
1258                 pilehead = foundLocal(qnode);
1259
1260                 /* Lock mutex of pool queue */
1261                 pthread_mutex_lock(&mcqueue.qlock);
1262                 /* Update the pool queue with the new remote machine piles generated per prefetch call */
1263                 mcpileenqueue(pilehead);
1264                 /* Broadcast signal on machine pile queue */
1265                 pthread_cond_broadcast(&mcqueue.qcond);
1266                 /* Unlock mutex of  machine pile queue */
1267                 pthread_mutex_unlock(&mcqueue.qlock);
1268                 /* Deallocate the prefetch queue pile node */
1269                 predealloc(qnode);
1270
1271         }
1272 }
1273
1274 /* Each thread in the  pool of threads calls this function to establish connection with
1275  * remote machines, send the prefetch requests and process the reponses from
1276  * the remote machines .
1277  * The thread is active throughout the period of runtime */
1278
1279 void *mcqProcess(void *threadid) {
1280         int tid;
1281         prefetchpile_t *mcpilenode;
1282
1283         tid = (int) threadid;
1284         while(1) {
1285                 /* Lock mutex of mc pile queue */
1286                 pthread_mutex_lock(&mcqueue.qlock);
1287                 /* When mc pile queue is empty, wait */
1288                 while((mcqueue.front == NULL) && (mcqueue.rear == NULL)) {
1289                         pthread_cond_wait(&mcqueue.qcond, &mcqueue.qlock);
1290                 }
1291                 /* Dequeue node to send remote machine connections*/
1292                 if((mcpilenode = mcpiledequeue()) == NULL) {
1293                         printf("Dequeue Error: No node returned %s %d\n", __FILE__, __LINE__);
1294                         return NULL;
1295                 }
1296                 /* Unlock mutex */
1297                 pthread_mutex_unlock(&mcqueue.qlock);
1298
1299                 /*Initiate connection to remote host and send request */
1300                 /* Process Request */
1301                 sendPrefetchReq(mcpilenode, tid);
1302
1303                 /* Deallocate the machine queue pile node */
1304                 mcdealloc(mcpilenode);
1305         }
1306 }
1307
1308 void sendPrefetchReq(prefetchpile_t *mcpilenode, int threadid) {
1309         int sd, i, offset, off, len, endpair, numoffsets, count = 0;
1310         struct sockaddr_in serv_addr;
1311         struct hostent *server;
1312         char machineip[16], control;
1313         objpile_t *tmp;
1314
1315
1316         /* Send Trans Prefetch Request */
1317         if ((sd = socket(AF_INET, SOCK_STREAM, 0)) < 0) {
1318                 perror("Error in socket for TRANS_REQUEST\n");
1319                 return;
1320         }
1321         bzero((char*) &serv_addr, sizeof(serv_addr));
1322         serv_addr.sin_family = AF_INET;
1323         serv_addr.sin_port = htons(LISTEN_PORT);
1324         //serv_addr.sin_addr.s_addr = inet_addr(MACHINE_IP);
1325         midtoIP(mcpilenode->mid ,machineip);
1326         machineip[15] = '\0';
1327         serv_addr.sin_addr.s_addr = inet_addr(machineip);
1328
1329         /* Open Connection */
1330         if (connect(sd, (struct sockaddr *) &serv_addr, sizeof(struct sockaddr)) < 0) {
1331                 perror("Error in connect for TRANS_REQUEST\n");
1332                 return;
1333         }
1334
1335         /* Send TRANS_PREFETCH control message */
1336         control = TRANS_PREFETCH;
1337         if(send(sd, &control, sizeof(char), MSG_NOSIGNAL) < sizeof(char)) {
1338                 perror("Error in sending prefetch control\n");
1339                 return;
1340         }
1341
1342         /* Send Oids and offsets in pairs */
1343         tmp = mcpilenode->objpiles;
1344         while(tmp != NULL) {
1345                 off = offset = 0;
1346                 count++;  // Keeps track of the number of oid and offset tuples sent per remote machine
1347                 len = sizeof(int) + sizeof(unsigned int) + ((tmp->numoffset) * sizeof(short));
1348                 char oidnoffset[len];
1349                 memcpy(oidnoffset, &len, sizeof(int));
1350                 off = sizeof(int);
1351                 memcpy(oidnoffset + off, &tmp->oid, sizeof(unsigned int));
1352                 off += sizeof(unsigned int);
1353                 for(i = 0; i < numoffsets; i++) {
1354                         offset = off +  (i * sizeof(short));
1355                         memcpy(oidnoffset + offset, tmp->offset, sizeof(short));
1356                 }
1357                 if (send(sd, &oidnoffset, sizeof(oidnoffset),MSG_NOSIGNAL) < sizeof(oidnoffset)) {
1358                         perror("Error sending fixed bytes for thread\n");
1359                         return;
1360                 }
1361                 tmp = tmp->next;
1362         }
1363
1364         /* Send a special char -1 to represent the end of sending oids + offset pair to remote machine */
1365         endpair = -1;
1366         if (send(sd, &endpair, sizeof(int), MSG_NOSIGNAL) < sizeof(int)) {
1367                 perror("Error sending fixed bytes for thread\n");
1368                 return;
1369         }
1370
1371         /* Get Response from the remote machine */
1372         getPrefetchResponse(count,sd);
1373         close(sd);
1374         return;
1375 }
1376
1377 void getPrefetchResponse(int count, int sd) {
1378         int i = 0, val, n, N, sum, index, objsize;
1379         unsigned int bufsize,oid;
1380         char buffer[RECEIVE_BUFFER_SIZE], control;
1381         char *ptr;
1382         void *modptr, *oldptr;
1383
1384         /* Read  prefetch response from the Remote machine */
1385         if((val = read(sd, &control, sizeof(char))) <= 0) {
1386                 perror("No control response for Prefetch request sent\n");
1387                 return;
1388         }
1389
1390         if(control == TRANS_PREFETCH_RESPONSE) {
1391                 /*For each oid and offset tuple sent as prefetch request to remote machine*/
1392                 while(i < count) {
1393                         /* Clear contents of buffer */
1394                         memset(buffer, 0, RECEIVE_BUFFER_SIZE);
1395                         sum = 0;
1396                         index = 0;
1397                         /* Read the size of buffer to be received */
1398                         if((N = read(sd, buffer, sizeof(unsigned int))) <= 0) {
1399                                 perror("Size of buffer not recv\n");
1400                                 return;
1401                         }
1402                         memcpy(&bufsize, buffer, sizeof(unsigned int));
1403                         ptr = buffer + sizeof(unsigned int);
1404                         /* Keep receiving the buffer containing oid info */
1405                         do {
1406                                 n = recv((int)sd, (void *)ptr+sum, bufsize-sum, 0);
1407                                 sum +=n;
1408                         } while(sum < bufsize && n != 0);
1409                         /* Decode the contents of the buffer */
1410                         index = sizeof(unsigned int);
1411                         while(index < (bufsize - sizeof(unsigned int))) {
1412                                 if(buffer[index] == OBJECT_FOUND) {
1413                                         /* Increment it to get the object */
1414                                         index += sizeof(char);
1415                                         memcpy(&oid, buffer + index, sizeof(unsigned int));
1416                                         index += sizeof(unsigned int);
1417                                         /* Lock the Prefetch Cache look up table*/
1418                                         pthread_mutex_lock(&pflookup.lock);
1419                                         /* For each object found add to Prefetch Cache */
1420                                         memcpy(&objsize, buffer + index, sizeof(int));
1421                                         if ((modptr = objstrAlloc(prefetchcache, objsize)) == NULL) {
1422                                                 printf("objstrAlloc error for copying into prefetch cache %s, %d\n", __FILE__, __LINE__);
1423                                                 return;
1424                                         }
1425                                         memcpy(modptr, buffer+index, objsize);
1426                                         index += sizeof(int);
1427                                         /* Insert the oid and its address into the prefetch hash lookup table */
1428                                         /* Do a version comparison if the oid exists */
1429                                         if((oldptr = prehashSearch(oid)) != NULL) {
1430                                                 /* If older version then update with new object ptr */
1431                                                 if(((objheader_t *)oldptr)->version < ((objheader_t *)modptr)->version) {
1432                                                         prehashRemove(oid);
1433                                                         prehashInsert(oid, modptr);
1434                                                 } else if(((objheader_t *)oldptr)->version == ((objheader_t *)modptr)->version) {
1435                                                         /* Add the new object ptr to hash table */
1436                                                         prehashInsert(oid, modptr);
1437                                                 } else { /* Do nothing */
1438                                                         ;
1439                                                 }
1440                                         } else {/*If doesn't no match found in hashtable, add the object ptr to hash table*/
1441                                                 prehashInsert(oid, modptr);
1442                                         }
1443                                         /* Broadcast signal on prefetch cache condition variable */
1444                                         pthread_cond_broadcast(&pflookup.cond);
1445                                         /* Unlock the Prefetch Cache look up table*/
1446                                         pthread_mutex_unlock(&pflookup.lock);
1447                                 } else if(buffer[index] == OBJECT_NOT_FOUND) {
1448                                         /* Increment it to get the object */
1449                                         /* TODO: For each object not found query DHT for new location and retrieve the object */
1450                                         index += sizeof(char);
1451                                         memcpy(&oid, buffer + index, sizeof(unsigned int));
1452                                         index += sizeof(unsigned int);
1453                                         /* Throw an error */
1454                                         printf("OBJECT NOT FOUND.... THIS SHOULD NOT HAPPEN...TERMINATE PROGRAM\n");
1455                                         exit(-1);
1456                                 } else
1457                                         printf("Error in decoding the index value %s, %d\n",__FILE__, __LINE__);
1458                         }
1459
1460                         i++;
1461                 }
1462         } else
1463                 printf("Error in receving response for prefetch request %s, %d\n",__FILE__, __LINE__);
1464         return;
1465 }