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 #if 0
 222                 printf("Inside remote machine\n");
 223                 pthread_mutex_lock(&pflookup.lock);
 224                 while(!found) {
 225                         rc = pthread_cond_timedwait(&pflookup.cond, &pflookup.lock, &ts);
 226                         if(rc == ETIMEDOUT) {
 227                                 printf("Wait timed out\n");
 228                                 /* Check Prefetch cache again */
 229                                 if((tmp = (objheader_t *) prehashSearch(oid)) != NULL) { /* Look up in prefetch cache */
 230                                         found = 1;
 231                                         size = sizeof(objheader_t)+classsize[TYPE(tmp)];
 232                                         objcopy = objstrAlloc(record->cache, size);
 233                                         memcpy(objcopy, (void *)tmp, size);
 234                                         /* Insert into cache's lookup table */
 235                                         chashInsert(record->lookupTable, OID(tmp), objcopy);
 236                                         return(objcopy);
 237                                 } else {
 238                                         pthread_mutex_unlock(&pflookup.lock);
 239                                         break;
 240                                 }
 241                                 pthread_mutex_unlock(&pflookup.lock);
 242                         }
 243                 }
 244 #endif
 245                 /* Get the object from the remote location */
 246                 machinenumber = lhashSearch(oid);
 247                 objcopy = getRemoteObj(record, machinenumber, oid);
 248                 if(objcopy == NULL) {
 249                         //If object is not found in Remote location
 250                         return NULL;
 251                 }
 252                 else {
 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                         //TODO Check is the object is newly created ...if not then lookup the location table
 296
 297                         if ((machinenum = lhashSearch(curr->key)) == 0) {
 298                                 printf("Error: No such machine %s, %d\n", __FILE__, __LINE__);
 299                                 return NULL;
 300                         }
 301
 302                         if ((headeraddr = chashSearch(record->lookupTable, curr->key)) == NULL) {
 303                                 printf("Error: No such oid %s, %d\n", __FILE__, __LINE__);
 304                                 return NULL;
 305                         }
 306                         //Make machine groups
 307                         if ((pile = pInsert(pile, headeraddr, machinenum, record->lookupTable->numelements)) == NULL) {
 308                                 printf("pInsert error %s, %d\n", __FILE__, __LINE__);
 309                                 return NULL;
 310                         }
 311
 312                         /* Check if local or not */
 313                         if((localmachinenum = mhashSearch(curr->key)) != NULL) {
 314                                 /* Set the pile->local flag*/
 315                                 pile->local = 1; //True i.e. local
 316                         }
 317
 318                         curr = next;
 319                 }
 320         }
 321         return pile;
 322 }
 323
 324 /* This function initiates the transaction commit process
 325  * Spawns threads for each of the new connections with Participants
 326  * and creates new piles by calling the createPiles(),
 327  * Sends a transrequest() to each remote machines for objects found remotely
 328  * and calls handleLocalReq() to process objects found locally */
 329 int transCommit(transrecord_t *record) {
 330         unsigned int tot_bytes_mod, *listmid;
 331         plistnode_t *pile, *pile_ptr;
 332         int i, rc, val;
 333         int pilecount = 0, offset, threadnum = 0, trecvcount = 0, tmachcount = 0;
 334         char buffer[RECEIVE_BUFFER_SIZE],control;
 335         char transid[TID_LEN];
 336         trans_req_data_t *tosend;
 337         trans_commit_data_t transinfo;
 338         static int newtid = 0;
 339         char treplyctrl = 0, treplyretry = 0; /* keeps track of the common response that needs to be sent */
 340         char localstat = 0;
 341
 342
 343         /* Look through all the objects in the transaction record and make piles
 344          * for each machine involved in the transaction*/
 345         pile_ptr = pile = createPiles(record);
 346
 347         /* Create the packet to be sent in TRANS_REQUEST */
 348
 349         /* Count the number of participants */
 350         pilecount = pCount(pile);
 351
 352         /* Create a list of machine ids(Participants) involved in transaction   */
 353         if((listmid = calloc(pilecount, sizeof(unsigned int))) == NULL) {
 354                 printf("Calloc error %s, %d\n", __FILE__, __LINE__);
 355                 return 1;
 356         }
 357         pListMid(pile, listmid);
 358
 359
 360         /* Initialize thread variables,
 361          * Spawn a thread for each Participant involved in a transaction */
 362         pthread_t thread[pilecount];
 363         pthread_attr_t attr;
 364         pthread_cond_t tcond;
 365         pthread_mutex_t tlock;
 366         pthread_mutex_t tlshrd;
 367
 368         thread_data_array_t *thread_data_array;
 369         thread_data_array = (thread_data_array_t *) malloc(sizeof(thread_data_array_t)*pilecount);
 370         local_thread_data_array_t *ltdata;
 371         if((ltdata = calloc(1, sizeof(local_thread_data_array_t))) == NULL) {
 372                 printf("Calloc error %s, %d\n", __FILE__, __LINE__);
 373                 return 1;
 374         }
 375
 376         thread_response_t rcvd_control_msg[pilecount];  /* Shared thread array that keeps track of responses of participants */
 377
 378         /* Initialize and set thread detach attribute */
 379         pthread_attr_init(&attr);
 380         pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE);
 381         pthread_mutex_init(&tlock, NULL);
 382         pthread_cond_init(&tcond, NULL);
 383
 384         /* Process each machine pile */
 385         while(pile != NULL) {
 386                 //Create transaction id
 387                 newtid++;
 388                 if ((tosend = calloc(1, sizeof(trans_req_data_t))) == NULL) {
 389                         printf("Calloc error %s, %d\n", __FILE__, __LINE__);
 390                         return 1;
 391                 }
 392                 tosend->f.control = TRANS_REQUEST;
 393                 sprintf(tosend->f.trans_id, "%x_%d", pile->mid, newtid);
 394                 tosend->f.mcount = pilecount;
 395                 tosend->f.numread = pile->numread;
 396                 tosend->f.nummod = pile->nummod;
 397                 tosend->f.sum_bytes = pile->sum_bytes;
 398                 tosend->listmid = listmid;
 399                 tosend->objread = pile->objread;
 400                 tosend->oidmod = pile->oidmod;
 401                 thread_data_array[threadnum].thread_id = threadnum;
 402                 thread_data_array[threadnum].mid = pile->mid;
 403                 thread_data_array[threadnum].pilecount = pilecount;
 404                 thread_data_array[threadnum].buffer = tosend;
 405                 thread_data_array[threadnum].recvmsg = rcvd_control_msg;
 406                 thread_data_array[threadnum].threshold = &tcond;
 407                 thread_data_array[threadnum].lock = &tlock;
 408                 thread_data_array[threadnum].count = &trecvcount;
 409                 thread_data_array[threadnum].replyctrl = &treplyctrl;
 410                 thread_data_array[threadnum].replyretry = &treplyretry;
 411                 thread_data_array[threadnum].rec = record;
 412                 /* If local do not create any extra connection */
 413                 if(pile->local != 1) { /* Not local */
 414                         rc = pthread_create(&thread[threadnum], NULL, transRequest, (void *) &thread_data_array[threadnum]);
 415                         if (rc) {
 416                                 perror("Error in pthread create\n");
 417                                 return 1;
 418                         }
 419                 } else { /*Local*/
 420                         /*Unset the pile->local flag*/
 421                         pile->local = 0;
 422                         /*Set flag to identify that Local machine is involved*/
 423                         ltdata->tdata = &thread_data_array[threadnum];
 424                         ltdata->transinfo = &transinfo;
 425                         val = pthread_create(&thread[threadnum], NULL, handleLocalReq, (void *) ltdata);
 426                         if (val) {
 427                                 perror("Error in pthread create\n");
 428                                 return 1;
 429                         }
 430                 }
 431                 threadnum++;
 432                 pile = pile->next;
 433         }
 434
 435         /* Free attribute and wait for the other threads */
 436         pthread_attr_destroy(&attr);
 437         for (i = 0 ;i < pilecount ; i++) {
 438                 rc = pthread_join(thread[i], NULL);
 439                 if (rc)
 440                 {
 441                         printf("ERROR return code from pthread_join() is %d\n", rc);
 442                         return 1;
 443                 }
 444                 free(thread_data_array[i].buffer);
 445         }
 446
 447         /* Free resources */
 448         pthread_cond_destroy(&tcond);
 449         pthread_mutex_destroy(&tlock);
 450         free(listmid);
 451         pDelete(pile_ptr);
 452         free(thread_data_array);
 453         free(ltdata);
 454
 455         /* Retry trans commit procedure if not sucessful in the first try */
 456         if(treplyretry == 1) {
 457                 /* wait a random amount of time */
 458                 randomdelay();
 459                 /* Retry the commiting transaction again */
 460                 transCommit(record);
 461         }
 462
 463         return 0;
 464 }
 465
 466 /* This function sends information involved in the transaction request
 467  * to participants and accepts a response from particpants.
 468  * It calls decideresponse() to decide on what control message
 469  * to send next to participants and sends the message using sendResponse()*/
 470 void *transRequest(void *threadarg) {
 471         int sd, i, n;
 472         struct sockaddr_in serv_addr;
 473         struct hostent *server;
 474         thread_data_array_t *tdata;
 475         objheader_t *headeraddr;
 476         char buffer[RECEIVE_BUFFER_SIZE], control, recvcontrol;
 477         char machineip[16], retval;
 478
 479         tdata = (thread_data_array_t *) threadarg;
 480
 481         /* Send Trans Request */
 482         if ((sd = socket(AF_INET, SOCK_STREAM, 0)) < 0) {
 483                 perror("Error in socket for TRANS_REQUEST\n");
 484                 return NULL;
 485         }
 486         bzero((char*) &serv_addr, sizeof(serv_addr));
 487         serv_addr.sin_family = AF_INET;
 488         serv_addr.sin_port = htons(LISTEN_PORT);
 489         midtoIP(tdata->mid,machineip);
 490         machineip[15] = '\0';
 491         serv_addr.sin_addr.s_addr = inet_addr(machineip);
 492         /* Open Connection */
 493         if (connect(sd, (struct sockaddr *) &serv_addr, sizeof(struct sockaddr)) < 0) {
 494                 perror("Error in connect for TRANS_REQUEST\n");
 495                 return NULL;
 496         }
 497
 498         printf("DEBUG-> trans.c Sending TRANS_REQUEST to mid %s\n", machineip);
 499         /* Send bytes of data with TRANS_REQUEST control message */
 500         if (send(sd, &(tdata->buffer->f), sizeof(fixed_data_t),MSG_NOSIGNAL) < sizeof(fixed_data_t)) {
 501                 perror("Error sending fixed bytes for thread\n");
 502                 return NULL;
 503         }
 504         /* Send list of machines involved in the transaction */
 505         {
 506                 int size=sizeof(unsigned int)*tdata->pilecount;
 507                 if (send(sd, tdata->buffer->listmid, size, MSG_NOSIGNAL) < size) {
 508                         perror("Error sending list of machines for thread\n");
 509                         return NULL;
 510                 }
 511         }
 512         /* Send oids and version number tuples for objects that are read */
 513         {
 514                 int size=(sizeof(unsigned int)+sizeof(short))*tdata->buffer->f.numread;
 515                 if (send(sd, tdata->buffer->objread, size, MSG_NOSIGNAL) < size) {
 516                         perror("Error sending tuples for thread\n");
 517                         return NULL;
 518                 }
 519         }
 520         /* Send objects that are modified */
 521         for(i = 0; i < tdata->buffer->f.nummod ; i++) {
 522                 int size;
 523                 headeraddr = chashSearch(tdata->rec->lookupTable, tdata->buffer->oidmod[i]);
 524                 size=sizeof(objheader_t)+classsize[TYPE(headeraddr)];
 525                 if (send(sd, headeraddr, size, MSG_NOSIGNAL)  < size) {
 526                         perror("Error sending obj modified for thread\n");
 527                         return NULL;
 528                 }
 529         }
 530
 531         /* Read control message from Participant */
 532         if((n = read(sd, &control, sizeof(char))) <= 0) {
 533                 perror("Error in reading control message from Participant\n");
 534                 return NULL;
 535         }
 536         recvcontrol = control;
 537
 538         /* Update common data structure and increment count */
 539         tdata->recvmsg[tdata->thread_id].rcv_status = recvcontrol;
 540
 541         /* Lock and update count */
 542         /* Thread sleeps until all messages from pariticipants are received by coordinator */
 543         pthread_mutex_lock(tdata->lock);
 544
 545         (*(tdata->count))++; /* keeps track of no of messages received by the coordinator */
 546
 547         /* Wake up the threads and invoke decideResponse (once) */
 548         if(*(tdata->count) == tdata->pilecount) {
 549                 if (decideResponse(tdata) != 0) {
 550                         printf("decideResponse returned error %s,%d\n", __FILE__, __LINE__);
 551                         pthread_mutex_unlock(tdata->lock);
 552                         close(sd);
 553                         return NULL;
 554                 }
 555                 pthread_cond_broadcast(tdata->threshold);
 556         } else {
 557                 pthread_cond_wait(tdata->threshold, tdata->lock);
 558         }
 559         pthread_mutex_unlock(tdata->lock);
 560
 561         /* Send the final response such as TRANS_COMMIT or TRANS_ABORT t
 562          * to all participants in their respective socket */
 563         if (sendResponse(tdata, sd) == 0) {
 564                 printf("sendResponse returned error %s,%d\n", __FILE__, __LINE__);
 565                 pthread_mutex_unlock(tdata->lock);
 566                 close(sd);
 567                 return NULL;
 568         }
 569
 570         /* Close connection */
 571         close(sd);
 572         pthread_exit(NULL);
 573 }
 574
 575 /* This function decides the reponse that needs to be sent to
 576  * all Participant machines after the TRANS_REQUEST protocol */
 577 int decideResponse(thread_data_array_t *tdata) {
 578         char control;
 579         int i, transagree = 0, transdisagree = 0, transsoftabort = 0; /* Counters to formulate decision of what
 580                                                                          message to send */
 581
 582         for (i = 0 ; i < tdata->pilecount ; i++) {
 583                 control = tdata->recvmsg[i].rcv_status; /* tdata: keeps track of all participant responses
 584                                                            written onto the shared array */
 585                 switch(control) {
 586                         case TRANS_DISAGREE:
 587                                 printf("DEBUG-> trans.c Recv TRANS_DISAGREE\n");
 588                                 transdisagree++;
 589                                 break;
 590
 591                         case TRANS_AGREE:
 592                                 printf("DEBUG-> trans.c Recv TRANS_AGREE\n");
 593                                 transagree++;
 594                                 break;
 595
 596                         case TRANS_SOFT_ABORT:
 597                                 printf("DEBUG-> trans.c Recv TRANS_SOFT_ABORT\n");
 598                                 transsoftabort++;
 599                                 break;
 600                         default:
 601                                 printf("Participant sent unknown message in %s, %d\n", __FILE__, __LINE__);
 602                                 return -1;
 603                 }
 604         }
 605
 606         /* Send Abort */
 607         if(transdisagree > 0) {
 608                 *(tdata->replyctrl) = TRANS_ABORT;
 609                 printf("DEBUG-> trans.c Sending TRANS_ABORT\n");
 610                 /* Free resources */
 611                 objstrDelete(tdata->rec->cache);
 612                 chashDelete(tdata->rec->lookupTable);
 613                 free(tdata->rec);
 614         } else if(transagree == tdata->pilecount){
 615                 /* Send Commit */
 616                 *(tdata->replyctrl) = TRANS_COMMIT;
 617                 printf("DEBUG-> trans.c Sending TRANS_COMMIT\n");
 618                 /* Free resources */
 619                 objstrDelete(tdata->rec->cache);
 620                 chashDelete(tdata->rec->lookupTable);
 621                 free(tdata->rec);
 622         } else if(transsoftabort > 0 && transdisagree == 0) {
 623                 /* Send Abort in soft abort case followed by retry commiting transaction again*/
 624                 *(tdata->replyctrl) = TRANS_ABORT;
 625                 *(tdata->replyretry) = 1;
 626                 printf("DEBUG-> trans.c Sending TRANS_ABORT\n");
 627         } else {
 628                 printf("DEBUG -> %s, %d: Error: undecided response\n", __FILE__, __LINE__);
 629                 return -1;
 630         }
 631
 632         return 0;
 633 }
 634 /* This function sends the final response to remote machines per thread in their respective socket id */
 635 char sendResponse(thread_data_array_t *tdata, int sd) {
 636         int n, N, sum, oidcount = 0;
 637         char *ptr, retval = 0;
 638         unsigned int *oidnotfound;
 639
 640         /* If the decided response is due to a soft abort and missing objects at the Participant's side */
 641         if(tdata->recvmsg[tdata->thread_id].rcv_status == TRANS_SOFT_ABORT) {
 642                 /* Read list of objects missing */
 643                 if((read(sd, &oidcount, sizeof(int)) != 0) && (oidcount != 0)) {
 644                         N = oidcount * sizeof(unsigned int);
 645                         if((oidnotfound = calloc(oidcount, sizeof(unsigned int))) == NULL) {
 646                                 printf("Calloc error %s, %d\n", __FILE__, __LINE__);
 647                                 return 0;
 648                         }
 649                         ptr = (char *) oidnotfound;
 650                         do {
 651                                 n = read(sd, ptr+sum, N-sum);
 652                                 sum += n;
 653                         } while(sum < N && n !=0);
 654                 }
 655                 retval =  TRANS_SOFT_ABORT;
 656         }
 657         /* If the decided response is TRANS_ABORT */
 658         if(*(tdata->replyctrl) == TRANS_ABORT) {
 659                 retval = TRANS_ABORT;
 660         } else if(*(tdata->replyctrl) == TRANS_COMMIT) { /* If the decided response is TRANS_COMMIT */
 661                 retval = TRANS_COMMIT;
 662         }
 663
 664         if (send(sd, tdata->replyctrl, sizeof(char),MSG_NOSIGNAL) < sizeof(char)) {
 665                 perror("Error sending ctrl message for participant\n");
 666         }
 667
 668         return retval;
 669 }
 670
 671 /* This function opens a connection, places an object read request to the
 672  * remote machine, reads the control message and object if available  and
 673  * copies the object and its header to the local cache.
 674  * TODO replace mnum and midtoIP() with MACHINE_IP address later */
 675
 676 void *getRemoteObj(transrecord_t *record, unsigned int mnum, unsigned int oid) {
 677         int sd, size, val;
 678         struct sockaddr_in serv_addr;
 679         struct hostent *server;
 680         char control;
 681         char machineip[16];
 682         objheader_t *h;
 683         void *objcopy;
 684
 685         if ((sd = socket(AF_INET, SOCK_STREAM, 0)) < 0) {
 686                 perror("Error in socket\n");
 687                 return NULL;
 688         }
 689         bzero((char*) &serv_addr, sizeof(serv_addr));
 690         serv_addr.sin_family = AF_INET;
 691         serv_addr.sin_port = htons(LISTEN_PORT);
 692         //serv_addr.sin_addr.s_addr = inet_addr(MACHINE_IP);
 693         midtoIP(mnum,machineip);
 694         machineip[15] = '\0';
 695         serv_addr.sin_addr.s_addr = inet_addr(machineip);
 696         /* Open connection */
 697         if (connect(sd, (struct sockaddr *) &serv_addr, sizeof(struct sockaddr)) < 0) {
 698                 perror("Error in connect\n");
 699                 return NULL;
 700         }
 701         char readrequest[sizeof(char)+sizeof(unsigned int)];
 702         readrequest[0] = READ_REQUEST;
 703         *((unsigned int *)(&readrequest[1])) = oid;
 704         if (send(sd, &readrequest, sizeof(readrequest), MSG_NOSIGNAL) < sizeof(readrequest)) {
 705                 perror("Error sending message\n");
 706                 return NULL;
 707         }
 708
 709 #ifdef DEBUG1
 710         printf("DEBUG -> ready to rcv ...\n");
 711 #endif
 712         /* Read response from the Participant */
 713         if((val = read(sd, &control, sizeof(char))) <= 0) {
 714                 perror("No control response for getRemoteObj sent\n");
 715                 return NULL;
 716         }
 717         switch(control) {
 718                 case OBJECT_NOT_FOUND:
 719                         printf("DEBUG -> Control OBJECT_NOT_FOUND received\n");
 720                         return NULL;
 721                 case OBJECT_FOUND:
 722                         /* Read object if found into local cache */
 723                         if((val = read(sd, &size, sizeof(int))) <= 0) {
 724                                 perror("No size is read from the participant\n");
 725                                 return NULL;
 726                         }
 727                         objcopy = objstrAlloc(record->cache, size);
 728                         if((val = read(sd, objcopy, size)) <= 0) {
 729                                 perror("No objects are read from the remote participant\n");
 730                                 return NULL;
 731                         }
 732                         /* Insert into cache's lookup table */
 733                         chashInsert(record->lookupTable, oid, objcopy);
 734                         break;
 735                 default:
 736                         printf("Error in recv request from participant on a READ_REQUEST %s, %d\n",__FILE__, __LINE__);
 737                         return NULL;
 738         }
 739         /* Close connection */
 740         close(sd);
 741         return objcopy;
 742 }
 743
 744 /* This function handles the local objects involved in a transaction commiting process.
 745  * It also makes a decision if this local machine sends AGREE or DISAGREE or SOFT_ABORT to coordinator.
 746  * Note Coordinator = local machine
 747  * It wakes up the other threads from remote participants that are waiting for the coordinator's decision and
 748  * based on common agreement it either commits or aborts the transaction.
 749  * It also frees the memory resources */
 750 void *handleLocalReq(void *threadarg) {
 751         int val, i = 0, size, offset = 0;
 752         short version;
 753         char control = 0, *ptr;
 754         unsigned int oid;
 755         unsigned int *oidnotfound = NULL, *oidlocked = NULL, *oidmod = NULL;
 756         void *mobj, *modptr;
 757         objheader_t *headptr, *headeraddr;
 758         local_thread_data_array_t *localtdata;
 759
 760         localtdata = (local_thread_data_array_t *) threadarg;
 761
 762         /* Counters and arrays to formulate decision on control message to be sent */
 763         oidnotfound = (unsigned int *) calloc((localtdata->tdata->buffer->f.numread + localtdata->tdata->buffer->f.nummod), sizeof(unsigned int));
 764         oidlocked = (unsigned int *) calloc((localtdata->tdata->buffer->f.numread + localtdata->tdata->buffer->f.nummod), sizeof(unsigned int));
 765         int objnotfound = 0, objlocked = 0;
 766         int v_nomatch = 0, v_matchlock = 0, v_matchnolock = 0;
 767
 768         /* modptr points to the beginning of the object store
 769          * created at the Pariticipant */
 770         if ((modptr = objstrAlloc(mainobjstore, localtdata->tdata->buffer->f.sum_bytes)) == NULL) {
 771                 printf("objstrAlloc error for modified objects %s, %d\n", __FILE__, __LINE__);
 772                 return NULL;
 773         }
 774         /* Write modified objects into the mainobject store */
 775         for(i = 0; i< localtdata->tdata->buffer->f.nummod; i++) {
 776                 headeraddr = chashSearch(localtdata->tdata->rec->lookupTable, localtdata->tdata->buffer->oidmod[i]);
 777                 size = sizeof(objheader_t) + classsize[TYPE(headeraddr)];
 778                 memcpy(modptr+offset, headeraddr, size);
 779                 offset += size;
 780         }
 781
 782         ptr = modptr;
 783         offset = 0; //Reset
 784
 785         /* Process each oid in the machine pile/ group per thread */
 786         for (i = 0; i < localtdata->tdata->buffer->f.numread + localtdata->tdata->buffer->f.nummod; i++) {
 787                 if (i < localtdata->tdata->buffer->f.numread) {//Objs only read and not modified
 788                         int incr = sizeof(unsigned int) + sizeof(short);// Offset that points to next position in the objread array
 789                         incr *= i;
 790                         oid = *((unsigned int *)(localtdata->tdata->buffer->objread + incr));
 791                         incr += sizeof(unsigned int);
 792                         version = *((short *)(localtdata->tdata->buffer->objread + incr));
 793                 } else {//Objs modified
 794                         headptr = (objheader_t *)ptr;
 795                         oid = OID(headptr);
 796                         version = headptr->version;
 797                         ptr += sizeof(objheader_t) + classsize[TYPE(headptr)];
 798                 }
 799
 800                 /* Check if object is still present in the machine since the beginning of TRANS_REQUEST */
 801
 802                 /* Save the oids not found and number of oids not found for later use */
 803                 if ((mobj = mhashSearch(oid)) == NULL) {/* Obj not found */
 804                         /* Save the oids not found and number of oids not found for later use */
 805                         oidnotfound[objnotfound] = oid;
 806                         objnotfound++;
 807                 } else { /* If Obj found in machine (i.e. has not moved) */
 808                         /* Check if Obj is locked by any previous transaction */
 809                         if (STATUS(((objheader_t *)mobj)) & LOCK) {
 810                                 if (version == ((objheader_t *)mobj)->version) {      /* If not locked then match versions */
 811                                         v_matchlock++;
 812                                 } else {/* If versions don't match ...HARD ABORT */
 813                                         v_nomatch++;
 814                                         /* Send TRANS_DISAGREE to Coordinator */
 815                                         localtdata->tdata->recvmsg[localtdata->tdata->thread_id].rcv_status = TRANS_DISAGREE;
 816                                         printf("DEBUG -> Sending TRANS_DISAGREE\n");
 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                                 }
 834                         }
 835                 }
 836         }
 837
 838         /* Condition to send TRANS_AGREE */
 839         if(v_matchnolock == localtdata->tdata->buffer->f.numread + localtdata->tdata->buffer->f.nummod) {
 840                 localtdata->tdata->recvmsg[localtdata->tdata->thread_id].rcv_status = TRANS_AGREE;
 841                 printf("DEBUG -> Sending TRANS_AGREE\n");
 842         }
 843         /* Condition to send TRANS_SOFT_ABORT */
 844         if((v_matchlock > 0 && v_nomatch == 0) || (objnotfound > 0 && v_nomatch == 0)) {
 845                 localtdata->tdata->recvmsg[localtdata->tdata->thread_id].rcv_status = TRANS_SOFT_ABORT;
 846                 printf("DEBUG -> Sending TRANS_SOFT_ABORT\n");
 847                 //TODO  currently the only soft abort case that is supported is when object locked by previous
 848                 //transaction => v_matchlock > 0
 849                 //The other case for SOFT ABORT i.e. when object is not found but versions match is not supported
 850                 /* Send number of oids not found and the missing oids if objects are missing in the machine */
 851                 /* TODO Remember to store the oidnotfound for later use
 852                    if(objnotfound != 0) {
 853                    int size = sizeof(unsigned int)* objnotfound;
 854                    }
 855                    */
 856         }
 857
 858         /* Fill out the trans_commit_data_t data structure. This is required for a trans commit process
 859          * if Participant receives a TRANS_COMMIT */
 860         localtdata->transinfo->objlocked = oidlocked;
 861         localtdata->transinfo->objnotfound = oidnotfound;
 862         localtdata->transinfo->modptr = modptr;
 863         localtdata->transinfo->numlocked = objlocked;
 864         localtdata->transinfo->numnotfound = objnotfound;
 865
 866         /* Lock and update count */
 867         //Thread sleeps until all messages from pariticipants are received by coordinator
 868         pthread_mutex_lock(localtdata->tdata->lock);
 869         (*(localtdata->tdata->count))++; /* keeps track of no of messages received by the coordinator */
 870
 871         /* Wake up the threads and invoke decideResponse (once) */
 872         if(*(localtdata->tdata->count) == localtdata->tdata->pilecount) {
 873                 if (decideResponse(localtdata->tdata) != 0) {
 874                         printf("decideResponse returned error %s,%d\n", __FILE__, __LINE__);
 875                         pthread_mutex_unlock(localtdata->tdata->lock);
 876                         return NULL;
 877                 }
 878                 pthread_cond_broadcast(localtdata->tdata->threshold);
 879         } else {
 880                 pthread_cond_wait(localtdata->tdata->threshold, localtdata->tdata->lock);
 881         }
 882         pthread_mutex_unlock(localtdata->tdata->lock);
 883
 884         /*Based on DecideResponse(), Either COMMIT or ABORT the operation*/
 885         if(*(localtdata->tdata->replyctrl) == TRANS_ABORT){
 886                 if(transAbortProcess(modptr,oidlocked, localtdata->transinfo->numlocked, localtdata->tdata->buffer->f.nummod) != 0) {
 887                         printf("Error in transAbortProcess() %s,%d\n", __FILE__, __LINE__);
 888                         return NULL;
 889                 }
 890         }else if(*(localtdata->tdata->replyctrl) == TRANS_COMMIT){
 891                 if(transComProcess(modptr, localtdata->tdata->buffer->oidmod, oidlocked, localtdata->tdata->buffer->f.nummod, localtdata->transinfo->numlocked) != 0) {
 892                         printf("Error in transComProcess() %s,%d\n", __FILE__, __LINE__);
 893                         return NULL;
 894                 }
 895         }
 896
 897         /* Free memory */
 898         printf("DEBUG -> Freeing...\n");
 899         fflush(stdout);
 900
 901         if (localtdata->transinfo->objlocked != NULL) {
 902                 free(localtdata->transinfo->objlocked);
 903                 localtdata->transinfo->objlocked = NULL;
 904         }
 905         if (localtdata->transinfo->objnotfound != NULL) {
 906                 free(localtdata->transinfo->objnotfound);
 907                 localtdata->transinfo->objnotfound = NULL;
 908         }
 909
 910         pthread_exit(NULL);
 911 }
 912 /* This function completes the ABORT process if the transaction is aborting
 913 */
 914 int transAbortProcess(void *modptr, unsigned int *objlocked, int numlocked, int nummod) {
 915         char *ptr;
 916         int i;
 917         objheader_t *tmp_header;
 918         void *header;
 919
 920         printf("DEBUG -> Recv TRANS_ABORT\n");
 921         /* Set all ref counts as 1 and do garbage collection */
 922         ptr = (char *)modptr;
 923         for(i = 0; i< nummod; i++) {
 924                 tmp_header = (objheader_t *)ptr;
 925                 tmp_header->rcount = 1;
 926                 ptr += sizeof(objheader_t) + classsize[TYPE(tmp_header)];
 927         }
 928         /* Unlock objects that was locked due to this transaction */
 929         for(i = 0; i< numlocked; i++) {
 930                 if((header = mhashSearch(objlocked[i])) == NULL) {
 931                         printf("mhashsearch returns NULL at %s, %d\n", __FILE__, __LINE__);
 932                         return 1;
 933                 }
 934                 STATUS(((objheader_t *)header)) &= ~(LOCK);
 935         }
 936
 937         /* Send ack to Coordinator */
 938         printf("DEBUG-> TRANS_SUCCESSFUL\n");
 939
 940         /*Free the pointer */
 941         ptr = NULL;
 942         return 0;
 943 }
 944
 945 /*This function completes the COMMIT process is the transaction is commiting
 946 */
 947 int transComProcess(void *modptr, unsigned int *oidmod, unsigned int *objlocked, int nummod, int numlocked) {
 948         objheader_t *header;
 949         int i = 0, offset = 0;
 950         char control;
 951
 952         /* Process each modified object saved in the mainobject store */
 953         for(i = 0; i < nummod; i++) {
 954                 if((header = (objheader_t *) mhashSearch(oidmod[i])) == NULL) {
 955                         printf("mhashsearch returns NULL at %s, %d\n", __FILE__, __LINE__);
 956                         return 1;
 957                 }
 958                 /* Change reference count of older address and free space in objstr ?? */
 959                 header->rcount = 1; //TODO Not sure what would be the val
 960
 961                 /* Change ptr address in mhash table */
 962                 mhashRemove(oidmod[i]);
 963                 mhashInsert(oidmod[i], (((char *)modptr) + offset));
 964                 offset += sizeof(objheader_t) + classsize[TYPE(header)];
 965
 966                 /* Update object version number */
 967                 header = (objheader_t *) mhashSearch(oidmod[i]);
 968                 header->version += 1;
 969         }
 970
 971         /* Unlock locked objects */
 972         for(i = 0; i < numlocked; i++) {
 973                 if((header = (objheader_t *) mhashSearch(objlocked[i])) == NULL) {
 974                         printf("mhashsearch returns NULL at %s, %d\n", __FILE__, __LINE__);
 975                         return 1;
 976                 }
 977                 STATUS(header) &= ~(LOCK);
 978         }
 979
 980         //TODO Update location lookup table
 981
 982         /* Send ack to Coordinator */
 983         printf("DEBUG-> TRANS_SUCESSFUL\n");
 984         return 0;
 985 }
 986
 987 /* This function checks if the prefetch oids are same and have same offsets
 988  * for case x.a.b and y.a.b where x and y have same oid's
 989  * or if a.b.c is a subset of x.b.c.d*/
 990 /* check for case where the generated request a.y.z or x.y.z.g then
 991  * prefetch needs to be generated for x.y.z.g  if oid of a and x are same*/
 992 void checkPrefetchTuples(prefetchqelem_t *node) {
 993         int i,j, count,k, sindex, index;
 994         char *ptr, *tmp;
 995         int ntuples, slength;
 996         unsigned int *oid;
 997         short *endoffsets, *arryfields;
 998
 999         /* Check for the case x.y.z and a.b.c are same oids */
1000         ptr = (char *) node;
1001         ntuples = *(GET_NTUPLES(ptr));
1002         oid = GET_PTR_OID(ptr);
1003         endoffsets = GET_PTR_EOFF(ptr, ntuples);
1004         arryfields = GET_PTR_ARRYFLD(ptr, ntuples);
1005         /* Find offset length for each tuple */
1006         int numoffset[ntuples];
1007         numoffset[0] = endoffsets[0];
1008         for(i = 1; i<ntuples; i++) {
1009                 numoffset[i] = endoffsets[i] - endoffsets[i-1];
1010         }
1011         /* Check for redundant tuples by comparing oids of each tuple */
1012         for(i = 0; i < ntuples; i++) {
1013                 if(oid[i] == -1)
1014                         continue;
1015                 for(j = i+1 ; j < ntuples; j++) {
1016                         if(oid[j] == -1)
1017                                 continue;
1018                         /*If oids of tuples match */
1019                         if (oid[i] == oid[j]) {
1020                                 /* Find the smallest offset length of two tuples*/
1021                                 if(numoffset[i] >  numoffset[j]){
1022                                         slength = numoffset[j];
1023                                         sindex = j;
1024                                 }
1025                                 else {
1026                                         slength = numoffset[i];
1027                                         sindex = i;
1028                                 }
1029
1030                                 /* Compare the offset values based on the current indices
1031                                  * break if they do not match
1032                                  * if all offset values match then pick the largest tuple*/
1033
1034                                 if(i == 0) {
1035                                         k = 0;
1036                                         index = endoffsets[j -1];
1037                                         for(count = 0; count < slength; count ++) {
1038                                                 if (arryfields[k] != arryfields[index]) {
1039                                                         break;
1040                                                 }
1041                                                 index++;
1042                                                 k++;
1043                                         }
1044                                 } else {
1045                                         k = endoffsets[i-1];
1046                                         index = endoffsets[j-1];
1047                                         printf("Value of slength = %d\n", slength);
1048                                         for(count = 0; count < slength; count++) {
1049                                                 if(arryfields[k] != arryfields[index]) {
1050                                                         break;
1051                                                 }
1052                                                 index++;
1053                                                 k++;
1054                                         }
1055                                 }
1056
1057                                 if(slength == count) {
1058                                         oid[sindex] = -1;
1059                                 }
1060                         }
1061                 }
1062         }
1063 }
1064
1065 void checkPreCache(prefetchqelem_t *node, int *numoffset, int counter, int loopcount, unsigned int objoid, int index, int iter, int oidnfound) {
1066         char *ptr, *tmp;
1067         int ntuples, i, k, flag;
1068         unsigned int * oid;
1069         short *endoffsets, *arryfields;
1070         objheader_t *header;
1071
1072         ptr = (char *) node;
1073         ntuples = *(GET_NTUPLES(ptr));
1074         oid = GET_PTR_OID(ptr);
1075         endoffsets = GET_PTR_EOFF(ptr, ntuples);
1076         arryfields = GET_PTR_ARRYFLD(ptr, ntuples);
1077
1078         if(oidnfound == 1) {
1079                 if((header = (objheader_t *) prehashSearch(objoid)) == NULL) {
1080                         return;
1081                 } else { //Found in Prefetch Cache
1082                         //TODO Decide if object is too old, if old remove from cache
1083                         tmp = (char *) header;
1084                         /* Check if any of the offset oid is available in the Prefetch cache */
1085                         for(i = counter; i < loopcount; i++) {
1086                                 objoid = *(tmp + sizeof(objheader_t) + arryfields[counter]);
1087                                 if((header = (objheader_t *)prehashSearch(objoid)) != NULL) {
1088                                         flag = 0;
1089                                 } else {
1090                                         flag = 1;
1091                                         break;
1092                                 }
1093                         }
1094                 }
1095         } else {
1096                 for(i = counter; i<loopcount; i++) {
1097                         if((header = (objheader_t *)prehashSearch(objoid)) != NULL) {
1098                                 tmp = (char *) header;
1099                                 objoid = *(tmp + sizeof(objheader_t) + arryfields[index]);
1100                                 flag = 0;
1101                                 index++;
1102                         } else {
1103                                 flag = 1;
1104                                 break;
1105                         }
1106                 }
1107         }
1108
1109         /* If oid not found locally or in prefetch cache then
1110          * assign the latest oid found as the new oid
1111          * and copy left over offsets into the arrayoffsetfieldarray*/
1112         oid[iter] = objoid;
1113         numoffset[iter] = numoffset[iter] - (i+1);
1114         for(k = 0; k < numoffset[iter] ; k++) {
1115                 arryfields[endoffsets[counter]+k] = arryfields[endoffsets[counter]+k+1];
1116         }
1117
1118         if(flag == 0) {
1119                 oid[iter] = -1;
1120                 numoffset[iter] = 0;
1121         }
1122 }
1123
1124 /* This function makes machine piles to be added into the machine pile queue for each prefetch call */
1125 prefetchpile_t *makePreGroups(prefetchqelem_t *node, int *numoffset) {
1126         char *ptr, *tmp;
1127         int ntuples, slength, i, machinenum;
1128         int maxoffset;
1129         unsigned int *oid;
1130         short *endoffsets, *arryfields, *offset;
1131         prefetchpile_t *head = NULL;
1132
1133         /* Check for the case x.y.z and a.b.c are same oids */
1134         ptr = (char *) node;
1135         ntuples = *(GET_NTUPLES(ptr));
1136         oid = GET_PTR_OID(ptr);
1137         endoffsets = GET_PTR_EOFF(ptr, ntuples);
1138         arryfields = GET_PTR_ARRYFLD(ptr, ntuples);
1139
1140         /* Check for redundant tuples by comparing oids of each tuple */
1141         for(i = 0; i < ntuples; i++) {
1142                 if(oid[i] == -1)
1143                         continue;
1144                 /* For each tuple make piles */
1145                 if ((machinenum = lhashSearch(oid[i])) == 0) {
1146                         printf("Error: No such Machine %s, %d\n", __FILE__, __LINE__);
1147                         return NULL;
1148                 }
1149                 /* Insert into machine pile */
1150                 offset = &arryfields[endoffsets[i-1]];
1151                 insertPile(machinenum, oid[i], numoffset[i], offset, head);
1152         }
1153
1154         return head;
1155 }
1156
1157
1158 /* This function checks if the oids within the prefetch tuples are available locally.
1159  * If yes then makes the tuple invalid. If no then rearranges oid and offset values in
1160  * the prefetchqelem_t node to represent a new prefetch tuple */
1161 prefetchpile_t *foundLocal(prefetchqelem_t *node) {
1162         int ntuples,i, j, k, oidnfound = 0, index, flag;
1163         unsigned int *oid;
1164         unsigned int  objoid;
1165         char *ptr, *tmp;
1166         objheader_t *objheader;
1167         short *endoffsets, *arryfields;
1168         prefetchpile_t *head = NULL;
1169
1170         ptr = (char *) node;
1171         ntuples = *(GET_NTUPLES(ptr));
1172         oid = GET_PTR_OID(ptr);
1173         endoffsets = GET_PTR_EOFF(ptr, ntuples);
1174         arryfields = GET_PTR_ARRYFLD(ptr, ntuples);
1175         /* Find offset length for each tuple */
1176         int numoffset[ntuples];//Number of offsets for each tuple
1177         numoffset[0] = endoffsets[0];
1178         for(i = 1; i<ntuples; i++) {
1179                 numoffset[i] = endoffsets[i] - endoffsets[i-1];
1180         }
1181         for(i = 0; i < ntuples; i++) {
1182                 if(oid[i] == -1)
1183                         continue;
1184                 /* If object found locally */
1185                 if((objheader = (objheader_t*) mhashSearch(oid[i])) != NULL) {
1186                         oidnfound = 0;
1187                         tmp = (char *) objheader;
1188                         /* Find the oid of its offset value */
1189                         if(i == 0)
1190                                 index = 0;
1191                         else
1192                                 index = endoffsets[i - 1];
1193                         for(j = 0 ; j < numoffset[i] ; j++) {
1194                                 objoid = *(tmp + sizeof(objheader_t) + arryfields[index]);
1195                                 /*If oid found locally then
1196                                  *assign the latest oid found as the new oid
1197                                  *and copy left over offsets into the arrayoffsetfieldarray*/
1198                                 oid[i] = objoid;
1199                                 numoffset[i] = numoffset[i] - (j+1);
1200                                 for(k = 0; k < numoffset[i]; k++)
1201                                         arryfields[endoffsets[j]+ k] = arryfields[endoffsets[j]+k+1];
1202                                 index++;
1203                                 /*New offset oid not found */
1204                                 if((objheader = (objheader_t*) mhashSearch(objoid)) == NULL) {
1205                                         flag = 1;
1206                                         checkPreCache(node, numoffset, j, numoffset[i], objoid, index, i, oidnfound);
1207                                         break;
1208                                 } else
1209                                         flag = 0;
1210                         }
1211
1212                         /*If all offset oids are found locally,make the prefetch tuple invalid */
1213                         if(flag == 0) {
1214                                 oid[i] = -1;
1215                                 numoffset[i] = 0;
1216                         }
1217                 } else {
1218                         oidnfound = 1;
1219                         /* Look in Prefetch cache */
1220                         checkPreCache(node, numoffset, 0, numoffset[i], oid[i], 0, i, oidnfound);
1221                 }
1222
1223         }
1224         /* Make machine groups */
1225         head = makePreGroups(node, numoffset);
1226         return head;
1227 }
1228
1229 /* This function is called by the thread calling transPrefetch */
1230 void *transPrefetch(void *t) {
1231         prefetchqelem_t *qnode;
1232         prefetchpile_t *pilehead = NULL;
1233
1234         while(1) {
1235                 /* lock mutex of primary prefetch queue */
1236                 pthread_mutex_lock(&pqueue.qlock);
1237                 /* while primary queue is empty, then wait */
1238                 while((pqueue.front == NULL) && (pqueue.rear == NULL)) {
1239                         pthread_cond_wait(&pqueue.qcond, &pqueue.qlock);
1240                 }
1241
1242                 /* dequeue node to create a machine piles and  finally unlock mutex */
1243                 if((qnode = pre_dequeue()) == NULL) {
1244                         printf("Error: No node returned %s, %d\n", __FILE__, __LINE__);
1245                         return NULL;
1246                 }
1247                 pthread_mutex_unlock(&pqueue.qlock);
1248                 /* Reduce redundant prefetch requests */
1249                 checkPrefetchTuples(qnode);
1250                 /* Check if the tuples are found locally, if yes then reduce them further*/
1251                 /* and group requests by remote machine ids by calling the makePreGroups() */
1252                 pilehead = foundLocal(qnode);
1253
1254                 /* Lock mutex of pool queue */
1255                 pthread_mutex_lock(&mcqueue.qlock);
1256                 /* Update the pool queue with the new remote machine piles generated per prefetch call */
1257                 mcpileenqueue(pilehead);
1258                 /* Broadcast signal on machine pile queue */
1259                 pthread_cond_broadcast(&mcqueue.qcond);
1260                 /* Unlock mutex of  machine pile queue */
1261                 pthread_mutex_unlock(&mcqueue.qlock);
1262                 /* Deallocate the prefetch queue pile node */
1263                 predealloc(qnode);
1264
1265         }
1266 }
1267
1268 /* Each thread in the  pool of threads calls this function to establish connection with
1269  * remote machines, send the prefetch requests and process the reponses from
1270  * the remote machines .
1271  * The thread is active throughout the period of runtime */
1272
1273 void *mcqProcess(void *threadid) {
1274         int tid;
1275         prefetchpile_t *mcpilenode;
1276
1277         tid = (int) threadid;
1278         while(1) {
1279                 /* Lock mutex of mc pile queue */
1280                 pthread_mutex_lock(&mcqueue.qlock);
1281                 /* When mc pile queue is empty, wait */
1282                 while((mcqueue.front == NULL) && (mcqueue.rear == NULL)) {
1283                         pthread_cond_wait(&mcqueue.qcond, &mcqueue.qlock);
1284                 }
1285                 /* Dequeue node to send remote machine connections*/
1286                 if((mcpilenode = mcpiledequeue()) == NULL) {
1287                         printf("Dequeue Error: No node returned %s %d\n", __FILE__, __LINE__);
1288                         return NULL;
1289                 }
1290                 /* Unlock mutex */
1291                 pthread_mutex_unlock(&mcqueue.qlock);
1292
1293                 /*Initiate connection to remote host and send request */
1294                 /* Process Request */
1295                 sendPrefetchReq(mcpilenode, tid);
1296
1297                 /* Deallocate the machine queue pile node */
1298                 mcdealloc(mcpilenode);
1299         }
1300 }
1301
1302 void sendPrefetchReq(prefetchpile_t *mcpilenode, int threadid) {
1303         int sd, i, offset, off, len, endpair, numoffsets, count = 0;
1304         struct sockaddr_in serv_addr;
1305         struct hostent *server;
1306         char machineip[16], control;
1307         objpile_t *tmp;
1308
1309
1310         /* Send Trans Prefetch Request */
1311         if ((sd = socket(AF_INET, SOCK_STREAM, 0)) < 0) {
1312                 perror("Error in socket for TRANS_REQUEST\n");
1313                 return;
1314         }
1315         bzero((char*) &serv_addr, sizeof(serv_addr));
1316         serv_addr.sin_family = AF_INET;
1317         serv_addr.sin_port = htons(LISTEN_PORT);
1318         //serv_addr.sin_addr.s_addr = inet_addr(MACHINE_IP);
1319         midtoIP(mcpilenode->mid ,machineip);
1320         machineip[15] = '\0';
1321         serv_addr.sin_addr.s_addr = inet_addr(machineip);
1322
1323         /* Open Connection */
1324         if (connect(sd, (struct sockaddr *) &serv_addr, sizeof(struct sockaddr)) < 0) {
1325                 perror("Error in connect for TRANS_REQUEST\n");
1326                 return;
1327         }
1328
1329         /* Send TRANS_PREFETCH control message */
1330         control = TRANS_PREFETCH;
1331         if(send(sd, &control, sizeof(char), MSG_NOSIGNAL) < sizeof(char)) {
1332                 perror("Error in sending prefetch control\n");
1333                 return;
1334         }
1335
1336         /* Send Oids and offsets in pairs */
1337         tmp = mcpilenode->objpiles;
1338         while(tmp != NULL) {
1339                 off = offset = 0;
1340                 count++;  /* Keeps track of the number of oid and offset tuples sent per remote machine */
1341                 len = sizeof(int) + sizeof(unsigned int) + ((tmp->numoffset) * sizeof(short));
1342                 char oidnoffset[len];
1343                 memcpy(oidnoffset, &len, sizeof(int));
1344                 off = sizeof(int);
1345                 memcpy(oidnoffset + off, &tmp->oid, sizeof(unsigned int));
1346                 off += sizeof(unsigned int);
1347                 for(i = 0; i < numoffsets; i++) {
1348                         offset = off +  (i * sizeof(short));
1349                         memcpy(oidnoffset + offset, tmp->offset, sizeof(short));
1350                 }
1351                 if (send(sd, &oidnoffset, sizeof(oidnoffset),MSG_NOSIGNAL) < sizeof(oidnoffset)) {
1352                         perror("Error sending fixed bytes for thread\n");
1353                         return;
1354                 }
1355                 tmp = tmp->next;
1356         }
1357
1358         /* Send a special char -1 to represent the end of sending oids + offset pair to remote machine */
1359         endpair = -1;
1360         if (send(sd, &endpair, sizeof(int), MSG_NOSIGNAL) < sizeof(int)) {
1361                 perror("Error sending fixed bytes for thread\n");
1362                 return;
1363         }
1364
1365         /* Get Response from the remote machine */
1366         getPrefetchResponse(count,sd);
1367         close(sd);
1368         return;
1369 }
1370
1371 void getPrefetchResponse(int count, int sd) {
1372         int i = 0, val, n, N, sum, index, objsize;
1373         unsigned int bufsize,oid;
1374         char buffer[RECEIVE_BUFFER_SIZE], control;
1375         char *ptr;
1376         void *modptr, *oldptr;
1377
1378         /* Read  prefetch response from the Remote machine */
1379         if((val = read(sd, &control, sizeof(char))) <= 0) {
1380                 perror("No control response for Prefetch request sent\n");
1381                 return;
1382         }
1383
1384         if(control == TRANS_PREFETCH_RESPONSE) {
1385                 /*For each oid and offset tuple sent as prefetch request to remote machine*/
1386                 while(i < count) {
1387                         /* Clear contents of buffer */
1388                         memset(buffer, 0, RECEIVE_BUFFER_SIZE);
1389                         sum = 0;
1390                         index = 0;
1391                         /* Read the size of buffer to be received */
1392                         if((N = read(sd, buffer, sizeof(unsigned int))) <= 0) {
1393                                 perror("Size of buffer not recv\n");
1394                                 return;
1395                         }
1396                         memcpy(&bufsize, buffer, sizeof(unsigned int));
1397                         ptr = buffer + sizeof(unsigned int);
1398                         /* Keep receiving the buffer containing oid info */
1399                         do {
1400                                 n = recv((int)sd, (void *)ptr+sum, bufsize-sum, 0);
1401                                 sum +=n;
1402                         } while(sum < bufsize && n != 0);
1403                         /* Decode the contents of the buffer */
1404                         index = sizeof(unsigned int);
1405                         while(index < (bufsize - sizeof(unsigned int))) {
1406                                 if(buffer[index] == OBJECT_FOUND) {
1407                                         /* Increment it to get the object */
1408                                         index += sizeof(char);
1409                                         memcpy(&oid, buffer + index, sizeof(unsigned int));
1410                                         index += sizeof(unsigned int);
1411                                         /* Lock the Prefetch Cache look up table*/
1412                                         pthread_mutex_lock(&pflookup.lock);
1413                                         /* For each object found add to Prefetch Cache */
1414                                         memcpy(&objsize, buffer + index, sizeof(int));
1415                                         if ((modptr = objstrAlloc(prefetchcache, objsize)) == NULL) {
1416                                                 printf("objstrAlloc error for copying into prefetch cache %s, %d\n", __FILE__, __LINE__);
1417                                                 return;
1418                                         }
1419                                         memcpy(modptr, buffer+index, objsize);
1420                                         index += sizeof(int);
1421                                         /* Insert the oid and its address into the prefetch hash lookup table */
1422                                         /* Do a version comparison if the oid exists */
1423                                         if((oldptr = prehashSearch(oid)) != NULL) {
1424                                                 /* If older version then update with new object ptr */
1425                                                 if(((objheader_t *)oldptr)->version < ((objheader_t *)modptr)->version) {
1426                                                         prehashRemove(oid);
1427                                                         prehashInsert(oid, modptr);
1428                                                 } else if(((objheader_t *)oldptr)->version == ((objheader_t *)modptr)->version) {
1429                                                         /* Add the new object ptr to hash table */
1430                                                         prehashInsert(oid, modptr);
1431                                                 } else { /* Do nothing */
1432                                                         ;
1433                                                 }
1434                                         } else {/*If doesn't no match found in hashtable, add the object ptr to hash table*/
1435                                                 prehashInsert(oid, modptr);
1436                                         }
1437                                         /* Broadcast signal on prefetch cache condition variable */
1438                                         pthread_cond_broadcast(&pflookup.cond);
1439                                         /* Unlock the Prefetch Cache look up table*/
1440                                         pthread_mutex_unlock(&pflookup.lock);
1441                                 } else if(buffer[index] == OBJECT_NOT_FOUND) {
1442                                         /* Increment it to get the object */
1443                                         /* TODO: For each object not found query DHT for new location and retrieve the object */
1444                                         index += sizeof(char);
1445                                         memcpy(&oid, buffer + index, sizeof(unsigned int));
1446                                         index += sizeof(unsigned int);
1447                                         /* Throw an error */
1448                                         printf("OBJECT NOT FOUND.... THIS SHOULD NOT HAPPEN...TERMINATE PROGRAM\n");
1449                                         exit(-1);
1450                                 } else
1451                                         printf("Error in decoding the index value %s, %d\n",__FILE__, __LINE__);
1452                         }
1453
1454                         i++;
1455                 }
1456         } else
1457                 printf("Error in receving response for prefetch request %s, %d\n",__FILE__, __LINE__);
1458         return;
1459 }
1460
1461 unsigned short getObjType(unsigned int oid)
1462 {
1463         objheader_t *objheader;
1464         unsigned short numoffsets = 0;
1465
1466         if ((objheader = (objheader_t *) mhashSearch(oid)) == NULL)
1467         {
1468                 if ((objheader = (objheader_t *) prehashSearch(oid)) == NULL)
1469                 {
1470                         prefetch(1, &oid, &numoffsets, NULL);
1471                         pthread_mutex_lock(&pflookup.lock);
1472                         while ((objheader = (objheader_t *) prehashSearch(oid)) == NULL)
1473                                 pthread_cond_wait(&pflookup.cond, &pflookup.lock);
1474                         pthread_mutex_unlock(&pflookup.lock);
1475                 }
1476         }
1477
1478         return TYPE(objheader);
1479 }
1480
1481 int startRemoteThread(unsigned int oid, unsigned int mid)
1482 {
1483         int sock;
1484         struct sockaddr_in remoteAddr;
1485         char msg[1 + sizeof(unsigned int)];
1486         int bytesSent;
1487         int status;
1488
1489         if ((sock = socket(AF_INET, SOCK_STREAM, 0)) < 0)
1490         {
1491                 perror("startRemoteThread():socket()");
1492                 return -1;
1493         }
1494
1495         bzero(&remoteAddr, sizeof(remoteAddr));
1496         remoteAddr.sin_family = AF_INET;
1497         remoteAddr.sin_port = htons(LISTEN_PORT);
1498         remoteAddr.sin_addr.s_addr = htonl(mid);
1499
1500         if (connect(sock, (struct sockaddr *)&remoteAddr, sizeof(remoteAddr)) < 0)
1501         {
1502                 printf("startRemoteThread():error %d connecting to %s:%d\n", errno,
1503                         inet_ntoa(remoteAddr.sin_addr), LISTEN_PORT);
1504                 status = -1;
1505         }
1506         else
1507         {
1508                 msg[0] = START_REMOTE_THREAD;
1509                 memcpy(&msg[1], &oid, sizeof(unsigned int));
1510
1511                 bytesSent = send(sock, msg, 1 + sizeof(unsigned int), 0);
1512                 if (bytesSent < 0)
1513                 {
1514                         perror("startRemoteThread():send()");
1515                         status = -1;
1516                 }
1517                 else if (bytesSent != 1 + sizeof(unsigned int))
1518                 {
1519                         printf("startRemoteThread(): error, sent %d bytes\n", bytesSent);
1520                         status = -1;
1521                 }
1522                 else
1523                 {
1524                         status = 0;
1525                 }
1526         }
1527
1528         close(sock);
1529         return status;
1530 }
1531