ch3u_rma_sync.c 133 KB
Newer Older
1 2 3 4 5 6 7 8 9
/* -*- Mode: C; c-basic-offset:4 ; -*- */
/*
 *  (C) 2001 by Argonne National Laboratory.
 *      See COPYRIGHT in top-level directory.
 */

#include "mpidimpl.h"
#include "mpidrma.h"

10 11 12 13 14 15 16
#ifdef USE_MPIU_INSTR
MPIU_INSTR_DURATION_DECL(winfence_clearlock);
MPIU_INSTR_DURATION_DECL(winfence_rs);
MPIU_INSTR_DURATION_DECL(winfence_issue);
MPIU_INSTR_DURATION_DECL(winfence_complete);
MPIU_INSTR_DURATION_DECL(winfence_wait);
MPIU_INSTR_DURATION_DECL(winfence_block);
17 18 19
MPIU_INSTR_COUNTER_DECL(winfence_reqs);
MPIU_INSTR_COUNTER_DECL(winunlock_reqs);
MPIU_INSTR_COUNTER_DECL(wincomplete_reqs);
20 21 22 23 24 25
MPIU_INSTR_DURATION_DECL(winpost_clearlock);
MPIU_INSTR_DURATION_DECL(winpost_sendsync);
MPIU_INSTR_DURATION_DECL(winstart_clearlock);
MPIU_INSTR_DURATION_DECL(wincomplete_issue);
MPIU_INSTR_DURATION_DECL(wincomplete_complete);
MPIU_INSTR_DURATION_DECL(wincomplete_recvsync);
26
MPIU_INSTR_DURATION_DECL(wincomplete_block);
27 28 29 30 31
MPIU_INSTR_DURATION_DECL(winwait_wait);
MPIU_INSTR_DURATION_DECL(winlock_getlocallock);
MPIU_INSTR_DURATION_DECL(winunlock_getlock);
MPIU_INSTR_DURATION_DECL(winunlock_issue);
MPIU_INSTR_DURATION_DECL(winunlock_complete);
32
MPIU_INSTR_DURATION_DECL(winunlock_block);
33 34 35 36
MPIU_INSTR_DURATION_DECL(lockqueue_alloc);
MPIU_INSTR_DURATION_DECL(rmapkt_acc);
MPIU_INSTR_DURATION_DECL(rmapkt_acc_predef);
MPIU_INSTR_DURATION_DECL(rmapkt_acc_immed);
37
MPIU_INSTR_DURATION_DECL(rmapkt_acc_immed_op);
38
MPIU_INSTR_DURATION_EXTERN_DECL(rmaqueue_alloc);
39
MPIU_INSTR_DURATION_EXTERN_DECL(rmaqueue_set);
40 41 42 43 44 45 46 47 48 49 50
void MPIDI_CH3_RMA_InitInstr(void);

void MPIDI_CH3_RMA_InitInstr(void)
{
    MPIU_INSTR_DURATION_INIT(lockqueue_alloc,0,"Allocate Lock Queue element");
    MPIU_INSTR_DURATION_INIT(winfence_clearlock,1,"WIN_FENCE:Clear prior lock");
    MPIU_INSTR_DURATION_INIT(winfence_rs,0,"WIN_FENCE:ReduceScatterBlock");
    MPIU_INSTR_DURATION_INIT(winfence_issue,2,"WIN_FENCE:Issue RMA ops");
    MPIU_INSTR_DURATION_INIT(winfence_complete,1,"WIN_FENCE:Complete RMA ops");
    MPIU_INSTR_DURATION_INIT(winfence_wait,1,"WIN_FENCE:Wait for ops from other processes");
    MPIU_INSTR_DURATION_INIT(winfence_block,0,"WIN_FENCE:Wait for any progress");
51 52 53
    MPIU_INSTR_COUNTER_INIT(winfence_reqs,"WIN_FENCE:Pending requests");
    MPIU_INSTR_COUNTER_INIT(winunlock_reqs,"WIN_UNLOCK:Pending requests");
    MPIU_INSTR_COUNTER_INIT(wincomplete_reqs,"WIN_COMPLETE:Pending requests");
54 55 56 57 58 59
    MPIU_INSTR_DURATION_INIT(winpost_clearlock,1,"WIN_POST:Clear prior lock");
    MPIU_INSTR_DURATION_INIT(winpost_sendsync,1,"WIN_POST:Senc sync messages");
    MPIU_INSTR_DURATION_INIT(winstart_clearlock,1,"WIN_START:Clear prior lock");
    MPIU_INSTR_DURATION_INIT(wincomplete_recvsync,1,"WIN_COMPLETE:Recv sync messages");
    MPIU_INSTR_DURATION_INIT(wincomplete_issue,2,"WIN_COMPLETE:Issue RMA ops");
    MPIU_INSTR_DURATION_INIT(wincomplete_complete,1,"WIN_COMPLETE:Complete RMA ops");
60
    MPIU_INSTR_DURATION_INIT(wincomplete_block,0,"WIN_COMPLETE:Wait for any progress");
61 62 63 64 65
    MPIU_INSTR_DURATION_INIT(winwait_wait,1,"WIN_WAIT:Wait for ops from other processes");
    MPIU_INSTR_DURATION_INIT(winlock_getlocallock,0,"WIN_LOCK:Get local lock");
    MPIU_INSTR_DURATION_INIT(winunlock_issue,2,"WIN_UNLOCK:Issue RMA ops");
    MPIU_INSTR_DURATION_INIT(winunlock_complete,1,"WIN_UNLOCK:Complete RMA ops");
    MPIU_INSTR_DURATION_INIT(winunlock_getlock,0,"WIN_UNLOCK:Acquire lock");
66
    MPIU_INSTR_DURATION_INIT(winunlock_block,0,"WIN_UNLOCK:Wait for any progress");
67 68 69
    MPIU_INSTR_DURATION_INIT(rmapkt_acc,0,"RMA:PKTHANDLER for Accumulate");
    MPIU_INSTR_DURATION_INIT(rmapkt_acc_predef,0,"RMA:PKTHANDLER for Accumulate: predef dtype");
    MPIU_INSTR_DURATION_INIT(rmapkt_acc_immed,0,"RMA:PKTHANDLER for Accum immed");
70
    MPIU_INSTR_DURATION_INIT(rmapkt_acc_immed_op,0,"RMA:PKTHANDLER for Accum immed operation");
71
}
72 73 74 75 76 77 78 79 80 81

/* These are used to use a common routine to complete lists of RMA 
   operations with a single routine, while collecting data that 
   distinguishes between different synchronization modes.  This is not
   thread-safe; the best choice for thread-safety is to eliminate this
   ability to discriminate between the different types of RMA synchronization.
*/
static MPIU_INSTR_Duration_count *list_complete;  /* outer */
static MPIU_INSTR_Duration_count *list_block;     /* Inner; while waiting */

82 83
#endif

84 85 86
/*
 * These routines provide a default implementation of the MPI RMA operations
 * in terms of the low-level, two-sided channel operations.  A channel
87 88
 * may override these functions, on a per-window basis, by overriding
 * the MPID functions in the RMAFns section of MPID_Win object.
89 90
 */

91 92
#define SYNC_POST_TAG 100

93 94 95 96 97 98 99 100 101 102
static int MPIDI_CH3I_Send_rma_msg(MPIDI_RMA_ops * rma_op, MPID_Win * win_ptr, 
				   MPI_Win source_win_handle, 
				   MPI_Win target_win_handle, 
				   MPIDI_RMA_dtype_info * dtype_info, 
				   void ** dataloop, MPID_Request ** request);
static int MPIDI_CH3I_Recv_rma_msg(MPIDI_RMA_ops * rma_op, MPID_Win * win_ptr, 
				   MPI_Win source_win_handle, 
				   MPI_Win target_win_handle, 
				   MPIDI_RMA_dtype_info * dtype_info, 
				   void ** dataloop, MPID_Request ** request); 
103 104 105 106 107 108 109 110
static int MPIDI_CH3I_Send_contig_acc_msg(MPIDI_RMA_ops *, MPID_Win *,
					  MPI_Win, MPI_Win, MPID_Request ** );
static int MPIDI_CH3I_Do_passive_target_rma(MPID_Win *, int *);
static int MPIDI_CH3I_Send_lock_put_or_acc(MPID_Win *);
static int MPIDI_CH3I_Send_lock_get(MPID_Win *);
static int MPIDI_CH3I_RMAListComplete(MPID_Win *);
static int MPIDI_CH3I_RMAListPartialComplete( MPID_Win *, MPIDI_RMA_ops *, 
					      int * );
111

112 113
static int create_datatype(const MPIDI_RMA_dtype_info *dtype_info,
                           const void *dataloop, MPI_Aint dataloop_sz,
114 115
                           const void *o_addr, int o_count,
			   MPI_Datatype o_datatype,
116
                           MPID_Datatype **combined_dtp);
117 118 119 120 121 122 123 124

#undef FUNCNAME
#define FUNCNAME MPIDI_Win_fence
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
int MPIDI_Win_fence(int assert, MPID_Win *win_ptr)
{
    int mpi_errno = MPI_SUCCESS;
125
    int comm_size;
126
    int *rma_target_proc, *nops_to_proc, i, total_op_count, *curr_ops_cnt;
127
    MPIDI_RMA_ops *curr_ptr;
128 129 130
    MPID_Comm *comm_ptr;
    MPI_Win source_win_handle, target_win_handle;
    MPID_Progress_state progress_state;
131
    int errflag = FALSE;
132
    MPIU_CHKLMEM_DECL(3);
133 134 135 136 137 138 139 140 141 142 143 144 145
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_FENCE);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_FENCE);

    /* In case this process was previously the target of passive target rma
     * operations, we need to take care of the following...
     * Since we allow MPI_Win_unlock to return without a done ack from
     * the target in the case of multiple rma ops and exclusive lock,
     * we need to check whether there is a lock on the window, and if
     * there is a lock, poke the progress engine until the operartions
     * have completed and the lock is released. */
    if (win_ptr->current_lock_type != MPID_LOCK_NONE)
    {
146
	MPIU_INSTR_DURATION_START(winfence_clearlock);
147 148 149 150 151 152
	MPID_Progress_start(&progress_state);
	while (win_ptr->current_lock_type != MPID_LOCK_NONE)
	{
	    /* poke the progress engine */
	    mpi_errno = MPID_Progress_wait(&progress_state);
	    /* --BEGIN ERROR HANDLING-- */
153
	    if (mpi_errno != MPI_SUCCESS) {
154
		MPID_Progress_end(&progress_state);
155
		MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**winnoprogress");
156 157
	    }
	    /* --END ERROR HANDLING-- */
158
	    MPIU_INSTR_DURATION_INCR(winfence_clearlock,0,1);
159 160
	}
	MPID_Progress_end(&progress_state);
161
	MPIU_INSTR_DURATION_END(winfence_clearlock);
162
    }
163 164 165

    /* Note that the NOPRECEDE and NOSUCCEED must be specified by all processes
       in the window's group if any specify it */
166 167 168 169 170 171
    if (assert & MPI_MODE_NOPRECEDE)
    {
	win_ptr->fence_cnt = (assert & MPI_MODE_NOSUCCEED) ? 0 : 1;
	goto fn_exit;
    }
    
172
    if (win_ptr->fence_cnt == 0)
173 174 175 176
    {
	/* win_ptr->fence_cnt == 0 means either this is the very first
	   call to fence or the preceding fence had the
	   MPI_MODE_NOSUCCEED assert. 
177 178 179 180 181

           If this fence has MPI_MODE_NOSUCCEED, do nothing and return.
	   Otherwise just increment the fence count and return. */

	if (!(assert & MPI_MODE_NOSUCCEED)) win_ptr->fence_cnt = 1;
182 183 184
    }
    else
    {
185
	MPIDI_RMA_ops **prevNextPtr, *tmpptr;
186 187
	int nRequest = 0;
	int nRequestNew = 0;
188
	MPIU_INSTR_DURATION_START(winfence_rs);
189
	/* This is the second or later fence. Do all the preceding RMA ops. */
190
	comm_ptr = win_ptr->comm_ptr;
191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209
	/* First inform every process whether it is a target of RMA
	   ops from this process */
	comm_size = comm_ptr->local_size;

	MPIU_CHKLMEM_MALLOC(rma_target_proc, int *, comm_size*sizeof(int),
			    mpi_errno, "rma_target_proc");
	for (i=0; i<comm_size; i++) rma_target_proc[i] = 0;
	
	/* keep track of no. of ops to each proc. Needed for knowing
	   whether or not to decrement the completion counter. The
	   completion counter is decremented only on the last
	   operation. */
	MPIU_CHKLMEM_MALLOC(nops_to_proc, int *, comm_size*sizeof(int),
			    mpi_errno, "nops_to_proc");
	for (i=0; i<comm_size; i++) nops_to_proc[i] = 0;

	/* set rma_target_proc[i] to 1 if rank i is a target of RMA
	   ops from this process */
	total_op_count = 0;
210
	curr_ptr = win_ptr->rma_ops_list_head;
211 212 213 214 215 216 217 218 219 220 221
	while (curr_ptr != NULL)
	{
	    total_op_count++;
	    rma_target_proc[curr_ptr->target_rank] = 1;
	    nops_to_proc[curr_ptr->target_rank]++;
	    curr_ptr = curr_ptr->next;
	}
	
	MPIU_CHKLMEM_MALLOC(curr_ops_cnt, int *, comm_size*sizeof(int),
			    mpi_errno, "curr_ops_cnt");
	for (i=0; i<comm_size; i++) curr_ops_cnt[i] = 0;
222 223
	/* do a reduce_scatter_block (with MPI_SUM) on rma_target_proc. 
	   As a result,
224 225 226 227 228 229
	   each process knows how many other processes will be doing
	   RMA ops on its window */  
            
	/* first initialize the completion counter. */
	win_ptr->my_counter = comm_size;
            
230
	mpi_errno = MPIR_Reduce_scatter_block_impl(MPI_IN_PLACE, rma_target_proc, 1,
231
                                                   MPI_INT, MPI_SUM, comm_ptr, &errflag);
232
	MPIU_INSTR_DURATION_END(winfence_rs);
233 234
	/* result is stored in rma_target_proc[0] */
	if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
235
        MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
236 237 238 239 240 241

	/* Set the completion counter */
	/* FIXME: MT: this needs to be done atomically because other
	   procs have the address and could decrement it. */
	win_ptr->my_counter = win_ptr->my_counter - comm_size + 
	    rma_target_proc[0];  
242 243 244 245

	MPIU_INSTR_DURATION_START(winfence_issue);
	MPIU_INSTR_DURATION_INCR(winfence_issue,0,total_op_count);
	MPIU_INSTR_DURATION_MAX(winfence_issue,1,total_op_count);
246
	MPIU_INSTR_COUNTER_RESET(winfence_reqs);
247
	i = 0;
248
	curr_ptr    = win_ptr->rma_ops_list_head;
249
	prevNextPtr = &win_ptr->rma_ops_list_head;
250 251 252 253 254 255 256 257 258 259 260 261 262
	while (curr_ptr != NULL)
	{
	    /* The completion counter at the target is decremented only on 
	       the last RMA operation. We indicate the last operation by 
	       passing the source_win_handle only on the last operation. 
	       Otherwise, we pass NULL */
	    if (curr_ops_cnt[curr_ptr->target_rank] ==
		nops_to_proc[curr_ptr->target_rank] - 1) 
		source_win_handle = win_ptr->handle;
	    else 
		source_win_handle = MPI_WIN_NULL;
	    
	    target_win_handle = win_ptr->all_win_handles[curr_ptr->target_rank];
263 264

	    curr_ptr->dataloop = 0;
265 266 267 268 269 270
	    switch (curr_ptr->type)
	    {
	    case (MPIDI_RMA_PUT):
	    case (MPIDI_RMA_ACCUMULATE):
		mpi_errno = MPIDI_CH3I_Send_rma_msg(curr_ptr, win_ptr,
					source_win_handle, target_win_handle, 
271 272
					&curr_ptr->dtype_info,
					&curr_ptr->dataloop, &curr_ptr->request);
273 274
		if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
		break;
275 276 277 278
	    case MPIDI_RMA_ACC_CONTIG:
		mpi_errno = MPIDI_CH3I_Send_contig_acc_msg(curr_ptr, win_ptr,
				   source_win_handle, target_win_handle, 
				   &curr_ptr->request );
279
		if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
280
		break;
281 282 283
	    case (MPIDI_RMA_GET):
		mpi_errno = MPIDI_CH3I_Recv_rma_msg(curr_ptr, win_ptr,
					source_win_handle, target_win_handle, 
284 285
					&curr_ptr->dtype_info, 
					&curr_ptr->dataloop, &curr_ptr->request);
286 287 288
		if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
		break;
	    default:
289
		MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**winInvalidOp");
290 291 292
	    }
	    i++;
	    curr_ops_cnt[curr_ptr->target_rank]++;
293 294 295 296 297 298 299 300 301 302 303 304
	    /* If the request is null, we can remove it immediately */
	    if (!curr_ptr->request) {
		if (curr_ptr->dataloop != NULL) {
		    MPIU_Free(curr_ptr->dataloop); /* allocated in send_rma_msg or 
						      recv_rma_msg */
		}
		tmpptr       = curr_ptr->next;
		*prevNextPtr = tmpptr;
		MPIU_Free( curr_ptr );
		curr_ptr     = tmpptr;
	    }
	    else  {
305 306
		nRequest++;
		MPIU_INSTR_COUNTER_INCR(winfence_reqs,1);
307
		prevNextPtr = &curr_ptr->next;
308
		curr_ptr    = curr_ptr->next;
309 310 311 312 313 314 315 316 317 318 319 320 321
		/* The test on the difference is to reduce the number
		   of times the partial complete routine is called. Without
		   this, significant overhead is added once the
		   number of requests exceeds the threshold, since the
		   number that are completed in a call may be small. */
		if (nRequest > MPIR_PARAM_RMA_NREQUEST_THRESHOLD && 
		    nRequest - nRequestNew > MPIR_PARAM_RMA_NREQUEST_NEW_THRESHOLD) {
		    int nDone = 0;
		    MPIDI_CH3I_RMAListPartialComplete( win_ptr, curr_ptr, &nDone );
		    /* if (nDone > 0) printf( "nDone = %d\n", nDone ); */
		    nRequest -= nDone;
		    nRequestNew = nRequest;
		}
322
	    }
323
	}
324 325 326 327 328 329 330 331 332 333 334 335 336
	MPIU_INSTR_DURATION_END(winfence_issue);

	/* We replaced a loop over an array of requests with a list of the
	   incomplete requests.  The reason to do 
	   that is for long lists - processing the entire list until
	   all are done introduces a potentially n^2 time.  In 
	   testing with test/mpi/perf/manyrma.c , the number of iterations
	   within the "while (total_op_count) was O(total_op_count).
	   
	   Another alternative is to create a more compressed list (storing
	   only the necessary information, reducing the number of cache lines
	   needed while looping through the requests.
	*/
337 338
	if (total_op_count)
	{ 
339 340 341
	    MPIU_INSTR_STMT(list_complete=MPIU_INSTR_GET_VAR(winfence_complete));
	    MPIU_INSTR_STMT(list_block=MPIU_INSTR_GET_VAR(winfence_block));
	    mpi_errno = MPIDI_CH3I_RMAListComplete(win_ptr);
342
	}
343

344
            
345 346
	win_ptr->rma_ops_list_head = NULL;
	win_ptr->rma_ops_list_tail = NULL;
347 348 349 350
	
	/* wait for all operations from other processes to finish */
	if (win_ptr->my_counter)
	{
351
	    MPIU_INSTR_DURATION_START(winfence_wait);
352 353 354 355 356
	    MPID_Progress_start(&progress_state);
	    while (win_ptr->my_counter)
	    {
		mpi_errno = MPID_Progress_wait(&progress_state);
		/* --BEGIN ERROR HANDLING-- */
357
		if (mpi_errno != MPI_SUCCESS) {
358
		    MPID_Progress_end(&progress_state);
359
		    MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**winnoprogress");
360 361
		}
		/* --END ERROR HANDLING-- */
362
		MPIU_INSTR_DURATION_INCR(winfence_wait,0,1);
363 364
	    }
	    MPID_Progress_end(&progress_state);
365
	    MPIU_INSTR_DURATION_END(winfence_wait);
366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383
	} 
	
	if (assert & MPI_MODE_NOSUCCEED)
	{
	    win_ptr->fence_cnt = 0;
	}
    }

 fn_exit:
    MPIU_CHKLMEM_FREEALL();
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_FENCE);
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
 fn_fail:
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}

384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427
/* create_datatype() creates a new struct datatype for the dtype_info
   and the dataloop of the target datatype together with the user data */
#undef FUNCNAME
#define FUNCNAME create_datatype
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
static int create_datatype(const MPIDI_RMA_dtype_info *dtype_info,
                           const void *dataloop, MPI_Aint dataloop_sz,
                           const void *o_addr, int o_count, MPI_Datatype o_datatype,
                           MPID_Datatype **combined_dtp)
{
    int mpi_errno = MPI_SUCCESS;
    /* datatype_set_contents wants an array 'ints' which is the
       blocklens array with count prepended to it.  So blocklens
       points to the 2nd element of ints to avoid having to copy
       blocklens into ints later. */
    int ints[4];
    int *blocklens = &ints[1];
    MPI_Aint displaces[3];
    MPI_Datatype datatypes[3];
    const int count = 3;
    MPI_Datatype combined_datatype;
    MPIDI_STATE_DECL(MPID_STATE_CREATE_DATATYPE);

    MPIDI_FUNC_ENTER(MPID_STATE_CREATE_DATATYPE);

    /* create datatype */
    displaces[0] = MPIU_PtrToAint(dtype_info);
    blocklens[0] = sizeof(*dtype_info);
    datatypes[0] = MPI_BYTE;
    
    displaces[1] = MPIU_PtrToAint(dataloop);
    blocklens[1] = dataloop_sz;
    datatypes[1] = MPI_BYTE;
    
    displaces[2] = MPIU_PtrToAint(o_addr);
    blocklens[2] = o_count;
    datatypes[2] = o_datatype;
    
    mpi_errno = MPID_Type_struct(count,
                                 blocklens,
                                 displaces,
                                 datatypes,
                                 &combined_datatype);
428 429
    if (mpi_errno) MPIU_ERR_POP(mpi_errno);
   
430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464
    ints[0] = count;

    MPID_Datatype_get_ptr(combined_datatype, *combined_dtp);    
    mpi_errno = MPID_Datatype_set_contents(*combined_dtp,
				           MPI_COMBINER_STRUCT,
				           count+1, /* ints (cnt,blklen) */
				           count, /* aints (disps) */
				           count, /* types */
				           ints,
				           displaces,
				           datatypes);
    if (mpi_errno) MPIU_ERR_POP(mpi_errno);

    /* Commit datatype */
    
    MPID_Dataloop_create(combined_datatype,
                         &(*combined_dtp)->dataloop,
                         &(*combined_dtp)->dataloop_size,
                         &(*combined_dtp)->dataloop_depth,
                         MPID_DATALOOP_HOMOGENEOUS);
    
    /* create heterogeneous dataloop */
    MPID_Dataloop_create(combined_datatype,
                         &(*combined_dtp)->hetero_dloop,
                         &(*combined_dtp)->hetero_dloop_size,
                         &(*combined_dtp)->hetero_dloop_depth,
                         MPID_DATALOOP_HETEROGENEOUS);
 
 fn_exit:
    MPIDI_FUNC_EXIT(MPID_STATE_CREATE_DATATYPE);
    return mpi_errno;
 fn_fail:
    goto fn_exit;
}

465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480

#undef FUNCNAME
#define FUNCNAME MPIDI_CH3I_Send_rma_msg
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
static int MPIDI_CH3I_Send_rma_msg(MPIDI_RMA_ops *rma_op, MPID_Win *win_ptr,
				   MPI_Win source_win_handle, 
				   MPI_Win target_win_handle, 
				   MPIDI_RMA_dtype_info *dtype_info, 
				   void **dataloop, MPID_Request **request) 
{
    MPIDI_CH3_Pkt_t upkt;
    MPIDI_CH3_Pkt_put_t *put_pkt = &upkt.put;
    MPIDI_CH3_Pkt_accum_t *accum_pkt = &upkt.accum;
    MPID_IOV iov[MPID_IOV_LIMIT];
    int mpi_errno=MPI_SUCCESS, predefined;
481
    int origin_dt_derived, target_dt_derived, origin_type_size, iovcnt; 
482 483 484 485 486 487 488 489 490
    MPIDI_VC_t * vc;
    MPID_Comm *comm_ptr;
    MPID_Datatype *target_dtp=NULL, *origin_dtp=NULL;
    MPIU_CHKPMEM_DECL(1);
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_SEND_RMA_MSG);
    MPIDI_STATE_DECL(MPID_STATE_MEMCPY);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_SEND_RMA_MSG);

491 492
    *request = NULL;

493 494 495 496 497 498 499 500 501 502 503
    if (rma_op->type == MPIDI_RMA_PUT)
    {
        MPIDI_Pkt_init(put_pkt, MPIDI_CH3_PKT_PUT);
        put_pkt->addr = (char *) win_ptr->base_addrs[rma_op->target_rank] +
            win_ptr->disp_units[rma_op->target_rank] * rma_op->target_disp;

        put_pkt->count = rma_op->target_count;
        put_pkt->datatype = rma_op->target_datatype;
        put_pkt->dataloop_size = 0;
        put_pkt->target_win_handle = target_win_handle;
        put_pkt->source_win_handle = source_win_handle;
504
        
505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523
        iov[0].MPID_IOV_BUF = (MPID_IOV_BUF_CAST) put_pkt;
        iov[0].MPID_IOV_LEN = sizeof(*put_pkt);
    }
    else
    {
        MPIDI_Pkt_init(accum_pkt, MPIDI_CH3_PKT_ACCUMULATE);
        accum_pkt->addr = (char *) win_ptr->base_addrs[rma_op->target_rank] +
            win_ptr->disp_units[rma_op->target_rank] * rma_op->target_disp;
        accum_pkt->count = rma_op->target_count;
        accum_pkt->datatype = rma_op->target_datatype;
        accum_pkt->dataloop_size = 0;
        accum_pkt->op = rma_op->op;
        accum_pkt->target_win_handle = target_win_handle;
        accum_pkt->source_win_handle = source_win_handle;

        iov[0].MPID_IOV_BUF = (MPID_IOV_BUF_CAST) accum_pkt;
        iov[0].MPID_IOV_LEN = sizeof(*accum_pkt);
    }

524 525 526 527
    /*    printf("send pkt: type %d, addr %d, count %d, base %d\n", rma_pkt->type,
          rma_pkt->addr, rma_pkt->count, win_ptr->base_addrs[rma_op->target_rank]);
          fflush(stdout);
    */
528

529
    comm_ptr = win_ptr->comm_ptr;
530
    MPIDI_Comm_get_vc_set_active(comm_ptr, rma_op->target_rank, &vc);
531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557

    MPIDI_CH3I_DATATYPE_IS_PREDEFINED(rma_op->origin_datatype, predefined);
    if (!predefined)
    {
        origin_dt_derived = 1;
        MPID_Datatype_get_ptr(rma_op->origin_datatype, origin_dtp);
    }
    else
    {
        origin_dt_derived = 0;
    }

    MPIDI_CH3I_DATATYPE_IS_PREDEFINED(rma_op->target_datatype, predefined);
    if (!predefined)
    {
        target_dt_derived = 1;
        MPID_Datatype_get_ptr(rma_op->target_datatype, target_dtp);
    }
    else
    {
        target_dt_derived = 0;
    }

    if (target_dt_derived)
    {
        /* derived datatype on target. fill derived datatype info */
        dtype_info->is_contig = target_dtp->is_contig;
558
        dtype_info->max_contig_blocks = target_dtp->max_contig_blocks;
559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575
        dtype_info->size = target_dtp->size;
        dtype_info->extent = target_dtp->extent;
        dtype_info->dataloop_size = target_dtp->dataloop_size;
        dtype_info->dataloop_depth = target_dtp->dataloop_depth;
        dtype_info->eltype = target_dtp->eltype;
        dtype_info->dataloop = target_dtp->dataloop;
        dtype_info->ub = target_dtp->ub;
        dtype_info->lb = target_dtp->lb;
        dtype_info->true_ub = target_dtp->true_ub;
        dtype_info->true_lb = target_dtp->true_lb;
        dtype_info->has_sticky_ub = target_dtp->has_sticky_ub;
        dtype_info->has_sticky_lb = target_dtp->has_sticky_lb;

	MPIU_CHKPMEM_MALLOC(*dataloop, void *, target_dtp->dataloop_size, 
			    mpi_errno, "dataloop");

	MPIDI_FUNC_ENTER(MPID_STATE_MEMCPY);
576
        MPIU_Memcpy(*dataloop, target_dtp->dataloop, target_dtp->dataloop_size);
577
	MPIDI_FUNC_EXIT(MPID_STATE_MEMCPY);
578 579
        /* the dataloop can have undefined padding sections, so we need to let
         * valgrind know that it is OK to pass this data to writev later on */
580
        MPL_VG_MAKE_MEM_DEFINED(*dataloop, target_dtp->dataloop_size);
581 582 583 584 585 586 587 588 589 590 591 592 593

        if (rma_op->type == MPIDI_RMA_PUT)
	{
            put_pkt->dataloop_size = target_dtp->dataloop_size;
	}
        else
	{
            accum_pkt->dataloop_size = target_dtp->dataloop_size;
	}
    }

    MPID_Datatype_get_size_macro(rma_op->origin_datatype, origin_type_size);

594
    if (!target_dt_derived)
595
    {
596 597 598 599
        /* basic datatype on target */
        if (!origin_dt_derived)
        {
            /* basic datatype on origin */
600 601 602
            iov[1].MPID_IOV_BUF = (MPID_IOV_BUF_CAST)rma_op->origin_addr;
            iov[1].MPID_IOV_LEN = rma_op->origin_count * origin_type_size;
            iovcnt = 2;
603
	    MPIU_THREAD_CS_ENTER(CH3COMM,vc);
604
            mpi_errno = MPIDI_CH3_iStartMsgv(vc, iov, iovcnt, request);
605
	    MPIU_THREAD_CS_EXIT(CH3COMM,vc);
606
            MPIU_ERR_CHKANDJUMP(mpi_errno, mpi_errno, MPI_ERR_OTHER, "**ch3|rmamsg");
607 608
        }
        else
609 610 611
        {
            /* derived datatype on origin */
            *request = MPID_Request_create();
612
            MPIU_ERR_CHKANDJUMP(*request == NULL,mpi_errno,MPI_ERR_OTHER,"**nomemreq");
613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631
            
            MPIU_Object_set_ref(*request, 2);
            (*request)->kind = MPID_REQUEST_SEND;
            
            (*request)->dev.segment_ptr = MPID_Segment_alloc( );
            MPIU_ERR_CHKANDJUMP1((*request)->dev.segment_ptr == NULL, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");

            (*request)->dev.datatype_ptr = origin_dtp;
            /* this will cause the datatype to be freed when the request
               is freed. */
            MPID_Segment_init(rma_op->origin_addr, rma_op->origin_count,
                              rma_op->origin_datatype,
                              (*request)->dev.segment_ptr, 0);
            (*request)->dev.segment_first = 0;
            (*request)->dev.segment_size = rma_op->origin_count * origin_type_size;

            (*request)->dev.OnFinal = 0;
            (*request)->dev.OnDataAvail = 0;

632
	    MPIU_THREAD_CS_ENTER(CH3COMM,vc);
633
            mpi_errno = vc->sendNoncontig_fn(vc, *request, iov[0].MPID_IOV_BUF, iov[0].MPID_IOV_LEN);
634
	    MPIU_THREAD_CS_EXIT(CH3COMM,vc);
635
            MPIU_ERR_CHKANDJUMP(mpi_errno, mpi_errno, MPI_ERR_OTHER, "**ch3|rmamsg");
636 637 638 639
        }
    }
    else
    {
640
        /* derived datatype on target */
641
        MPID_Datatype *combined_dtp = NULL;
642 643 644

        *request = MPID_Request_create();
        if (*request == NULL) {
645
	    MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomemreq");
646 647 648 649 650 651
        }

        MPIU_Object_set_ref(*request, 2);
        (*request)->kind = MPID_REQUEST_SEND;

	(*request)->dev.segment_ptr = MPID_Segment_alloc( );
652 653 654 655
        MPIU_ERR_CHKANDJUMP1((*request)->dev.segment_ptr == NULL, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");

        /* create a new datatype containing the dtype_info, dataloop, and origin data */

656 657
        mpi_errno = create_datatype(dtype_info, *dataloop, target_dtp->dataloop_size, rma_op->origin_addr,
                                    rma_op->origin_count, rma_op->origin_datatype, &combined_dtp);
658
        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
659 660 661 662 663

        (*request)->dev.datatype_ptr = combined_dtp;
        /* combined_datatype will be freed when request is freed */

        MPID_Segment_init(MPI_BOTTOM, 1, combined_dtp->handle,
664 665
                          (*request)->dev.segment_ptr, 0);
        (*request)->dev.segment_first = 0;
666 667 668 669 670
        (*request)->dev.segment_size = combined_dtp->size;

        (*request)->dev.OnFinal = 0;
        (*request)->dev.OnDataAvail = 0;

671
	MPIU_THREAD_CS_ENTER(CH3COMM,vc);
672
        mpi_errno = vc->sendNoncontig_fn(vc, *request, iov[0].MPID_IOV_BUF, iov[0].MPID_IOV_LEN);
673
	MPIU_THREAD_CS_EXIT(CH3COMM,vc);
674
        MPIU_ERR_CHKANDJUMP(mpi_errno, mpi_errno, MPI_ERR_OTHER, "**ch3|rmamsg");
675

676 677 678
        /* we're done with the datatypes */
        if (origin_dt_derived)
            MPID_Datatype_release(origin_dtp);
679
        MPID_Datatype_release(target_dtp);
680
    }    
681 682 683 684 685 686

 fn_exit:
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_SEND_RMA_MSG);
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
 fn_fail:
687 688 689
    if (*request)
    {
        MPIU_CHKPMEM_REAP();
690 691
        if ((*request)->dev.datatype_ptr)
            MPID_Datatype_release((*request)->dev.datatype_ptr);
692 693 694 695
        MPIU_Object_set_ref(*request, 0);
        MPIDI_CH3_Request_destroy(*request);
    }
    *request = NULL;
696 697 698 699
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}

700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730
/*
 * Use this for contiguous accumulate operations
 */
#undef FUNCNAME
#define FUNCNAME MPIDI_CH3I_Send_contig_acc_msg
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
static int MPIDI_CH3I_Send_contig_acc_msg(MPIDI_RMA_ops *rma_op, 
					  MPID_Win *win_ptr,
					  MPI_Win source_win_handle, 
					  MPI_Win target_win_handle, 
					  MPID_Request **request) 
{
    MPIDI_CH3_Pkt_t upkt;
    MPIDI_CH3_Pkt_accum_t *accum_pkt = &upkt.accum;
    MPID_IOV iov[MPID_IOV_LIMIT];
    int mpi_errno=MPI_SUCCESS;
    int origin_type_size, iovcnt; 
    MPIDI_VC_t * vc;
    MPID_Comm *comm_ptr;
    int len;
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_SEND_CONTIG_ACC_MSG);
    MPIDI_STATE_DECL(MPID_STATE_MEMCPY);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_SEND_CONTIG_ACC_MSG);

    *request = NULL;

    MPID_Datatype_get_size_macro(rma_op->origin_datatype, origin_type_size);
    /* FIXME: Make this size check efficient and match the packet type */
    len = rma_op->origin_count * origin_type_size;
731
    if (MPIR_PARAM_RMA_ACC_IMMED && len <= MPIDI_RMA_IMMED_INTS*sizeof(int)) {
732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754
	MPIDI_CH3_Pkt_accum_immed_t * accumi_pkt = &upkt.accum_immed;
	void *dest = accumi_pkt->data, *src = rma_op->origin_addr;
	
	MPIDI_Pkt_init(accumi_pkt, MPIDI_CH3_PKT_ACCUM_IMMED);
	accumi_pkt->addr = (char *) win_ptr->base_addrs[rma_op->target_rank] +
	    win_ptr->disp_units[rma_op->target_rank] * rma_op->target_disp;
	accumi_pkt->count = rma_op->target_count;
	accumi_pkt->datatype = rma_op->target_datatype;
	accumi_pkt->op = rma_op->op;
	accumi_pkt->target_win_handle = target_win_handle;
	accumi_pkt->source_win_handle = source_win_handle;
	
	switch (len) {
	case 1: *(uint8_t *)dest  = *(uint8_t *)src;  break;
	case 2: *(uint16_t *)dest = *(uint16_t *)src; break;
	case 4: *(uint32_t *)dest = *(uint32_t *)src; break;
	case 8: *(uint64_t *)dest = *(uint64_t *)src; break;
	default:
	    MPIU_Memcpy( accumi_pkt->data, (void *)rma_op->origin_addr, len );
	}
	comm_ptr = win_ptr->comm_ptr;
	MPIDI_Comm_get_vc_set_active(comm_ptr, rma_op->target_rank, &vc);
	MPIU_THREAD_CS_ENTER(CH3COMM,vc);
755
	mpi_errno = MPIDI_CH3_iStartMsg(vc, accumi_pkt, sizeof(*accumi_pkt), request);
756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792
	MPIU_THREAD_CS_EXIT(CH3COMM,vc);
	MPIU_ERR_CHKANDJUMP(mpi_errno, mpi_errno, MPI_ERR_OTHER, "**ch3|rmamsg");
	goto fn_exit;
    }

    MPIDI_Pkt_init(accum_pkt, MPIDI_CH3_PKT_ACCUMULATE);
    accum_pkt->addr = (char *) win_ptr->base_addrs[rma_op->target_rank] +
	win_ptr->disp_units[rma_op->target_rank] * rma_op->target_disp;
    accum_pkt->count = rma_op->target_count;
    accum_pkt->datatype = rma_op->target_datatype;
    accum_pkt->dataloop_size = 0;
    accum_pkt->op = rma_op->op;
    accum_pkt->target_win_handle = target_win_handle;
    accum_pkt->source_win_handle = source_win_handle;
    
    iov[0].MPID_IOV_BUF = (MPID_IOV_BUF_CAST) accum_pkt;
    iov[0].MPID_IOV_LEN = sizeof(*accum_pkt);

    /*    printf("send pkt: type %d, addr %d, count %d, base %d\n", rma_pkt->type,
          rma_pkt->addr, rma_pkt->count, win_ptr->base_addrs[rma_op->target_rank]);
          fflush(stdout);
    */

    comm_ptr = win_ptr->comm_ptr;
    MPIDI_Comm_get_vc_set_active(comm_ptr, rma_op->target_rank, &vc);


    /* basic datatype on target */
    /* basic datatype on origin */
    /* FIXME: This is still very heavyweight for a small message operation,
       such as a single word update */
    /* One possibility is to use iStartMsg with a buffer that is just large 
       enough, though note that nemesis has an optimization for this */
    iov[1].MPID_IOV_BUF = (MPID_IOV_BUF_CAST)rma_op->origin_addr;
    iov[1].MPID_IOV_LEN = rma_op->origin_count * origin_type_size;
    iovcnt = 2;
    MPIU_THREAD_CS_ENTER(CH3COMM,vc);
793
    mpi_errno = MPIDI_CH3_iStartMsgv(vc, iov, iovcnt, request);
794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811
    MPIU_THREAD_CS_EXIT(CH3COMM,vc);
    MPIU_ERR_CHKANDJUMP(mpi_errno, mpi_errno, MPI_ERR_OTHER, "**ch3|rmamsg");

 fn_exit:
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_SEND_CONTIG_ACC_MSG);
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
 fn_fail:
    if (*request)
    {
        MPIU_Object_set_ref(*request, 0);
        MPIDI_CH3_Request_destroy(*request);
    }
    *request = NULL;
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}

812 813 814 815 816 817 818 819 820 821 822 823 824 825


#undef FUNCNAME
#define FUNCNAME MPIDI_CH3I_Recv_rma_msg
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
static int MPIDI_CH3I_Recv_rma_msg(MPIDI_RMA_ops *rma_op, MPID_Win *win_ptr,
				   MPI_Win source_win_handle, 
				   MPI_Win target_win_handle, 
				   MPIDI_RMA_dtype_info *dtype_info, 
				   void **dataloop, MPID_Request **request) 
{
    MPIDI_CH3_Pkt_t upkt;
    MPIDI_CH3_Pkt_get_t *get_pkt = &upkt.get;
826
    int mpi_errno=MPI_SUCCESS, predefined;
827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843
    MPIDI_VC_t * vc;
    MPID_Comm *comm_ptr;
    MPID_Request *req = NULL;
    MPID_Datatype *dtp;
    MPID_IOV iov[MPID_IOV_LIMIT];
    MPIU_CHKPMEM_DECL(1);
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_RECV_RMA_MSG);
    MPIDI_STATE_DECL(MPID_STATE_MEMCPY);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_RECV_RMA_MSG);

    /* create a request, store the origin buf, cnt, datatype in it,
       and pass a handle to it in the get packet. When the get
       response comes from the target, it will contain the request
       handle. */  
    req = MPID_Request_create();
    if (req == NULL) {
844
	MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomemreq");
845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878
    }

    *request = req;

    MPIU_Object_set_ref(req, 2);

    req->dev.user_buf = rma_op->origin_addr;
    req->dev.user_count = rma_op->origin_count;
    req->dev.datatype = rma_op->origin_datatype;
    req->dev.target_win_handle = MPI_WIN_NULL;
    req->dev.source_win_handle = source_win_handle;
    MPIDI_CH3I_DATATYPE_IS_PREDEFINED(req->dev.datatype, predefined);
    if (!predefined)
    {
        MPID_Datatype_get_ptr(req->dev.datatype, dtp);
        req->dev.datatype_ptr = dtp;
        /* this will cause the datatype to be freed when the
           request is freed. */  
    }

    MPIDI_Pkt_init(get_pkt, MPIDI_CH3_PKT_GET);
    get_pkt->addr = (char *) win_ptr->base_addrs[rma_op->target_rank] +
        win_ptr->disp_units[rma_op->target_rank] * rma_op->target_disp;
    get_pkt->count = rma_op->target_count;
    get_pkt->datatype = rma_op->target_datatype;
    get_pkt->request_handle = req->handle;
    get_pkt->target_win_handle = target_win_handle;
    get_pkt->source_win_handle = source_win_handle;

/*    printf("send pkt: type %d, addr %d, count %d, base %d\n", rma_pkt->type,
           rma_pkt->addr, rma_pkt->count, win_ptr->base_addrs[rma_op->target_rank]);
    fflush(stdout);
*/
	    
879
    comm_ptr = win_ptr->comm_ptr;
880
    MPIDI_Comm_get_vc_set_active(comm_ptr, rma_op->target_rank, &vc);
881 882 883 884 885

    MPIDI_CH3I_DATATYPE_IS_PREDEFINED(rma_op->target_datatype, predefined);
    if (predefined)
    {
        /* basic datatype on target. simply send the get_pkt. */
886
	MPIU_THREAD_CS_ENTER(CH3COMM,vc);
887
        mpi_errno = MPIDI_CH3_iStartMsg(vc, get_pkt, sizeof(*get_pkt), &req);
888
	MPIU_THREAD_CS_EXIT(CH3COMM,vc);
889 890 891 892 893 894 895 896
    }
    else
    {
        /* derived datatype on target. fill derived datatype info and
           send it along with get_pkt. */

        MPID_Datatype_get_ptr(rma_op->target_datatype, dtp);
        dtype_info->is_contig = dtp->is_contig;
897
        dtype_info->max_contig_blocks = dtp->max_contig_blocks;
898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914
        dtype_info->size = dtp->size;
        dtype_info->extent = dtp->extent;
        dtype_info->dataloop_size = dtp->dataloop_size;
        dtype_info->dataloop_depth = dtp->dataloop_depth;
        dtype_info->eltype = dtp->eltype;
        dtype_info->dataloop = dtp->dataloop;
        dtype_info->ub = dtp->ub;
        dtype_info->lb = dtp->lb;
        dtype_info->true_ub = dtp->true_ub;
        dtype_info->true_lb = dtp->true_lb;
        dtype_info->has_sticky_ub = dtp->has_sticky_ub;
        dtype_info->has_sticky_lb = dtp->has_sticky_lb;

	MPIU_CHKPMEM_MALLOC(*dataloop, void *, dtp->dataloop_size, 
			    mpi_errno, "dataloop");

	MPIDI_FUNC_ENTER(MPID_STATE_MEMCPY);
915
        MPIU_Memcpy(*dataloop, dtp->dataloop, dtp->dataloop_size);
916 917
	MPIDI_FUNC_EXIT(MPID_STATE_MEMCPY);

918 919
        /* the dataloop can have undefined padding sections, so we need to let
         * valgrind know that it is OK to pass this data to writev later on */
920
        MPL_VG_MAKE_MEM_DEFINED(*dataloop, dtp->dataloop_size);
921

922 923 924 925 926 927 928 929
        get_pkt->dataloop_size = dtp->dataloop_size;

        iov[0].MPID_IOV_BUF = (MPID_IOV_BUF_CAST)get_pkt;
        iov[0].MPID_IOV_LEN = sizeof(*get_pkt);
        iov[1].MPID_IOV_BUF = (MPID_IOV_BUF_CAST)dtype_info;
        iov[1].MPID_IOV_LEN = sizeof(*dtype_info);
        iov[2].MPID_IOV_BUF = (MPID_IOV_BUF_CAST)*dataloop;
        iov[2].MPID_IOV_LEN = dtp->dataloop_size;
930

931
	MPIU_THREAD_CS_ENTER(CH3COMM,vc);
932
        mpi_errno = MPIDI_CH3_iStartMsgv(vc, iov, 3, &req);
933
	MPIU_THREAD_CS_EXIT(CH3COMM,vc);
934 935 936 937 938

        /* release the target datatype */
        MPID_Datatype_release(dtp);
    }

939 940
    if (mpi_errno != MPI_SUCCESS) {
	MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**ch3|rmamsg");
941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964
    }

    /* release the request returned by iStartMsg or iStartMsgv */
    if (req != NULL)
    {
        MPID_Request_release(req);
    }

 fn_exit:
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_RECV_RMA_MSG);
    return mpi_errno;

    /* --BEGIN ERROR HANDLING-- */
 fn_fail:
    MPIU_CHKPMEM_REAP();
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}


#undef FUNCNAME
#define FUNCNAME MPIDI_Win_post
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
965
int MPIDI_Win_post(MPID_Group *post_grp_ptr, int assert, MPID_Win *win_ptr)
966 967
{
    int mpi_errno=MPI_SUCCESS;
968
    MPID_Group *win_grp_ptr;
969
    int i, post_grp_size, *ranks_in_post_grp, *ranks_in_win_grp, dst, rank;
970
    MPID_Comm *win_comm_ptr;
971
    MPIU_CHKLMEM_DECL(4);
972 973 974 975
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_POST);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_POST);

976 977 978 979 980
    /* Even though we would want to reset the fence counter to keep
     * the user from using the previous fence to mark the beginning of
     * a fence epoch if he switched from fence to lock-unlock
     * synchronization, we cannot do this because fence_cnt must be
     * updated collectively */
981 982 983 984 985 986 987 988 989 990 991 992

    /* In case this process was previously the target of passive target rma
     * operations, we need to take care of the following...
     * Since we allow MPI_Win_unlock to return without a done ack from
     * the target in the case of multiple rma ops and exclusive lock,
     * we need to check whether there is a lock on the window, and if
     * there is a lock, poke the progress engine until the operations
     * have completed and the lock is therefore released. */
    if (win_ptr->current_lock_type != MPID_LOCK_NONE)
    {
	MPID_Progress_state progress_state;
	
993
	MPIU_INSTR_DURATION_START(winpost_clearlock);
994 995 996 997 998 999
	/* poke the progress engine */
	MPID_Progress_start(&progress_state);
	while (win_ptr->current_lock_type != MPID_LOCK_NONE)
	{
	    mpi_errno = MPID_Progress_wait(&progress_state);
	    /* --BEGIN ERROR HANDLING-- */
1000
	    if (mpi_errno != MPI_SUCCESS) {
1001
		MPID_Progress_end(&progress_state);
1002
		MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**winnoprogress");
1003 1004
	    }
	    /* --END ERROR HANDLING-- */
1005
	    MPIU_INSTR_DURATION_INCR(winpost_clearlock,0,1);
1006 1007
	}
	MPID_Progress_end(&progress_state);
1008
	MPIU_INSTR_DURATION_END(winpost_clearlock);
1009 1010
    }
        
1011
    post_grp_size = post_grp_ptr->size;
1012 1013 1014 1015 1016 1017
        
    /* initialize the completion counter */
    win_ptr->my_counter = post_grp_size;
        
    if ((assert & MPI_MODE_NOCHECK) == 0)
    {
1018 1019
        MPI_Request *req;
        MPI_Status *status;
1020 1021

	MPIU_INSTR_DURATION_START(winpost_sendsync);
1022
 
1023 1024 1025 1026
	/* NOCHECK not specified. We need to notify the source
	   processes that Post has been called. */  
	
	/* We need to translate the ranks of the processes in
1027
	   post_group to ranks in win_ptr->comm_ptr, so that we
1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041
	   can do communication */
            
	MPIU_CHKLMEM_MALLOC(ranks_in_post_grp, int *, 
			    post_grp_size * sizeof(int),
			    mpi_errno, "ranks_in_post_grp");
	MPIU_CHKLMEM_MALLOC(ranks_in_win_grp, int *, 
			    post_grp_size * sizeof(int),
			    mpi_errno, "ranks_in_win_grp");
        
	for (i=0; i<post_grp_size; i++)
	{
	    ranks_in_post_grp[i] = i;
	}
        
1042
	win_comm_ptr = win_ptr->comm_ptr;
1043 1044 1045

        mpi_errno = MPIR_Comm_group_impl(win_comm_ptr, &win_grp_ptr);
	if (mpi_errno) MPIU_ERR_POP(mpi_errno);
1046
	
1047

1048 1049 1050
        mpi_errno = MPIR_Group_translate_ranks_impl(post_grp_ptr, post_grp_size, ranks_in_post_grp,
                                                    win_grp_ptr, ranks_in_win_grp);
        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
1051
	
1052
        rank = win_ptr->myrank;
1053
	
1054 1055 1056
	MPIU_CHKLMEM_MALLOC(req, MPI_Request *, post_grp_size * sizeof(MPI_Request), mpi_errno, "req");
        MPIU_CHKLMEM_MALLOC(status, MPI_Status *, post_grp_size*sizeof(MPI_Status), mpi_errno, "status");

1057
	/* Send a 0-byte message to the source processes */
1058
	MPIU_INSTR_DURATION_INCR(winpost_sendsync,0,post_grp_size);
1059
	for (i = 0; i < post_grp_size; i++) {
1060
	    dst = ranks_in_win_grp[i];
1061 1062 1063 1064 1065

	    /* FIXME: Short messages like this shouldn't normally need a 
	       request - this should consider using the ch3 call to send
	       a short message and return a request only if the message is
	       not delivered. */
1066
	    if (dst != rank) {
1067 1068 1069 1070 1071 1072 1073 1074
                MPID_Request *req_ptr;
		mpi_errno = MPID_Isend(&i, 0, MPI_INT, dst, SYNC_POST_TAG, win_comm_ptr,
                                       MPID_CONTEXT_INTRA_PT2PT, &req_ptr);
		if (mpi_errno) MPIU_ERR_POP(mpi_errno);
                req[i] = req_ptr->handle;
	    } else {
                req[i] = MPI_REQUEST_NULL;
            }
1075
	}
1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088
        mpi_errno = MPIR_Waitall_impl(post_grp_size, req, status);
        if (mpi_errno && mpi_errno != MPI_ERR_IN_STATUS) MPIU_ERR_POP(mpi_errno);

        /* --BEGIN ERROR HANDLING-- */
        if (mpi_errno == MPI_ERR_IN_STATUS) {
            for (i = 0; i < post_grp_size; i++) {
                if (status[i].MPI_ERROR != MPI_SUCCESS) {
                    mpi_errno = status[i].MPI_ERROR;
                    MPIU_ERR_POP(mpi_errno);
                }
            }
        }
        /* --END ERROR HANDLING-- */
1089 1090 1091

        mpi_errno = MPIR_Group_free_impl(win_grp_ptr);
	if (mpi_errno) MPIU_ERR_POP(mpi_errno);
1092
	MPIU_INSTR_DURATION_END(winpost_sendsync);
1093
    }
1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117

 fn_exit:
    MPIU_CHKLMEM_FREEALL();
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_POST);
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
 fn_fail:
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}



#undef FUNCNAME
#define FUNCNAME MPIDI_Win_start
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
int MPIDI_Win_start(MPID_Group *group_ptr, int assert, MPID_Win *win_ptr)
{
    int mpi_errno=MPI_SUCCESS;
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_START);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_START);

1118 1119 1120 1121 1122
    /* Even though we would want to reset the fence counter to keep
     * the user from using the previous fence to mark the beginning of
     * a fence epoch if he switched from fence to lock-unlock
     * synchronization, we cannot do this because fence_cnt must be
     * updated collectively */
1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134

    /* In case this process was previously the target of passive target rma
     * operations, we need to take care of the following...
     * Since we allow MPI_Win_unlock to return without a done ack from
     * the target in the case of multiple rma ops and exclusive lock,
     * we need to check whether there is a lock on the window, and if
     * there is a lock, poke the progress engine until the operations
     * have completed and the lock is therefore released. */
    if (win_ptr->current_lock_type != MPID_LOCK_NONE)
    {
	MPID_Progress_state progress_state;
	
1135
	MPIU_INSTR_DURATION_START(winstart_clearlock);
1136 1137 1138 1139 1140 1141
	/* poke the progress engine */
	MPID_Progress_start(&progress_state);
	while (win_ptr->current_lock_type != MPID_LOCK_NONE)
	{
	    mpi_errno = MPID_Progress_wait(&progress_state);
	    /* --BEGIN ERROR HANDLING-- */
1142
	    if (mpi_errno != MPI_SUCCESS) {
1143
		MPID_Progress_end(&progress_state);
1144
		MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**winnoprogress");
1145 1146
	    }
	    /* --END ERROR HANDLING-- */
1147
	    MPIU_INSTR_DURATION_INCR(winstart_clearlock,0,1);
1148 1149
	}
	MPID_Progress_end(&progress_state);
1150
	MPIU_INSTR_DURATION_END(winstart_clearlock);
1151 1152 1153 1154 1155
    }
    
    win_ptr->start_group_ptr = group_ptr;
    MPIR_Group_add_ref( group_ptr );
    win_ptr->start_assert = assert;
1156 1157

 fn_fail:
1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_START);
    return mpi_errno;
}



#undef FUNCNAME
#define FUNCNAME MPIDI_Win_complete
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
int MPIDI_Win_complete(MPID_Win *win_ptr)
{
    int mpi_errno = MPI_SUCCESS;
    int comm_size, *nops_to_proc, src, new_total_op_count;
1172
    int i, j, dst, total_op_count, *curr_ops_cnt;
1173
    MPIDI_RMA_ops *curr_ptr, *tmpptr, **prevNextPtr;
1174 1175
    MPID_Comm *comm_ptr;
    MPI_Win source_win_handle, target_win_handle;
1176
    MPID_Group *win_grp_ptr;
1177
    int start_grp_size, *ranks_in_start_grp, *ranks_in_win_grp, rank;
1178 1179
    int nRequest = 0;
    int nRequestNew = 0;
1180
    MPIU_CHKLMEM_DECL(9);
1181 1182 1183 1184
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_COMPLETE);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_COMPLETE);

1185
    comm_ptr = win_ptr->comm_ptr;
1186 1187 1188
    comm_size = comm_ptr->local_size;
        
    /* Translate the ranks of the processes in
1189
       start_group to ranks in win_ptr->comm_ptr */
1190 1191
    
    start_grp_size = win_ptr->start_group_ptr->size;
1192 1193

    MPIU_INSTR_DURATION_START(wincomplete_recvsync);
1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204
    MPIU_CHKLMEM_MALLOC(ranks_in_start_grp, int *, start_grp_size*sizeof(int), 
			mpi_errno, "ranks_in_start_grp");
        
    MPIU_CHKLMEM_MALLOC(ranks_in_win_grp, int *, start_grp_size*sizeof(int), 
			mpi_errno, "ranks_in_win_grp");
        
    for (i=0; i<start_grp_size; i++)
    {
	ranks_in_start_grp[i] = i;
    }
        
1205 1206 1207
    mpi_errno = MPIR_Comm_group_impl(comm_ptr, &win_grp_ptr);
    if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }

1208 1209 1210 1211 1212
    mpi_errno = MPIR_Group_translate_ranks_impl(win_ptr->start_group_ptr, start_grp_size,
                                                ranks_in_start_grp,
                                                win_grp_ptr, ranks_in_win_grp);
    if (mpi_errno) MPIU_ERR_POP(mpi_errno);

1213
    rank = win_ptr->myrank;
1214

1215 1216 1217 1218 1219
    /* If MPI_MODE_NOCHECK was not specified, we need to check if
       Win_post was called on the target processes. Wait for a 0-byte sync
       message from each target process */
    if ((win_ptr->start_assert & MPI_MODE_NOCHECK) == 0)
    {
1220 1221
        MPI_Request *req;
        MPI_Status *status