margo.c 24.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11

/*
 * (C) 2015 The University of Chicago
 * 
 * See COPYRIGHT in top-level directory.
 */

#include <assert.h>
#include <unistd.h>
#include <errno.h>
#include <abt.h>
12
#include <abt-snoozer.h>
13
#include <time.h>
Philip Carns's avatar
Philip Carns committed
14
#include <math.h>
15 16

#include "margo.h"
17
#include "margo-timer.h"
Philip Carns's avatar
Philip Carns committed
18
#include "utlist.h"
19
#include "uthash.h"
20

21 22
#define MERCURY_PROGRESS_TIMEOUT_UB 100 /* 100 milliseconds */

23 24 25 26 27 28 29 30 31 32 33 34 35
struct mplex_key
{
    hg_id_t id;
    uint32_t mplex_id;
};

struct mplex_element
{
    struct mplex_key key;
    ABT_pool pool;
    UT_hash_handle hh;
};

Philip Carns's avatar
Philip Carns committed
36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51
struct diag_data
{
    double min;
    double max;
    double cumulative;
    int count;
};

#define __DIAG_UPDATE(__data, __time)\
do {\
    __data.count++; \
    __data.cumulative += (__time); \
    if((__time) > __data.max) __data.max = (__time); \
    if((__time) < __data.min) __data.min = (__time); \
} while(0)

52 53
struct margo_instance
{
54
    /* provided by caller */
55 56
    hg_context_t *hg_context;
    hg_class_t *hg_class;
57 58 59
    ABT_pool handler_pool;
    ABT_pool progress_pool;

60
    /* internal to margo for this particular instance */
61 62
    ABT_thread hg_progress_tid;
    int hg_progress_shutdown_flag;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
63
    ABT_xstream progress_xstream;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
64 65 66
    int owns_progress_pool;
    ABT_xstream *rpc_xstreams;
    int num_handler_pool_threads;
67 68 69

    /* control logic for callers waiting on margo to be finalized */
    int finalize_flag;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
70
    int refcount;
71 72 73
    ABT_mutex finalize_mutex;
    ABT_cond finalize_cond;

74 75
    /* hash table to track multiplexed rpcs registered with margo */
    struct mplex_element *mplex_table;
Philip Carns's avatar
Philip Carns committed
76 77 78 79 80 81 82 83 84 85 86 87

    /* optional diagnostics data tracking */
    /* NOTE: technically the following fields are subject to races if they
     * are updated from more than one thread at a time.  We will be careful
     * to only update the counters from the progress_fn,
     * which will serialize access.
     */
    int diag_enabled;
    struct diag_data diag_trigger_elapsed;
    struct diag_data diag_progress_elapsed_zero_timeout;
    struct diag_data diag_progress_elapsed_nonzero_timeout;
    struct diag_data diag_progress_timeout_value;
88 89
};

90 91 92 93 94 95
struct margo_cb_arg
{
    ABT_eventual *eventual;
    margo_instance_id mid;
};

96 97 98 99 100 101
struct margo_rpc_data
{
	margo_instance_id mid;
	void* user_data;
	void (*user_free_callback)(void *);
};
102

103
static void hg_progress_fn(void* foo);
104
static int margo_xstream_is_in_progress_pool(margo_instance_id mid);
105
static void margo_rpc_data_free(void* ptr);
106 107 108 109 110 111 112 113

struct handler_entry
{
    void* fn;
    hg_handle_t handle;
    struct handler_entry *next; 
};

Jonathan Jenkins's avatar
Jonathan Jenkins committed
114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161
margo_instance_id margo_init(int use_progress_thread, int rpc_thread_count,
    hg_context_t *hg_context)
{
    struct margo_instance *mid = MARGO_INSTANCE_NULL;
    ABT_xstream progress_xstream = ABT_XSTREAM_NULL;
    ABT_pool progress_pool = ABT_POOL_NULL;
    ABT_xstream *rpc_xstreams = NULL;
    ABT_xstream rpc_xstream = ABT_XSTREAM_NULL;
    ABT_pool rpc_pool = ABT_POOL_NULL;
    int ret;
    int i;

    if (use_progress_thread)
    {
        ret = ABT_snoozer_xstream_create(1, &progress_pool, &progress_xstream);
        if (ret != ABT_SUCCESS) goto err;
    }
    else
    {
        ret = ABT_xstream_self(&progress_xstream);
        if (ret != ABT_SUCCESS) goto err;
        ret = ABT_xstream_get_main_pools(progress_xstream, 1, &progress_pool);
        if (ret != ABT_SUCCESS) goto err;
    }

    if (rpc_thread_count > 0)
    {
        rpc_xstreams = malloc(rpc_thread_count * sizeof(*rpc_xstreams));
        if (rpc_xstreams == NULL) goto err;
        ret = ABT_snoozer_xstream_create(rpc_thread_count, &rpc_pool,
                rpc_xstreams);
        if (ret != ABT_SUCCESS) goto err;
    }
    else if (rpc_thread_count == 0)
    {
        ret = ABT_xstream_self(&rpc_xstream);
        if (ret != ABT_SUCCESS) goto err;
        ret = ABT_xstream_get_main_pools(rpc_xstream, 1, &rpc_pool);
        if (ret != ABT_SUCCESS) goto err;
    }
    else
    {
        rpc_pool = progress_pool;
    }

    mid = margo_init_pool(progress_pool, rpc_pool, hg_context);
    if (mid == MARGO_INSTANCE_NULL) goto err;

Jonathan Jenkins's avatar
Jonathan Jenkins committed
162 163 164
    mid->owns_progress_pool = use_progress_thread;
    mid->progress_xstream = progress_xstream;
    mid->num_handler_pool_threads = rpc_thread_count < 0 ? 0 : rpc_thread_count;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186
    mid->rpc_xstreams = rpc_xstreams;
    return mid;

err:
    if (use_progress_thread && progress_xstream != ABT_XSTREAM_NULL)
    {
        ABT_xstream_join(progress_xstream);
        ABT_xstream_free(&progress_xstream);
    }
    if (rpc_thread_count > 0 && rpc_xstreams != NULL)
    {
        for (i = 0; i < rpc_thread_count; i++)
        {
            ABT_xstream_join(rpc_xstreams[i]);
            ABT_xstream_free(&rpc_xstreams[i]);
        }
        free(rpc_xstreams);
    }
    return MARGO_INSTANCE_NULL;
}

margo_instance_id margo_init_pool(ABT_pool progress_pool, ABT_pool handler_pool,
Jonathan Jenkins's avatar
Jonathan Jenkins committed
187
    hg_context_t *hg_context)
188 189
{
    int ret;
190 191 192 193
    struct margo_instance *mid;

    mid = malloc(sizeof(*mid));
    if(!mid)
194
        return(MARGO_INSTANCE_NULL);
195
    memset(mid, 0, sizeof(*mid));
196

197 198 199
    ABT_mutex_create(&mid->finalize_mutex);
    ABT_cond_create(&mid->finalize_cond);

200 201
    mid->progress_pool = progress_pool;
    mid->handler_pool = handler_pool;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
202
    mid->hg_class = HG_Context_get_class(hg_context);
203
    mid->hg_context = hg_context;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
204
    mid->refcount = 1;
205

206 207 208 209 210 211 212 213
    ret = margo_timer_instance_init(mid);
    if(ret != 0)
    {
        fprintf(stderr, "Error: margo_timer_instance_init()\n");
        free(mid);
        return(MARGO_INSTANCE_NULL);
    }

214
    ret = ABT_thread_create(mid->progress_pool, hg_progress_fn, mid, 
215
        ABT_THREAD_ATTR_NULL, &mid->hg_progress_tid);
216 217 218
    if(ret != 0)
    {
        fprintf(stderr, "Error: ABT_thread_create()\n");
219
        free(mid);
220
        return(MARGO_INSTANCE_NULL);
221 222
    }

223
    return mid;
224 225
}

Jonathan Jenkins's avatar
Jonathan Jenkins committed
226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253
static void margo_cleanup(margo_instance_id mid)
{
    int i;

    margo_timer_instance_finalize(mid);

    ABT_mutex_free(&mid->finalize_mutex);
    ABT_cond_free(&mid->finalize_cond);

    if (mid->owns_progress_pool)
    {
        ABT_xstream_join(mid->progress_xstream);
        ABT_xstream_free(&mid->progress_xstream);
    }

    if (mid->num_handler_pool_threads > 0)
    {
        for (i = 0; i < mid->num_handler_pool_threads; i++)
        {
            ABT_xstream_join(mid->rpc_xstreams[i]);
            ABT_xstream_free(&mid->rpc_xstreams[i]);
        }
        free(mid->rpc_xstreams);
    }

    free(mid);
}

254
void margo_finalize(margo_instance_id mid)
255
{
256
    int i;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
257
    int do_cleanup;
258

259
    /* tell progress thread to wrap things up */
260
    mid->hg_progress_shutdown_flag = 1;
261 262

    /* wait for it to shutdown cleanly */
263 264
    ABT_thread_join(mid->hg_progress_tid);
    ABT_thread_free(&mid->hg_progress_tid);
265

266 267 268 269
    ABT_mutex_lock(mid->finalize_mutex);
    mid->finalize_flag = 1;
    ABT_cond_broadcast(mid->finalize_cond);

Jonathan Jenkins's avatar
Jonathan Jenkins committed
270 271
    mid->refcount--;
    do_cleanup = mid->refcount == 0;
272

Jonathan Jenkins's avatar
Jonathan Jenkins committed
273 274 275 276 277 278 279
    ABT_mutex_unlock(mid->finalize_mutex);

    /* if there was noone waiting on the finalize at the time of the finalize
     * broadcast, then we're safe to clean up. Otherwise, let the finalizer do
     * it */
    if (do_cleanup)
        margo_cleanup(mid);
280 281 282 283 284 285

    return;
}

void margo_wait_for_finalize(margo_instance_id mid)
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
286
    int do_cleanup;
287 288 289

    ABT_mutex_lock(mid->finalize_mutex);

Jonathan Jenkins's avatar
Jonathan Jenkins committed
290
        mid->refcount++;
291 292 293 294
            
        while(!mid->finalize_flag)
            ABT_cond_wait(mid->finalize_cond, mid->finalize_mutex);

Jonathan Jenkins's avatar
Jonathan Jenkins committed
295 296 297
        mid->refcount--;
        do_cleanup = mid->refcount == 0;

298
    ABT_mutex_unlock(mid->finalize_mutex);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
299 300 301 302

    if (do_cleanup)
        margo_cleanup(mid);

303 304 305 306
    return;
}

/* dedicated thread function to drive Mercury progress */
307
static void hg_progress_fn(void* foo)
308 309 310
{
    int ret;
    unsigned int actual_count;
311
    struct margo_instance *mid = (struct margo_instance *)foo;
312
    size_t size;
313 314
    unsigned int hg_progress_timeout = MERCURY_PROGRESS_TIMEOUT_UB;
    double next_timer_exp;
315
    int trigger_happened;
Philip Carns's avatar
Philip Carns committed
316 317
    double tm1, tm2;
    int diag_enabled = 0;
318

319
    while(!mid->hg_progress_shutdown_flag)
320
    {
321
        trigger_happened = 0;
322
        do {
Philip Carns's avatar
Philip Carns committed
323 324 325 326 327 328
            /* save value of instance diag variable, in case it is modified
             * while we are in loop 
             */
            diag_enabled = mid->diag_enabled;

            if(diag_enabled) tm1 = ABT_get_wtime();
329
            ret = HG_Trigger(mid->hg_context, 0, 1, &actual_count);
Philip Carns's avatar
Philip Carns committed
330 331 332 333 334 335
            if(diag_enabled)
            {
                tm2 = ABT_get_wtime();
                __DIAG_UPDATE(mid->diag_trigger_elapsed, (tm2-tm1));
            }

336 337
            if(ret == HG_SUCCESS && actual_count > 0)
                trigger_happened = 1;
338
        } while((ret == HG_SUCCESS) && actual_count && !mid->hg_progress_shutdown_flag);
339

340 341
        if(trigger_happened)
            ABT_thread_yield();
342

343
        ABT_pool_get_size(mid->progress_pool, &size);
344
        /* Are there any other threads executing in this pool that are *not*
345 346 347 348
         * blocked ?  If so then, we can't sleep here or else those threads 
         * will not get a chance to execute.
         * TODO: check is ABT_pool_get_size returns the number of ULT/tasks
         * that can be executed including this one, or not including this one.
349
         */
350
        if(size > 0)
351 352 353 354 355 356 357
        {
            /* TODO: this is being executed more than is necessary (i.e.
             * in cases where there are other legitimate ULTs eligible
             * for execution that are not blocking on any events, Margo
             * or otherwise). Maybe we need an abt scheduling tweak here
             * to make sure that this ULT is the lowest priority in that
             * scenario.
358
             */
359
            if(diag_enabled) tm1 = ABT_get_wtime();
360
            ret = HG_Progress(mid->hg_context, 0);
361 362 363 364 365 366
            if(diag_enabled)
            {
                tm2 = ABT_get_wtime();
                __DIAG_UPDATE(mid->diag_progress_elapsed_zero_timeout, (tm2-tm1));
                __DIAG_UPDATE(mid->diag_progress_timeout_value, 0);
            }
367 368 369 370 371 372 373 374 375
            if(ret == HG_SUCCESS)
            {
                /* Mercury completed something; loop around to trigger
                 * callbacks 
                 */
            }
            else if(ret == HG_TIMEOUT)
            {
                /* No completion; yield here to allow other ULTs to run */
Philip Carns's avatar
Philip Carns committed
376
                ABT_thread_yield();
377 378 379 380 381 382
            }
            else
            {
                /* TODO: error handling */
                fprintf(stderr, "WARNING: unexpected return code (%d) from HG_Progress()\n", ret);
            }
383 384 385
        }
        else
        {
386
            hg_progress_timeout = MERCURY_PROGRESS_TIMEOUT_UB;
387 388
            ret = margo_timer_get_next_expiration(mid, &next_timer_exp);
            if(ret == 0)
389
            {
390 391 392 393
                /* there is a queued timer, don't block long enough
                 * to keep this timer waiting
                 */
                if(next_timer_exp >= 0.0)
394
                {
395 396 397 398 399 400 401
                    next_timer_exp *= 1000; /* convert to milliseconds */
                    if(next_timer_exp < MERCURY_PROGRESS_TIMEOUT_UB)
                        hg_progress_timeout = (unsigned int)next_timer_exp;
                }
                else
                {
                    hg_progress_timeout = 0;
402
                }
403
            }
404
            if(diag_enabled) tm1 = ABT_get_wtime();
405
            ret = HG_Progress(mid->hg_context, hg_progress_timeout);
406 407 408 409 410 411 412 413 414 415
            if(diag_enabled)
            {
                tm2 = ABT_get_wtime();
                if(hg_progress_timeout == 0)
                    __DIAG_UPDATE(mid->diag_progress_elapsed_zero_timeout, (tm2-tm1));
                else
                    __DIAG_UPDATE(mid->diag_progress_elapsed_nonzero_timeout, (tm2-tm1));
                    
                __DIAG_UPDATE(mid->diag_progress_timeout_value, hg_progress_timeout);
            }
416 417 418 419 420
            if(ret != HG_SUCCESS && ret != HG_TIMEOUT)
            {
                /* TODO: error handling */
                fprintf(stderr, "WARNING: unexpected return code (%d) from HG_Progress()\n", ret);
            }
421
        }
422

423
        /* check for any expired timers */
424
        margo_check_timers(mid);
425 426
    }

427
    return;
428 429
}

430
ABT_pool* margo_get_handler_pool(margo_instance_id mid)
431
{
432
    return(&mid->handler_pool);
433 434
}

435 436 437 438 439 440 441 442 443 444
hg_context_t* margo_get_context(margo_instance_id mid)
{
    return(mid->hg_context);
}

hg_class_t* margo_get_class(margo_instance_id mid)
{
    return(mid->hg_class);
}

445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475
hg_return_t margo_register_data(
    margo_instance_id mid,
    hg_id_t id,
    void *data,
    void (*free_callback)(void *)) 
{
	struct margo_rpc_data* margo_data 
		= (struct margo_rpc_data*) HG_Registered_data(margo_get_class(mid), id);
	if(!margo_data) return HG_OTHER_ERROR;
	margo_data->user_data = data;
	margo_data->user_free_callback = free_callback;
	return HG_SUCCESS;
}

void* margo_registered_data(margo_instance_id mid, hg_id_t id)
{
	struct margo_rpc_data* data
		= (struct margo_rpc_data*) HG_Registered_data(margo_get_class(mid), id);
	if(!data) return NULL;
	else return data->user_data;
}

margo_instance_id margo_hg_handle_get_instance(hg_handle_t h)
{
	const struct hg_info* info = HG_Get_info(h);
	if(!info) return MARGO_INSTANCE_NULL;
	struct margo_rpc_data* data = 
		(struct margo_rpc_data*) HG_Registered_data(info->hg_class, info->id);
	if(!data) return MARGO_INSTANCE_NULL;
	return data->mid;
}
476

Jonathan Jenkins's avatar
Jonathan Jenkins committed
477
static hg_return_t margo_cb(const struct hg_cb_info *info)
478 479
{
    hg_return_t hret = info->ret;
480
    struct margo_cb_arg* arg = info->arg;
481 482

    /* propagate return code out through eventual */
483
    ABT_eventual_set(*(arg->eventual), &hret, sizeof(hret));
484 485 486 487
    
    return(HG_SUCCESS);
}

488 489 490
typedef struct
{
    hg_handle_t handle;
Shane Snyder's avatar
Shane Snyder committed
491
} margo_forward_timeout_cb_dat;
492 493 494

static void margo_forward_timeout_cb(void *arg)
{
Shane Snyder's avatar
Shane Snyder committed
495 496
    margo_forward_timeout_cb_dat *timeout_cb_dat =
        (margo_forward_timeout_cb_dat *)arg;
497 498

    /* cancel the Mercury op if the forward timed out */
499
    HG_Cancel(timeout_cb_dat->handle);
500 501 502
    return;
}

503 504 505 506 507 508
hg_return_t margo_forward_timed(
    margo_instance_id mid,
    hg_handle_t handle,
    void *in_struct,
    double timeout_ms)
{
Shane Snyder's avatar
Shane Snyder committed
509
    int ret;
510
    hg_return_t hret;
511 512
    ABT_eventual eventual;
    hg_return_t* waited_hret;
Shane Snyder's avatar
Shane Snyder committed
513 514
    margo_timer_t forward_timer;
    margo_forward_timeout_cb_dat timeout_cb_dat;
515
    struct margo_cb_arg arg;
516 517 518 519 520 521 522

    ret = ABT_eventual_create(sizeof(hret), &eventual);
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);        
    }

Shane Snyder's avatar
Shane Snyder committed
523 524
    /* set a timer object to expire when this forward times out */
    timeout_cb_dat.handle = handle;
525
    margo_timer_init(mid, &forward_timer, margo_forward_timeout_cb,
Shane Snyder's avatar
Shane Snyder committed
526
        &timeout_cb_dat, timeout_ms);
527

528 529
    arg.eventual = &eventual;
    arg.mid = mid;
530

531
    hret = HG_Forward(handle, margo_cb, &arg, in_struct);
532 533 534 535 536 537
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

538 539 540 541
    /* convert HG_CANCELED to HG_TIMEOUT to indicate op timed out */
    if(hret == HG_CANCELED)
        hret = HG_TIMEOUT;

542 543
    /* remove timer if it is still in place (i.e., not timed out) */
    if(hret != HG_TIMEOUT)
544
        margo_timer_destroy(mid, &forward_timer);
545 546 547 548

    ABT_eventual_free(&eventual);

    return(hret);
549 550 551
}


552
hg_return_t margo_forward(
553
    margo_instance_id mid,
554 555 556 557 558 559 560
    hg_handle_t handle,
    void *in_struct)
{
    hg_return_t hret = HG_TIMEOUT;
    ABT_eventual eventual;
    int ret;
    hg_return_t* waited_hret;
561
    struct margo_cb_arg arg;
562 563 564 565 566 567 568

    ret = ABT_eventual_create(sizeof(hret), &eventual);
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);        
    }

569 570
    arg.eventual = &eventual;
    arg.mid = mid;
571

572
    hret = HG_Forward(handle, margo_cb, &arg, in_struct);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

hg_return_t margo_respond(
    margo_instance_id mid,
    hg_handle_t handle,
    void *out_struct)
{
    hg_return_t hret = HG_TIMEOUT;
    ABT_eventual eventual;
    int ret;
    hg_return_t* waited_hret;
593
    struct margo_cb_arg arg;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
594 595 596 597 598 599 600

    ret = ABT_eventual_create(sizeof(hret), &eventual);
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);
    }

601 602
    arg.eventual = &eventual;
    arg.mid = mid;
603

604
    hret = HG_Respond(handle, margo_cb, &arg, out_struct);
605 606 607 608 609 610 611 612 613 614 615
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

616

617
static hg_return_t margo_bulk_transfer_cb(const struct hg_cb_info *info)
618
{
619
    hg_return_t hret = info->ret;
620
    struct margo_cb_arg* arg = info->arg;
621 622

    /* propagate return code out through eventual */
623
    ABT_eventual_set(*(arg->eventual), &hret, sizeof(hret));
624 625 626 627
    
    return(HG_SUCCESS);
}

Philip Carns's avatar
Philip Carns committed
628 629
struct lookup_cb_evt
{
630 631
    hg_return_t nret;
    hg_addr_t addr;
Philip Carns's avatar
Philip Carns committed
632 633
};

634
static hg_return_t margo_addr_lookup_cb(const struct hg_cb_info *info)
Philip Carns's avatar
Philip Carns committed
635 636
{
    struct lookup_cb_evt evt;
637 638
    evt.nret = info->ret;
    evt.addr = info->info.lookup.addr;
639
    struct margo_cb_arg* arg = info->arg;
Philip Carns's avatar
Philip Carns committed
640 641

    /* propagate return code out through eventual */
642 643
    ABT_eventual_set(*(arg->eventual), &evt, sizeof(evt));

644
    return(HG_SUCCESS);
Philip Carns's avatar
Philip Carns committed
645 646 647
}


648
hg_return_t margo_addr_lookup(
649
    margo_instance_id mid,
Philip Carns's avatar
Philip Carns committed
650
    const char   *name,
651
    hg_addr_t    *addr)
652
{
653
    hg_return_t nret;
Philip Carns's avatar
Philip Carns committed
654
    struct lookup_cb_evt *evt;
655 656
    ABT_eventual eventual;
    int ret;
657
    struct margo_cb_arg arg;
658

Philip Carns's avatar
Philip Carns committed
659
    ret = ABT_eventual_create(sizeof(*evt), &eventual);
660 661 662 663 664
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);        
    }

665 666
    arg.eventual = &eventual;
    arg.mid = mid;
667

Jonathan Jenkins's avatar
Jonathan Jenkins committed
668
    nret = HG_Addr_lookup(mid->hg_context, margo_addr_lookup_cb,
669
        &arg, name, HG_OP_ID_IGNORE);
670 671
    if(nret == 0)
    {
Philip Carns's avatar
Philip Carns committed
672 673 674
        ABT_eventual_wait(eventual, (void**)&evt);
        *addr = evt->addr;
        nret = evt->nret;
675 676 677 678 679 680 681
    }

    ABT_eventual_free(&eventual);

    return(nret);
}

682
hg_return_t margo_bulk_transfer(
683
    margo_instance_id mid,
684
    hg_bulk_op_t op,
685
    hg_addr_t origin_addr,
686 687 688 689 690 691 692 693 694 695
    hg_bulk_t origin_handle,
    size_t origin_offset,
    hg_bulk_t local_handle,
    size_t local_offset,
    size_t size)
{
    hg_return_t hret = HG_TIMEOUT;
    hg_return_t *waited_hret;
    ABT_eventual eventual;
    int ret;
696
    struct margo_cb_arg arg;
697 698 699 700 701 702 703

    ret = ABT_eventual_create(sizeof(hret), &eventual);
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);        
    }

704 705
    arg.eventual = &eventual;
    arg.mid = mid;
706

Jonathan Jenkins's avatar
Jonathan Jenkins committed
707
    hret = HG_Bulk_transfer(mid->hg_context, margo_bulk_transfer_cb,
708
        &arg, op, origin_addr, origin_handle, origin_offset, local_handle,
Jonathan Jenkins's avatar
Jonathan Jenkins committed
709
        local_offset, size, HG_OP_ID_IGNORE);
710 711 712 713 714 715 716 717 718 719 720
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

721 722
typedef struct
{
723
    margo_instance_id mid;
724 725
    ABT_mutex mutex;
    ABT_cond cond;
Shane Snyder's avatar
Shane Snyder committed
726
    char is_asleep;
727 728 729 730 731 732 733 734 735
} margo_thread_sleep_cb_dat;

static void margo_thread_sleep_cb(void *arg)
{
    margo_thread_sleep_cb_dat *sleep_cb_dat =
        (margo_thread_sleep_cb_dat *)arg;

    /* wake up the sleeping thread */
    ABT_mutex_lock(sleep_cb_dat->mutex);
736
    sleep_cb_dat->is_asleep = 0;
737 738 739 740 741 742 743
    ABT_cond_signal(sleep_cb_dat->cond);
    ABT_mutex_unlock(sleep_cb_dat->mutex);

    return;
}

void margo_thread_sleep(
744
    margo_instance_id mid,
745 746 747 748 749 750
    double timeout_ms)
{
    margo_timer_t sleep_timer;
    margo_thread_sleep_cb_dat sleep_cb_dat;

    /* set data needed for sleep callback */
751
    sleep_cb_dat.mid = mid;
752 753
    ABT_mutex_create(&(sleep_cb_dat.mutex));
    ABT_cond_create(&(sleep_cb_dat.cond));
754
    sleep_cb_dat.is_asleep = 1;
755 756

    /* initialize the sleep timer */
757
    margo_timer_init(mid, &sleep_timer, margo_thread_sleep_cb,
758 759 760 761
        &sleep_cb_dat, timeout_ms);

    /* yield thread for specified timeout */
    ABT_mutex_lock(sleep_cb_dat.mutex);
762 763
    while(sleep_cb_dat.is_asleep)
        ABT_cond_wait(sleep_cb_dat.cond, sleep_cb_dat.mutex);
764 765
    ABT_mutex_unlock(sleep_cb_dat.mutex);

766 767 768 769
    /* clean up */
    ABT_mutex_free(&sleep_cb_dat.mutex);
    ABT_cond_free(&sleep_cb_dat.cond);

770 771 772
    return;
}

773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789
/* returns 1 if current xstream is in the progress pool, 0 if not */
static int margo_xstream_is_in_progress_pool(margo_instance_id mid)
{
    int ret;
    ABT_xstream xstream;
    ABT_pool pool;

    ret = ABT_xstream_self(&xstream);
    assert(ret == ABT_SUCCESS);
    ret = ABT_xstream_get_main_pools(xstream, 1, &pool);
    assert(ret == ABT_SUCCESS);

    if(pool == mid->progress_pool)
        return(1);
    else
        return(0);
}
Philip Carns's avatar
Philip Carns committed
790

791 792 793 794 795 796 797 798 799
static void margo_rpc_data_free(void* ptr)
{
	struct margo_rpc_data* data = (struct margo_rpc_data*) ptr;
	if(data->user_data && data->user_free_callback) {
		data->user_free_callback(data->user_data);
	}
	free(ptr);
}

800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818
int margo_lookup_mplex(margo_instance_id mid, hg_id_t id, uint32_t mplex_id, ABT_pool *pool)
{
    struct mplex_key key;
    struct mplex_element *element;

    if(!mplex_id)
    {
        *pool = mid->handler_pool;
        return(0);
    }

    memset(&key, 0, sizeof(key));
    key.id = id;
    key.mplex_id = mplex_id;

    HASH_FIND(hh, mid->mplex_table, &key, sizeof(key), element);
    if(!element)
        return(-1);

Philip Carns's avatar
Philip Carns committed
819 820
    assert(element->key.id == id && element->key.mplex_id == mplex_id);

821 822 823 824 825
    *pool = element->pool;

    return(0);
}

826 827 828 829 830 831 832 833 834 835 836
int margo_register(margo_instance_id mid, hg_id_t id)
{
	/* register the margo data with the RPC */
	struct margo_rpc_data* margo_data = (struct margo_rpc_data*)malloc(sizeof(struct margo_rpc_data));
	margo_data->mid = mid;
	margo_data->user_data = NULL;
	margo_data->user_free_callback = NULL;
	hg_return_t ret = HG_Register_data(margo_get_class(mid), id, margo_data, margo_rpc_data_free);
	return ret;
}

Philip Carns's avatar
Philip Carns committed
837 838
int margo_register_mplex(margo_instance_id mid, hg_id_t id, uint32_t mplex_id, ABT_pool pool)
{
839 840 841
    struct mplex_key key;
    struct mplex_element *element;

842 843 844 845 846 847 848 849 850
	/* register the margo data with the RPC */
	struct margo_rpc_data* margo_data = (struct margo_rpc_data*)malloc(sizeof(struct margo_rpc_data));
	margo_data->mid = mid;
	margo_data->user_data = NULL;
	margo_data->user_free_callback = NULL;
	hg_return_t ret = HG_Register_data(margo_get_class(mid), id, margo_data, margo_rpc_data_free);
	if(ret != HG_SUCCESS) 
		return ret;

851 852 853
    /* nothing to do, we'll let the handler pool take this directly */
    if(mplex_id == MARGO_DEFAULT_MPLEX_ID)
        return(0);
854

855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871
    memset(&key, 0, sizeof(key));
    key.id = id;
    key.mplex_id = mplex_id;

    HASH_FIND(hh, mid->mplex_table, &key, sizeof(key), element);
    if(element)
        return(0);

    element = malloc(sizeof(*element));
    if(!element)
        return(-1);
    element->key = key;
    element->pool = pool;

    HASH_ADD(hh, mid->mplex_table, key, sizeof(key), element);

    return(0);
Philip Carns's avatar
Philip Carns committed
872
}
Philip Carns's avatar
Philip Carns committed
873 874 875 876 877 878 879


void margo_diag_start(margo_instance_id mid)
{
    mid->diag_enabled = 1;
}

880
static void print_diag_data(FILE *file, const char* name, const char *description, struct diag_data *data)
Philip Carns's avatar
Philip Carns committed
881
{
882 883 884 885 886 887 888 889
    double avg;

    fprintf(file, "# %s\n", description);
    if(data->count != 0)
        avg = data->cumulative/data->count;
    else
        avg = 0;
    fprintf(file, "%s\t%.9f\t%.9f\t%.9f\t%d\n", name, avg, data->min, data->max, data->count);
Philip Carns's avatar
Philip Carns committed
890 891 892
    return;
}

893
void margo_diag_dump(margo_instance_id mid, const char* file, int uniquify)
Philip Carns's avatar
Philip Carns committed
894 895 896
{
    FILE *outfile;
    time_t ltime;
897
    char revised_file_name[256] = {0};
Philip Carns's avatar
Philip Carns committed
898 899 900

    assert(mid->diag_enabled);

901 902 903 904 905 906 907 908 909 910 911 912 913 914 915
    if(uniquify)
    {
        char hostname[128] = {0};
        int pid;

        gethostname(hostname, 128);
        pid = getpid();

        sprintf(revised_file_name, "%s-%s-%d", file, hostname, pid);
    }
    else
    {
        sprintf(revised_file_name, "%s", file);
    }

Philip Carns's avatar
Philip Carns committed
916 917 918 919 920 921
    if(strcmp("-", file) == 0)
    {
        outfile = stdout;
    }
    else
    {
922
        outfile = fopen(revised_file_name, "a");
Philip Carns's avatar
Philip Carns committed
923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938
        if(!outfile)
        {
            perror("fopen");
            return;
        }
    }

    /* TODO: retrieve self addr and include in output */
    /* TODO: support pattern substitution in file name to create unique
     * output files per process
     */

    time(&ltime);
    fprintf(outfile, "# Margo diagnostics\n");
    fprintf(outfile, "# %s\n", ctime(&ltime));
    fprintf(outfile, "# <stat>\t<avg>\t<min>\t<max>\t<count>\n");
939 940 941 942 943 944 945 946 947 948 949 950
    print_diag_data(outfile, "trigger_elapsed", 
        "Time consumed by HG_Trigger()", 
        &mid->diag_trigger_elapsed);
    print_diag_data(outfile, "progress_elapsed_zero_timeout", 
        "Time consumed by HG_Progress() when called with timeout==0", 
        &mid->diag_progress_elapsed_zero_timeout);
    print_diag_data(outfile, "progress_elapsed_nonzero_timeout", 
        "Time consumed by HG_Progress() when called with timeout!=0", 
        &mid->diag_progress_elapsed_nonzero_timeout);
    print_diag_data(outfile, "progress_timeout_value", 
        "Timeout values passed to HG_Progress()", 
        &mid->diag_progress_timeout_value);
Philip Carns's avatar
Philip Carns committed
951 952 953 954 955 956

    if(outfile != stdout)
        fclose(outfile);
    
    return;
}