margo.c 30.1 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
#define DEFAULT_MERCURY_PROGRESS_TIMEOUT_UB 100 /* 100 milliseconds */
22

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
    int margo_init;
62 63
    ABT_thread hg_progress_tid;
    int hg_progress_shutdown_flag;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
64
    ABT_xstream progress_xstream;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
65 66 67
    int owns_progress_pool;
    ABT_xstream *rpc_xstreams;
    int num_handler_pool_threads;
68
    unsigned int hg_progress_timeout_ub;
69 70 71

    /* control logic for callers waiting on margo to be finalized */
    int finalize_flag;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
72
    int refcount;
73 74 75
    ABT_mutex finalize_mutex;
    ABT_cond finalize_cond;

76 77
    /* hash table to track multiplexed rpcs registered with margo */
    struct mplex_element *mplex_table;
Philip Carns's avatar
Philip Carns committed
78 79 80 81 82 83 84 85 86 87 88 89

    /* 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;
90 91
};

92 93 94 95 96 97
struct margo_cb_arg
{
    ABT_eventual *eventual;
    margo_instance_id mid;
};

98 99 100 101 102 103
struct margo_rpc_data
{
	margo_instance_id mid;
	void* user_data;
	void (*user_free_callback)(void *);
};
104

105
static void hg_progress_fn(void* foo);
106
static void margo_rpc_data_free(void* ptr);
107

108
margo_instance_id margo_init(const char *addr_str, int mode,
109
    int use_progress_thread, int rpc_thread_count)
110
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
111 112 113 114 115
    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;
116 117
    hg_class_t *hg_class = NULL;
    hg_context_t *hg_context = NULL;
118
    int listen_flag = (mode == MARGO_CLIENT_MODE) ? HG_FALSE : HG_TRUE;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
119
    int i;
120 121 122
    int ret;
    struct margo_instance *mid = MARGO_INSTANCE_NULL;

Shane Snyder's avatar
Shane Snyder committed
123
    if(mode != MARGO_CLIENT_MODE && mode != MARGO_SERVER_MODE) goto err;
124

125 126 127
    ret = ABT_init(0, NULL); /* XXX: argc/argv not currently used by ABT ... */
    if(ret != 0) goto err;

128
    /* set caller (self) ES to idle without polling */
129 130
    ret = ABT_snoozer_xstream_self_set();
    if(ret != 0) goto err;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
131 132 133 134 135 136 137 138 139 140 141 142 143 144

    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;
    }

145
    if (mode == MARGO_SERVER_MODE)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
146
    {
147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165
        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;
        }
Jonathan Jenkins's avatar
Jonathan Jenkins committed
166 167
    }

168 169 170 171 172 173
    hg_class = HG_Init(addr_str, listen_flag);
    if(!hg_class) goto err;

    hg_context = HG_Context_create(hg_class);
    if(!hg_context) goto err;

Jonathan Jenkins's avatar
Jonathan Jenkins committed
174 175 176
    mid = margo_init_pool(progress_pool, rpc_pool, hg_context);
    if (mid == MARGO_INSTANCE_NULL) goto err;

177
    mid->margo_init = 1;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
178 179 180
    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
181
    mid->rpc_xstreams = rpc_xstreams;
182

Jonathan Jenkins's avatar
Jonathan Jenkins committed
183 184 185
    return mid;

err:
186 187 188 189 190 191 192
    if(mid)
    {
        margo_timer_instance_finalize(mid);
        ABT_mutex_free(&mid->finalize_mutex);
        ABT_cond_free(&mid->finalize_cond);
        free(mid);
    }
Jonathan Jenkins's avatar
Jonathan Jenkins committed
193 194 195 196 197 198 199 200 201 202 203 204 205 206
    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);
    }
207 208 209 210 211
    if(hg_context)
        HG_Context_destroy(hg_context);
    if(hg_class)
        HG_Finalize(hg_class);
    ABT_finalize();
Jonathan Jenkins's avatar
Jonathan Jenkins committed
212 213 214 215
    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
216
    hg_context_t *hg_context)
217 218
{
    int ret;
219 220 221
    struct margo_instance *mid;

    mid = malloc(sizeof(*mid));
222
    if(!mid) goto err;
223
    memset(mid, 0, sizeof(*mid));
224

225 226 227
    ABT_mutex_create(&mid->finalize_mutex);
    ABT_cond_create(&mid->finalize_cond);

228 229
    mid->progress_pool = progress_pool;
    mid->handler_pool = handler_pool;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
230
    mid->hg_class = HG_Context_get_class(hg_context);
231
    mid->hg_context = hg_context;
232
    mid->hg_progress_timeout_ub = DEFAULT_MERCURY_PROGRESS_TIMEOUT_UB;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
233
    mid->refcount = 1;
234

235
    ret = margo_timer_instance_init(mid);
236
    if(ret != 0) goto err;
237

238
    ret = ABT_thread_create(mid->progress_pool, hg_progress_fn, mid, 
239
        ABT_THREAD_ATTR_NULL, &mid->hg_progress_tid);
240 241
    if(ret != 0) goto err;

242 243
    return mid;

244 245
err:
    if(mid)
246
    {
247 248 249
        margo_timer_instance_finalize(mid);
        ABT_mutex_free(&mid->finalize_mutex);
        ABT_cond_free(&mid->finalize_cond);
250
        free(mid);
251
    }
252
    return MARGO_INSTANCE_NULL;
253 254
}

Jonathan Jenkins's avatar
Jonathan Jenkins committed
255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279
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);
    }

280 281 282 283 284 285 286 287 288
    if (mid->margo_init)
    {
        if (mid->hg_context)
            HG_Context_destroy(mid->hg_context);
        if (mid->hg_class)
            HG_Finalize(mid->hg_class);
        ABT_finalize();
    }

Jonathan Jenkins's avatar
Jonathan Jenkins committed
289 290 291
    free(mid);
}

292
void margo_finalize(margo_instance_id mid)
293
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
294
    int do_cleanup;
295

296
    /* tell progress thread to wrap things up */
297
    mid->hg_progress_shutdown_flag = 1;
298 299

    /* wait for it to shutdown cleanly */
300 301
    ABT_thread_join(mid->hg_progress_tid);
    ABT_thread_free(&mid->hg_progress_tid);
302

303 304 305 306
    ABT_mutex_lock(mid->finalize_mutex);
    mid->finalize_flag = 1;
    ABT_cond_broadcast(mid->finalize_cond);

Jonathan Jenkins's avatar
Jonathan Jenkins committed
307 308
    mid->refcount--;
    do_cleanup = mid->refcount == 0;
309

Jonathan Jenkins's avatar
Jonathan Jenkins committed
310 311 312 313 314 315 316
    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);
317 318 319 320 321 322

    return;
}

void margo_wait_for_finalize(margo_instance_id mid)
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
323
    int do_cleanup;
324 325 326

    ABT_mutex_lock(mid->finalize_mutex);

Jonathan Jenkins's avatar
Jonathan Jenkins committed
327
        mid->refcount++;
328 329 330 331
            
        while(!mid->finalize_flag)
            ABT_cond_wait(mid->finalize_cond, mid->finalize_mutex);

Jonathan Jenkins's avatar
Jonathan Jenkins committed
332 333 334
        mid->refcount--;
        do_cleanup = mid->refcount == 0;

335
    ABT_mutex_unlock(mid->finalize_mutex);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
336 337 338 339

    if (do_cleanup)
        margo_cleanup(mid);

340 341 342
    return;
}

343 344
hg_id_t margo_register_name(margo_instance_id mid, const char *func_name,
    hg_proc_cb_t in_proc_cb, hg_proc_cb_t out_proc_cb, hg_rpc_cb_t rpc_cb)
345
{
346 347 348
	struct margo_rpc_data* margo_data;
    hg_return_t hret;
    hg_id_t id;
349

350 351 352
    id = HG_Register_name(mid->hg_class, func_name, in_proc_cb, out_proc_cb, rpc_cb);
    if(id <= 0)
        return(0);
353

354 355 356 357 358 359 360 361 362 363 364 365
	/* register the margo data with the RPC */
    margo_data = (struct margo_rpc_data*)HG_Registered_data(mid->hg_class, id);
    if(!margo_data)
    {
        margo_data = (struct margo_rpc_data*)malloc(sizeof(struct margo_rpc_data));
        if(!margo_data)
            return(0);
        margo_data->mid = mid;
        margo_data->user_data = NULL;
        margo_data->user_free_callback = NULL;
        hret = HG_Register_data(mid->hg_class, id, margo_data, margo_rpc_data_free);
        if(hret != HG_SUCCESS)
366
        {
367 368
            free(margo_data);
            return(0);
369
        }
370 371
    }

372
	return(id);
373 374
}

375 376 377
hg_id_t margo_register_name_mplex(margo_instance_id mid, const char *func_name,
    hg_proc_cb_t in_proc_cb, hg_proc_cb_t out_proc_cb, hg_rpc_cb_t rpc_cb,
    uint32_t mplex_id, ABT_pool pool)
378
{
379 380 381
    struct mplex_key key;
    struct mplex_element *element;
    hg_id_t id;
382

383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407
    id = margo_register_name(mid, func_name, in_proc_cb, out_proc_cb, rpc_cb);
    if(id <= 0)
        return(0);

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

    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(id);

    element = malloc(sizeof(*element));
    if(!element)
        return(0);
    element->key = key;
    element->pool = pool;

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

    return(id);
408 409
}

410 411
hg_return_t margo_registered_name(margo_instance_id mid, const char *func_name,
    hg_id_t *id, hg_bool_t *flag)
412
{
413
    return(HG_Registered_name(mid->hg_class, func_name, id, flag));
414 415
}

416 417 418 419 420 421 422
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 
423
		= (struct margo_rpc_data*) HG_Registered_data(mid->hg_class, id);
424 425 426 427 428 429 430 431 432 433 434 435 436 437
	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;
}

438 439 440 441
hg_return_t margo_registered_disable_response(
    margo_instance_id mid,
    hg_id_t id,
    int disable_flag)
442
{
443
    return(HG_Registered_disable_response(mid->hg_class, id, disable_flag));
444
}
445

446
struct lookup_cb_evt
447
{
448
    hg_return_t hret;
449 450 451 452 453 454
    hg_addr_t addr;
};

static hg_return_t margo_addr_lookup_cb(const struct hg_cb_info *info)
{
    struct lookup_cb_evt evt;
455
    evt.hret = info->ret;
456
    evt.addr = info->info.lookup.addr;
457
    struct margo_cb_arg* arg = info->arg;
458 459

    /* propagate return code out through eventual */
460 461
    ABT_eventual_set(*(arg->eventual), &evt, sizeof(evt));

462 463 464
    return(HG_SUCCESS);
}

465 466 467 468
hg_return_t margo_addr_lookup(
    margo_instance_id mid,
    const char   *name,
    hg_addr_t    *addr)
469
{
470
    hg_return_t hret;
471 472 473 474
    struct lookup_cb_evt *evt;
    ABT_eventual eventual;
    int ret;
    struct margo_cb_arg arg;
475

476 477 478 479 480 481 482 483
    ret = ABT_eventual_create(sizeof(*evt), &eventual);
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);        
    }

    arg.eventual = &eventual;
    arg.mid = mid;
484

485
    hret = HG_Addr_lookup(mid->hg_context, margo_addr_lookup_cb,
486
        &arg, name, HG_OP_ID_IGNORE);
487
    if(hret == HG_SUCCESS)
488 489 490
    {
        ABT_eventual_wait(eventual, (void**)&evt);
        *addr = evt->addr;
491
        hret = evt->hret;
492 493 494 495
    }

    ABT_eventual_free(&eventual);

496
    return(hret);
497 498 499 500 501
}

hg_return_t margo_addr_free(
    margo_instance_id mid,
    hg_addr_t addr)
502
{
503 504
    return(HG_Addr_free(mid->hg_class, addr));
}
505

506 507 508 509 510
hg_return_t margo_addr_self(
    margo_instance_id mid,
    hg_addr_t *addr)
{
    return(HG_Addr_self(mid->hg_class, addr));
511 512
}

513 514 515 516 517 518 519 520 521
hg_return_t margo_addr_dup(
    margo_instance_id mid,
    hg_addr_t addr,
    hg_addr_t *new_addr)
{
    return(HG_Addr_dup(mid->hg_class, addr, new_addr));
}

hg_return_t margo_addr_to_string(
522
    margo_instance_id mid,
523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557
    char *buf,
    hg_size_t *buf_size,
    hg_addr_t addr)
{
    return(HG_Addr_to_string(mid->hg_class, buf, buf_size, addr));
}

hg_return_t margo_create(margo_instance_id mid, hg_addr_t addr,
    hg_id_t id, hg_handle_t *handle)
{
    /* TODO: handle caching logic? */

    return(HG_Create(mid->hg_context, addr, id, handle));
}

hg_return_t margo_destroy(hg_handle_t handle)
{
    /* TODO handle caching logic? */

    return(HG_Destroy(handle));
}

hg_return_t margo_ref_incr(
    hg_handle_t handle)
{
    return(HG_Ref_incr(handle));
}

const struct hg_info *margo_get_info(
    hg_handle_t handle)
{
    return(HG_Get_info(handle));
}

hg_return_t margo_get_input(
558
    hg_handle_t handle,
559
    void *in_struct)
560
{
561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584
    return(HG_Get_input(handle, in_struct));
}

hg_return_t margo_free_input(
        hg_handle_t handle,
        void *in_struct)
{
    return(HG_Free_input(handle, in_struct));
}

hg_return_t margo_get_output(
    hg_handle_t handle,
    void *out_struct)
{
    return(HG_Get_output(handle, out_struct));
}

hg_return_t margo_free_output(
    hg_handle_t handle,
    void *out_struct)
{
    return(HG_Free_output(handle, out_struct));
}

Shane Snyder's avatar
Shane Snyder committed
585 586 587 588 589 590 591
hg_return_t margo_set_target_id(
    hg_handle_t handle,
    hg_uint8_t target_id)
{
    return(HG_Set_target_id(handle, target_id));
}

592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608
static hg_return_t margo_cb(const struct hg_cb_info *info)
{
    hg_return_t hret = info->ret;
    struct margo_cb_arg* arg = info->arg;

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

hg_return_t margo_forward(
    margo_instance_id mid,
    hg_handle_t handle,
    void *in_struct)
{
    hg_return_t hret = HG_TIMEOUT;
609
    ABT_eventual eventual;
610
    int ret;
611
    hg_return_t* waited_hret;
612
    struct margo_cb_arg arg;
613 614 615 616 617 618 619

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

620 621
    arg.eventual = &eventual;
    arg.mid = mid;
622

623
    hret = HG_Forward(handle, margo_cb, &arg, in_struct);
624
    if(hret == HG_SUCCESS)
625 626 627 628 629 630 631 632
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
633 634
}

635 636 637 638
typedef struct
{
    hg_handle_t handle;
} margo_forward_timeout_cb_dat;
639

640 641 642 643 644 645 646 647 648 649 650
static void margo_forward_timeout_cb(void *arg)
{
    margo_forward_timeout_cb_dat *timeout_cb_dat =
        (margo_forward_timeout_cb_dat *)arg;

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

hg_return_t margo_forward_timed(
651
    margo_instance_id mid,
652
    hg_handle_t handle,
653 654
    void *in_struct,
    double timeout_ms)
655 656
{
    int ret;
657 658
    hg_return_t hret;
    ABT_eventual eventual;
659
    hg_return_t* waited_hret;
660 661
    margo_timer_t forward_timer;
    margo_forward_timeout_cb_dat timeout_cb_dat;
662
    struct margo_cb_arg arg;
663 664 665 666 667 668 669

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

670 671 672 673 674
    /* set a timer object to expire when this forward times out */
    timeout_cb_dat.handle = handle;
    margo_timer_init(mid, &forward_timer, margo_forward_timeout_cb,
        &timeout_cb_dat, timeout_ms);

675 676
    arg.eventual = &eventual;
    arg.mid = mid;
677

678
    hret = HG_Forward(handle, margo_cb, &arg, in_struct);
679
    if(hret == HG_SUCCESS)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
680 681 682 683 684
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

685 686 687 688 689 690 691 692
    /* convert HG_CANCELED to HG_TIMEOUT to indicate op timed out */
    if(hret == HG_CANCELED)
        hret = HG_TIMEOUT;

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

Jonathan Jenkins's avatar
Jonathan Jenkins committed
693 694 695 696 697 698 699 700 701 702 703 704 705 706
    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;
707
    struct margo_cb_arg arg;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
708 709 710 711 712 713 714

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

715 716
    arg.eventual = &eventual;
    arg.mid = mid;
717

718
    hret = HG_Respond(handle, margo_cb, &arg, out_struct);
719
    if(hret == HG_SUCCESS)
720 721 722 723 724 725 726 727 728 729
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

730 731 732 733 734
hg_return_t margo_cancel(
    hg_handle_t handle)
{
    return(HG_Cancel(handle));
}
735

736 737 738 739 740 741 742
hg_return_t margo_bulk_create(
    margo_instance_id mid,
    hg_uint32_t count,
    void **buf_ptrs,
    const hg_size_t *buf_sizes,
    hg_uint8_t flags,
    hg_bulk_t *handle)
743
{
744
    /* XXX: handle caching logic? */
745

746 747 748
    return(HG_Bulk_create(mid->hg_class, count,
        buf_ptrs, buf_sizes, flags, handle));
}
749

750 751 752 753 754 755
hg_return_t margo_bulk_free(
    hg_bulk_t handle)
{
    /* XXX: handle caching logic? */

    return(HG_Bulk_free(handle));
756 757
}

758 759
hg_return_t margo_bulk_ref_incr(
    hg_bulk_t handle)
Philip Carns's avatar
Philip Carns committed
760
{
761 762
    return(HG_Bulk_ref_incr(handle));
}
Philip Carns's avatar
Philip Carns committed
763

764 765 766 767 768 769 770 771 772
hg_return_t margo_bulk_access(
    hg_bulk_t handle,
    hg_size_t offset,
    hg_size_t size,
    hg_uint8_t flags,
    hg_uint32_t max_count,
    void **buf_ptrs,
    hg_size_t *buf_sizes,
    hg_uint32_t *actual_count)
Philip Carns's avatar
Philip Carns committed
773
{
774 775 776
    return(HG_Bulk_access(handle, offset, size, flags, max_count,
        buf_ptrs, buf_sizes, actual_count));
}
Philip Carns's avatar
Philip Carns committed
777

778 779 780 781 782
hg_size_t margo_bulk_get_size(
    hg_bulk_t handle)
{
    return(HG_Bulk_get_size(handle));
}
783

784 785 786 787
hg_uint32_t margo_bulk_get_segment_count(
    hg_bulk_t handle)
{
    return(HG_Bulk_get_segment_count(handle));
Philip Carns's avatar
Philip Carns committed
788 789
}

790 791 792 793 794 795
hg_size_t margo_bulk_get_serialize_size(
    hg_bulk_t handle,
    hg_bool_t request_eager)
{
    return(HG_Bulk_get_serialize_size(handle, request_eager));
}
Philip Carns's avatar
Philip Carns committed
796

797 798 799 800 801
hg_return_t margo_bulk_serialize(
    void *buf,
    hg_size_t buf_size,
    hg_bool_t request_eager,
    hg_bulk_t handle)
802
{
803 804
    return(HG_Bulk_serialize(buf, buf_size, request_eager, handle));
}
805

806 807 808 809 810 811 812 813
hg_return_t margo_bulk_deserialize(
    margo_instance_id mid,
    hg_bulk_t *handle,
    const void *buf,
    hg_size_t buf_size)
{
    return(HG_Bulk_deserialize(mid->hg_class, handle, buf, buf_size));
}
814

815 816 817 818
static hg_return_t margo_bulk_transfer_cb(const struct hg_cb_info *info)
{
    hg_return_t hret = info->ret;
    struct margo_cb_arg* arg = info->arg;
819

820 821 822 823
    /* propagate return code out through eventual */
    ABT_eventual_set(*(arg->eventual), &hret, sizeof(hret));
    
    return(HG_SUCCESS);
824 825
}

826
hg_return_t margo_bulk_transfer(
827
    margo_instance_id mid,
828
    hg_bulk_op_t op,
829
    hg_addr_t origin_addr,
830 831 832 833
    hg_bulk_t origin_handle,
    size_t origin_offset,
    hg_bulk_t local_handle,
    size_t local_offset,
834 835
    size_t size,
    hg_op_id_t *op_id)
836 837 838 839 840
{
    hg_return_t hret = HG_TIMEOUT;
    hg_return_t *waited_hret;
    ABT_eventual eventual;
    int ret;
841
    struct margo_cb_arg arg;
842 843 844 845 846 847 848

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

849 850
    arg.eventual = &eventual;
    arg.mid = mid;
851

Jonathan Jenkins's avatar
Jonathan Jenkins committed
852
    hret = HG_Bulk_transfer(mid->hg_context, margo_bulk_transfer_cb,
853
        &arg, op, origin_addr, origin_handle, origin_offset, local_handle,
854
        local_offset, size, op_id);
855
    if(hret == HG_SUCCESS)
856 857 858 859 860 861 862 863 864 865
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

866 867 868 869 870 871
hg_return_t margo_bulk_cancel(
    hg_op_id_t op_id)
{
    return(HG_Bulk_cancel(op_id));
}

872 873 874 875
typedef struct
{
    ABT_mutex mutex;
    ABT_cond cond;
Shane Snyder's avatar
Shane Snyder committed
876
    char is_asleep;
877 878 879 880 881 882 883 884 885
} 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);
886
    sleep_cb_dat->is_asleep = 0;
887 888 889 890 891 892 893
    ABT_cond_signal(sleep_cb_dat->cond);
    ABT_mutex_unlock(sleep_cb_dat->mutex);

    return;
}

void margo_thread_sleep(
894
    margo_instance_id mid,
895 896 897 898 899 900 901 902
    double timeout_ms)
{
    margo_timer_t sleep_timer;
    margo_thread_sleep_cb_dat sleep_cb_dat;

    /* set data needed for sleep callback */
    ABT_mutex_create(&(sleep_cb_dat.mutex));
    ABT_cond_create(&(sleep_cb_dat.cond));
903
    sleep_cb_dat.is_asleep = 1;
904 905

    /* initialize the sleep timer */
906
    margo_timer_init(mid, &sleep_timer, margo_thread_sleep_cb,
907 908 909 910
        &sleep_cb_dat, timeout_ms);

    /* yield thread for specified timeout */
    ABT_mutex_lock(sleep_cb_dat.mutex);
911 912
    while(sleep_cb_dat.is_asleep)
        ABT_cond_wait(sleep_cb_dat.cond, sleep_cb_dat.mutex);
913 914
    ABT_mutex_unlock(sleep_cb_dat.mutex);

915 916 917 918
    /* clean up */
    ABT_mutex_free(&sleep_cb_dat.mutex);
    ABT_cond_free(&sleep_cb_dat.cond);

919 920 921
    return;
}

922
ABT_pool* margo_get_handler_pool(margo_instance_id mid)
923
{
924 925
    return(&mid->handler_pool);
}
926

927 928 929 930
hg_context_t* margo_get_context(margo_instance_id mid)
{
    return(mid->hg_context);
}
931

932 933 934
hg_class_t* margo_get_class(margo_instance_id mid)
{
    return(mid->hg_class);
935
}
Philip Carns's avatar
Philip Carns committed
936

937
margo_instance_id margo_hg_handle_get_instance(hg_handle_t h)
938
{
939 940
	const struct hg_info* info = HG_Get_info(h);
	if(!info) return MARGO_INSTANCE_NULL;
941 942 943 944 945
    return margo_hg_info_get_instance(info);
}

margo_instance_id margo_hg_info_get_instance(const struct hg_info *info)
{
946 947 948 949
	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;
950 951
}

952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970
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
971 972
    assert(element->key.id == id && element->key.mplex_id == mplex_id);

973 974 975 976 977
    *pool = element->pool;

    return(0);
}

978
static void margo_rpc_data_free(void* ptr)
Philip Carns's avatar
Philip Carns committed
979
{
980 981 982 983 984 985
	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);
}
986

987 988 989 990 991 992 993
/* dedicated thread function to drive Mercury progress */
static void hg_progress_fn(void* foo)
{
    int ret;
    unsigned int actual_count;
    struct margo_instance *mid = (struct margo_instance *)foo;
    size_t size;
994
    unsigned int hg_progress_timeout = mid->hg_progress_timeout_ub;
995 996
    double next_timer_exp;
    int trigger_happened;
997 998
    double tm1, tm2;
    int diag_enabled = 0;
999

1000 1001 1002 1003
    while(!mid->hg_progress_shutdown_flag)
    {
        trigger_happened = 0;
        do {
1004 1005 1006 1007
            /* save value of instance diag variable, in case it is modified
             * while we are in loop 
             */
            diag_enabled = mid->diag_enabled;
1008

1009
            if(diag_enabled) tm1 = ABT_get_wtime();
1010
            ret = HG_Trigger(mid->hg_context, 0, 1, &actual_count);
1011 1012 1013 1014 1015
            if(diag_enabled)
            {
                tm2 = ABT_get_wtime();
                __DIAG_UPDATE(mid->diag_trigger_elapsed, (tm2-tm1));
            }
1016

1017 1018 1019
            if(ret == HG_SUCCESS && actual_count > 0)
                trigger_happened = 1;
        } while((ret == HG_SUCCESS) && actual_count && !mid->hg_progress_shutdown_flag);
1020

1021 1022
        if(trigger_happened)
            ABT_thread_yield();
1023

1024
        ABT_pool_get_size(mid->progress_pool, &size);
1025
        /* Are there any other threads executing in this pool that are *not*
1026 1027 1028 1029
         * 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.
1030
         */
1031
        if(size > 0)
1032 1033 1034 1035 1036 1037 1038 1039
        {
            /* 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.
             */
1040
            if(diag_enabled) tm1 = ABT_get_wtime();
1041
            ret = HG_Progress(mid->hg_context, 0);
1042 1043 1044 1045 1046 1047
            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);
            }
1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066
            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 */
                ABT_thread_yield();
            }
            else
            {
                /* TODO: error handling */
                fprintf(stderr, "WARNING: unexpected return code (%d) from HG_Progress()\n", ret);
            }
        }
        else
        {
1067
            hg_progress_timeout = mid->hg_progress_timeout_ub;
1068 1069 1070 1071 1072 1073 1074 1075 1076
            ret = margo_timer_get_next_expiration(mid, &next_timer_exp);
            if(ret == 0)
            {
                /* there is a queued timer, don't block long enough
                 * to keep this timer waiting
                 */
                if(next_timer_exp >= 0.0)
                {
                    next_timer_exp *= 1000; /* convert to milliseconds */
1077
                    if(next_timer_exp < mid->hg_progress_timeout_ub)
1078 1079 1080 1081 1082 1083 1084
                        hg_progress_timeout = (unsigned int)next_timer_exp;
                }
                else
                {
                    hg_progress_timeout = 0;
                }
            }
1085
            if(diag_enabled) tm1 = ABT_get_wtime();
1086
            ret = HG_Progress(mid->hg_context, hg_progress_timeout);
1087 1088 1089 1090 1091 1092 1093 1094 1095 1096
            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);
            }
1097 1098 1099 1100 1101 1102
            if(ret != HG_SUCCESS && ret != HG_TIMEOUT)
            {
                /* TODO: error handling */
                fprintf(stderr, "WARNING: unexpected return code (%d) from HG_Progress()\n", ret);
            }
        }
1103

1104 1105 1106
        /* check for any expired timers */
        margo_check_timers(mid);
    }
1107

1108
    return;
Philip Carns's avatar
Philip Carns committed
1109
}
Philip Carns's avatar
Philip Carns committed
1110 1111 1112 1113 1114 1115 1116


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

1117
static void print_diag_data(FILE *file, const char* name, const char *description, struct diag_data *data)
Philip Carns's avatar
Philip Carns committed
1118
{
1119 1120 1121 1122 1123 1124 1125 1126
    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
1127 1128 1129
    return;
}

1130
void margo_diag_dump(margo_instance_id mid, const char* file, int uniquify)
Philip Carns's avatar
Philip Carns committed
1131 1132 1133
{
    FILE *outfile;
    time_t ltime;
1134
    char revised_file_name[256] = {0};
Philip Carns's avatar
Philip Carns committed
1135 1136 1137

    assert(mid->diag_enabled);

1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152
    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
1153 1154 1155 1156 1157 1158
    if(strcmp("-", file) == 0)
    {
        outfile = stdout;
    }
    else
    {
1159
        outfile = fopen(revised_file_name, "a");
Philip Carns's avatar
Philip Carns committed
1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175
        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");
1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187
    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
1188 1189 1190 1191 1192 1193

    if(outfile != stdout)
        fclose(outfile);
    
    return;
}
1194

1195
void margo_set_param(margo_instance_id mid, int option, const void *param)
1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206
{
    switch(option)
    {
        case MARGO_INFO_PROGRESS_TIMEOUT_UB:
            mid->hg_progress_timeout_ub = (*((const unsigned int*)param));
            break;
    }

    return;
}

1207
void margo_get_param(margo_instance_id mid, int option, void *param)
1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218
{

    switch(option)
    {
        case MARGO_INFO_PROGRESS_TIMEOUT_UB:
            (*((unsigned int*)param)) = mid->hg_progress_timeout_ub;
            break;
    }

    return;
}