margo.c 27.3 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;
};

36 37
struct margo_instance
{
38
    /* provided by caller */
39 40
    hg_context_t *hg_context;
    hg_class_t *hg_class;
41 42 43
    ABT_pool handler_pool;
    ABT_pool progress_pool;

44
    /* internal to margo for this particular instance */
45
    int margo_init;
46 47
    ABT_thread hg_progress_tid;
    int hg_progress_shutdown_flag;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
48
    ABT_xstream progress_xstream;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
49 50 51
    int owns_progress_pool;
    ABT_xstream *rpc_xstreams;
    int num_handler_pool_threads;
52 53 54

    /* control logic for callers waiting on margo to be finalized */
    int finalize_flag;
55
    int waiters_in_progress_pool;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
56
    int refcount;
57 58 59
    ABT_mutex finalize_mutex;
    ABT_cond finalize_cond;

60 61
    /* hash table to track multiplexed rpcs registered with margo */
    struct mplex_element *mplex_table;
62 63
};

64 65 66 67 68 69 70
struct margo_cb_arg
{
    ABT_eventual *eventual;
    margo_instance_id mid;
    char in_pool;
};

71 72 73 74 75 76
struct margo_rpc_data
{
	margo_instance_id mid;
	void* user_data;
	void (*user_free_callback)(void *);
};
77

78
static void hg_progress_fn(void* foo);
79
static int margo_xstream_is_in_progress_pool(margo_instance_id mid);
80
static void margo_rpc_data_free(void* ptr);
81

82 83 84
/* XXX: maybe instead of listen_flag, we can specify either CLIENT or SERVER mode? */
margo_instance_id margo_init(const char *addr_str, int listen_flag,
    int use_progress_thread, int rpc_thread_count)
85
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
86 87 88 89 90
    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;
91 92
    hg_class_t *hg_class = NULL;
    hg_context_t *hg_context = NULL;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
93
    int i;
94 95 96 97 98 99 100 101 102 103
    int ret;
    struct margo_instance *mid = MARGO_INSTANCE_NULL;

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

    /* set primary ES to idle without polling */
    /* XXX: is this right? always set snoozer scheduler on the calling xstream? */
    ret = ABT_snoozer_xstream_self_set();
    if(ret != 0) goto err;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
104 105 106 107 108 109 110 111 112 113 114 115 116 117

    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;
    }

118
    if (listen_flag)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
119
    {
120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138
        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
139 140
    }

141 142 143 144 145 146
    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
147 148 149
    mid = margo_init_pool(progress_pool, rpc_pool, hg_context);
    if (mid == MARGO_INSTANCE_NULL) goto err;

150
    mid->margo_init = 1;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
151 152 153
    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
154 155 156 157
    mid->rpc_xstreams = rpc_xstreams;
    return mid;

err:
158 159 160 161 162 163 164
    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
165 166 167 168 169 170 171 172 173 174 175 176 177 178
    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);
    }
179 180 181 182 183
    if(hg_context)
        HG_Context_destroy(hg_context);
    if(hg_class)
        HG_Finalize(hg_class);
    ABT_finalize();
Jonathan Jenkins's avatar
Jonathan Jenkins committed
184 185 186 187
    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
188
    hg_context_t *hg_context)
189 190
{
    int ret;
191 192 193
    struct margo_instance *mid;

    mid = malloc(sizeof(*mid));
194
    if(!mid) goto err;
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
    ret = margo_timer_instance_init(mid);
207
    if(ret != 0) goto err;
208

209
    ret = ABT_thread_create(mid->progress_pool, hg_progress_fn, mid, 
210
        ABT_THREAD_ATTR_NULL, &mid->hg_progress_tid);
211 212 213 214
    if(ret != 0) goto err;

err:
    if(mid)
215
    {
216 217 218
        margo_timer_instance_finalize(mid);
        ABT_mutex_free(&mid->finalize_mutex);
        ABT_cond_free(&mid->finalize_cond);
219
        free(mid);
220
    }
221
    return MARGO_INSTANCE_NULL;
222 223
}

Jonathan Jenkins's avatar
Jonathan Jenkins committed
224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248
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);
    }

249 250 251 252 253 254 255 256 257
    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
258 259 260
    free(mid);
}

261
void margo_finalize(margo_instance_id mid)
262
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
263
    int do_cleanup;
264

265
    /* tell progress thread to wrap things up */
266
    mid->hg_progress_shutdown_flag = 1;
267 268

    /* wait for it to shutdown cleanly */
269 270
    ABT_thread_join(mid->hg_progress_tid);
    ABT_thread_free(&mid->hg_progress_tid);
271

272 273 274 275
    ABT_mutex_lock(mid->finalize_mutex);
    mid->finalize_flag = 1;
    ABT_cond_broadcast(mid->finalize_cond);

Jonathan Jenkins's avatar
Jonathan Jenkins committed
276 277
    mid->refcount--;
    do_cleanup = mid->refcount == 0;
278

Jonathan Jenkins's avatar
Jonathan Jenkins committed
279 280 281 282 283 284 285
    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);
286 287 288 289 290 291 292

    return;
}

void margo_wait_for_finalize(margo_instance_id mid)
{
    int in_pool = 0;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
293
    int do_cleanup;
294 295 296 297

    /* Is this waiter in the same pool as the pool running the progress
     * thread?
     */
298
    if(margo_xstream_is_in_progress_pool(mid))
299 300 301 302
        in_pool = 1;

    ABT_mutex_lock(mid->finalize_mutex);

303
        mid->waiters_in_progress_pool += in_pool;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
304
        mid->refcount++;
305 306 307 308
            
        while(!mid->finalize_flag)
            ABT_cond_wait(mid->finalize_cond, mid->finalize_mutex);

Jonathan Jenkins's avatar
Jonathan Jenkins committed
309 310 311
        mid->refcount--;
        do_cleanup = mid->refcount == 0;

312
    ABT_mutex_unlock(mid->finalize_mutex);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
313 314 315 316

    if (do_cleanup)
        margo_cleanup(mid);

317 318 319
    return;
}

320 321
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)
322
{
323 324 325
	struct margo_rpc_data* margo_data;
    hg_return_t hret;
    hg_id_t id;
326

327 328 329
    id = HG_Register_name(mid->hg_class, func_name, in_proc_cb, out_proc_cb, rpc_cb);
    if(id <= 0)
        return(0);
330

331 332 333 334 335 336 337 338 339 340 341 342
	/* 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)
343
        {
344 345
            free(margo_data);
            return(0);
346
        }
347 348
    }

349
	return(id);
350 351
}

352 353 354
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)
355
{
356 357 358
    struct mplex_key key;
    struct mplex_element *element;
    hg_id_t id;
359

360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384
    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);
385 386
}

387 388
hg_return_t margo_registered_name(margo_instance_id mid, const char *func_name,
    hg_id_t *id, hg_bool_t *flag)
389
{
390
    return(HG_Registered_name(mid->hg_class, func_name, id, flag));
391 392
}

393 394 395 396 397 398 399
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 
400
		= (struct margo_rpc_data*) HG_Registered_data(mid->hg_class, id);
401 402 403 404 405 406 407 408 409 410 411 412 413 414
	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;
}

415 416 417 418
hg_return_t margo_registered_disable_response(
    margo_instance_id mid,
    hg_id_t id,
    int disable_flag)
419
{
420
    return(HG_Registered_disable_response(mid->hg_class, id, disable_flag));
421
}
422

423
struct lookup_cb_evt
424
{
425 426 427 428 429 430 431 432 433
    hg_return_t nret;
    hg_addr_t addr;
};

static hg_return_t margo_addr_lookup_cb(const struct hg_cb_info *info)
{
    struct lookup_cb_evt evt;
    evt.nret = info->ret;
    evt.addr = info->info.lookup.addr;
434
    struct margo_cb_arg* arg = info->arg;
435 436

    /* propagate return code out through eventual */
437 438
    ABT_eventual_set(*(arg->eventual), &evt, sizeof(evt));

439
#if 0
440 441
    if(arg->in_pool)
        arg->mid->waiters_in_progress_pool--;
442
#endif
443
    
444 445 446
    return(HG_SUCCESS);
}

447 448 449 450
hg_return_t margo_addr_lookup(
    margo_instance_id mid,
    const char   *name,
    hg_addr_t    *addr)
451
{
452 453 454 455 456
    hg_return_t nret;
    struct lookup_cb_evt *evt;
    ABT_eventual eventual;
    int ret;
    struct margo_cb_arg arg;
457

458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491
    ret = ABT_eventual_create(sizeof(*evt), &eventual);
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);        
    }

    arg.eventual = &eventual;
    arg.mid = mid;
#if 0
    if(margo_xstream_is_in_progress_pool(mid))
    {
        arg.in_pool = 1;
        mid->waiters_in_progress_pool++;
    }
    else
        arg.in_pool = 0;
#endif
    nret = HG_Addr_lookup(mid->hg_context, margo_addr_lookup_cb,
        &arg, name, HG_OP_ID_IGNORE);
    if(nret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&evt);
        *addr = evt->addr;
        nret = evt->nret;
    }

    ABT_eventual_free(&eventual);

    return(nret);
}

hg_return_t margo_addr_free(
    margo_instance_id mid,
    hg_addr_t addr)
492
{
493 494
    return(HG_Addr_free(mid->hg_class, addr));
}
495

496 497 498 499 500
hg_return_t margo_addr_self(
    margo_instance_id mid,
    hg_addr_t *addr)
{
    return(HG_Addr_self(mid->hg_class, addr));
501 502
}

503 504 505 506 507 508 509 510 511
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(
512
    margo_instance_id mid,
513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547
    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(
548
    hg_handle_t handle,
549
    void *in_struct)
550
{
551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597
    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));
}


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));
    
#if 0
    if(arg->in_pool)
        arg->mid->waiters_in_progress_pool--;
#endif

    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;
598
    ABT_eventual eventual;
599
    int ret;
600
    hg_return_t* waited_hret;
601
    struct margo_cb_arg arg;
602 603 604 605 606 607 608

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

609 610
    arg.eventual = &eventual;
    arg.mid = mid;
611
#if 0
612 613 614 615 616 617 618
    if(margo_xstream_is_in_progress_pool(mid))
    {
        arg.in_pool = 1;
        mid->waiters_in_progress_pool++;
    }
    else
        arg.in_pool = 0;
619
#endif
620
    hret = HG_Forward(handle, margo_cb, &arg, in_struct);
621 622 623 624 625 626 627 628 629
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
630 631
}

632 633 634 635
typedef struct
{
    hg_handle_t handle;
} margo_forward_timeout_cb_dat;
636

637 638 639 640 641 642 643 644 645 646 647
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(
648
    margo_instance_id mid,
649
    hg_handle_t handle,
650 651
    void *in_struct,
    double timeout_ms)
652 653
{
    int ret;
654 655
    hg_return_t hret;
    ABT_eventual eventual;
656
    hg_return_t* waited_hret;
657 658
    margo_timer_t forward_timer;
    margo_forward_timeout_cb_dat timeout_cb_dat;
659
    struct margo_cb_arg arg;
660 661 662 663 664 665 666

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

667 668 669 670 671
    /* 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);

672 673
    arg.eventual = &eventual;
    arg.mid = mid;
674
#if 0
675 676 677 678 679 680 681
    if(margo_xstream_is_in_progress_pool(mid))
    {
        arg.in_pool = 1;
        mid->waiters_in_progress_pool++;
    }
    else
        arg.in_pool = 0;
682
#endif
683
    hret = HG_Forward(handle, margo_cb, &arg, in_struct);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
684 685 686 687 688 689
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

690 691 692 693 694 695 696 697
    /* 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
698 699 700 701 702 703 704 705 706 707 708 709 710 711
    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;
712
    struct margo_cb_arg arg;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
713 714 715 716 717 718 719

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

720 721
    arg.eventual = &eventual;
    arg.mid = mid;
722
#if 0
723 724 725 726 727 728 729
    if(margo_xstream_is_in_progress_pool(mid))
    {
        arg.in_pool = 1;
        mid->waiters_in_progress_pool++;
    }
    else
        arg.in_pool = 0;
730
#endif
731
    hret = HG_Respond(handle, margo_cb, &arg, out_struct);
732 733 734 735 736 737 738 739 740 741 742
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

743 744 745 746 747
hg_return_t margo_cancel(
    hg_handle_t handle)
{
    return(HG_Cancel(handle));
}
748

749 750 751 752 753 754 755
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)
756
{
757
    /* XXX: handle caching logic? */
758

759 760 761
    return(HG_Bulk_create(mid->hg_class, count,
        buf_ptrs, buf_sizes, flags, handle));
}
762

763 764 765 766 767 768
hg_return_t margo_bulk_free(
    hg_bulk_t handle)
{
    /* XXX: handle caching logic? */

    return(HG_Bulk_free(handle));
769 770
}

771 772
hg_return_t margo_bulk_ref_incr(
    hg_bulk_t handle)
Philip Carns's avatar
Philip Carns committed
773
{
774 775
    return(HG_Bulk_ref_incr(handle));
}
Philip Carns's avatar
Philip Carns committed
776

777 778 779 780 781 782 783 784 785
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
786
{
787 788 789
    return(HG_Bulk_access(handle, offset, size, flags, max_count,
        buf_ptrs, buf_sizes, actual_count));
}
Philip Carns's avatar
Philip Carns committed
790

791 792 793 794 795
hg_size_t margo_bulk_get_size(
    hg_bulk_t handle)
{
    return(HG_Bulk_get_size(handle));
}
796

797 798 799 800
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
801 802
}

803 804 805 806 807 808
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
809

810 811 812 813 814
hg_return_t margo_bulk_serialize(
    void *buf,
    hg_size_t buf_size,
    hg_bool_t request_eager,
    hg_bulk_t handle)
815
{
816 817
    return(HG_Bulk_serialize(buf, buf_size, request_eager, handle));
}
818

819 820 821 822 823 824 825 826
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));
}
827

828 829 830 831
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;
832

833 834 835 836 837
    /* propagate return code out through eventual */
    ABT_eventual_set(*(arg->eventual), &hret, sizeof(hret));
    
    if(arg->in_pool)
        arg->mid->waiters_in_progress_pool--;
838

839
    return(HG_SUCCESS);
840 841
}

842
hg_return_t margo_bulk_transfer(
843
    margo_instance_id mid,
844
    hg_bulk_op_t op,
845
    hg_addr_t origin_addr,
846 847 848 849
    hg_bulk_t origin_handle,
    size_t origin_offset,
    hg_bulk_t local_handle,
    size_t local_offset,
850 851
    size_t size,
    hg_op_id_t *op_id)
852 853 854 855 856
{
    hg_return_t hret = HG_TIMEOUT;
    hg_return_t *waited_hret;
    ABT_eventual eventual;
    int ret;
857
    struct margo_cb_arg arg;
858 859 860 861 862 863 864

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

865 866 867 868 869 870 871 872 873
    arg.eventual = &eventual;
    arg.mid = mid;
    if(margo_xstream_is_in_progress_pool(mid))
    {
        arg.in_pool = 1;
        mid->waiters_in_progress_pool++;
    }
    else
        arg.in_pool = 0;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
874
    hret = HG_Bulk_transfer(mid->hg_context, margo_bulk_transfer_cb,
875
        &arg, op, origin_addr, origin_handle, origin_offset, local_handle,
876
        local_offset, size, op_id);
877 878 879 880 881 882 883 884 885 886 887
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

888 889 890 891 892 893 894
hg_return_t margo_bulk_cancel(
    hg_op_id_t op_id)
{
    return(HG_Bulk_cancel(op_id));
}

/* returns 1 if current xstream is in the progress pool, 0 if not */
895 896
typedef struct
{
897
    margo_instance_id mid;
898 899
    ABT_mutex mutex;
    ABT_cond cond;
Shane Snyder's avatar
Shane Snyder committed
900 901
    char is_asleep;
    char in_pool;
902 903 904 905 906 907 908
} 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;

909
    /* decrement number of waiting threads */
910
    sleep_cb_dat->mid->waiters_in_progress_pool -=
Shane Snyder's avatar
Shane Snyder committed
911
        sleep_cb_dat->in_pool;
912

913 914
    /* wake up the sleeping thread */
    ABT_mutex_lock(sleep_cb_dat->mutex);
915
    sleep_cb_dat->is_asleep = 0;
916 917 918 919 920 921 922
    ABT_cond_signal(sleep_cb_dat->cond);
    ABT_mutex_unlock(sleep_cb_dat->mutex);

    return;
}

void margo_thread_sleep(
923
    margo_instance_id mid,
924 925
    double timeout_ms)
{
926
    int in_pool = 0;
927 928 929
    margo_timer_t sleep_timer;
    margo_thread_sleep_cb_dat sleep_cb_dat;

Shane Snyder's avatar
Shane Snyder committed
930 931 932
    if(margo_xstream_is_in_progress_pool(mid))
        in_pool = 1;

933
    /* set data needed for sleep callback */
934
    sleep_cb_dat.mid = mid;
935 936
    ABT_mutex_create(&(sleep_cb_dat.mutex));
    ABT_cond_create(&(sleep_cb_dat.cond));
937
    sleep_cb_dat.is_asleep = 1;
Shane Snyder's avatar
Shane Snyder committed
938
    sleep_cb_dat.in_pool = in_pool;
939 940

    /* initialize the sleep timer */
941
    margo_timer_init(mid, &sleep_timer, margo_thread_sleep_cb,
942 943
        &sleep_cb_dat, timeout_ms);

944
    /* increment number of waiting threads */
945
    mid->waiters_in_progress_pool += in_pool;
946

947 948
    /* yield thread for specified timeout */
    ABT_mutex_lock(sleep_cb_dat.mutex);
949 950
    while(sleep_cb_dat.is_asleep)
        ABT_cond_wait(sleep_cb_dat.cond, sleep_cb_dat.mutex);
951 952
    ABT_mutex_unlock(sleep_cb_dat.mutex);

953 954 955 956
    /* clean up */
    ABT_mutex_free(&sleep_cb_dat.mutex);
    ABT_cond_free(&sleep_cb_dat.cond);

957 958 959
    return;
}

960
ABT_pool* margo_get_handler_pool(margo_instance_id mid)
961
{
962 963
    return(&mid->handler_pool);
}
964

965 966 967 968
hg_context_t* margo_get_context(margo_instance_id mid)
{
    return(mid->hg_context);
}
969

970 971 972
hg_class_t* margo_get_class(margo_instance_id mid)
{
    return(mid->hg_class);
973
}
Philip Carns's avatar
Philip Carns committed
974

975
margo_instance_id margo_hg_handle_get_instance(hg_handle_t h)
976
{
977 978 979 980 981 982
	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;
983 984
}

985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003
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
1004 1005
    assert(element->key.id == id && element->key.mplex_id == mplex_id);

1006 1007 1008 1009 1010
    *pool = element->pool;

    return(0);
}

1011
static int margo_xstream_is_in_progress_pool(margo_instance_id mid)
1012
{
1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025
    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);
1026 1027
}

1028
static void margo_rpc_data_free(void* ptr)
Philip Carns's avatar
Philip Carns committed
1029
{
1030 1031 1032 1033 1034 1035
	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);
}
1036

1037 1038 1039 1040 1041 1042 1043 1044 1045 1046
/* 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;
    unsigned int hg_progress_timeout = MERCURY_PROGRESS_TIMEOUT_UB;
    double next_timer_exp;
    int trigger_happened;
1047

1048 1049 1050 1051 1052 1053 1054 1055
    while(!mid->hg_progress_shutdown_flag)
    {
        trigger_happened = 0;
        do {
            ret = HG_Trigger(mid->hg_context, 0, 1, &actual_count);
            if(ret == HG_SUCCESS && actual_count > 0)
                trigger_happened = 1;
        } while((ret == HG_SUCCESS) && actual_count && !mid->hg_progress_shutdown_flag);
1056

1057 1058
        if(trigger_happened)
            ABT_thread_yield();
1059

1060 1061 1062 1063 1064 1065 1066 1067 1068
        ABT_pool_get_total_size(mid->progress_pool, &size);
        /* Are there any other threads executing in this pool that are *not*
         * blocked on margo_wait_for_finalize()?  If so then, we can't
         * sleep here or else those threads will not get a chance to
         * execute.
         */
        if(size > mid->waiters_in_progress_pool)
        {
            //printf("DEBUG: Margo progress function running while other ULTs are eligible for execution (size: %d, waiters: %d.\n", size, mid->waiters_in_progress_pool);
1069

1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 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 1118 1119 1120 1121
            /* 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.
             */
            ret = HG_Progress(mid->hg_context, 0);
            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
        {
            hg_progress_timeout = MERCURY_PROGRESS_TIMEOUT_UB;
            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 */
                    if(next_timer_exp < MERCURY_PROGRESS_TIMEOUT_UB)
                        hg_progress_timeout = (unsigned int)next_timer_exp;
                }
                else
                {
                    hg_progress_timeout = 0;
                }
            }
            ret = HG_Progress(mid->hg_context, hg_progress_timeout);
            if(ret != HG_SUCCESS && ret != HG_TIMEOUT)
            {
                /* TODO: error handling */
                fprintf(stderr, "WARNING: unexpected return code (%d) from HG_Progress()\n", ret);
            }
        }
1122

1123 1124 1125
        /* check for any expired timers */
        margo_check_timers(mid);
    }
1126

1127
    return;
Philip Carns's avatar
Philip Carns committed
1128
}