margo.c 33.8 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 */
Shane Snyder's avatar
Shane Snyder committed
22
#define DEFAULT_MERCURY_HANDLE_CACHE_SIZE 32
23

24 25 26 27 28 29 30 31 32 33 34 35 36
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
37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52
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)

Shane Snyder's avatar
Shane Snyder committed
53 54 55 56 57 58 59
struct margo_handle_cache_el
{
    hg_handle_t handle;
    UT_hash_handle hh; /* in-use hash link */
    struct margo_handle_cache_el *next; /* free list link */
};

60 61
struct margo_instance
{
Shane Snyder's avatar
Shane Snyder committed
62
    /* mercury/argobots state */
63 64
    hg_context_t *hg_context;
    hg_class_t *hg_class;
65 66 67
    ABT_pool handler_pool;
    ABT_pool progress_pool;

68
    /* internal to margo for this particular instance */
69
    int margo_init;
70 71
    ABT_thread hg_progress_tid;
    int hg_progress_shutdown_flag;
72
    ABT_xstream progress_xstream;
73 74 75
    int owns_progress_pool;
    ABT_xstream *rpc_xstreams;
    int num_handler_pool_threads;
76
    unsigned int hg_progress_timeout_ub;
77 78 79

    /* control logic for callers waiting on margo to be finalized */
    int finalize_flag;
80
    int refcount;
81 82 83
    ABT_mutex finalize_mutex;
    ABT_cond finalize_cond;

84 85
    /* hash table to track multiplexed rpcs registered with margo */
    struct mplex_element *mplex_table;
Philip Carns's avatar
Philip Carns committed
86

Shane Snyder's avatar
Shane Snyder committed
87 88 89 90
    /* linked list of free hg handles and a hash of in-use handles */
    struct margo_handle_cache_el *free_handle_list;
    struct margo_handle_cache_el *used_handle_hash;

Philip Carns's avatar
Philip Carns committed
91 92 93 94 95 96 97 98 99 100 101
    /* 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;
102 103
};

104 105 106 107 108 109
struct margo_cb_arg
{
    ABT_eventual *eventual;
    margo_instance_id mid;
};

110 111 112 113 114 115
struct margo_rpc_data
{
	margo_instance_id mid;
	void* user_data;
	void (*user_free_callback)(void *);
};
116

117
static void hg_progress_fn(void* foo);
118
static void margo_rpc_data_free(void* ptr);
119

Shane Snyder's avatar
Shane Snyder committed
120 121 122 123 124 125 126
static hg_return_t margo_handle_cache_init(margo_instance_id mid);
static void margo_handle_cache_destroy(margo_instance_id mid);
static hg_return_t margo_handle_cache_get(margo_instance_id mid,
    hg_addr_t addr, hg_id_t id, hg_handle_t *handle);
static hg_return_t margo_handle_cache_put(margo_instance_id mid,
    hg_handle_t handle);

127
margo_instance_id margo_init(const char *addr_str, int mode,
128
    int use_progress_thread, int rpc_thread_count)
129
{
130 131 132 133 134
    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;
135 136
    hg_class_t *hg_class = NULL;
    hg_context_t *hg_context = NULL;
137
    int listen_flag = (mode == MARGO_CLIENT_MODE) ? HG_FALSE : HG_TRUE;
138
    int i;
139 140 141
    int ret;
    struct margo_instance *mid = MARGO_INSTANCE_NULL;

142
    if(mode != MARGO_CLIENT_MODE && mode != MARGO_SERVER_MODE) goto err;
143

144 145 146
    ret = ABT_init(0, NULL); /* XXX: argc/argv not currently used by ABT ... */
    if(ret != 0) goto err;

147
    /* set caller (self) ES to idle without polling */
148 149
    ret = ABT_snoozer_xstream_self_set();
    if(ret != 0) goto err;
150 151 152 153 154 155 156 157 158 159 160 161 162 163

    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;
    }

164
    if (mode == MARGO_SERVER_MODE)
165
    {
166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184
        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;
        }
185 186
    }

187 188 189 190 191 192
    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;

193 194 195
    mid = margo_init_pool(progress_pool, rpc_pool, hg_context);
    if (mid == MARGO_INSTANCE_NULL) goto err;

196
    mid->margo_init = 1;
197 198 199
    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;
200
    mid->rpc_xstreams = rpc_xstreams;
201

202 203 204
    return mid;

err:
205 206 207 208 209 210 211
    if(mid)
    {
        margo_timer_instance_finalize(mid);
        ABT_mutex_free(&mid->finalize_mutex);
        ABT_cond_free(&mid->finalize_cond);
        free(mid);
    }
212 213 214 215 216 217 218 219 220 221 222 223 224 225
    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);
    }
226 227 228 229 230
    if(hg_context)
        HG_Context_destroy(hg_context);
    if(hg_class)
        HG_Finalize(hg_class);
    ABT_finalize();
231 232 233 234
    return MARGO_INSTANCE_NULL;
}

margo_instance_id margo_init_pool(ABT_pool progress_pool, ABT_pool handler_pool,
235
    hg_context_t *hg_context)
236 237
{
    int ret;
Shane Snyder's avatar
Shane Snyder committed
238
    hg_return_t hret;
239 240 241
    struct margo_instance *mid;

    mid = malloc(sizeof(*mid));
242
    if(!mid) goto err;
243
    memset(mid, 0, sizeof(*mid));
244

245 246 247
    ABT_mutex_create(&mid->finalize_mutex);
    ABT_cond_create(&mid->finalize_cond);

248 249
    mid->progress_pool = progress_pool;
    mid->handler_pool = handler_pool;
250
    mid->hg_class = HG_Context_get_class(hg_context);
251
    mid->hg_context = hg_context;
252
    mid->hg_progress_timeout_ub = DEFAULT_MERCURY_PROGRESS_TIMEOUT_UB;
253
    mid->refcount = 1;
254

255
    ret = margo_timer_instance_init(mid);
256
    if(ret != 0) goto err;
257

Shane Snyder's avatar
Shane Snyder committed
258 259 260 261
    /* initialize the handle cache */
    hret = margo_handle_cache_init(mid);
    if(hret != HG_SUCCESS) goto err;

262
    ret = ABT_thread_create(mid->progress_pool, hg_progress_fn, mid, 
263
        ABT_THREAD_ATTR_NULL, &mid->hg_progress_tid);
264 265
    if(ret != 0) goto err;

266 267
    return mid;

268 269
err:
    if(mid)
270
    {
Shane Snyder's avatar
Shane Snyder committed
271
        margo_handle_cache_destroy(mid);
272 273 274
        margo_timer_instance_finalize(mid);
        ABT_mutex_free(&mid->finalize_mutex);
        ABT_cond_free(&mid->finalize_cond);
275
        free(mid);
276
    }
277
    return MARGO_INSTANCE_NULL;
278 279
}

280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304
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);
    }

Shane Snyder's avatar
Shane Snyder committed
305 306
    margo_handle_cache_destroy(mid);

307 308 309 310 311 312 313 314 315
    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();
    }

316 317 318
    free(mid);
}

319
void margo_finalize(margo_instance_id mid)
320
{
321
    int do_cleanup;
322

323
    /* tell progress thread to wrap things up */
324
    mid->hg_progress_shutdown_flag = 1;
325 326

    /* wait for it to shutdown cleanly */
327 328
    ABT_thread_join(mid->hg_progress_tid);
    ABT_thread_free(&mid->hg_progress_tid);
329

330 331 332 333
    ABT_mutex_lock(mid->finalize_mutex);
    mid->finalize_flag = 1;
    ABT_cond_broadcast(mid->finalize_cond);

334 335
    mid->refcount--;
    do_cleanup = mid->refcount == 0;
336

337 338 339 340 341 342 343
    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);
344 345 346 347 348 349

    return;
}

void margo_wait_for_finalize(margo_instance_id mid)
{
350
    int do_cleanup;
351 352 353

    ABT_mutex_lock(mid->finalize_mutex);

354
        mid->refcount++;
355 356 357 358
            
        while(!mid->finalize_flag)
            ABT_cond_wait(mid->finalize_cond, mid->finalize_mutex);

359 360 361
        mid->refcount--;
        do_cleanup = mid->refcount == 0;

362
    ABT_mutex_unlock(mid->finalize_mutex);
363 364 365 366

    if (do_cleanup)
        margo_cleanup(mid);

367 368 369
    return;
}

370 371
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)
372
{
373 374 375
	struct margo_rpc_data* margo_data;
    hg_return_t hret;
    hg_id_t id;
376

377 378 379
    id = HG_Register_name(mid->hg_class, func_name, in_proc_cb, out_proc_cb, rpc_cb);
    if(id <= 0)
        return(0);
380

381 382 383 384 385 386 387 388 389 390 391 392
	/* 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)
393
        {
394 395
            free(margo_data);
            return(0);
396
        }
397 398
    }

399
	return(id);
400 401
}

402 403 404
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)
405
{
406 407 408
    struct mplex_key key;
    struct mplex_element *element;
    hg_id_t id;
409

410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434
    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);
435 436
}

437 438
hg_return_t margo_registered_name(margo_instance_id mid, const char *func_name,
    hg_id_t *id, hg_bool_t *flag)
439
{
440
    return(HG_Registered_name(mid->hg_class, func_name, id, flag));
441 442
}

443 444 445 446 447 448 449
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 
450
		= (struct margo_rpc_data*) HG_Registered_data(mid->hg_class, id);
451 452 453 454 455 456 457 458 459 460 461 462 463 464
	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;
}

465 466 467 468
hg_return_t margo_registered_disable_response(
    margo_instance_id mid,
    hg_id_t id,
    int disable_flag)
469
{
470
    return(HG_Registered_disable_response(mid->hg_class, id, disable_flag));
471
}
472

473
struct lookup_cb_evt
474
{
475
    hg_return_t hret;
476 477 478 479 480 481
    hg_addr_t addr;
};

static hg_return_t margo_addr_lookup_cb(const struct hg_cb_info *info)
{
    struct lookup_cb_evt evt;
482
    evt.hret = info->ret;
483
    evt.addr = info->info.lookup.addr;
484
    struct margo_cb_arg* arg = info->arg;
485 486

    /* propagate return code out through eventual */
487 488
    ABT_eventual_set(*(arg->eventual), &evt, sizeof(evt));

489 490 491
    return(HG_SUCCESS);
}

492 493 494 495
hg_return_t margo_addr_lookup(
    margo_instance_id mid,
    const char   *name,
    hg_addr_t    *addr)
496
{
497
    hg_return_t hret;
498 499 500 501
    struct lookup_cb_evt *evt;
    ABT_eventual eventual;
    int ret;
    struct margo_cb_arg arg;
502

503 504 505 506 507 508 509 510
    ret = ABT_eventual_create(sizeof(*evt), &eventual);
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);        
    }

    arg.eventual = &eventual;
    arg.mid = mid;
511

512
    hret = HG_Addr_lookup(mid->hg_context, margo_addr_lookup_cb,
513
        &arg, name, HG_OP_ID_IGNORE);
514
    if(hret == HG_SUCCESS)
515 516 517
    {
        ABT_eventual_wait(eventual, (void**)&evt);
        *addr = evt->addr;
518
        hret = evt->hret;
519 520 521 522
    }

    ABT_eventual_free(&eventual);

523
    return(hret);
524 525 526 527 528
}

hg_return_t margo_addr_free(
    margo_instance_id mid,
    hg_addr_t addr)
529
{
530 531
    return(HG_Addr_free(mid->hg_class, addr));
}
532

533 534 535 536 537
hg_return_t margo_addr_self(
    margo_instance_id mid,
    hg_addr_t *addr)
{
    return(HG_Addr_self(mid->hg_class, addr));
538 539
}

540 541 542 543 544 545 546 547 548
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(
549
    margo_instance_id mid,
550 551 552 553 554 555 556 557 558 559
    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)
{
Shane Snyder's avatar
Shane Snyder committed
560 561 562 563 564 565 566 567 568
    hg_return_t hret;

    /* look for a handle to reuse */
    hret = margo_handle_cache_get(mid, addr, id, handle);
    if(hret != HG_SUCCESS)
    {
        /* else try creating a new handle */
        hret = HG_Create(mid->hg_context, addr, id, handle);
    }
569

Shane Snyder's avatar
Shane Snyder committed
570
    return hret;
571 572
}

Shane Snyder's avatar
Shane Snyder committed
573
hg_return_t margo_destroy(margo_instance_id mid, hg_handle_t handle)
574
{
Shane Snyder's avatar
Shane Snyder committed
575 576 577 578 579 580 581 582 583
    hg_return_t hret;

    /* recycle this handle if it came from the handle cache */
    hret = margo_handle_cache_put(mid, handle);
    if(hret != HG_SUCCESS)
    {
        /* else destroy the handle manually */
        hret = HG_Destroy(handle);
    }
584

Shane Snyder's avatar
Shane Snyder committed
585
    return hret;
586 587 588 589 590 591 592 593 594 595 596 597 598 599 600
}

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(
601
    hg_handle_t handle,
602
    void *in_struct)
603
{
604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627
    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));
}

628 629 630 631 632 633 634
hg_return_t margo_set_target_id(
    hg_handle_t handle,
    hg_uint8_t target_id)
{
    return(HG_Set_target_id(handle, target_id));
}

635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651
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;
652
    ABT_eventual eventual;
653
    int ret;
654
    hg_return_t* waited_hret;
655
    struct margo_cb_arg arg;
656 657 658 659 660 661 662

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

663 664
    arg.eventual = &eventual;
    arg.mid = mid;
665

666
    hret = HG_Forward(handle, margo_cb, &arg, in_struct);
667
    if(hret == HG_SUCCESS)
668 669 670 671 672 673 674 675
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
676 677
}

678 679 680 681
typedef struct
{
    hg_handle_t handle;
} margo_forward_timeout_cb_dat;
682

683 684 685 686 687 688 689 690 691 692 693
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(
694
    margo_instance_id mid,
695
    hg_handle_t handle,
696 697
    void *in_struct,
    double timeout_ms)
698 699
{
    int ret;
700 701
    hg_return_t hret;
    ABT_eventual eventual;
702
    hg_return_t* waited_hret;
703 704
    margo_timer_t forward_timer;
    margo_forward_timeout_cb_dat timeout_cb_dat;
705
    struct margo_cb_arg arg;
706 707 708 709 710 711 712

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

713 714 715 716 717
    /* 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);

718 719
    arg.eventual = &eventual;
    arg.mid = mid;
720

721
    hret = HG_Forward(handle, margo_cb, &arg, in_struct);
722
    if(hret == HG_SUCCESS)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
723 724 725 726 727
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

728 729 730 731 732 733 734 735
    /* 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
736 737 738 739 740 741 742 743 744 745 746 747 748 749
    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;
750
    struct margo_cb_arg arg;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
751 752 753 754 755 756 757

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

758 759
    arg.eventual = &eventual;
    arg.mid = mid;
760

761
    hret = HG_Respond(handle, margo_cb, &arg, out_struct);
762
    if(hret == HG_SUCCESS)
763 764 765 766 767 768 769 770 771 772
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

773 774 775 776 777
hg_return_t margo_cancel(
    hg_handle_t handle)
{
    return(HG_Cancel(handle));
}
778

779 780 781 782 783 784 785
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)
786
{
787
    /* XXX: handle caching logic? */
788

789 790 791
    return(HG_Bulk_create(mid->hg_class, count,
        buf_ptrs, buf_sizes, flags, handle));
}
792

793 794 795 796 797 798
hg_return_t margo_bulk_free(
    hg_bulk_t handle)
{
    /* XXX: handle caching logic? */

    return(HG_Bulk_free(handle));
799 800
}

801 802
hg_return_t margo_bulk_ref_incr(
    hg_bulk_t handle)
803
{
804 805
    return(HG_Bulk_ref_incr(handle));
}
806

807 808 809 810 811 812 813 814 815
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)
816
{
817 818 819
    return(HG_Bulk_access(handle, offset, size, flags, max_count,
        buf_ptrs, buf_sizes, actual_count));
}
820

821 822 823 824 825
hg_size_t margo_bulk_get_size(
    hg_bulk_t handle)
{
    return(HG_Bulk_get_size(handle));
}
826

827 828 829 830
hg_uint32_t margo_bulk_get_segment_count(
    hg_bulk_t handle)
{
    return(HG_Bulk_get_segment_count(handle));
831 832
}

833 834 835 836 837 838
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));
}
839

840 841 842 843 844
hg_return_t margo_bulk_serialize(
    void *buf,
    hg_size_t buf_size,
    hg_bool_t request_eager,
    hg_bulk_t handle)
845
{
846 847
    return(HG_Bulk_serialize(buf, buf_size, request_eager, handle));
}
848

849 850 851 852 853 854 855 856
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));
}
857

858 859 860 861
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;
862

863 864 865 866
    /* propagate return code out through eventual */
    ABT_eventual_set(*(arg->eventual), &hret, sizeof(hret));
    
    return(HG_SUCCESS);
867 868
}

869
hg_return_t margo_bulk_transfer(
870
    margo_instance_id mid,
871
    hg_bulk_op_t op,
872
    hg_addr_t origin_addr,
873 874 875 876
    hg_bulk_t origin_handle,
    size_t origin_offset,
    hg_bulk_t local_handle,
    size_t local_offset,
877 878
    size_t size,
    hg_op_id_t *op_id)
879 880 881 882 883
{
    hg_return_t hret = HG_TIMEOUT;
    hg_return_t *waited_hret;
    ABT_eventual eventual;
    int ret;
884
    struct margo_cb_arg arg;
885 886 887 888 889 890 891

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

892 893
    arg.eventual = &eventual;
    arg.mid = mid;
894

895
    hret = HG_Bulk_transfer(mid->hg_context, margo_bulk_transfer_cb,
896
        &arg, op, origin_addr, origin_handle, origin_offset, local_handle,
897
        local_offset, size, op_id);
898
    if(hret == HG_SUCCESS)
899 900 901 902 903 904 905 906 907 908
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

909 910 911 912 913 914
hg_return_t margo_bulk_cancel(
    hg_op_id_t op_id)
{
    return(HG_Bulk_cancel(op_id));
}

915 916 917 918
typedef struct
{
    ABT_mutex mutex;
    ABT_cond cond;
919
    char is_asleep;
920 921 922 923 924 925 926 927 928
} 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);
929
    sleep_cb_dat->is_asleep = 0;
930 931 932 933 934 935 936
    ABT_cond_signal(sleep_cb_dat->cond);
    ABT_mutex_unlock(sleep_cb_dat->mutex);

    return;
}

void margo_thread_sleep(
937
    margo_instance_id mid,
938 939 940 941 942 943 944 945
    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));
946
    sleep_cb_dat.is_asleep = 1;
947 948

    /* initialize the sleep timer */
949
    margo_timer_init(mid, &sleep_timer, margo_thread_sleep_cb,
950 951 952 953
        &sleep_cb_dat, timeout_ms);

    /* yield thread for specified timeout */
    ABT_mutex_lock(sleep_cb_dat.mutex);
954 955
    while(sleep_cb_dat.is_asleep)
        ABT_cond_wait(sleep_cb_dat.cond, sleep_cb_dat.mutex);
956 957
    ABT_mutex_unlock(sleep_cb_dat.mutex);

958 959 960 961
    /* clean up */
    ABT_mutex_free(&sleep_cb_dat.mutex);
    ABT_cond_free(&sleep_cb_dat.cond);

962 963 964
    return;
}

965
ABT_pool* margo_get_handler_pool(margo_instance_id mid)
966
{
967 968
    return(&mid->handler_pool);
}
969

970 971 972 973
hg_context_t* margo_get_context(margo_instance_id mid)
{
    return(mid->hg_context);
}
974

975 976 977
hg_class_t* margo_get_class(margo_instance_id mid)
{
    return(mid->hg_class);
978
}
Philip Carns's avatar
Philip Carns committed
979

980
margo_instance_id margo_hg_handle_get_instance(hg_handle_t h)
981
{
982 983
	const struct hg_info* info = HG_Get_info(h);
	if(!info) return MARGO_INSTANCE_NULL;
984 985 986 987 988
    return margo_hg_info_get_instance(info);
}

margo_instance_id margo_hg_info_get_instance(const struct hg_info *info)
{
989 990 991 992
	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;
993 994
}

995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013
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
1014 1015
    assert(element->key.id == id && element->key.mplex_id == mplex_id);

1016 1017 1018 1019 1020
    *pool = element->pool;

    return(0);
}

1021
static void margo_rpc_data_free(void* ptr)
Philip Carns's avatar
Philip Carns committed
1022
{
1023 1024 1025 1026 1027 1028
	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);
}
1029

1030 1031 1032 1033 1034 1035 1036
/* 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;
1037
    unsigned int hg_progress_timeout = mid->hg_progress_timeout_ub;
1038 1039
    double next_timer_exp;
    int trigger_happened;
1040 1041
    double tm1, tm2;
    int diag_enabled = 0;
1042

1043 1044 1045 1046
    while(!mid->hg_progress_shutdown_flag)
    {
        trigger_happened = 0;
        do {
1047 1048 1049 1050
            /* save value of instance diag variable, in case it is modified
             * while we are in loop 
             */
            diag_enabled = mid->diag_enabled;
1051

1052
            if(diag_enabled) tm1 = ABT_get_wtime();
1053
            ret = HG_Trigger(mid->hg_context, 0, 1, &actual_count);
1054 1055 1056 1057 1058
            if(diag_enabled)
            {
                tm2 = ABT_get_wtime();
                __DIAG_UPDATE(mid->diag_trigger_elapsed, (tm2-tm1));
            }
1059

1060 1061 1062
            if(ret == HG_SUCCESS && actual_count > 0)
                trigger_happened = 1;
        } while((ret == HG_SUCCESS) && actual_count && !mid->hg_progress_shutdown_flag);
1063

1064 1065
        if(trigger_happened)
            ABT_thread_yield();
1066

1067
        ABT_pool_get_size(mid->progress_pool, &size);
1068
        /* Are there any other threads executing in this pool that are *not*
1069 1070 1071 1072
         * 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.
1073
         */
1074
        if(size > 0)
1075 1076 1077 1078 1079 1080 1081 1082
        {
            /* 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.
             */
1083
            if(diag_enabled) tm1 = ABT_get_wtime();
1084
            ret = HG_Progress(mid->hg_context, 0);
1085 1086 1087 1088 1089 1090
            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);
            }
1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109
            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
        {
1110
            hg_progress_timeout = mid->hg_progress_timeout_ub;
1111 1112 1113 1114 1115 1116 1117 1118 1119
            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 */
1120
                    if(next_timer_exp < mid->hg_progress_timeout_ub)
1121 1122 1123 1124 1125 1126 1127
                        hg_progress_timeout = (unsigned int)next_timer_exp;
                }
                else
                {
                    hg_progress_timeout = 0;
                }
            }
1128
            if(diag_enabled) tm1 = ABT_get_wtime();
1129
            ret = HG_Progress(mid->hg_context, hg_progress_timeout);
1130 1131 1132 1133 1134 1135 1136 1137 1138 1139
            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);
            }
1140 1141 1142 1143 1144 1145
            if(ret != HG_SUCCESS && ret != HG_TIMEOUT)
            {
                /* TODO: error handling */
                fprintf(stderr, "WARNING: unexpected return code (%d) from HG_Progress()\n", ret);
            }
        }
1146

1147 1148 1149
        /* check for any expired timers */
        margo_check_timers(mid);
    }
1150

1151
    return;
Philip Carns's avatar
Philip Carns committed
1152
}
Philip Carns's avatar
Philip Carns committed
1153 1154 1155 1156 1157 1158 1159


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

1160
static void print_diag_data(FILE *file, const char* name, const char *description, struct diag_data *data)
Philip Carns's avatar
Philip Carns committed
1161
{
1162 1163 1164 1165 1166 1167 1168 1169
    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
1170 1171 1172
    return;
}

1173
void margo_diag_dump(margo_instance_id mid, const char* file, int uniquify)
Philip Carns's avatar
Philip Carns committed
1174 1175 1176
{
    FILE *outfile;
    time_t ltime;
1177
    char revised_file_name[256] = {0};
Philip Carns's avatar
Philip Carns committed
1178 1179 1180

    assert(mid->diag_enabled);

1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195
    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
1196 1197 1198 1199 1200 1201
    if(strcmp("-", file) == 0)
    {
        outfile = stdout;
    }
    else
    {
1202
        outfile = fopen(revised_file_name, "a");
Philip Carns's avatar
Philip Carns committed
1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218
        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");
1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230
    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
1231 1232 1233 1234 1235 1236

    if(outfile != stdout)
        fclose(outfile);
    
    return;
}
1237

1238
void margo_set_param(margo_instance_id mid, int option, const void *param)
1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249
{
    switch(option)
    {
        case MARGO_INFO_PROGRESS_TIMEOUT_UB:
            mid->hg_progress_timeout_ub = (*((const unsigned int*)param));
            break;
    }

    return;
}

1250
void margo_get_param(margo_instance_id mid, int option, void *param)
1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261
{

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

    return;
}
Shane Snyder's avatar
Shane Snyder committed
1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358

static hg_return_t margo_handle_cache_init(margo_instance_id mid)
{
    int i;
    struct margo_handle_cache_el *el;
    hg_return_t hret = HG_SUCCESS;

    for(i = 0; i < DEFAULT_MERCURY_HANDLE_CACHE_SIZE; i++)
    {
        el = malloc(sizeof(*el));
        if(!el)
        {
            hret = HG_NOMEM_ERROR;
            margo_handle_cache_destroy(mid);
            break;
        }

        /* create handle with NULL_ADDRs, we will reset later to valid addrs */
        hret = HG_Create(mid->hg_context, HG_ADDR_NULL, 0, &el->handle);
        if(hret != HG_SUCCESS)
        {
            free(el);
            margo_handle_cache_destroy(mid);
            break;
        }

        /* add to the free list */
        LL_PREPEND(mid->free_handle_list, el);
    }

    return hret;
}

static void margo_handle_cache_destroy(margo_instance_id mid)
{
    struct margo_handle_cache_el *el, *tmp;

    /* only free handle list elements -- handles in hash are still in use */
    LL_FOREACH_SAFE(mid->free_handle_list, el, tmp)
    {
        LL_DELETE(mid->free_handle_list, el);
        HG_Destroy(el->handle);
        free(el);
    }

    return;
}

static hg_return_t margo_handle_cache_get(margo_instance_id mid,
    hg_addr_t addr, hg_id_t id, hg_handle_t *handle)
{
    struct margo_handle_cache_el *el;
    hg_return_t hret;

    if(!mid->free_handle_list)
    {
        /* if no available handles, just fall through */
        return HG_OTHER_ERROR;
    }

    /* pop first element from the free handle list */
    el = mid->free_handle_list;
    LL_DELETE(mid->free_handle_list, el);

    /* reset handle */
    hret = HG_Reset(el->handle, addr, id);
    if(hret == HG_SUCCESS)
    {
        /* put on in-use list and pass back handle */
        HASH_ADD(hh, mid->used_handle_hash, handle, sizeof(hg_handle_t), el);
        *handle = el->handle;
    }

    return hret;
}

static hg_return_t margo_handle_cache_put(margo_instance_id mid,
    hg_handle_t handle)
{
    struct margo_handle_cache_el *el;

    /* look for handle in the in-use hash */
    HASH_FIND(hh, mid->used_handle_hash, &handle, sizeof(hg_handle_t), el);
    if(!el)
    {
        /* this handle was manually allocated -- just fall through */
        return HG_OTHER_ERROR;
    }

    /* remove from the in-use hash */
    HASH_DELETE(hh, mid->used_handle_hash, el);

    /* add to the tail of the free handle list */
    LL_APPEND(mid->free_handle_list, el);

    return HG_SUCCESS;
}