margo.c 39.5 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 <stdlib.h>
13 14 15

#include <margo-config.h>
#ifdef HAVE_ABT_SNOOZER
16
#include <abt-snoozer.h>
17
#endif
18
#include <time.h>
Philip Carns's avatar
Philip Carns committed
19
#include <math.h>
20 21

#include "margo.h"
22
#include "margo-timer.h"
Philip Carns's avatar
Philip Carns committed
23
#include "utlist.h"
24
#include "uthash.h"
25

26
#define DEFAULT_MERCURY_PROGRESS_TIMEOUT_UB 100 /* 100 milliseconds */
Shane Snyder's avatar
Shane Snyder committed
27
#define DEFAULT_MERCURY_HANDLE_CACHE_SIZE 32
28

29 30 31
struct mplex_key
{
    hg_id_t id;
32
    uint8_t mplex_id;
33 34 35 36 37 38
};

struct mplex_element
{
    struct mplex_key key;
    ABT_pool pool;
39 40
    void* user_data;
    void(*user_free_callback)(void*);
41 42 43
    UT_hash_handle hh;
};

Philip Carns's avatar
Philip Carns committed
44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59
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
60 61 62 63 64 65 66
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 */
};

67 68 69 70 71 72 73
struct margo_finalize_cb
{
    void(*callback)(void*);
    void* uargs;
    struct margo_finalize_cb* next;
};

74 75
struct margo_timer_list; /* defined in margo-timer.c */

76 77
struct margo_instance
{
Shane Snyder's avatar
Shane Snyder committed
78
    /* mercury/argobots state */
79 80
    hg_context_t *hg_context;
    hg_class_t *hg_class;
81 82 83
    ABT_pool handler_pool;
    ABT_pool progress_pool;

84
    /* internal to margo for this particular instance */
85
    int margo_init;
86
    int abt_init;
87 88
    ABT_thread hg_progress_tid;
    int hg_progress_shutdown_flag;
89
    ABT_xstream progress_xstream;
90 91 92
    int owns_progress_pool;
    ABT_xstream *rpc_xstreams;
    int num_handler_pool_threads;
93
    unsigned int hg_progress_timeout_ub;
94 95 96

    /* control logic for callers waiting on margo to be finalized */
    int finalize_flag;
97
    int refcount;
98 99
    ABT_mutex finalize_mutex;
    ABT_cond finalize_cond;
100
    struct margo_finalize_cb* finalize_cb;
101

Matthieu Dorier's avatar
Matthieu Dorier committed
102 103 104 105
    /* control logic for shutting down */
    hg_id_t shutdown_rpc_id;
    int enable_remote_shutdown;

106 107 108
    /* timer data */
    struct margo_timer_list* timer_list;

109 110
    /* hash table to track multiplexed rpcs registered with margo */
    struct mplex_element *mplex_table;
Philip Carns's avatar
Philip Carns committed
111

Shane Snyder's avatar
Shane Snyder committed
112 113 114
    /* 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;
115
    ABT_mutex handle_cache_mtx; /* mutex protecting access to above caches */
Shane Snyder's avatar
Shane Snyder committed
116

Philip Carns's avatar
Philip Carns committed
117 118 119 120 121 122 123 124 125 126 127
    /* 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;
128 129
};

130 131 132 133 134 135
struct margo_rpc_data
{
	margo_instance_id mid;
	void* user_data;
	void (*user_free_callback)(void *);
};
136

Matthieu Dorier's avatar
Matthieu Dorier committed
137 138
MERCURY_GEN_PROC(margo_shutdown_out_t, ((int32_t)(ret)))

139
static void hg_progress_fn(void* foo);
140
static void margo_rpc_data_free(void* ptr);
Matthieu Dorier's avatar
Matthieu Dorier committed
141 142
static void remote_shutdown_ult(hg_handle_t handle);
DECLARE_MARGO_RPC_HANDLER(remote_shutdown_ult);
143

Shane Snyder's avatar
Shane Snyder committed
144 145 146 147 148 149
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);
150 151
static void delete_multiplexing_hash(margo_instance_id mid);

Shane Snyder's avatar
Shane Snyder committed
152

153
margo_instance_id margo_init(const char *addr_str, int mode,
154
    int use_progress_thread, int rpc_thread_count)
155
{
156 157 158 159 160
    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;
161 162
    hg_class_t *hg_class = NULL;
    hg_context_t *hg_context = NULL;
163
    int listen_flag = (mode == MARGO_CLIENT_MODE) ? HG_FALSE : HG_TRUE;
164
    int abt_init = 0;
165
    int i;
166 167 168
    int ret;
    struct margo_instance *mid = MARGO_INSTANCE_NULL;

169
    if(mode != MARGO_CLIENT_MODE && mode != MARGO_SERVER_MODE) goto err;
170

171 172 173 174 175 176 177 178 179 180 181 182
    /* NOTE: Margo is very likely to create a single producer (the
     * progress function), multiple consumer usage pattern that
     * causes excess memory consumption in some versions of
     * Argobots.  See
     * https://xgitlab.cels.anl.gov/sds/margo/issues/40 for details.
     * We therefore manually set the ABT_MEM_MAX_NUM_STACKS parameter 
     * for Argobots to a low value so that RPC handler threads do not
     * queue large numbers of stacks for reuse in per-ES data 
     * structures.
     */
    putenv("ABT_MEM_MAX_NUM_STACKS=8");

183 184 185 186 187 188
    if (ABT_initialized() == ABT_ERR_UNINITIALIZED)
    {
        ret = ABT_init(0, NULL); /* XXX: argc/argv not currently used by ABT ... */
        if(ret != 0) goto err;
        abt_init = 1;
    }
189

190
    /* set caller (self) ES to idle without polling */
191
#ifdef HAVE_ABT_SNOOZER
192 193
    ret = ABT_snoozer_xstream_self_set();
    if(ret != 0) goto err;
194
#endif
195 196 197

    if (use_progress_thread)
    {
198
#ifdef HAVE_ABT_SNOOZER
199
        ret = ABT_snoozer_xstream_create(1, &progress_pool, &progress_xstream);
200 201 202 203 204 205 206
		if (ret != ABT_SUCCESS) goto err;
#else
		ret = ABT_xstream_create(ABT_SCHED_NULL, &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;
#endif
207 208 209 210 211 212 213 214 215
    }
    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;
    }

216
    if (rpc_thread_count > 0)
217
    {
218 219
        rpc_xstreams = calloc(rpc_thread_count, sizeof(*rpc_xstreams));
        if (rpc_xstreams == NULL) goto err;
220
#ifdef HAVE_ABT_SNOOZER
221 222 223
        ret = ABT_snoozer_xstream_create(rpc_thread_count, &rpc_pool,
                rpc_xstreams);
        if (ret != ABT_SUCCESS) goto err;
224
#else
225 226 227 228 229
        int j;
        ret = ABT_pool_create_basic(ABT_POOL_FIFO, ABT_POOL_ACCESS_MPMC, ABT_TRUE, &rpc_pool);
        if (ret != ABT_SUCCESS) goto err;
        for(j=0; j<rpc_thread_count; j++) {
            ret = ABT_xstream_create(ABT_SCHED_NULL, rpc_xstreams+j);
230 231
            if (ret != ABT_SUCCESS) goto err;
        }
232 233 234 235 236 237 238 239 240 241 242 243
#endif
    }
    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;
244 245
    }

246 247 248 249 250 251
    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;

252 253 254
    mid = margo_init_pool(progress_pool, rpc_pool, hg_context);
    if (mid == MARGO_INSTANCE_NULL) goto err;

255
    mid->margo_init = 1;
256
    mid->abt_init = abt_init;
257 258 259
    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;
260
    mid->rpc_xstreams = rpc_xstreams;
261

262 263 264
    return mid;

err:
265 266
    if(mid)
    {
267
        margo_timer_list_free(mid->timer_list);
268 269 270 271
        ABT_mutex_free(&mid->finalize_mutex);
        ABT_cond_free(&mid->finalize_cond);
        free(mid);
    }
272 273 274 275 276 277 278 279 280 281 282 283 284 285
    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);
    }
286 287 288 289
    if(hg_context)
        HG_Context_destroy(hg_context);
    if(hg_class)
        HG_Finalize(hg_class);
290 291
    if(abt_init)
        ABT_finalize();
292 293 294 295
    return MARGO_INSTANCE_NULL;
}

margo_instance_id margo_init_pool(ABT_pool progress_pool, ABT_pool handler_pool,
296
    hg_context_t *hg_context)
297 298
{
    int ret;
Shane Snyder's avatar
Shane Snyder committed
299
    hg_return_t hret;
300 301
    struct margo_instance *mid;

Matthieu Dorier's avatar
Matthieu Dorier committed
302
    mid = calloc(1,sizeof(*mid));
303
    if(!mid) goto err;
304
    memset(mid, 0, sizeof(*mid));
305

306 307 308
    ABT_mutex_create(&mid->finalize_mutex);
    ABT_cond_create(&mid->finalize_cond);

309 310
    mid->progress_pool = progress_pool;
    mid->handler_pool = handler_pool;
311
    mid->hg_class = HG_Context_get_class(hg_context);
312
    mid->hg_context = hg_context;
313
    mid->hg_progress_timeout_ub = DEFAULT_MERCURY_PROGRESS_TIMEOUT_UB;
314
    mid->mplex_table = NULL;
315
    mid->refcount = 1;
316
    mid->finalize_cb = NULL;
Matthieu Dorier's avatar
Matthieu Dorier committed
317
    mid->enable_remote_shutdown = 0;
318

319 320
    mid->timer_list = margo_timer_list_create();
    if(mid->timer_list == NULL) goto err;
321

Shane Snyder's avatar
Shane Snyder committed
322 323 324 325
    /* initialize the handle cache */
    hret = margo_handle_cache_init(mid);
    if(hret != HG_SUCCESS) goto err;

326
    ret = ABT_thread_create(mid->progress_pool, hg_progress_fn, mid, 
327
        ABT_THREAD_ATTR_NULL, &mid->hg_progress_tid);
328 329
    if(ret != 0) goto err;

Matthieu Dorier's avatar
Matthieu Dorier committed
330 331 332
    mid->shutdown_rpc_id = MARGO_REGISTER(mid, "__shutdown__", 
            void, margo_shutdown_out_t, remote_shutdown_ult);

333 334
    return mid;

335 336
err:
    if(mid)
337
    {
Shane Snyder's avatar
Shane Snyder committed
338
        margo_handle_cache_destroy(mid);
339
        margo_timer_list_free(mid->timer_list);
340 341
        ABT_mutex_free(&mid->finalize_mutex);
        ABT_cond_free(&mid->finalize_cond);
342
        free(mid);
343
    }
344
    return MARGO_INSTANCE_NULL;
345 346
}

347 348 349 350
static void margo_cleanup(margo_instance_id mid)
{
    int i;

351 352 353 354 355 356 357 358 359
    /* call finalize callbacks */
    struct margo_finalize_cb* fcb = mid->finalize_cb;
    while(fcb) {
        (fcb->callback)(fcb->uargs);
        struct margo_finalize_cb* tmp = fcb;
        fcb = fcb->next;
        free(tmp);
    }

360
    margo_timer_list_free(mid->timer_list);
361

362 363 364
    /* delete the hash used for multiplexing */
    delete_multiplexing_hash(mid);

365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383
    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
384 385
    margo_handle_cache_destroy(mid);

386 387 388 389 390 391
    if (mid->margo_init)
    {
        if (mid->hg_context)
            HG_Context_destroy(mid->hg_context);
        if (mid->hg_class)
            HG_Finalize(mid->hg_class);
392 393
        if (mid->abt_init)
            ABT_finalize();
394 395
    }

396 397 398
    free(mid);
}

399
void margo_finalize(margo_instance_id mid)
400
{
401
    int do_cleanup;
402

403
    /* tell progress thread to wrap things up */
404
    mid->hg_progress_shutdown_flag = 1;
405 406

    /* wait for it to shutdown cleanly */
407 408
    ABT_thread_join(mid->hg_progress_tid);
    ABT_thread_free(&mid->hg_progress_tid);
409

410 411 412 413
    ABT_mutex_lock(mid->finalize_mutex);
    mid->finalize_flag = 1;
    ABT_cond_broadcast(mid->finalize_cond);

414 415
    mid->refcount--;
    do_cleanup = mid->refcount == 0;
416

417 418 419 420 421 422 423
    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);
424 425 426 427 428 429

    return;
}

void margo_wait_for_finalize(margo_instance_id mid)
{
430
    int do_cleanup;
431 432 433

    ABT_mutex_lock(mid->finalize_mutex);

434
        mid->refcount++;
435 436 437 438
            
        while(!mid->finalize_flag)
            ABT_cond_wait(mid->finalize_cond, mid->finalize_mutex);

439 440 441
        mid->refcount--;
        do_cleanup = mid->refcount == 0;

442
    ABT_mutex_unlock(mid->finalize_mutex);
443 444 445 446

    if (do_cleanup)
        margo_cleanup(mid);

447 448 449
    return;
}

450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466
void margo_push_finalize_callback(
            margo_instance_id mid,
            void(*cb)(void*),                  
            void* uargs)
{
    if(cb == NULL) return;

    struct margo_finalize_cb* fcb = 
        (struct margo_finalize_cb*)malloc(sizeof(*fcb));
    fcb->callback = cb;
    fcb->uargs = uargs;

    struct margo_finalize_cb* next = mid->finalize_cb;
    fcb->next = next;
    mid->finalize_cb = fcb;
}

467 468
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)
469
{
470 471 472
	struct margo_rpc_data* margo_data;
    hg_return_t hret;
    hg_id_t id;
473

474 475 476
    id = HG_Register_name(mid->hg_class, func_name, in_proc_cb, out_proc_cb, rpc_cb);
    if(id <= 0)
        return(0);
477

478 479 480 481 482 483 484 485 486 487 488 489
	/* 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)
490
        {
491 492
            free(margo_data);
            return(0);
493
        }
494 495
    }

496
	return(id);
497 498
}

Matthieu Dorier's avatar
Matthieu Dorier committed
499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529
void margo_enable_remote_shutdown(margo_instance_id mid)
{
    mid->enable_remote_shutdown = 1;
}

int margo_shutdown_remote_instance(
        margo_instance_id mid,
        hg_addr_t remote_addr)
{
    hg_return_t hret;
    hg_handle_t handle;

    hret = margo_create(mid, remote_addr,
                        mid->shutdown_rpc_id, &handle);
    if(hret != HG_SUCCESS) return -1;

    hret = margo_forward(handle, NULL);
    if(hret != HG_SUCCESS)
    {
        margo_destroy(handle);
        return -1;
    }

    margo_shutdown_out_t out;
    hret = margo_get_output(handle, &out);
    if(hret != HG_SUCCESS)
    {
        margo_destroy(handle);
        return -1;
    }

530
    margo_free_output(handle, &out);
Matthieu Dorier's avatar
Matthieu Dorier committed
531 532 533 534 535
    margo_destroy(handle);

    return out.ret;
}

536 537
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,
538
    uint8_t mplex_id, ABT_pool pool)
539
{
540 541 542
    struct mplex_key key;
    struct mplex_element *element;
    hg_id_t id;
543

544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559
    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);

560
    element = calloc(1,sizeof(*element));
561 562 563 564 565 566 567 568
    if(!element)
        return(0);
    element->key = key;
    element->pool = pool;

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

    return(id);
569 570
}

571 572
hg_return_t margo_registered_name(margo_instance_id mid, const char *func_name,
    hg_id_t *id, hg_bool_t *flag)
573
{
574
    return(HG_Registered_name(mid->hg_class, func_name, id, flag));
575 576
}

577
hg_return_t margo_registered_name_mplex(margo_instance_id mid, const char *func_name,
578
    uint8_t mplex_id, hg_id_t *id, hg_bool_t *flag)
579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607
{
    hg_bool_t b;
    hg_return_t ret = margo_registered_name(mid, func_name, id, &b);
    if(ret != HG_SUCCESS) 
        return ret;
    if((!b) || (!mplex_id)) {
        *flag = b;
        return ret;
    }

    struct mplex_key key;
    struct mplex_element *element;

    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) {
        *flag = 0;
        return HG_SUCCESS;
    }

    assert(element->key.id == *id && element->key.mplex_id == mplex_id);

    *flag = 1;
    return HG_SUCCESS;
}

608 609 610 611 612 613 614
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 
615
		= (struct margo_rpc_data*) HG_Registered_data(mid->hg_class, id);
616
	if(!margo_data) return HG_OTHER_ERROR;
617 618 619
    if(margo_data->user_data && margo_data->user_free_callback) {
        (margo_data->user_free_callback)(margo_data->user_data);
    }
620 621 622 623 624 625 626 627 628 629 630 631 632
	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;
}

633 634 635 636
hg_return_t margo_registered_disable_response(
    margo_instance_id mid,
    hg_id_t id,
    int disable_flag)
637
{
638
    return(HG_Registered_disable_response(mid->hg_class, id, disable_flag));
639
}
640

641
struct lookup_cb_evt
642
{
643
    hg_return_t hret;
644 645 646 647 648 649
    hg_addr_t addr;
};

static hg_return_t margo_addr_lookup_cb(const struct hg_cb_info *info)
{
    struct lookup_cb_evt evt;
650
    evt.hret = info->ret;
651
    evt.addr = info->info.lookup.addr;
Matthieu Dorier's avatar
Matthieu Dorier committed
652
    ABT_eventual eventual = (ABT_eventual)(info->arg);
653 654

    /* propagate return code out through eventual */
Matthieu Dorier's avatar
Matthieu Dorier committed
655
    ABT_eventual_set(eventual, &evt, sizeof(evt));
656

657 658 659
    return(HG_SUCCESS);
}

660 661 662 663
hg_return_t margo_addr_lookup(
    margo_instance_id mid,
    const char   *name,
    hg_addr_t    *addr)
664
{
665
    hg_return_t hret;
666 667 668
    struct lookup_cb_evt *evt;
    ABT_eventual eventual;
    int ret;
669

670 671 672 673 674 675
    ret = ABT_eventual_create(sizeof(*evt), &eventual);
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);        
    }

676
    hret = HG_Addr_lookup(mid->hg_context, margo_addr_lookup_cb,
Matthieu Dorier's avatar
Matthieu Dorier committed
677
        (void*)eventual, name, HG_OP_ID_IGNORE);
678
    if(hret == HG_SUCCESS)
679 680 681
    {
        ABT_eventual_wait(eventual, (void**)&evt);
        *addr = evt->addr;
682
        hret = evt->hret;
683 684 685 686
    }

    ABT_eventual_free(&eventual);

687
    return(hret);
688 689 690 691 692
}

hg_return_t margo_addr_free(
    margo_instance_id mid,
    hg_addr_t addr)
693
{
694 695
    return(HG_Addr_free(mid->hg_class, addr));
}
696

697 698 699 700 701
hg_return_t margo_addr_self(
    margo_instance_id mid,
    hg_addr_t *addr)
{
    return(HG_Addr_self(mid->hg_class, addr));
702 703
}

704 705 706 707 708 709 710 711 712
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(
713
    margo_instance_id mid,
714 715 716 717 718 719 720 721 722 723
    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)
{
724
    hg_return_t hret = HG_OTHER_ERROR;
Shane Snyder's avatar
Shane Snyder committed
725 726 727 728 729 730 731 732

    /* 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);
    }
733

Shane Snyder's avatar
Shane Snyder committed
734
    return hret;
735 736
}

737
hg_return_t margo_destroy(hg_handle_t handle)
738
{
739
    margo_instance_id mid;
740
    hg_return_t hret = HG_OTHER_ERROR;
Shane Snyder's avatar
Shane Snyder committed
741

742 743 744
    /* use the handle to get the associated mid */
    mid = margo_hg_handle_get_instance(handle);

Shane Snyder's avatar
Shane Snyder committed
745 746 747 748 749 750 751
    /* 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);
    }
752

Shane Snyder's avatar
Shane Snyder committed
753
    return hret;
754 755 756 757 758
}

static hg_return_t margo_cb(const struct hg_cb_info *info)
{
    hg_return_t hret = info->ret;
Matthieu Dorier's avatar
Matthieu Dorier committed
759
    ABT_eventual eventual = (ABT_eventual)(info->arg);
760 761

    /* propagate return code out through eventual */
Matthieu Dorier's avatar
Matthieu Dorier committed
762
    ABT_eventual_set(eventual, &hret, sizeof(hret));
763 764 765 766 767 768 769
    
    return(HG_SUCCESS);
}

hg_return_t margo_forward(
    hg_handle_t handle,
    void *in_struct)
770 771 772 773 774 775 776 777 778 779 780 781 782
{
	hg_return_t hret;
	margo_request req;
	hret = margo_iforward(handle, in_struct, &req);
	if(hret != HG_SUCCESS) 
		return hret;
	return margo_wait(req);
}

hg_return_t margo_iforward(
    hg_handle_t handle,
    void *in_struct,
    margo_request* req)
783 784
{
    hg_return_t hret = HG_TIMEOUT;
785
    ABT_eventual eventual;
786
    int ret;
787 788 789 790 791 792 793

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

794
    *req = eventual;
795

Matthieu Dorier's avatar
Matthieu Dorier committed
796
    return HG_Forward(handle, margo_cb, (void*)eventual, in_struct);
797
}
798

799 800 801 802
hg_return_t margo_wait(margo_request req)
{
	hg_return_t* waited_hret;
	hg_return_t  hret;
803

804 805 806 807
    ABT_eventual_wait(req, (void**)&waited_hret);
	hret = *waited_hret;
    ABT_eventual_free(&req);
	
808
    return(hret);
809 810
}

Matthieu Dorier's avatar
Matthieu Dorier committed
811 812 813 814 815
int margo_test(margo_request req, int* flag)
{
    return ABT_eventual_test(req, NULL, flag);
}

816 817 818 819
typedef struct
{
    hg_handle_t handle;
} margo_forward_timeout_cb_dat;
820

821 822 823 824 825 826 827 828 829 830 831
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(
832
    hg_handle_t handle,
833 834
    void *in_struct,
    double timeout_ms)
835 836
{
    int ret;
837
    hg_return_t hret;
838
    margo_instance_id mid;
839
    ABT_eventual eventual;
840
    hg_return_t* waited_hret;
841 842
    margo_timer_t forward_timer;
    margo_forward_timeout_cb_dat timeout_cb_dat;
843 844 845 846 847 848 849

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

850 851 852
    /* use the handle to get the associated mid */
    mid = margo_hg_handle_get_instance(handle);

853 854 855 856 857
    /* 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);

Matthieu Dorier's avatar
Matthieu Dorier committed
858
    hret = HG_Forward(handle, margo_cb, (void*)eventual, in_struct);
859
    if(hret == HG_SUCCESS)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
860 861 862 863 864
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

865 866 867 868 869 870 871 872
    /* 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
873 874 875 876 877 878 879 880
    ABT_eventual_free(&eventual);

    return(hret);
}

hg_return_t margo_respond(
    hg_handle_t handle,
    void *out_struct)
881 882 883 884 885 886 887 888 889 890 891 892 893
{
    hg_return_t hret;
    margo_request req;
    hret = margo_irespond(handle,out_struct,&req);
    if(hret != HG_SUCCESS)
        return hret;
    return margo_wait(req);
}

hg_return_t margo_irespond(
    hg_handle_t handle,
    void *out_struct,
    margo_request* req)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
894 895 896 897
{
    ABT_eventual eventual;
    int ret;

Matthieu Dorier's avatar
Matthieu Dorier committed
898
    ret = ABT_eventual_create(sizeof(hg_return_t), &eventual);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
899 900 901 902 903
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);
    }

904
    *req = eventual;
905

Matthieu Dorier's avatar
Matthieu Dorier committed
906
    return HG_Respond(handle, margo_cb, (void*)eventual, out_struct);
907 908
}

909 910 911 912 913 914 915
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)
916
{
917 918 919
    return(HG_Bulk_create(mid->hg_class, count,
        buf_ptrs, buf_sizes, flags, handle));
}
920

921 922 923 924
hg_return_t margo_bulk_free(
    hg_bulk_t handle)
{
    return(HG_Bulk_free(handle));
925 926
}

927 928 929 930 931 932 933 934
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));
}
935

936
hg_return_t margo_bulk_transfer(
937
    margo_instance_id mid,
938
    hg_bulk_op_t op,
939
    hg_addr_t origin_addr,
940 941 942 943
    hg_bulk_t origin_handle,
    size_t origin_offset,
    hg_bulk_t local_handle,
    size_t local_offset,
944
    size_t size)
945 946 947 948
{  
    margo_request req;
    hg_return_t hret = margo_bulk_itransfer(mid,op,origin_addr,
                          origin_handle, origin_offset, local_handle,
Matthieu Dorier's avatar
Matthieu Dorier committed
949
                          local_offset, size, &req);
950 951 952 953 954 955 956 957 958 959 960 961 962 963 964
    if(hret != HG_SUCCESS)
        return hret;
    return margo_wait(req);
}

hg_return_t margo_bulk_itransfer(
    margo_instance_id mid,
    hg_bulk_op_t op,
    hg_addr_t origin_addr,
    hg_bulk_t origin_handle,
    size_t origin_offset,
    hg_bulk_t local_handle,
    size_t local_offset,
    size_t size,
    margo_request* req)
965 966 967 968 969 970 971 972 973 974 975
{
    hg_return_t hret = HG_TIMEOUT;
    ABT_eventual eventual;
    int ret;

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

976
    *req = eventual;
977

Matthieu Dorier's avatar
Matthieu Dorier committed
978 979
    hret = HG_Bulk_transfer(mid->hg_context, margo_cb,
        (void*)eventual, op, origin_addr, origin_handle, origin_offset, local_handle,
980
        local_offset, size, HG_OP_ID_IGNORE);
981 982 983 984

    return(hret);
}

985 986 987 988
typedef struct
{
    ABT_mutex mutex;
    ABT_cond cond;
989
    char is_asleep;
990 991 992 993 994 995 996 997 998
} 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);
999
    sleep_cb_dat->is_asleep = 0;
1000 1001 1002 1003 1004 1005 1006
    ABT_cond_signal(sleep_cb_dat->cond);
    ABT_mutex_unlock(sleep_cb_dat->mutex);

    return;
}

void margo_thread_sleep(
1007
    margo_instance_id mid,
1008 1009 1010 1011 1012 1013 1014 1015
    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));
1016
    sleep_cb_dat.is_asleep = 1;
1017 1018

    /* initialize the sleep timer */
1019
    margo_timer_init(mid, &sleep_timer, margo_thread_sleep_cb,
1020 1021 1022 1023
        &sleep_cb_dat, timeout_ms);

    /* yield thread for specified timeout */
    ABT_mutex_lock(sleep_cb_dat.mutex);
1024 1025
    while(sleep_cb_dat.is_asleep)
        ABT_cond_wait(sleep_cb_dat.cond, sleep_cb_dat.mutex);
1026 1027
    ABT_mutex_unlock(sleep_cb_dat.mutex);

1028 1029 1030 1031
    /* clean up */
    ABT_mutex_free(&sleep_cb_dat.mutex);
    ABT_cond_free(&sleep_cb_dat.cond);

1032 1033 1034
    return;
}

1035
int margo_get_handler_pool(margo_instance_id mid, ABT_pool* pool)
1036
{
1037 1038 1039 1040 1041 1042
    if(mid) {
        *pool = mid->handler_pool;
        return 0;
    } else {
        return -1;
    }
1043
}
1044

1045 1046 1047 1048
hg_context_t* margo_get_context(margo_instance_id mid)
{
    return(mid->hg_context);
}
1049

1050 1051 1052
hg_class_t* margo_get_class(margo_instance_id mid)
{
    return(mid->hg_class);
1053
}
Philip Carns's avatar
Philip Carns committed
1054

1055
margo_instance_id margo_hg_handle_get_instance(hg_handle_t h)
1056
{
1057 1058
	const struct hg_info* info = HG_Get_info(h);
	if(!info) return MARGO_INSTANCE_NULL;
1059 1060 1061 1062 1063
    return margo_hg_info_get_instance(info);
}

margo_instance_id margo_hg_info_get_instance(const struct hg_info *info)
{
1064 1065 1066 1067
	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;
1068 1069
}

1070
int margo_lookup_mplex(margo_instance_id mid, hg_id_t id, uint8_t mplex_id, ABT_pool *pool)
1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085
{
    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);
1086 1087 1088 1089 1090 1091
    if(!element) {
        if(mplex_id == 0) // element does not exist and mplex is 0, return default handler
            *pool = mid->handler_pool;
        else // otherwise it is an error
            return(-1);
    }
1092

Philip Carns's avatar
Philip Carns committed
1093 1094
    assert(element->key.id == id && element->key.mplex_id == mplex_id);

1095 1096 1097 1098 1099
    *pool = element->pool;

    return(0);
}

1100
int margo_register_data_mplex(margo_instance_id mid, hg_id_t id, uint8_t mplex_id, void* data, void (*free_callback)(void *))
1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123
{
    struct mplex_key key;
    struct mplex_element *element;

    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;

    assert(element->key.id == id && element->key.mplex_id == mplex_id);

    if(element->user_data && element->user_free_callback)
        (element->user_free_callback)(element->user_data);

    element->user_data = data;
    element->user_free_callback = free_callback;

    return(0);
}

1124
void* margo_registered_data_mplex(margo_instance_id mid, hg_id_t id, uint8_t mplex_id)
1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140
{
    struct mplex_key key;
    struct mplex_element *element;

    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 NULL;

    assert(element->key.id == id && element->key.mplex_id == mplex_id);

    return element->user_data;
}
1141
static void margo_rpc_data_free(void* ptr)
Philip Carns's avatar
Philip Carns committed
1142
{
1143 1144 1145 1146 1147 1148
	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);
}
1149

1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161
static void delete_multiplexing_hash(margo_instance_id mid)
{
    struct mplex_element *current_element, *tmp;

    HASH_ITER(hh, mid->mplex_table, current_element, tmp) {
        if(current_element->user_data && current_element->user_free_callback)
            (current_element->user_free_callback)(current_element->user_data);
        HASH_DEL(mid->mplex_table, current_element);
        free(current_element);
    }
}

1162 1163 1164 1165 1166 1167 1168
/* 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;
1169
    unsigned int hg_progress_timeout = mid->hg_progress_timeout_ub;
1170 1171
    double next_timer_exp;
    int trigger_happened;
1172 1173
    double tm1, tm2;
    int diag_enabled = 0;
1174

1175 1176 1177 1178
    while(!mid->hg_progress_shutdown_flag)
    {
        trigger_happened = 0;
        do {
1179 1180 1181 1182
            /* save value of instance diag variable, in case it is modified
             * while we are in loop 
             */
            diag_enabled = mid->diag_enabled;
1183

1184
            if(diag_enabled) tm1 = ABT_get_wtime();
1185
            ret = HG_Trigger(mid->hg_context, 0, 1, &actual_count);
1186 1187 1188 1189 1190
            if(diag_enabled)
            {
                tm2 = ABT_get_wtime();
                __DIAG_UPDATE(mid->diag_trigger_elapsed, (tm2-tm1));
            }
1191

1192 1193 1194
            if(ret == HG_SUCCESS && actual_count > 0)
                trigger_happened = 1;
        } while((ret == HG_SUCCESS) && actual_count && !mid->hg_progress_shutdown_flag);
1195

1196 1197
        if(trigger_happened)
            ABT_thread_yield();
1198

1199
        ABT_pool_get_size(mid->progress_pool, &size);
1200
        /* Are there any other threads executing in this pool that are *not*
1201 1202 1203 1204
         * 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.
1205
         */
1206
        if(size > 0)
1207 1208 1209 1210 1211 1212 1213 1214
        {
            /* 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.
             */
1215
            if(diag_enabled) tm1 = ABT_get_wtime();
1216
            ret = HG_Progress(mid->hg_context, 0);
1217 1218 1219 1220 1221 1222
            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);
            }
1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241
            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
        {
1242
            hg_progress_timeout = mid->hg_progress_timeout_ub;
1243 1244 1245 1246 1247 1248 1249 1250 1251
            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 */
1252
                    if(next_timer_exp < mid->hg_progress_timeout_ub)
1253 1254 1255 1256 1257 1258 1259
                        hg_progress_timeout = (unsigned int)next_timer_exp;
                }
                else
                {
                    hg_progress_timeout = 0;
                }
            }
1260
            if(diag_enabled) tm1 = ABT_get_wtime();
1261
            ret = HG_Progress(mid->hg_context, hg_progress_timeout);
1262 1263 1264 1265 1266 1267 1268 1269 1270 1271
            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);
            }
1272 1273 1274 1275 1276 1277
            if(ret != HG_SUCCESS && ret != HG_TIMEOUT)
            {
                /* TODO: error handling */
                fprintf(stderr, "WARNING: unexpected return code (%d) from HG_Progress()\n", ret);
            }
        }
1278

1279 1280 1281
        /* check for any expired timers */
        margo_check_timers(mid);
    }
1282

1283
    return;
Philip Carns's avatar
Philip Carns committed
1284
}
Philip Carns's avatar
Philip Carns committed
1285 1286 1287 1288 1289 1290 1291


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

1292
static void print_diag_data(FILE *file, const char* name, const char *description, struct diag_data *data)
Philip Carns's avatar
Philip Carns committed
1293
{
1294 1295 1296 1297 1298 1299 1300 1301
    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
1302 1303 1304
    return;
}

1305
void margo_diag_dump(margo_instance_id mid, const char* file, int uniquify)
Philip Carns's avatar
Philip Carns committed
1306 1307 1308
{
    FILE *outfile;
    time_t ltime;
1309
    char revised_file_name[256] = {0};
Philip Carns's avatar
Philip Carns committed
1310 1311 1312

    assert(mid->diag_enabled);

1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327
    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
1328 1329 1330 1331 1332 1333
    if(strcmp("-", file) == 0)
    {
        outfile = stdout;
    }
    else
    {
1334
        outfile = fopen(revised_file_name, "a");
Philip Carns's avatar
Philip Carns committed
1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350
        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");
1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362
    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
1363 1364 1365 1366 1367 1368

    if(outfile != stdout)
        fclose(outfile);
    
    return;
}
1369

1370
void margo_set_param(margo_instance_id mid, int option, const void *param)
1371 1372 1373
{
    switch(option)
    {
1374
        case MARGO_PARAM_PROGRESS_TIMEOUT_UB:
1375 1376 1377 1378 1379 1380 1381
            mid->hg_progress_timeout_ub = (*((const unsigned int*)param));
            break;
    }

    return;
}

1382
void margo_get_param(margo_instance_id mid, int option, void *param)
1383 1384 1385 1386
{

    switch(option)
    {
1387
        case MARGO_PARAM_PROGRESS_TIMEOUT_UB:
1388 1389 1390 1391 1392 1393
            (*((unsigned int*)param)) = mid->hg_progress_timeout_ub;
            break;
    }

    return;
}
Shane Snyder's avatar
Shane Snyder committed
1394 1395 1396 1397 1398 1399

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;
1400

1401
    ABT_mutex_create(&(mid->handle_cache_mtx));
Shane Snyder's avatar
Shane Snyder committed
1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440

    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);
    }

1441
    ABT_mutex_free(&mid->handle_cache_mtx);
1442

Shane Snyder's avatar
Shane Snyder committed
1443 1444 1445 1446 1447 1448 1449
    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;
1450 1451 1452
    hg_return_t hret = HG_SUCCESS;

    ABT_mutex_lock(mid->handle_cache_mtx);
Shane Snyder's avatar
Shane Snyder committed
1453 1454 1455 1456

    if(!mid->free_handle_list)
    {
        /* if no available handles, just fall through */
1457 1458
        hret = HG_OTHER_ERROR;
        goto finish;
Shane Snyder's avatar
Shane Snyder committed
1459 1460 1461 1462 1463 1464 1465 1466 1467 1468
    }

    /* 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)
    {
1469 1470 1471
        /* XXX: Mercury doesn't reset the target_id so we need to do that manually for now */
        HG_Set_target_id(el->handle, 0);

Shane Snyder's avatar
Shane Snyder committed
1472 1473 1474 1475
        /* 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;
    }
1476 1477 1478 1479 1480
    else
    {
        /* reset failed, add handle back to the free list */
        LL_APPEND(mid->free_handle_list, el);
    }
Shane Snyder's avatar
Shane Snyder committed
1481

1482 1483
finish:
    ABT_mutex_unlock(mid->handle_cache_mtx);
Shane Snyder's avatar
Shane Snyder committed
1484 1485 1486 1487 1488 1489 1490
    return hret;
}

static hg_return_t margo_handle_cache_put(margo_instance_id mid,
    hg_handle_t handle)
{
    struct margo_handle_cache_el *el;
1491 1492 1493
    hg_return_t hret = HG_SUCCESS;

    ABT_mutex_lock(mid->handle_cache_mtx);
Shane Snyder's avatar
Shane Snyder committed
1494 1495 1496 1497 1498 1499

    /* 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 */
1500 1501
        hret = HG_OTHER_ERROR;
        goto finish;
Shane Snyder's avatar
Shane Snyder committed
1502 1503 1504 1505 1506 1507 1508 1509
    }

    /* 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);

1510 1511 1512
finish:
    ABT_mutex_unlock(mid->handle_cache_mtx);
    return hret;
Shane Snyder's avatar
Shane Snyder committed
1513
}
1514 1515 1516 1517 1518

struct margo_timer_list *margo_get_timer_list(margo_instance_id mid)
{
        return mid->timer_list;
}
Matthieu Dorier's avatar
Matthieu Dorier committed
1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535

static void remote_shutdown_ult(hg_handle_t handle)
{
    margo_instance_id mid = margo_hg_handle_get_instance(handle);
    margo_shutdown_out_t out;
    if(!(mid->enable_remote_shutdown)) {
        out.ret = -1;
    } else {
        out.ret = 0;
    }
    margo_respond(handle, &out);
    margo_destroy(handle);
    if(mid->enable_remote_shutdown) {
        margo_finalize(mid);
    }
}
DEFINE_MARGO_RPC_HANDLER(remote_shutdown_ult)