margo.c 32.6 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>
Matthieu Dorier's avatar
Matthieu Dorier committed
12 13 14

#include <margo-config.h>
#ifdef HAVE_ABT_SNOOZER
15
#include <abt-snoozer.h>
Matthieu Dorier's avatar
Matthieu Dorier committed
16
#endif
17
#include <time.h>
Philip Carns's avatar
Philip Carns committed
18
#include <math.h>
19 20

#include "margo.h"
21
#include "margo-timer.h"
Philip Carns's avatar
Philip Carns committed
22
#include "utlist.h"
23
#include "uthash.h"
24

25
#define DEFAULT_MERCURY_PROGRESS_TIMEOUT_UB 100 /* 100 milliseconds */
Shane Snyder's avatar
Shane Snyder committed
26
#define DEFAULT_MERCURY_HANDLE_CACHE_SIZE 32
27

28 29 30 31 32 33 34 35 36 37 38 39 40
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
41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56
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
57 58 59 60 61 62 63
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 */
};

64 65
struct margo_instance
{
Shane Snyder's avatar
Shane Snyder committed
66
    /* mercury/argobots state */
67 68
    hg_context_t *hg_context;
    hg_class_t *hg_class;
69 70 71
    ABT_pool handler_pool;
    ABT_pool progress_pool;

72
    /* internal to margo for this particular instance */
73
    int margo_init;
74 75
    ABT_thread hg_progress_tid;
    int hg_progress_shutdown_flag;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
76
    ABT_xstream progress_xstream;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
77 78 79
    int owns_progress_pool;
    ABT_xstream *rpc_xstreams;
    int num_handler_pool_threads;
80
    unsigned int hg_progress_timeout_ub;
81 82 83

    /* control logic for callers waiting on margo to be finalized */
    int finalize_flag;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
84
    int refcount;
85 86 87
    ABT_mutex finalize_mutex;
    ABT_cond finalize_cond;

88 89
    /* hash table to track multiplexed rpcs registered with margo */
    struct mplex_element *mplex_table;
Philip Carns's avatar
Philip Carns committed
90

Shane Snyder's avatar
Shane Snyder committed
91 92 93 94
    /* 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
95 96 97 98 99 100 101 102 103 104 105
    /* 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;
106 107
};

108 109 110 111 112
struct margo_cb_arg
{
    ABT_eventual *eventual;
};

113 114 115 116 117 118
struct margo_rpc_data
{
	margo_instance_id mid;
	void* user_data;
	void (*user_free_callback)(void *);
};
119

120
static void hg_progress_fn(void* foo);
121
static void margo_rpc_data_free(void* ptr);
122

Shane Snyder's avatar
Shane Snyder committed
123 124 125 126 127 128 129
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);

130
margo_instance_id margo_init(const char *addr_str, int mode,
131
    int use_progress_thread, int rpc_thread_count)
132
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
133 134 135 136 137
    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;
138 139
    hg_class_t *hg_class = NULL;
    hg_context_t *hg_context = NULL;
140
    int listen_flag = (mode == MARGO_CLIENT_MODE) ? HG_FALSE : HG_TRUE;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
141
    int i;
142 143 144
    int ret;
    struct margo_instance *mid = MARGO_INSTANCE_NULL;

Shane Snyder's avatar
Shane Snyder committed
145
    if(mode != MARGO_CLIENT_MODE && mode != MARGO_SERVER_MODE) goto err;
146

147 148 149
    ret = ABT_init(0, NULL); /* XXX: argc/argv not currently used by ABT ... */
    if(ret != 0) goto err;

150
    /* set caller (self) ES to idle without polling */
Matthieu Dorier's avatar
Matthieu Dorier committed
151
#ifdef HAVE_ABT_SNOOZER
152 153
    ret = ABT_snoozer_xstream_self_set();
    if(ret != 0) goto err;
Matthieu Dorier's avatar
Matthieu Dorier committed
154
#endif
Jonathan Jenkins's avatar
Jonathan Jenkins committed
155 156 157

    if (use_progress_thread)
    {
Matthieu Dorier's avatar
Matthieu Dorier committed
158
#ifdef HAVE_ABT_SNOOZER
Jonathan Jenkins's avatar
Jonathan Jenkins committed
159
        ret = ABT_snoozer_xstream_create(1, &progress_pool, &progress_xstream);
Matthieu Dorier's avatar
Matthieu Dorier committed
160 161 162 163 164 165 166
		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
Jonathan Jenkins's avatar
Jonathan Jenkins committed
167 168 169 170 171 172 173 174 175
    }
    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;
    }

176
    if (mode == MARGO_SERVER_MODE)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
177
    {
178 179
        if (rpc_thread_count > 0)
        {
Matthieu Dorier's avatar
Matthieu Dorier committed
180
            rpc_xstreams = calloc(rpc_thread_count, sizeof(*rpc_xstreams));
181
            if (rpc_xstreams == NULL) goto err;
Matthieu Dorier's avatar
Matthieu Dorier committed
182
#ifdef HAVE_ABT_SNOOZER
183 184 185
            ret = ABT_snoozer_xstream_create(rpc_thread_count, &rpc_pool,
                    rpc_xstreams);
            if (ret != ABT_SUCCESS) goto err;
Matthieu Dorier's avatar
Matthieu Dorier committed
186 187 188 189 190 191 192 193 194
#else
			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);
				if (ret != ABT_SUCCESS) goto err;
			}
#endif
195 196 197 198 199 200 201 202 203 204 205 206
        }
        else if (rpc_thread_count == 0)
        {
            ret = ABT_xstream_self(&rpc_xstream);
            if (ret != ABT_SUCCESS) goto err;
            ret = ABT_xstream_get_main_pools(rpc_xstream, 1, &rpc_pool);
            if (ret != ABT_SUCCESS) goto err;
        }
        else
        {
            rpc_pool = progress_pool;
        }
Jonathan Jenkins's avatar
Jonathan Jenkins committed
207 208
    }

209 210 211 212 213 214
    hg_class = HG_Init(addr_str, listen_flag);
    if(!hg_class) goto err;

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

Jonathan Jenkins's avatar
Jonathan Jenkins committed
215 216 217
    mid = margo_init_pool(progress_pool, rpc_pool, hg_context);
    if (mid == MARGO_INSTANCE_NULL) goto err;

218
    mid->margo_init = 1;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
219 220 221
    mid->owns_progress_pool = use_progress_thread;
    mid->progress_xstream = progress_xstream;
    mid->num_handler_pool_threads = rpc_thread_count < 0 ? 0 : rpc_thread_count;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
222
    mid->rpc_xstreams = rpc_xstreams;
223

Jonathan Jenkins's avatar
Jonathan Jenkins committed
224 225 226
    return mid;

err:
227 228 229 230 231 232 233
    if(mid)
    {
        margo_timer_instance_finalize(mid);
        ABT_mutex_free(&mid->finalize_mutex);
        ABT_cond_free(&mid->finalize_cond);
        free(mid);
    }
Jonathan Jenkins's avatar
Jonathan Jenkins committed
234 235 236 237 238 239 240 241 242 243 244 245 246 247
    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);
    }
248 249 250 251 252
    if(hg_context)
        HG_Context_destroy(hg_context);
    if(hg_class)
        HG_Finalize(hg_class);
    ABT_finalize();
Jonathan Jenkins's avatar
Jonathan Jenkins committed
253 254 255 256
    return MARGO_INSTANCE_NULL;
}

margo_instance_id margo_init_pool(ABT_pool progress_pool, ABT_pool handler_pool,
Jonathan Jenkins's avatar
Jonathan Jenkins committed
257
    hg_context_t *hg_context)
258 259
{
    int ret;
Shane Snyder's avatar
Shane Snyder committed
260
    hg_return_t hret;
261 262 263
    struct margo_instance *mid;

    mid = malloc(sizeof(*mid));
264
    if(!mid) goto err;
265
    memset(mid, 0, sizeof(*mid));
266

267 268 269
    ABT_mutex_create(&mid->finalize_mutex);
    ABT_cond_create(&mid->finalize_cond);

270 271
    mid->progress_pool = progress_pool;
    mid->handler_pool = handler_pool;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
272
    mid->hg_class = HG_Context_get_class(hg_context);
273
    mid->hg_context = hg_context;
274
    mid->hg_progress_timeout_ub = DEFAULT_MERCURY_PROGRESS_TIMEOUT_UB;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
275
    mid->refcount = 1;
276

277
    ret = margo_timer_instance_init(mid);
278
    if(ret != 0) goto err;
279

Shane Snyder's avatar
Shane Snyder committed
280 281 282 283
    /* initialize the handle cache */
    hret = margo_handle_cache_init(mid);
    if(hret != HG_SUCCESS) goto err;

284
    ret = ABT_thread_create(mid->progress_pool, hg_progress_fn, mid, 
285
        ABT_THREAD_ATTR_NULL, &mid->hg_progress_tid);
286 287
    if(ret != 0) goto err;

288 289
    return mid;

290 291
err:
    if(mid)
292
    {
Shane Snyder's avatar
Shane Snyder committed
293
        margo_handle_cache_destroy(mid);
294 295 296
        margo_timer_instance_finalize(mid);
        ABT_mutex_free(&mid->finalize_mutex);
        ABT_cond_free(&mid->finalize_cond);
297
        free(mid);
298
    }
299
    return MARGO_INSTANCE_NULL;
300 301
}

Jonathan Jenkins's avatar
Jonathan Jenkins committed
302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326
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
327 328
    margo_handle_cache_destroy(mid);

329 330 331 332 333 334 335 336 337
    if (mid->margo_init)
    {
        if (mid->hg_context)
            HG_Context_destroy(mid->hg_context);
        if (mid->hg_class)
            HG_Finalize(mid->hg_class);
        ABT_finalize();
    }

Jonathan Jenkins's avatar
Jonathan Jenkins committed
338 339 340
    free(mid);
}

341
void margo_finalize(margo_instance_id mid)
342
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
343
    int do_cleanup;
344

345
    /* tell progress thread to wrap things up */
346
    mid->hg_progress_shutdown_flag = 1;
347 348

    /* wait for it to shutdown cleanly */
349 350
    ABT_thread_join(mid->hg_progress_tid);
    ABT_thread_free(&mid->hg_progress_tid);
351

352 353 354 355
    ABT_mutex_lock(mid->finalize_mutex);
    mid->finalize_flag = 1;
    ABT_cond_broadcast(mid->finalize_cond);

Jonathan Jenkins's avatar
Jonathan Jenkins committed
356 357
    mid->refcount--;
    do_cleanup = mid->refcount == 0;
358

Jonathan Jenkins's avatar
Jonathan Jenkins committed
359 360 361 362 363 364 365
    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);
366 367 368 369 370 371

    return;
}

void margo_wait_for_finalize(margo_instance_id mid)
{
Jonathan Jenkins's avatar
Jonathan Jenkins committed
372
    int do_cleanup;
373 374 375

    ABT_mutex_lock(mid->finalize_mutex);

Jonathan Jenkins's avatar
Jonathan Jenkins committed
376
        mid->refcount++;
377 378 379 380
            
        while(!mid->finalize_flag)
            ABT_cond_wait(mid->finalize_cond, mid->finalize_mutex);

Jonathan Jenkins's avatar
Jonathan Jenkins committed
381 382 383
        mid->refcount--;
        do_cleanup = mid->refcount == 0;

384
    ABT_mutex_unlock(mid->finalize_mutex);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
385 386 387 388

    if (do_cleanup)
        margo_cleanup(mid);

389 390 391
    return;
}

392 393
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)
394
{
395 396 397
	struct margo_rpc_data* margo_data;
    hg_return_t hret;
    hg_id_t id;
398

399 400 401
    id = HG_Register_name(mid->hg_class, func_name, in_proc_cb, out_proc_cb, rpc_cb);
    if(id <= 0)
        return(0);
402

403 404 405 406 407 408 409 410 411 412 413 414
	/* 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)
415
        {
416 417
            free(margo_data);
            return(0);
418
        }
419 420
    }

421
	return(id);
422 423
}

424 425 426
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)
427
{
428 429 430
    struct mplex_key key;
    struct mplex_element *element;
    hg_id_t id;
431

432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456
    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);
457 458
}

459 460
hg_return_t margo_registered_name(margo_instance_id mid, const char *func_name,
    hg_id_t *id, hg_bool_t *flag)
461
{
462
    return(HG_Registered_name(mid->hg_class, func_name, id, flag));
463 464
}

465 466 467 468 469 470 471
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 
472
		= (struct margo_rpc_data*) HG_Registered_data(mid->hg_class, id);
473 474 475 476 477 478 479 480 481 482 483 484 485 486
	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;
}

487 488 489 490
hg_return_t margo_registered_disable_response(
    margo_instance_id mid,
    hg_id_t id,
    int disable_flag)
491
{
492
    return(HG_Registered_disable_response(mid->hg_class, id, disable_flag));
493
}
494

495
struct lookup_cb_evt
496
{
497
    hg_return_t hret;
498 499 500 501 502 503
    hg_addr_t addr;
};

static hg_return_t margo_addr_lookup_cb(const struct hg_cb_info *info)
{
    struct lookup_cb_evt evt;
504
    evt.hret = info->ret;
505
    evt.addr = info->info.lookup.addr;
506
    struct margo_cb_arg* arg = info->arg;
507 508

    /* propagate return code out through eventual */
509 510
    ABT_eventual_set(*(arg->eventual), &evt, sizeof(evt));

511 512 513
    return(HG_SUCCESS);
}

514 515 516 517
hg_return_t margo_addr_lookup(
    margo_instance_id mid,
    const char   *name,
    hg_addr_t    *addr)
518
{
519
    hg_return_t hret;
520 521 522 523
    struct lookup_cb_evt *evt;
    ABT_eventual eventual;
    int ret;
    struct margo_cb_arg arg;
524

525 526 527 528 529 530 531
    ret = ABT_eventual_create(sizeof(*evt), &eventual);
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);        
    }

    arg.eventual = &eventual;
532

533
    hret = HG_Addr_lookup(mid->hg_context, margo_addr_lookup_cb,
534
        &arg, name, HG_OP_ID_IGNORE);
535
    if(hret == HG_SUCCESS)
536 537 538
    {
        ABT_eventual_wait(eventual, (void**)&evt);
        *addr = evt->addr;
539
        hret = evt->hret;
540 541 542 543
    }

    ABT_eventual_free(&eventual);

544
    return(hret);
545 546 547 548 549
}

hg_return_t margo_addr_free(
    margo_instance_id mid,
    hg_addr_t addr)
550
{
551 552
    return(HG_Addr_free(mid->hg_class, addr));
}
553

554 555 556 557 558
hg_return_t margo_addr_self(
    margo_instance_id mid,
    hg_addr_t *addr)
{
    return(HG_Addr_self(mid->hg_class, addr));
559 560
}

561 562 563 564 565 566 567 568 569
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(
570
    margo_instance_id mid,
571 572 573 574 575 576 577 578 579 580
    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
581 582 583 584 585 586 587 588 589
    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);
    }
590

Shane Snyder's avatar
Shane Snyder committed
591
    return hret;
592 593
}

594
hg_return_t margo_destroy(hg_handle_t handle)
595
{
596
    margo_instance_id mid;
Shane Snyder's avatar
Shane Snyder committed
597 598
    hg_return_t hret;

599 600 601
    /* use the handle to get the associated mid */
    mid = margo_hg_handle_get_instance(handle);

Shane Snyder's avatar
Shane Snyder committed
602 603 604 605 606 607 608
    /* 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);
    }
609

Shane Snyder's avatar
Shane Snyder committed
610
    return hret;
611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628
}

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(
    hg_handle_t handle,
    void *in_struct)
{
    hg_return_t hret = HG_TIMEOUT;
629
    ABT_eventual eventual;
630
    int ret;
631
    hg_return_t* waited_hret;
632
    struct margo_cb_arg arg;
633 634 635 636 637 638 639

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

640
    arg.eventual = &eventual;
641

642
    hret = HG_Forward(handle, margo_cb, &arg, in_struct);
643
    if(hret == HG_SUCCESS)
644 645 646 647 648 649 650 651
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
652 653
}

654 655 656 657
typedef struct
{
    hg_handle_t handle;
} margo_forward_timeout_cb_dat;
658

659 660 661 662 663 664 665 666 667 668 669
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(
670
    hg_handle_t handle,
671 672
    void *in_struct,
    double timeout_ms)
673 674
{
    int ret;
675
    hg_return_t hret;
676
    margo_instance_id mid;
677
    ABT_eventual eventual;
678
    hg_return_t* waited_hret;
679 680
    margo_timer_t forward_timer;
    margo_forward_timeout_cb_dat timeout_cb_dat;
681
    struct margo_cb_arg arg;
682 683 684 685 686 687 688

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

689 690 691
    /* use the handle to get the associated mid */
    mid = margo_hg_handle_get_instance(handle);

692 693 694 695 696
    /* 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);

697
    arg.eventual = &eventual;
698

699
    hret = HG_Forward(handle, margo_cb, &arg, in_struct);
700
    if(hret == HG_SUCCESS)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
701 702 703 704 705
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

706 707 708 709 710 711 712 713
    /* 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
714 715 716 717 718 719 720 721 722 723 724 725 726
    ABT_eventual_free(&eventual);

    return(hret);
}

hg_return_t margo_respond(
    hg_handle_t handle,
    void *out_struct)
{
    hg_return_t hret = HG_TIMEOUT;
    ABT_eventual eventual;
    int ret;
    hg_return_t* waited_hret;
727
    struct margo_cb_arg arg;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
728 729 730 731 732 733 734

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

735
    arg.eventual = &eventual;
736

737
    hret = HG_Respond(handle, margo_cb, &arg, out_struct);
738
    if(hret == HG_SUCCESS)
739 740 741 742 743 744 745 746 747 748
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

749 750 751 752 753 754 755
hg_return_t margo_bulk_create(
    margo_instance_id mid,
    hg_uint32_t count,
    void **buf_ptrs,
    const hg_size_t *buf_sizes,
    hg_uint8_t flags,
    hg_bulk_t *handle)
756
{
757 758 759
    return(HG_Bulk_create(mid->hg_class, count,
        buf_ptrs, buf_sizes, flags, handle));
}
760

761 762 763 764
hg_return_t margo_bulk_free(
    hg_bulk_t handle)
{
    return(HG_Bulk_free(handle));
765 766
}

767 768 769 770 771 772 773 774
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));
}
775

776
/* TODO: currently identical to a vanilla margo_cb -- consider reusing that */
777 778 779 780
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;
781

782

783 784 785 786
    /* propagate return code out through eventual */
    ABT_eventual_set(*(arg->eventual), &hret, sizeof(hret));
    
    return(HG_SUCCESS);
787 788
}

789
hg_return_t margo_bulk_transfer(
790
    margo_instance_id mid,
791
    hg_bulk_op_t op,
792
    hg_addr_t origin_addr,
793 794 795 796
    hg_bulk_t origin_handle,
    size_t origin_offset,
    hg_bulk_t local_handle,
    size_t local_offset,
797
    size_t size)
798 799 800 801 802
{
    hg_return_t hret = HG_TIMEOUT;
    hg_return_t *waited_hret;
    ABT_eventual eventual;
    int ret;
803
    struct margo_cb_arg arg;
804 805 806 807 808 809 810

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

811
    arg.eventual = &eventual;
812

Jonathan Jenkins's avatar
Jonathan Jenkins committed
813
    hret = HG_Bulk_transfer(mid->hg_context, margo_bulk_transfer_cb,
814
        &arg, op, origin_addr, origin_handle, origin_offset, local_handle,
815
        local_offset, size, HG_OP_ID_IGNORE);
816
    if(hret == HG_SUCCESS)
817 818 819 820 821 822 823 824 825 826
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

827 828 829 830
typedef struct
{
    ABT_mutex mutex;
    ABT_cond cond;
Shane Snyder's avatar
Shane Snyder committed
831
    char is_asleep;
832 833 834 835 836 837 838 839 840
} 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);
841
    sleep_cb_dat->is_asleep = 0;
842 843 844 845 846 847 848
    ABT_cond_signal(sleep_cb_dat->cond);
    ABT_mutex_unlock(sleep_cb_dat->mutex);

    return;
}

void margo_thread_sleep(
849
    margo_instance_id mid,
850 851 852 853 854 855 856 857
    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));
858
    sleep_cb_dat.is_asleep = 1;
859 860

    /* initialize the sleep timer */
861
    margo_timer_init(mid, &sleep_timer, margo_thread_sleep_cb,
862 863 864 865
        &sleep_cb_dat, timeout_ms);

    /* yield thread for specified timeout */
    ABT_mutex_lock(sleep_cb_dat.mutex);
866 867
    while(sleep_cb_dat.is_asleep)
        ABT_cond_wait(sleep_cb_dat.cond, sleep_cb_dat.mutex);
868 869
    ABT_mutex_unlock(sleep_cb_dat.mutex);

870 871 872 873
    /* clean up */
    ABT_mutex_free(&sleep_cb_dat.mutex);
    ABT_cond_free(&sleep_cb_dat.cond);

874 875 876
    return;
}

877
ABT_pool* margo_get_handler_pool(margo_instance_id mid)
878
{
879 880
    return(&mid->handler_pool);
}
881

882 883 884 885
hg_context_t* margo_get_context(margo_instance_id mid)
{
    return(mid->hg_context);
}
886

887 888 889
hg_class_t* margo_get_class(margo_instance_id mid)
{
    return(mid->hg_class);
890
}
Philip Carns's avatar
Philip Carns committed
891

892
margo_instance_id margo_hg_handle_get_instance(hg_handle_t h)
893
{
894 895
	const struct hg_info* info = HG_Get_info(h);
	if(!info) return MARGO_INSTANCE_NULL;
896 897 898 899 900
    return margo_hg_info_get_instance(info);
}

margo_instance_id margo_hg_info_get_instance(const struct hg_info *info)
{
901 902 903 904
	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;
905 906
}

907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925
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
926 927
    assert(element->key.id == id && element->key.mplex_id == mplex_id);

928 929 930 931 932
    *pool = element->pool;

    return(0);
}

933
static void margo_rpc_data_free(void* ptr)
Philip Carns's avatar
Philip Carns committed
934
{
935 936 937 938 939 940
	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);
}
941

942 943 944 945 946 947 948
/* 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;
949
    unsigned int hg_progress_timeout = mid->hg_progress_timeout_ub;
950 951
    double next_timer_exp;
    int trigger_happened;
952 953
    double tm1, tm2;
    int diag_enabled = 0;
954

955 956 957 958
    while(!mid->hg_progress_shutdown_flag)
    {
        trigger_happened = 0;
        do {
959 960 961 962
            /* save value of instance diag variable, in case it is modified
             * while we are in loop 
             */
            diag_enabled = mid->diag_enabled;
963

964
            if(diag_enabled) tm1 = ABT_get_wtime();
965
            ret = HG_Trigger(mid->hg_context, 0, 1, &actual_count);
966 967 968 969 970
            if(diag_enabled)
            {
                tm2 = ABT_get_wtime();
                __DIAG_UPDATE(mid->diag_trigger_elapsed, (tm2-tm1));
            }
971

972 973 974
            if(ret == HG_SUCCESS && actual_count > 0)
                trigger_happened = 1;
        } while((ret == HG_SUCCESS) && actual_count && !mid->hg_progress_shutdown_flag);
975

976 977
        if(trigger_happened)
            ABT_thread_yield();
978

979
        ABT_pool_get_size(mid->progress_pool, &size);
980
        /* Are there any other threads executing in this pool that are *not*
981 982 983 984
         * 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.
985
         */
986
        if(size > 0)
987 988 989 990 991 992 993 994
        {
            /* 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.
             */
995
            if(diag_enabled) tm1 = ABT_get_wtime();
996
            ret = HG_Progress(mid->hg_context, 0);
997 998 999 1000 1001 1002
            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);
            }
1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021
            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
        {
1022
            hg_progress_timeout = mid->hg_progress_timeout_ub;
1023 1024 1025 1026 1027 1028 1029 1030 1031
            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 */
1032
                    if(next_timer_exp < mid->hg_progress_timeout_ub)
1033 1034 1035 1036 1037 1038 1039
                        hg_progress_timeout = (unsigned int)next_timer_exp;
                }
                else
                {
                    hg_progress_timeout = 0;
                }
            }
1040
            if(diag_enabled) tm1 = ABT_get_wtime();
1041
            ret = HG_Progress(mid->hg_context, hg_progress_timeout);
1042 1043 1044 1045 1046 1047 1048 1049 1050 1051
            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);
            }
1052 1053 1054 1055 1056 1057
            if(ret != HG_SUCCESS && ret != HG_TIMEOUT)
            {
                /* TODO: error handling */
                fprintf(stderr, "WARNING: unexpected return code (%d) from HG_Progress()\n", ret);
            }
        }
1058

1059 1060 1061
        /* check for any expired timers */
        margo_check_timers(mid);
    }
1062

1063
    return;
Philip Carns's avatar
Philip Carns committed
1064
}
Philip Carns's avatar
Philip Carns committed
1065 1066 1067 1068 1069 1070 1071


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

1072
static void print_diag_data(FILE *file, const char* name, const char *description, struct diag_data *data)
Philip Carns's avatar
Philip Carns committed
1073
{
1074 1075 1076 1077 1078 1079 1080 1081
    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
1082 1083 1084
    return;
}

1085
void margo_diag_dump(margo_instance_id mid, const char* file, int uniquify)
Philip Carns's avatar
Philip Carns committed
1086 1087 1088
{
    FILE *outfile;
    time_t ltime;
1089
    char revised_file_name[256] = {0};
Philip Carns's avatar
Philip Carns committed
1090 1091 1092

    assert(mid->diag_enabled);

1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107
    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
1108 1109 1110 1111 1112 1113
    if(strcmp("-", file) == 0)
    {
        outfile = stdout;
    }
    else
    {
1114
        outfile = fopen(revised_file_name, "a");
Philip Carns's avatar
Philip Carns committed
1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130
        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");
1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142
    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
1143 1144 1145 1146 1147 1148

    if(outfile != stdout)
        fclose(outfile);
    
    return;
}
1149

1150
void margo_set_param(margo_instance_id mid, int option, const void *param)
1151 1152 1153
{
    switch(option)
    {
1154
        case MARGO_PARAM_PROGRESS_TIMEOUT_UB:
1155 1156 1157 1158 1159 1160 1161
            mid->hg_progress_timeout_ub = (*((const unsigned int*)param));
            break;
    }

    return;
}

1162
void margo_get_param(margo_instance_id mid, int option, void *param)
1163 1164 1165 1166
{

    switch(option)
    {
1167
        case MARGO_PARAM_PROGRESS_TIMEOUT_UB:
1168 1169 1170 1171 1172 1173
            (*((unsigned int*)param)) = mid->hg_progress_timeout_ub;
            break;
    }

    return;
}
Shane Snyder's avatar
Shane Snyder committed
1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245

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;
    }
1246 1247 1248 1249 1250
    else
    {
        /* reset failed, add handle back to the free list */
        LL_APPEND(mid->free_handle_list, el);
    }
Shane Snyder's avatar
Shane Snyder committed
1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275

    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;
}