margo.c 42 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
struct provider_element
30 31 32
{
    hg_id_t id;
    ABT_pool pool;
33 34
    void* user_data;
    void(*user_free_callback)(void*);
35 36 37
    UT_hash_handle hh;
};

Philip Carns's avatar
Philip Carns committed
38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53
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
54 55 56 57 58 59 60
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 */
};

61 62 63 64 65 66 67
struct margo_finalize_cb
{
    void(*callback)(void*);
    void* uargs;
    struct margo_finalize_cb* next;
};

68 69
struct margo_timer_list; /* defined in margo-timer.c */

70 71
struct margo_instance
{
Shane Snyder's avatar
Shane Snyder committed
72
    /* mercury/argobots state */
73 74
    hg_context_t *hg_context;
    hg_class_t *hg_class;
75 76 77
    ABT_pool handler_pool;
    ABT_pool progress_pool;

78
    /* internal to margo for this particular instance */
79
    int margo_init;
80
    int abt_init;
81 82
    ABT_thread hg_progress_tid;
    int hg_progress_shutdown_flag;
83
    ABT_xstream progress_xstream;
84 85 86
    int owns_progress_pool;
    ABT_xstream *rpc_xstreams;
    int num_handler_pool_threads;
87
    unsigned int hg_progress_timeout_ub;
88 89 90

    /* control logic for callers waiting on margo to be finalized */
    int finalize_flag;
91
    int refcount;
92 93
    ABT_mutex finalize_mutex;
    ABT_cond finalize_cond;
94
    struct margo_finalize_cb* finalize_cb;
95

Matthieu Dorier's avatar
Matthieu Dorier committed
96 97 98 99
    /* control logic for shutting down */
    hg_id_t shutdown_rpc_id;
    int enable_remote_shutdown;

100 101 102
    /* timer data */
    struct margo_timer_list* timer_list;

103 104
    /* hash table to track provider IDs registered with margo */
    struct provider_element *provider_table;
Philip Carns's avatar
Philip Carns committed
105

Shane Snyder's avatar
Shane Snyder committed
106 107 108
    /* 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;
109
    ABT_mutex handle_cache_mtx; /* mutex protecting access to above caches */
Shane Snyder's avatar
Shane Snyder committed
110

Philip Carns's avatar
Philip Carns committed
111 112 113 114 115 116 117 118 119 120 121
    /* 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;
122 123
};

124 125 126 127 128 129
struct margo_rpc_data
{
	margo_instance_id mid;
	void* user_data;
	void (*user_free_callback)(void *);
};
130

Matthieu Dorier's avatar
Matthieu Dorier committed
131 132
MERCURY_GEN_PROC(margo_shutdown_out_t, ((int32_t)(ret)))

133
static void hg_progress_fn(void* foo);
134
static void margo_rpc_data_free(void* ptr);
Matthieu Dorier's avatar
Matthieu Dorier committed
135 136
static void remote_shutdown_ult(hg_handle_t handle);
DECLARE_MARGO_RPC_HANDLER(remote_shutdown_ult);
137

138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173
static inline void demux_id(hg_id_t in, hg_id_t* base_id, uint16_t *provider_id)
{
    /* retrieve low bits for provider */
    *provider_id = 0;
    *provider_id += (in & (((1<<(__MARGO_PROVIDER_ID_SIZE*8))-1)));

    /* clear low order bits */
    *base_id = (in >> (__MARGO_PROVIDER_ID_SIZE*8)) <<
        (__MARGO_PROVIDER_ID_SIZE*8);

    return;
}

static inline hg_id_t mux_id(hg_id_t base_id, uint16_t provider_id)
{
    hg_id_t id;

    id = (base_id >> (__MARGO_PROVIDER_ID_SIZE*8)) <<
       (__MARGO_PROVIDER_ID_SIZE*8);
    id |= provider_id;

    return id;
}

static inline hg_id_t gen_id(const char* func_name, uint16_t provider_id)
{
    hg_id_t id;
    unsigned hashval;

    HASH_JEN(func_name, strlen(func_name), hashval);
    id = hashval << (__MARGO_PROVIDER_ID_SIZE*8);
    id |= provider_id;

    return id;
}

Shane Snyder's avatar
Shane Snyder committed
174 175 176 177 178 179
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);
180 181 182 183
static void delete_provider_hash(margo_instance_id mid);
static int margo_lookup_provider(margo_instance_id mid, hg_id_t id, uint16_t provider_id, ABT_pool *pool);
static hg_id_t margo_register_internal(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, uint16_t provider_id);
Shane Snyder's avatar
Shane Snyder committed
184

185
margo_instance_id margo_init(const char *addr_str, int mode,
186
    int use_progress_thread, int rpc_thread_count)
187
{
188 189 190 191 192
    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;
193 194
    hg_class_t *hg_class = NULL;
    hg_context_t *hg_context = NULL;
195
    int listen_flag = (mode == MARGO_CLIENT_MODE) ? HG_FALSE : HG_TRUE;
196
    int abt_init = 0;
197
    int i;
198 199 200
    int ret;
    struct margo_instance *mid = MARGO_INSTANCE_NULL;

201
    if(mode != MARGO_CLIENT_MODE && mode != MARGO_SERVER_MODE) goto err;
202

203 204 205 206 207 208 209 210 211 212 213 214
    /* 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");

215 216 217 218 219 220
    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;
    }
221

222
    /* set caller (self) ES to idle without polling */
223
#ifdef HAVE_ABT_SNOOZER
224 225
    ret = ABT_snoozer_xstream_self_set();
    if(ret != 0) goto err;
226
#endif
227 228 229

    if (use_progress_thread)
    {
230
#ifdef HAVE_ABT_SNOOZER
231
        ret = ABT_snoozer_xstream_create(1, &progress_pool, &progress_xstream);
232 233 234 235 236 237 238
		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
239 240 241 242 243 244 245 246 247
    }
    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;
    }

248
    if (rpc_thread_count > 0)
249
    {
250 251
        rpc_xstreams = calloc(rpc_thread_count, sizeof(*rpc_xstreams));
        if (rpc_xstreams == NULL) goto err;
252
#ifdef HAVE_ABT_SNOOZER
253 254 255
        ret = ABT_snoozer_xstream_create(rpc_thread_count, &rpc_pool,
                rpc_xstreams);
        if (ret != ABT_SUCCESS) goto err;
256
#else
257 258 259 260 261
        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);
262 263
            if (ret != ABT_SUCCESS) goto err;
        }
264 265 266 267 268 269 270 271 272 273 274 275
#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;
276 277
    }

278 279 280 281 282 283
    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;

284 285 286
    mid = margo_init_pool(progress_pool, rpc_pool, hg_context);
    if (mid == MARGO_INSTANCE_NULL) goto err;

287
    mid->margo_init = 1;
288
    mid->abt_init = abt_init;
289 290 291
    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;
292
    mid->rpc_xstreams = rpc_xstreams;
293

294 295 296
    return mid;

err:
297 298
    if(mid)
    {
299
        margo_timer_list_free(mid->timer_list);
300 301 302 303
        ABT_mutex_free(&mid->finalize_mutex);
        ABT_cond_free(&mid->finalize_cond);
        free(mid);
    }
304 305 306 307 308 309 310 311 312 313 314 315 316 317
    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);
    }
318 319 320 321
    if(hg_context)
        HG_Context_destroy(hg_context);
    if(hg_class)
        HG_Finalize(hg_class);
322 323
    if(abt_init)
        ABT_finalize();
324 325 326 327
    return MARGO_INSTANCE_NULL;
}

margo_instance_id margo_init_pool(ABT_pool progress_pool, ABT_pool handler_pool,
328
    hg_context_t *hg_context)
329 330
{
    int ret;
Shane Snyder's avatar
Shane Snyder committed
331
    hg_return_t hret;
332 333
    struct margo_instance *mid;

Matthieu Dorier's avatar
Matthieu Dorier committed
334
    mid = calloc(1,sizeof(*mid));
335
    if(!mid) goto err;
336
    memset(mid, 0, sizeof(*mid));
337

338 339 340
    ABT_mutex_create(&mid->finalize_mutex);
    ABT_cond_create(&mid->finalize_cond);

341 342
    mid->progress_pool = progress_pool;
    mid->handler_pool = handler_pool;
343
    mid->hg_class = HG_Context_get_class(hg_context);
344
    mid->hg_context = hg_context;
345
    mid->hg_progress_timeout_ub = DEFAULT_MERCURY_PROGRESS_TIMEOUT_UB;
346
    mid->provider_table = NULL;
347
    mid->refcount = 1;
348
    mid->finalize_cb = NULL;
Matthieu Dorier's avatar
Matthieu Dorier committed
349
    mid->enable_remote_shutdown = 0;
350

351 352
    mid->timer_list = margo_timer_list_create();
    if(mid->timer_list == NULL) goto err;
353

Shane Snyder's avatar
Shane Snyder committed
354 355 356 357
    /* initialize the handle cache */
    hret = margo_handle_cache_init(mid);
    if(hret != HG_SUCCESS) goto err;

358
    ret = ABT_thread_create(mid->progress_pool, hg_progress_fn, mid, 
359
        ABT_THREAD_ATTR_NULL, &mid->hg_progress_tid);
360 361
    if(ret != 0) goto err;

Matthieu Dorier's avatar
Matthieu Dorier committed
362 363 364
    mid->shutdown_rpc_id = MARGO_REGISTER(mid, "__shutdown__", 
            void, margo_shutdown_out_t, remote_shutdown_ult);

365 366
    return mid;

367 368
err:
    if(mid)
369
    {
Shane Snyder's avatar
Shane Snyder committed
370
        margo_handle_cache_destroy(mid);
371
        margo_timer_list_free(mid->timer_list);
372 373
        ABT_mutex_free(&mid->finalize_mutex);
        ABT_cond_free(&mid->finalize_cond);
374
        free(mid);
375
    }
376
    return MARGO_INSTANCE_NULL;
377 378
}

379 380 381 382
static void margo_cleanup(margo_instance_id mid)
{
    int i;

383 384 385 386 387 388 389 390 391
    /* 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);
    }

392
    margo_timer_list_free(mid->timer_list);
393

394 395
    /* delete the hash used for provider IDs */
    delete_provider_hash(mid);
396

397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415
    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
416 417
    margo_handle_cache_destroy(mid);

418 419 420 421 422 423
    if (mid->margo_init)
    {
        if (mid->hg_context)
            HG_Context_destroy(mid->hg_context);
        if (mid->hg_class)
            HG_Finalize(mid->hg_class);
424 425
        if (mid->abt_init)
            ABT_finalize();
426 427
    }

428 429 430
    free(mid);
}

431
void margo_finalize(margo_instance_id mid)
432
{
433
    int do_cleanup;
434

435
    /* tell progress thread to wrap things up */
436
    mid->hg_progress_shutdown_flag = 1;
437 438

    /* wait for it to shutdown cleanly */
439 440
    ABT_thread_join(mid->hg_progress_tid);
    ABT_thread_free(&mid->hg_progress_tid);
441

442 443 444 445
    ABT_mutex_lock(mid->finalize_mutex);
    mid->finalize_flag = 1;
    ABT_cond_broadcast(mid->finalize_cond);

446 447
    mid->refcount--;
    do_cleanup = mid->refcount == 0;
448

449 450 451 452 453 454 455
    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);
456 457 458 459 460 461

    return;
}

void margo_wait_for_finalize(margo_instance_id mid)
{
462
    int do_cleanup;
463 464 465

    ABT_mutex_lock(mid->finalize_mutex);

466
        mid->refcount++;
467 468 469 470
            
        while(!mid->finalize_flag)
            ABT_cond_wait(mid->finalize_cond, mid->finalize_mutex);

471 472 473
        mid->refcount--;
        do_cleanup = mid->refcount == 0;

474
    ABT_mutex_unlock(mid->finalize_mutex);
475 476 477 478

    if (do_cleanup)
        margo_cleanup(mid);

479 480 481
    return;
}

482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498
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;
}

Matthieu Dorier's avatar
Matthieu Dorier committed
499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514
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;

515
    hret = margo_forward(handle, NULL);
Matthieu Dorier's avatar
Matthieu Dorier committed
516 517 518 519 520 521 522 523 524 525 526 527 528 529
    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
hg_id_t margo_register_name_provider(margo_instance_id mid, const char *func_name,
537
    hg_proc_cb_t in_proc_cb, hg_proc_cb_t out_proc_cb, hg_rpc_cb_t rpc_cb,
538
    uint16_t provider_id, ABT_pool pool)
539
{
540
    struct provider_element *element;
541
    hg_id_t id;
542

543
    id = margo_register_internal(mid, func_name, in_proc_cb, out_proc_cb, rpc_cb, provider_id);
544 545 546 547
    if(id <= 0)
        return(0);

    /* nothing to do, we'll let the handler pool take this directly */
548
    if(provider_id == MARGO_DEFAULT_PROVIDER_ID)
549 550
        return(id);

551
    HASH_FIND(hh, mid->provider_table, &id, sizeof(id), element);
552 553 554
    if(element)
        return(id);

555
    element = calloc(1,sizeof(*element));
556 557
    if(!element)
        return(0);
558
    element->id = id;
559 560
    element->pool = pool;

561
    HASH_ADD(hh, mid->provider_table, id, sizeof(id), element);
562 563

    return(id);
564 565
}

566 567
hg_return_t margo_registered_name(margo_instance_id mid, const char *func_name,
    hg_id_t *id, hg_bool_t *flag)
568
{
569 570
    *id = gen_id(func_name, 0);
    return(HG_Registered(mid->hg_class, *id, flag));
571 572
}

573 574
hg_return_t margo_registered_name_provider(margo_instance_id mid, const char *func_name,
    uint16_t provider_id, hg_id_t *id, hg_bool_t *flag)
575 576
{
    hg_bool_t b;
577 578 579 580 581
    hg_return_t ret;

    *id = gen_id(func_name, provider_id);

    ret = HG_Registered(mid->hg_class, *id, &b);
582 583
    if(ret != HG_SUCCESS) 
        return ret;
584
    if((!b) || (!provider_id)) {
585 586 587 588
        *flag = b;
        return ret;
    }

589
    struct provider_element *element;
590

591
    HASH_FIND(hh, mid->provider_table, id, sizeof(*id), element);
592 593 594 595 596
    if(!element) {
        *flag = 0;
        return HG_SUCCESS;
    }

597
    assert(element->id == *id);
598 599 600 601 602

    *flag = 1;
    return HG_SUCCESS;
}

603 604 605 606 607 608 609
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 
610
		= (struct margo_rpc_data*) HG_Registered_data(mid->hg_class, id);
611
	if(!margo_data) return HG_OTHER_ERROR;
612 613 614
    if(margo_data->user_data && margo_data->user_free_callback) {
        (margo_data->user_free_callback)(margo_data->user_data);
    }
615 616 617 618 619 620 621 622 623 624 625 626 627
	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;
}

628 629 630 631
hg_return_t margo_registered_disable_response(
    margo_instance_id mid,
    hg_id_t id,
    int disable_flag)
632
{
633
    return(HG_Registered_disable_response(mid->hg_class, id, disable_flag));
634
}
635

636
struct lookup_cb_evt
637
{
638
    hg_return_t hret;
639 640 641 642 643 644
    hg_addr_t addr;
};

static hg_return_t margo_addr_lookup_cb(const struct hg_cb_info *info)
{
    struct lookup_cb_evt evt;
645
    evt.hret = info->ret;
646
    evt.addr = info->info.lookup.addr;
Matthieu Dorier's avatar
Matthieu Dorier committed
647
    ABT_eventual eventual = (ABT_eventual)(info->arg);
648 649

    /* propagate return code out through eventual */
Matthieu Dorier's avatar
Matthieu Dorier committed
650
    ABT_eventual_set(eventual, &evt, sizeof(evt));
651

652 653 654
    return(HG_SUCCESS);
}

655 656 657 658
hg_return_t margo_addr_lookup(
    margo_instance_id mid,
    const char   *name,
    hg_addr_t    *addr)
659
{
660
    hg_return_t hret;
661 662 663
    struct lookup_cb_evt *evt;
    ABT_eventual eventual;
    int ret;
664

665 666 667 668 669 670
    ret = ABT_eventual_create(sizeof(*evt), &eventual);
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);        
    }

671
    hret = HG_Addr_lookup(mid->hg_context, margo_addr_lookup_cb,
Matthieu Dorier's avatar
Matthieu Dorier committed
672
        (void*)eventual, name, HG_OP_ID_IGNORE);
673
    if(hret == HG_SUCCESS)
674 675 676
    {
        ABT_eventual_wait(eventual, (void**)&evt);
        *addr = evt->addr;
677
        hret = evt->hret;
678 679 680 681
    }

    ABT_eventual_free(&eventual);

682
    return(hret);
683 684 685 686 687
}

hg_return_t margo_addr_free(
    margo_instance_id mid,
    hg_addr_t addr)
688
{
689 690
    return(HG_Addr_free(mid->hg_class, addr));
}
691

692 693 694 695 696
hg_return_t margo_addr_self(
    margo_instance_id mid,
    hg_addr_t *addr)
{
    return(HG_Addr_self(mid->hg_class, addr));
697 698
}

699 700 701 702 703 704 705 706 707
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(
708
    margo_instance_id mid,
709 710 711 712 713 714 715 716 717 718
    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)
{
719
    hg_return_t hret = HG_OTHER_ERROR;
Shane Snyder's avatar
Shane Snyder committed
720 721 722 723 724 725 726 727

    /* 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);
    }
728

Shane Snyder's avatar
Shane Snyder committed
729
    return hret;
730 731
}

732
hg_return_t margo_destroy(hg_handle_t handle)
733
{
734
    margo_instance_id mid;
735
    hg_return_t hret = HG_OTHER_ERROR;
Shane Snyder's avatar
Shane Snyder committed
736

737 738 739
    /* use the handle to get the associated mid */
    mid = margo_hg_handle_get_instance(handle);

Shane Snyder's avatar
Shane Snyder committed
740 741 742 743 744 745 746
    /* 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);
    }
747

Shane Snyder's avatar
Shane Snyder committed
748
    return hret;
749 750 751 752 753
}

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
754
    ABT_eventual eventual = (ABT_eventual)(info->arg);
755 756

    /* propagate return code out through eventual */
Matthieu Dorier's avatar
Matthieu Dorier committed
757
    ABT_eventual_set(eventual, &hret, sizeof(hret));
758 759 760 761
    
    return(HG_SUCCESS);
}

762
hg_return_t margo_forward_provider_id(
763
    uint16_t provider_id,
764 765
    hg_handle_t handle,
    void *in_struct)
766 767 768
{
	hg_return_t hret;
	margo_request req;
769
	hret = margo_iforward_provider_id(provider_id, handle, in_struct, &req);
770 771 772 773 774
	if(hret != HG_SUCCESS) 
		return hret;
	return margo_wait(req);
}

775
hg_return_t margo_iforward_provider_id(
776
    uint16_t provider_id,
777 778 779
    hg_handle_t handle,
    void *in_struct,
    margo_request* req)
780 781
{
    hg_return_t hret = HG_TIMEOUT;
782
    ABT_eventual eventual;
783
    int ret;
784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817
    const struct hg_info* hgi; 
    hg_id_t id;
    hg_proc_cb_t in_cb, out_cb;
    hg_bool_t flag;

    assert(provider_id <= MARGO_MAX_PROVIDER_ID);

    hgi = HG_Get_info(handle);
    /* make sure bottom bits of id are clear */
    id = (hgi->id >> (__MARGO_PROVIDER_ID_SIZE*8)) << 
        (__MARGO_PROVIDER_ID_SIZE*8);
    id |= provider_id;

    /* TODO: if we reset the handle here, is there any reason to do so in
     * the handle cache?
     */
    ret = HG_Reset(handle, hgi->addr, id);
    if(ret == HG_NO_MATCH)
    {
        /* if Mercury does not recognize this ID (with provider id included)
         * then register it now
         */
        /* find encoders for base ID */
        ret = HG_Registered_proc_cb(hgi->hg_class, hgi->id, &flag, &in_cb, &out_cb);
        if(ret != HG_SUCCESS)
            return(ret);
        if(!flag)
            return(HG_NO_MATCH);

        /* register new ID that includes provider id */
        ret = HG_Register(hgi->hg_class, id, in_cb, out_cb, NULL);
        if(ret != HG_SUCCESS)
            return(ret);
    }
818 819 820 821 822 823 824

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

825
    *req = eventual;
826

Matthieu Dorier's avatar
Matthieu Dorier committed
827
    return HG_Forward(handle, margo_cb, (void*)eventual, in_struct);
828
}
829

830 831 832 833
hg_return_t margo_wait(margo_request req)
{
	hg_return_t* waited_hret;
	hg_return_t  hret;
834

835 836 837 838
    ABT_eventual_wait(req, (void**)&waited_hret);
	hret = *waited_hret;
    ABT_eventual_free(&req);
	
839
    return(hret);
840 841
}

Matthieu Dorier's avatar
Matthieu Dorier committed
842 843 844 845 846
int margo_test(margo_request req, int* flag)
{
    return ABT_eventual_test(req, NULL, flag);
}

847 848 849 850
typedef struct
{
    hg_handle_t handle;
} margo_forward_timeout_cb_dat;
851

852 853 854 855 856 857 858 859 860 861 862
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(
863
    hg_handle_t handle,
864 865
    void *in_struct,
    double timeout_ms)
866 867
{
    int ret;
868
    hg_return_t hret;
869
    margo_instance_id mid;
870
    ABT_eventual eventual;
871
    hg_return_t* waited_hret;
872 873
    margo_timer_t forward_timer;
    margo_forward_timeout_cb_dat timeout_cb_dat;
874 875 876 877 878 879 880

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

881 882 883
    /* use the handle to get the associated mid */
    mid = margo_hg_handle_get_instance(handle);

884 885 886 887 888
    /* 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
889
    hret = HG_Forward(handle, margo_cb, (void*)eventual, in_struct);
890
    if(hret == HG_SUCCESS)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
891 892 893 894 895
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

896 897 898 899 900 901 902 903
    /* 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
904 905 906 907 908 909 910 911
    ABT_eventual_free(&eventual);

    return(hret);
}

hg_return_t margo_respond(
    hg_handle_t handle,
    void *out_struct)
912 913 914 915 916 917 918 919 920 921 922 923 924
{
    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
925 926 927 928
{
    ABT_eventual eventual;
    int ret;

Matthieu Dorier's avatar
Matthieu Dorier committed
929
    ret = ABT_eventual_create(sizeof(hg_return_t), &eventual);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
930 931 932 933 934
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);
    }

935
    *req = eventual;
936

Matthieu Dorier's avatar
Matthieu Dorier committed
937
    return HG_Respond(handle, margo_cb, (void*)eventual, out_struct);
938 939
}

940 941 942 943 944 945 946
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)
947
{
948 949 950
    return(HG_Bulk_create(mid->hg_class, count,
        buf_ptrs, buf_sizes, flags, handle));
}
951

952 953 954 955
hg_return_t margo_bulk_free(
    hg_bulk_t handle)
{
    return(HG_Bulk_free(handle));
956 957
}

958 959 960 961 962 963 964 965
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));
}
966

967
hg_return_t margo_bulk_transfer(
968
    margo_instance_id mid,
969
    hg_bulk_op_t op,
970
    hg_addr_t origin_addr,
971 972 973 974
    hg_bulk_t origin_handle,
    size_t origin_offset,
    hg_bulk_t local_handle,
    size_t local_offset,
975
    size_t size)
976 977 978 979
{  
    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
980
                          local_offset, size, &req);
981 982 983 984 985 986 987 988 989 990 991 992 993 994 995
    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)
996 997 998 999 1000 1001 1002 1003 1004 1005 1006
{
    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);        
    }

1007
    *req = eventual;
1008

Matthieu Dorier's avatar
Matthieu Dorier committed
1009 1010
    hret = HG_Bulk_transfer(mid->hg_context, margo_cb,
        (void*)eventual, op, origin_addr, origin_handle, origin_offset, local_handle,
1011
        local_offset, size, HG_OP_ID_IGNORE);
1012 1013 1014 1015

    return(hret);
}

1016 1017 1018 1019
typedef struct
{
    ABT_mutex mutex;
    ABT_cond cond;
1020
    char is_asleep;
1021 1022 1023 1024 1025 1026 1027 1028 1029
} 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);
1030
    sleep_cb_dat->is_asleep = 0;
1031 1032 1033 1034 1035 1036 1037
    ABT_cond_signal(sleep_cb_dat->cond);
    ABT_mutex_unlock(sleep_cb_dat->mutex);

    return;
}

void margo_thread_sleep(
1038
    margo_instance_id mid,
1039 1040 1041 1042 1043 1044 1045 1046
    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));
1047
    sleep_cb_dat.is_asleep = 1;
1048 1049

    /* initialize the sleep timer */
1050
    margo_timer_init(mid, &sleep_timer, margo_thread_sleep_cb,
1051 1052 1053 1054
        &sleep_cb_dat, timeout_ms);

    /* yield thread for specified timeout */
    ABT_mutex_lock(sleep_cb_dat.mutex);
1055 1056
    while(sleep_cb_dat.is_asleep)
        ABT_cond_wait(sleep_cb_dat.cond, sleep_cb_dat.mutex);
1057 1058
    ABT_mutex_unlock(sleep_cb_dat.mutex);

1059 1060 1061 1062
    /* clean up */
    ABT_mutex_free(&sleep_cb_dat.mutex);
    ABT_cond_free(&sleep_cb_dat.cond);

1063 1064 1065
    return;
}

1066
int margo_get_handler_pool(margo_instance_id mid, ABT_pool* pool)
1067
{
1068 1069 1070 1071 1072 1073
    if(mid) {
        *pool = mid->handler_pool;
        return 0;
    } else {
        return -1;
    }
1074
}
1075

1076 1077 1078 1079
hg_context_t* margo_get_context(margo_instance_id mid)
{
    return(mid->hg_context);
}
1080

1081 1082 1083
hg_class_t* margo_get_class(margo_instance_id mid)
{
    return(mid->hg_class);
1084
}
Philip Carns's avatar
Philip Carns committed
1085

1086
ABT_pool margo_hg_handle_get_handler_pool(hg_handle_t h)
1087
{
1088 1089 1090 1091 1092 1093 1094 1095 1096
    struct margo_rpc_data* data;
    const struct hg_info* info;
    hg_id_t base_id; 
    uint16_t provider_id;
    int ret;
    ABT_pool pool;
    
    info = HG_Get_info(h);
    if(!info) return ABT_POOL_NULL;
1097

1098 1099
    data = (struct margo_rpc_data*) HG_Registered_data(info->hg_class, info->id);
    if(!data) return ABT_POOL_NULL;
1100

1101
    demux_id(info->id, &base_id, &provider_id);
1102

1103 1104
    ret = margo_lookup_provider(data->mid, base_id, provider_id, &pool);
    if(ret != 0) return ABT_POOL_NULL;
1105

1106 1107
    return pool;
}
1108

1109 1110 1111 1112 1113 1114 1115 1116
margo_instance_id margo_hg_info_get_instance(const struct hg_info *info)
{
    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;
}

1117 1118 1119 1120 1121 1122 1123
margo_instance_id margo_hg_handle_get_instance(hg_handle_t h)
{
    struct margo_rpc_data* data;
    const struct hg_info* info;
    
    info = HG_Get_info(h);
    if(!info) return MARGO_INSTANCE_NULL;
Philip Carns's avatar
Philip Carns committed
1124

1125 1126
    data = (struct margo_rpc_data*) HG_Registered_data(info->hg_class, info->id);
    if(!data) return MARGO_INSTANCE_NULL;
1127

1128
    return data->mid;
1129 1130
}

1131
int margo_register_data_provider(margo_instance_id mid, hg_id_t id, uint16_t provider_id, void* data, void (*free_callback)(void *))
1132
{
1133 1134
    struct provider_element *element;
    hg_id_t muxed_id;
1135

1136
    muxed_id = mux_id(id, provider_id);
1137

1138
    HASH_FIND(hh, mid->provider_table, &muxed_id, sizeof(muxed_id), element);
1139 1140 1141
    if(!element)
        return -1;

1142
    assert(element->id == muxed_id);
1143 1144 1145 1146 1147 1148 1149 1150 1151 1152

    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);
}

1153
void* margo_registered_data_provider(margo_instance_id mid, hg_id_t id, uint16_t provider_id)
1154
{
1155 1156
    struct provider_element *element;
    hg_id_t muxed_id;
1157

1158
    muxed_id = mux_id(id, provider_id);
1159

1160
    HASH_FIND(hh, mid->provider_table, &muxed_id, sizeof(muxed_id), element);
1161 1162 1163
    if(!element)
        return NULL;

1164
    assert(element->id == muxed_id);
1165 1166 1167

    return element->user_data;
}
1168
static void margo_rpc_data_free(void* ptr)
Philip Carns's avatar
Philip Carns committed
1169
{
1170 1171 1172 1173 1174 1175
	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);
}
1176

1177
static void delete_provider_hash(margo_instance_id mid)
1178
{
1179
    struct provider_element *current_element, *tmp;
1180

1181
    HASH_ITER(hh, mid->provider_table, current_element, tmp) {
1182 1183
        if(current_element->user_data && current_element->user_free_callback)
            (current_element->user_free_callback)(current_element->user_data);
1184
        HASH_DEL(mid->provider_table, current_element);
1185 1186 1187 1188
        free(current_element);
    }
}

1189 1190 1191 1192 1193 1194 1195
/* 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;
1196
    unsigned int hg_progress_timeout = mid->hg_progress_timeout_ub;
1197 1198
    double next_timer_exp;
    int trigger_happened;
1199 1200
    double tm1, tm2;
    int diag_enabled = 0;
1201

1202 1203 1204 1205
    while(!mid->hg_progress_shutdown_flag)
    {
        trigger_happened = 0;
        do {
1206 1207 1208 1209
            /* save value of instance diag variable, in case it is modified
             * while we are in loop 
             */
            diag_enabled = mid->diag_enabled;
1210

1211
            if(diag_enabled) tm1 = ABT_get_wtime();
1212
            ret = HG_Trigger(mid->hg_context, 0, 1, &actual_count);
1213 1214 1215 1216 1217
            if(diag_enabled)
            {
                tm2 = ABT_get_wtime();
                __DIAG_UPDATE(mid->diag_trigger_elapsed, (tm2-tm1));
            }
1218

1219 1220 1221
            if(ret == HG_SUCCESS && actual_count > 0)
                trigger_happened = 1;
        } while((ret == HG_SUCCESS) && actual_count && !mid->hg_progress_shutdown_flag);
1222

1223 1224
        if(trigger_happened)
            ABT_thread_yield();
1225

1226
        ABT_pool_get_size(mid->progress_pool, &size);
1227
        /* Are there any other threads executing in this pool that are *not*
1228 1229 1230 1231
         * 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.
1232
         */
1233
        if(size > 0)
1234 1235 1236 1237 1238 1239 1240 1241
        {
            /* 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.
             */
1242
            if(diag_enabled) tm1 = ABT_get_wtime();
1243
            ret = HG_Progress(mid->hg_context, 0);
1244 1245 1246 1247 1248 1249
            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);
            }
1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268
            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
        {
1269
            hg_progress_timeout = mid->hg_progress_timeout_ub;
1270 1271 1272 1273 1274 1275 1276 1277 1278
            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 */
1279
                    if(next_timer_exp < mid->hg_progress_timeout_ub)
1280 1281 1282 1283 1284 1285 1286
                        hg_progress_timeout = (unsigned int)next_timer_exp;
                }
                else
                {
                    hg_progress_timeout = 0;
                }
            }
1287
            if(diag_enabled) tm1 = ABT_get_wtime();
1288
            ret = HG_Progress(mid->hg_context, hg_progress_timeout);
1289 1290 1291 1292 1293 1294 1295 1296 1297 1298
            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);
            }
1299 1300 1301 1302 1303 1304
            if(ret != HG_SUCCESS && ret != HG_TIMEOUT)
            {
                /* TODO: error handling */
                fprintf(stderr, "WARNING: unexpected return code (%d) from HG_Progress()\n", ret);
            }
        }
1305

1306 1307 1308
        /* check for any expired timers */
        margo_check_timers(mid);
    }
1309

1310
    return;
Philip Carns's avatar
Philip Carns committed
1311
}
Philip Carns's avatar
Philip Carns committed
1312 1313 1314 1315 1316 1317 1318


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

1319
static void print_diag_data(FILE *file, const char* name, const char *description, struct diag_data *data)
Philip Carns's avatar
Philip Carns committed
1320
{
1321 1322 1323 1324 1325 1326 1327 1328
    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
1329 1330 1331
    return;
}

1332
void margo_diag_dump(margo_instance_id mid, const char* file, int uniquify)
Philip Carns's avatar
Philip Carns committed
1333 1334 1335
{
    FILE *outfile;
    time_t ltime;
1336
    char revised_file_name[256] = {0};
Philip Carns's avatar
Philip Carns committed
1337 1338 1339

    assert(mid->diag_enabled);

1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354
    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
1355 1356 1357 1358 1359 1360
    if(strcmp("-", file) == 0)
    {
        outfile = stdout;
    }
    else
    {
1361
        outfile = fopen(revised_file_name, "a");
Philip Carns's avatar
Philip Carns committed
1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377
        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");
1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389
    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
1390 1391 1392 1393 1394 1395

    if(outfile != stdout)
        fclose(outfile);
    
    return;
}
1396

1397
void margo_set_param(margo_instance_id mid, int option, const void *param)
1398 1399 1400
{
    switch(option)
    {
1401
        case MARGO_PARAM_PROGRESS_TIMEOUT_UB:
1402 1403 1404 1405 1406 1407 1408
            mid->hg_progress_timeout_ub = (*((const unsigned int*)param));
            break;
    }

    return;
}

1409
void margo_get_param(margo_instance_id mid, int option, void *param)
1410 1411 1412 1413
{

    switch(option)
    {
1414
        case MARGO_PARAM_PROGRESS_TIMEOUT_UB:
1415 1416 1417 1418 1419 1420
            (*((unsigned int*)param)) = mid->hg_progress_timeout_ub;
            break;
    }

    return;
}
Shane Snyder's avatar
Shane Snyder committed
1421 1422 1423 1424 1425 1426

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;
1427

1428
    ABT_mutex_create(&(mid->handle_cache_mtx));
Shane Snyder's avatar
Shane Snyder committed
1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467

    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);
    }

1468
    ABT_mutex_free(&mid->handle_cache_mtx);
1469

Shane Snyder's avatar
Shane Snyder committed
1470 1471 1472 1473 1474 1475 1476
    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;
1477 1478 1479
    hg_return_t hret = HG_SUCCESS;

    ABT_mutex_lock(mid->handle_cache_mtx);
Shane Snyder's avatar
Shane Snyder committed
1480 1481 1482 1483

    if(!mid->free_handle_list)
    {
        /* if no available handles, just fall through */
1484 1485
        hret = HG_OTHER_ERROR;
        goto finish;
Shane Snyder's avatar
Shane Snyder committed
1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499
    }

    /* 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;
    }
1500 1501 1502 1503 1504
    else
    {
        /* reset failed, add handle back to the free list */
        LL_APPEND(mid->free_handle_list, el);
    }
Shane Snyder's avatar
Shane Snyder committed
1505