margo.c 14 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11

/*
 * (C) 2015 The University of Chicago
 * 
 * See COPYRIGHT in top-level directory.
 */

#include <assert.h>
#include <unistd.h>
#include <errno.h>
#include <abt.h>
12
#include <abt-snoozer.h>
13
#include <time.h>
Philip Carns's avatar
bug fix  
Philip Carns committed
14
#include <math.h>
15 16

#include "margo.h"
17
#include "margo-timer.h"
Philip Carns's avatar
Philip Carns committed
18
#include "utlist.h"
19

20 21
#define MERCURY_PROGRESS_TIMEOUT_UB 100 /* 100 milliseconds */

22 23
struct margo_instance
{
24
    /* provided by caller */
25 26
    hg_context_t *hg_context;
    hg_class_t *hg_class;
27 28 29
    ABT_pool handler_pool;
    ABT_pool progress_pool;

30
    /* internal to margo for this particular instance */
31 32
    ABT_thread hg_progress_tid;
    int hg_progress_shutdown_flag;
33 34 35 36 37 38 39

    /* control logic for callers waiting on margo to be finalized */
    int finalize_flag;
    int finalize_waiters_in_progress_pool;
    ABT_mutex finalize_mutex;
    ABT_cond finalize_cond;

40 41 42 43 44 45 46 47 48 49 50 51
    int table_index;
};

struct margo_handler_mapping
{
    hg_class_t *class;
    margo_instance_id mid;
};

#define MAX_HANDLER_MAPPING 8
static int handler_mapping_table_size = 0;
static struct margo_handler_mapping handler_mapping_table[MAX_HANDLER_MAPPING] = {0};
52

53
static void hg_progress_fn(void* foo);
54
static int margo_xstream_is_in_progress_pool(margo_instance_id mid);
55 56 57 58 59 60 61 62

struct handler_entry
{
    void* fn;
    hg_handle_t handle;
    struct handler_entry *next; 
};

63 64
margo_instance_id margo_init(ABT_pool progress_pool, ABT_pool handler_pool,
    hg_context_t *hg_context, hg_class_t *hg_class)
65 66
{
    int ret;
67 68 69
    struct margo_instance *mid;

    if(handler_mapping_table_size >= MAX_HANDLER_MAPPING)
70
        return(MARGO_INSTANCE_NULL);
71 72 73

    mid = malloc(sizeof(*mid));
    if(!mid)
74
        return(MARGO_INSTANCE_NULL);
75
    memset(mid, 0, sizeof(*mid));
76

77 78 79
    ABT_mutex_create(&mid->finalize_mutex);
    ABT_cond_create(&mid->finalize_cond);

80 81
    mid->progress_pool = progress_pool;
    mid->handler_pool = handler_pool;
82 83
    mid->hg_class = hg_class;
    mid->hg_context = hg_context;
84

85 86 87 88 89 90 91 92
    ret = margo_timer_instance_init(mid);
    if(ret != 0)
    {
        fprintf(stderr, "Error: margo_timer_instance_init()\n");
        free(mid);
        return(MARGO_INSTANCE_NULL);
    }

93
    ret = ABT_thread_create(mid->progress_pool, hg_progress_fn, mid, 
94
        ABT_THREAD_ATTR_NULL, &mid->hg_progress_tid);
95 96 97
    if(ret != 0)
    {
        fprintf(stderr, "Error: ABT_thread_create()\n");
98
        free(mid);
99
        return(MARGO_INSTANCE_NULL);
100 101
    }

102 103 104 105 106 107
    handler_mapping_table[handler_mapping_table_size].mid = mid;
    handler_mapping_table[handler_mapping_table_size].class = mid->hg_class;
    mid->table_index = handler_mapping_table_size;
    handler_mapping_table_size++;

    return mid;
108 109
}

110
void margo_finalize(margo_instance_id mid)
111
{
112 113
    int i;

114
    /* tell progress thread to wrap things up */
115
    mid->hg_progress_shutdown_flag = 1;
116 117

    /* wait for it to shutdown cleanly */
118 119
    ABT_thread_join(mid->hg_progress_tid);
    ABT_thread_free(&mid->hg_progress_tid);
120

121 122 123 124 125
    for(i=mid->table_index; i<(handler_mapping_table_size-1); i++)
    {
        handler_mapping_table[i] = handler_mapping_table[i+1];
    }
    handler_mapping_table_size--;
126

127 128 129 130 131 132 133 134 135 136 137 138 139
    ABT_mutex_lock(mid->finalize_mutex);
    mid->finalize_flag = 1;
    ABT_cond_broadcast(mid->finalize_cond);
    ABT_mutex_unlock(mid->finalize_mutex);

    /* TODO: yuck, there is a race here if someone was really waiting for
     * finalize; we can't destroy the data structures out from under them.
     * We could fix this by reference counting so that the last caller
     * (whether a finalize() caller or wait_for_finalize() caller) knows it
     * is safe to turn off the lights on their way out.  For now we just leak 
     * a small amount of memory.
     */
#if 0
140
    margo_timer_instance_finalize(mid);
141

142 143 144 145 146 147 148 149 150 151 152 153 154 155 156
    ABT_mutex_free(&mid->finalize_mutex);
    ABT_cond_free(&mid->finalize_cond);
    free(mid);
#endif

    return;
}

void margo_wait_for_finalize(margo_instance_id mid)
{
    int in_pool = 0;

    /* Is this waiter in the same pool as the pool running the progress
     * thread?
     */
157
    if(margo_xstream_is_in_progress_pool(mid))
158 159 160 161 162 163 164 165 166 167 168
        in_pool = 1;

    ABT_mutex_lock(mid->finalize_mutex);

        mid->finalize_waiters_in_progress_pool += in_pool;
            
        while(!mid->finalize_flag)
            ABT_cond_wait(mid->finalize_cond, mid->finalize_mutex);

    ABT_mutex_unlock(mid->finalize_mutex);
    
169 170 171 172
    return;
}

/* dedicated thread function to drive Mercury progress */
173
static void hg_progress_fn(void* foo)
174 175 176
{
    int ret;
    unsigned int actual_count;
177
    struct margo_instance *mid = (struct margo_instance *)foo;
178
    size_t size;
179 180
    unsigned int hg_progress_timeout = MERCURY_PROGRESS_TIMEOUT_UB;
    double next_timer_exp;
181

182
    while(!mid->hg_progress_shutdown_flag)
183 184
    {
        do {
185
            ret = HG_Trigger(mid->hg_context, 0, 1, &actual_count);
186
        } while((ret == HG_SUCCESS) && actual_count && !mid->hg_progress_shutdown_flag);
187

188
        if(!mid->hg_progress_shutdown_flag)
189
        {
190 191
            ABT_mutex_lock(mid->finalize_mutex);

192
            ABT_pool_get_total_size(mid->progress_pool, &size);
193 194 195 196 197 198
            /* Are there any other threads executing in this pool that are *not*
             * blocked on margo_wait_for_finalize()?  If so then, we can't
             * sleep here or else those threads will not get a chance to
             * execute.
             */
            if(size > mid->finalize_waiters_in_progress_pool)
199
            {
200
                ABT_mutex_unlock(mid->finalize_mutex);
201
                HG_Progress(mid->hg_context, 0);
202 203 204 205
                ABT_thread_yield();
            }
            else
            {
206
                ABT_mutex_unlock(mid->finalize_mutex);
207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225

                ret = margo_timer_get_next_expiration(mid, &next_timer_exp);
                if(ret == 0)
                {
                    /* there is a queued timer, don't block long enough
                     * to keep this timer waiting
                     */
                    if(next_timer_exp >= 0.0)
                    {
                        next_timer_exp *= 1000; /* convert to milliseconds */
                        if(next_timer_exp < MERCURY_PROGRESS_TIMEOUT_UB)
                            hg_progress_timeout = (unsigned int)next_timer_exp;
                    }
                    else
                    {
                        hg_progress_timeout = 0;
                    }
                }
                HG_Progress(mid->hg_context, hg_progress_timeout);
226 227
            }
        }
228

229
        /* check for any expired timers */
230
        margo_check_timers(mid);
231 232
    }

233
    return;
234 235
}

236
ABT_pool* margo_get_handler_pool(margo_instance_id mid)
237
{
238
    return(&mid->handler_pool);
239 240
}

241 242 243 244 245 246 247 248 249 250 251
hg_context_t* margo_get_context(margo_instance_id mid)
{
    return(mid->hg_context);
}

hg_class_t* margo_get_class(margo_instance_id mid)
{
    return(mid->hg_class);
}


Jonathan Jenkins's avatar
Jonathan Jenkins committed
252
static hg_return_t margo_cb(const struct hg_cb_info *info)
253 254 255 256 257 258 259 260 261 262
{
    hg_return_t hret = info->ret;

    ABT_eventual *eventual = info->arg;
    /* propagate return code out through eventual */
    ABT_eventual_set(*eventual, &hret, sizeof(hret));
    
    return(HG_SUCCESS);
}

263 264 265
typedef struct
{
    hg_handle_t handle;
Shane Snyder's avatar
Shane Snyder committed
266
} margo_forward_timeout_cb_dat;
267 268 269

static void margo_forward_timeout_cb(void *arg)
{
Shane Snyder's avatar
Shane Snyder committed
270 271
    margo_forward_timeout_cb_dat *timeout_cb_dat =
        (margo_forward_timeout_cb_dat *)arg;
272 273

    /* cancel the Mercury op if the forward timed out */
274
    HG_Cancel(timeout_cb_dat->handle);
275 276 277
    return;
}

278 279 280 281 282 283
hg_return_t margo_forward_timed(
    margo_instance_id mid,
    hg_handle_t handle,
    void *in_struct,
    double timeout_ms)
{
Shane Snyder's avatar
Shane Snyder committed
284
    int ret;
285 286 287
    hg_return_t hret = HG_TIMEOUT;
    ABT_eventual eventual;
    hg_return_t* waited_hret;
Shane Snyder's avatar
Shane Snyder committed
288 289
    margo_timer_t forward_timer;
    margo_forward_timeout_cb_dat timeout_cb_dat;
290 291 292 293 294 295 296

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

Shane Snyder's avatar
Shane Snyder committed
297 298
    /* set a timer object to expire when this forward times out */
    timeout_cb_dat.handle = handle;
299
    margo_timer_init(mid, &forward_timer, margo_forward_timeout_cb,
Shane Snyder's avatar
Shane Snyder committed
300
        &timeout_cb_dat, timeout_ms);
301 302 303 304 305 306 307 308

    hret = HG_Forward(handle, margo_cb, &eventual, in_struct);
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

309 310
    /* remove timer if it is still in place (i.e., not timed out) */
    if(hret != HG_TIMEOUT)
311
        margo_timer_destroy(mid, &forward_timer);
312 313 314 315

    ABT_eventual_free(&eventual);

    return(hret);
316 317 318 319

}


320
hg_return_t margo_forward(
321
    margo_instance_id mid,
322 323 324 325 326 327 328 329 330 331 332 333 334 335
    hg_handle_t handle,
    void *in_struct)
{
    hg_return_t hret = HG_TIMEOUT;
    ABT_eventual eventual;
    int ret;
    hg_return_t* waited_hret;

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

Jonathan Jenkins's avatar
Jonathan Jenkins committed
336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364
    hret = HG_Forward(handle, margo_cb, &eventual, in_struct);
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

hg_return_t margo_respond(
    margo_instance_id mid,
    hg_handle_t handle,
    void *out_struct)
{
    hg_return_t hret = HG_TIMEOUT;
    ABT_eventual eventual;
    int ret;
    hg_return_t* waited_hret;

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

    hret = HG_Respond(handle, margo_cb, &eventual, out_struct);
365 366 367 368 369 370 371 372 373 374 375
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

376

377
static hg_return_t margo_bulk_transfer_cb(const struct hg_cb_info *info)
378
{
379 380
    hg_return_t hret = info->ret;
    ABT_eventual *eventual = info->arg;
381 382 383 384 385 386 387

    /* propagate return code out through eventual */
    ABT_eventual_set(*eventual, &hret, sizeof(hret));
    
    return(HG_SUCCESS);
}

Philip Carns's avatar
Philip Carns committed
388 389
struct lookup_cb_evt
{
390 391
    hg_return_t nret;
    hg_addr_t addr;
Philip Carns's avatar
Philip Carns committed
392 393
};

394
static hg_return_t margo_addr_lookup_cb(const struct hg_cb_info *info)
Philip Carns's avatar
Philip Carns committed
395 396
{
    struct lookup_cb_evt evt;
397 398
    evt.nret = info->ret;
    evt.addr = info->info.lookup.addr;
Philip Carns's avatar
Philip Carns committed
399

400
    ABT_eventual *eventual = info->arg;
Philip Carns's avatar
Philip Carns committed
401 402 403 404

    /* propagate return code out through eventual */
    ABT_eventual_set(*eventual, &evt, sizeof(evt));
    
405
    return(HG_SUCCESS);
Philip Carns's avatar
Philip Carns committed
406 407 408
}


409
hg_return_t margo_addr_lookup(
410
    margo_instance_id mid,
411
    hg_context_t *context,
Philip Carns's avatar
Philip Carns committed
412
    const char   *name,
413
    hg_addr_t    *addr)
414
{
415
    hg_return_t nret;
Philip Carns's avatar
Philip Carns committed
416
    struct lookup_cb_evt *evt;
417 418 419
    ABT_eventual eventual;
    int ret;

Philip Carns's avatar
Philip Carns committed
420
    ret = ABT_eventual_create(sizeof(*evt), &eventual);
421 422 423 424 425
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);        
    }

426 427
    nret = HG_Addr_lookup(context, margo_addr_lookup_cb,
        &eventual, name, HG_OP_ID_IGNORE);
428 429
    if(nret == 0)
    {
Philip Carns's avatar
Philip Carns committed
430 431 432
        ABT_eventual_wait(eventual, (void**)&evt);
        *addr = evt->addr;
        nret = evt->nret;
433 434 435 436 437 438 439
    }

    ABT_eventual_free(&eventual);

    return(nret);
}

440
hg_return_t margo_bulk_transfer(
441
    margo_instance_id mid,
442
    hg_context_t *context,
443
    hg_bulk_op_t op,
444
    hg_addr_t origin_addr,
445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475
    hg_bulk_t origin_handle,
    size_t origin_offset,
    hg_bulk_t local_handle,
    size_t local_offset,
    size_t size)
{
    hg_return_t hret = HG_TIMEOUT;
    hg_return_t *waited_hret;
    ABT_eventual eventual;
    int ret;

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

    hret = HG_Bulk_transfer(context, margo_bulk_transfer_cb, &eventual, op, 
        origin_addr, origin_handle, origin_offset, local_handle, local_offset,
        size, HG_OP_ID_IGNORE);
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

476 477
typedef struct
{
478
    margo_instance_id mid;
479 480
    ABT_mutex mutex;
    ABT_cond cond;
Shane Snyder's avatar
Shane Snyder committed
481 482
    char is_asleep;
    char in_pool;
483 484 485 486 487 488 489
} 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;

490 491
    /* decrement number of waiting threads */
    ABT_mutex_lock(sleep_cb_dat->mid->finalize_mutex);
Shane Snyder's avatar
Shane Snyder committed
492 493
    sleep_cb_dat->mid->finalize_waiters_in_progress_pool -=
        sleep_cb_dat->in_pool;
494 495
    ABT_mutex_unlock(sleep_cb_dat->mid->finalize_mutex);

496 497
    /* wake up the sleeping thread */
    ABT_mutex_lock(sleep_cb_dat->mutex);
498
    sleep_cb_dat->is_asleep = 0;
499 500 501 502 503 504 505
    ABT_cond_signal(sleep_cb_dat->cond);
    ABT_mutex_unlock(sleep_cb_dat->mutex);

    return;
}

void margo_thread_sleep(
506
    margo_instance_id mid,
507 508
    double timeout_ms)
{
509
    int in_pool = 0;
510 511 512
    margo_timer_t sleep_timer;
    margo_thread_sleep_cb_dat sleep_cb_dat;

Shane Snyder's avatar
Shane Snyder committed
513 514 515
    if(margo_xstream_is_in_progress_pool(mid))
        in_pool = 1;

516
    /* set data needed for sleep callback */
517
    sleep_cb_dat.mid = mid;
518 519
    ABT_mutex_create(&(sleep_cb_dat.mutex));
    ABT_cond_create(&(sleep_cb_dat.cond));
520
    sleep_cb_dat.is_asleep = 1;
Shane Snyder's avatar
Shane Snyder committed
521
    sleep_cb_dat.in_pool = in_pool;
522 523

    /* initialize the sleep timer */
524
    margo_timer_init(mid, &sleep_timer, margo_thread_sleep_cb,
525 526
        &sleep_cb_dat, timeout_ms);

527 528
    /* increment number of waiting threads */
    ABT_mutex_lock(mid->finalize_mutex);
529
    mid->finalize_waiters_in_progress_pool += in_pool;
530 531
    ABT_mutex_unlock(mid->finalize_mutex);

532 533
    /* yield thread for specified timeout */
    ABT_mutex_lock(sleep_cb_dat.mutex);
534 535
    while(sleep_cb_dat.is_asleep)
        ABT_cond_wait(sleep_cb_dat.cond, sleep_cb_dat.mutex);
536 537 538 539 540
    ABT_mutex_unlock(sleep_cb_dat.mutex);

    return;
}

541
margo_instance_id margo_hg_class_to_instance(hg_class_t *cl)
542 543 544 545 546
{
    int i;

    for(i=0; i<handler_mapping_table_size; i++)
    {
547
        if(handler_mapping_table[i].class == cl)
548 549 550 551
            return(handler_mapping_table[i].mid);
    }
    return(NULL);
}
552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569

/* returns 1 if current xstream is in the progress pool, 0 if not */
static int margo_xstream_is_in_progress_pool(margo_instance_id mid)
{
    int ret;
    ABT_xstream xstream;
    ABT_pool pool;

    ret = ABT_xstream_self(&xstream);
    assert(ret == ABT_SUCCESS);
    ret = ABT_xstream_get_main_pools(xstream, 1, &pool);
    assert(ret == ABT_SUCCESS);

    if(pool == mid->progress_pool)
        return(1);
    else
        return(0);
}