margo.c 16.2 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;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
33 34 35
    int owns_progress_pool;
    ABT_xstream *rpc_xstreams;
    int num_handler_pool_threads;
36 37 38 39 40 41 42

    /* 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;

43 44 45 46 47 48 49 50 51 52 53 54
    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};
55

56
static void hg_progress_fn(void* foo);
57
static int margo_xstream_is_in_progress_pool(margo_instance_id mid);
58 59 60 61 62 63 64 65

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

Jonathan Jenkins's avatar
Jonathan Jenkins committed
66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137
margo_instance_id margo_init(int use_progress_thread, int rpc_thread_count,
    hg_context_t *hg_context)
{
    struct margo_instance *mid = MARGO_INSTANCE_NULL;
    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;
    int ret;
    int i;

    if (use_progress_thread)
    {
        ret = ABT_snoozer_xstream_create(1, &progress_pool, &progress_xstream);
        if (ret != ABT_SUCCESS) goto err;
    }
    else
    {
        ret = ABT_xstream_self(&progress_xstream);
        if (ret != ABT_SUCCESS) goto err;
        ret = ABT_xstream_get_main_pools(progress_xstream, 1, &progress_pool);
        if (ret != ABT_SUCCESS) goto err;
    }

    if (rpc_thread_count > 0)
    {
        rpc_xstreams = malloc(rpc_thread_count * sizeof(*rpc_xstreams));
        if (rpc_xstreams == NULL) goto err;
        ret = ABT_snoozer_xstream_create(rpc_thread_count, &rpc_pool,
                rpc_xstreams);
        if (ret != ABT_SUCCESS) goto err;
    }
    else if (rpc_thread_count == 0)
    {
        ret = ABT_xstream_self(&rpc_xstream);
        if (ret != ABT_SUCCESS) goto err;
        ret = ABT_xstream_get_main_pools(rpc_xstream, 1, &rpc_pool);
        if (ret != ABT_SUCCESS) goto err;
    }
    else
    {
        rpc_pool = progress_pool;
    }

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

    mid->owns_progress_pool = !use_progress_thread;
    mid->num_handler_pool_threads = rpc_thread_count;
    mid->rpc_xstreams = rpc_xstreams;
    return mid;

err:
    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);
    }
    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
138
    hg_context_t *hg_context)
139 140
{
    int ret;
141 142 143
    struct margo_instance *mid;

    if(handler_mapping_table_size >= MAX_HANDLER_MAPPING)
144
        return(MARGO_INSTANCE_NULL);
145 146 147

    mid = malloc(sizeof(*mid));
    if(!mid)
148
        return(MARGO_INSTANCE_NULL);
149
    memset(mid, 0, sizeof(*mid));
150

151 152 153
    ABT_mutex_create(&mid->finalize_mutex);
    ABT_cond_create(&mid->finalize_cond);

154 155
    mid->progress_pool = progress_pool;
    mid->handler_pool = handler_pool;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
156
    mid->hg_class = HG_Context_get_class(hg_context);
157
    mid->hg_context = hg_context;
158

159 160 161 162 163 164 165 166
    ret = margo_timer_instance_init(mid);
    if(ret != 0)
    {
        fprintf(stderr, "Error: margo_timer_instance_init()\n");
        free(mid);
        return(MARGO_INSTANCE_NULL);
    }

167
    ret = ABT_thread_create(mid->progress_pool, hg_progress_fn, mid, 
168
        ABT_THREAD_ATTR_NULL, &mid->hg_progress_tid);
169 170 171
    if(ret != 0)
    {
        fprintf(stderr, "Error: ABT_thread_create()\n");
172
        free(mid);
173
        return(MARGO_INSTANCE_NULL);
174 175
    }

176 177 178 179 180 181
    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;
182 183
}

184
void margo_finalize(margo_instance_id mid)
185
{
186 187
    int i;

188
    /* tell progress thread to wrap things up */
189
    mid->hg_progress_shutdown_flag = 1;
190 191

    /* wait for it to shutdown cleanly */
192 193
    ABT_thread_join(mid->hg_progress_tid);
    ABT_thread_free(&mid->hg_progress_tid);
194

195 196 197 198 199
    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--;
200

201 202 203 204 205 206 207 208 209 210 211 212 213
    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
214
    margo_timer_instance_finalize(mid);
215

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230
    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?
     */
231
    if(margo_xstream_is_in_progress_pool(mid))
232 233 234 235 236 237 238 239 240 241 242
        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);
    
243 244 245 246
    return;
}

/* dedicated thread function to drive Mercury progress */
247
static void hg_progress_fn(void* foo)
248 249 250
{
    int ret;
    unsigned int actual_count;
251
    struct margo_instance *mid = (struct margo_instance *)foo;
252
    size_t size;
253 254
    unsigned int hg_progress_timeout = MERCURY_PROGRESS_TIMEOUT_UB;
    double next_timer_exp;
255

256
    while(!mid->hg_progress_shutdown_flag)
257 258
    {
        do {
259
            ret = HG_Trigger(mid->hg_context, 0, 1, &actual_count);
260
        } while((ret == HG_SUCCESS) && actual_count && !mid->hg_progress_shutdown_flag);
261

262
        if(!mid->hg_progress_shutdown_flag)
263
        {
264 265
            ABT_mutex_lock(mid->finalize_mutex);

266
            ABT_pool_get_total_size(mid->progress_pool, &size);
267 268 269 270 271 272
            /* 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)
273
            {
274
                ABT_mutex_unlock(mid->finalize_mutex);
275
                HG_Progress(mid->hg_context, 0);
276 277 278 279
                ABT_thread_yield();
            }
            else
            {
280
                ABT_mutex_unlock(mid->finalize_mutex);
281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299

                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);
300 301
            }
        }
302

303
        /* check for any expired timers */
304
        margo_check_timers(mid);
305 306
    }

307
    return;
308 309
}

310
ABT_pool* margo_get_handler_pool(margo_instance_id mid)
311
{
312
    return(&mid->handler_pool);
313 314
}

315 316 317 318 319 320 321 322 323 324 325
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
326
static hg_return_t margo_cb(const struct hg_cb_info *info)
327 328 329 330 331 332 333 334 335 336
{
    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);
}

337 338 339
typedef struct
{
    hg_handle_t handle;
Shane Snyder's avatar
Shane Snyder committed
340
} margo_forward_timeout_cb_dat;
341 342 343

static void margo_forward_timeout_cb(void *arg)
{
Shane Snyder's avatar
Shane Snyder committed
344 345
    margo_forward_timeout_cb_dat *timeout_cb_dat =
        (margo_forward_timeout_cb_dat *)arg;
346 347

    /* cancel the Mercury op if the forward timed out */
348
    HG_Cancel(timeout_cb_dat->handle);
349 350 351
    return;
}

352 353 354 355 356 357
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
358
    int ret;
359
    hg_return_t hret;
360 361
    ABT_eventual eventual;
    hg_return_t* waited_hret;
Shane Snyder's avatar
Shane Snyder committed
362 363
    margo_timer_t forward_timer;
    margo_forward_timeout_cb_dat timeout_cb_dat;
364 365 366 367 368 369 370

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

Shane Snyder's avatar
Shane Snyder committed
371 372
    /* set a timer object to expire when this forward times out */
    timeout_cb_dat.handle = handle;
373
    margo_timer_init(mid, &forward_timer, margo_forward_timeout_cb,
Shane Snyder's avatar
Shane Snyder committed
374
        &timeout_cb_dat, timeout_ms);
375 376 377 378 379 380 381 382

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

383 384 385 386
    /* convert HG_CANCELED to HG_TIMEOUT to indicate op timed out */
    if(hret == HG_CANCELED)
        hret = HG_TIMEOUT;

387 388
    /* remove timer if it is still in place (i.e., not timed out) */
    if(hret != HG_TIMEOUT)
389
        margo_timer_destroy(mid, &forward_timer);
390 391 392 393

    ABT_eventual_free(&eventual);

    return(hret);
394 395 396
}


397
hg_return_t margo_forward(
398
    margo_instance_id mid,
399 400 401 402 403 404 405 406 407 408 409 410 411 412
    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
413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441
    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);
442 443 444 445 446 447 448 449 450 451 452
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

453

454
static hg_return_t margo_bulk_transfer_cb(const struct hg_cb_info *info)
455
{
456 457
    hg_return_t hret = info->ret;
    ABT_eventual *eventual = info->arg;
458 459 460 461 462 463 464

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

Philip Carns's avatar
Philip Carns committed
465 466
struct lookup_cb_evt
{
467 468
    hg_return_t nret;
    hg_addr_t addr;
Philip Carns's avatar
Philip Carns committed
469 470
};

471
static hg_return_t margo_addr_lookup_cb(const struct hg_cb_info *info)
Philip Carns's avatar
Philip Carns committed
472 473
{
    struct lookup_cb_evt evt;
474 475
    evt.nret = info->ret;
    evt.addr = info->info.lookup.addr;
Philip Carns's avatar
Philip Carns committed
476

477
    ABT_eventual *eventual = info->arg;
Philip Carns's avatar
Philip Carns committed
478 479 480 481

    /* propagate return code out through eventual */
    ABT_eventual_set(*eventual, &evt, sizeof(evt));
    
482
    return(HG_SUCCESS);
Philip Carns's avatar
Philip Carns committed
483 484 485
}


486
hg_return_t margo_addr_lookup(
487
    margo_instance_id mid,
Philip Carns's avatar
Philip Carns committed
488
    const char   *name,
489
    hg_addr_t    *addr)
490
{
491
    hg_return_t nret;
Philip Carns's avatar
Philip Carns committed
492
    struct lookup_cb_evt *evt;
493 494 495
    ABT_eventual eventual;
    int ret;

Philip Carns's avatar
Philip Carns committed
496
    ret = ABT_eventual_create(sizeof(*evt), &eventual);
497 498 499 500 501
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);        
    }

Jonathan Jenkins's avatar
Jonathan Jenkins committed
502
    nret = HG_Addr_lookup(mid->hg_context, margo_addr_lookup_cb,
503
        &eventual, name, HG_OP_ID_IGNORE);
504 505
    if(nret == 0)
    {
Philip Carns's avatar
Philip Carns committed
506 507 508
        ABT_eventual_wait(eventual, (void**)&evt);
        *addr = evt->addr;
        nret = evt->nret;
509 510 511 512 513 514 515
    }

    ABT_eventual_free(&eventual);

    return(nret);
}

516
hg_return_t margo_bulk_transfer(
517
    margo_instance_id mid,
518
    hg_bulk_op_t op,
519
    hg_addr_t origin_addr,
520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536
    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);        
    }

Jonathan Jenkins's avatar
Jonathan Jenkins committed
537 538 539
    hret = HG_Bulk_transfer(mid->hg_context, margo_bulk_transfer_cb,
        &eventual, op, origin_addr, origin_handle, origin_offset, local_handle,
        local_offset, size, HG_OP_ID_IGNORE);
540 541 542 543 544 545 546 547 548 549 550
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

551 552
typedef struct
{
553
    margo_instance_id mid;
554 555
    ABT_mutex mutex;
    ABT_cond cond;
Shane Snyder's avatar
Shane Snyder committed
556 557
    char is_asleep;
    char in_pool;
558 559 560 561 562 563 564
} 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;

565 566
    /* decrement number of waiting threads */
    ABT_mutex_lock(sleep_cb_dat->mid->finalize_mutex);
Shane Snyder's avatar
Shane Snyder committed
567 568
    sleep_cb_dat->mid->finalize_waiters_in_progress_pool -=
        sleep_cb_dat->in_pool;
569 570
    ABT_mutex_unlock(sleep_cb_dat->mid->finalize_mutex);

571 572
    /* wake up the sleeping thread */
    ABT_mutex_lock(sleep_cb_dat->mutex);
573
    sleep_cb_dat->is_asleep = 0;
574 575 576 577 578 579 580
    ABT_cond_signal(sleep_cb_dat->cond);
    ABT_mutex_unlock(sleep_cb_dat->mutex);

    return;
}

void margo_thread_sleep(
581
    margo_instance_id mid,
582 583
    double timeout_ms)
{
584
    int in_pool = 0;
585 586 587
    margo_timer_t sleep_timer;
    margo_thread_sleep_cb_dat sleep_cb_dat;

Shane Snyder's avatar
Shane Snyder committed
588 589 590
    if(margo_xstream_is_in_progress_pool(mid))
        in_pool = 1;

591
    /* set data needed for sleep callback */
592
    sleep_cb_dat.mid = mid;
593 594
    ABT_mutex_create(&(sleep_cb_dat.mutex));
    ABT_cond_create(&(sleep_cb_dat.cond));
595
    sleep_cb_dat.is_asleep = 1;
Shane Snyder's avatar
Shane Snyder committed
596
    sleep_cb_dat.in_pool = in_pool;
597 598

    /* initialize the sleep timer */
599
    margo_timer_init(mid, &sleep_timer, margo_thread_sleep_cb,
600 601
        &sleep_cb_dat, timeout_ms);

602 603
    /* increment number of waiting threads */
    ABT_mutex_lock(mid->finalize_mutex);
604
    mid->finalize_waiters_in_progress_pool += in_pool;
605 606
    ABT_mutex_unlock(mid->finalize_mutex);

607 608
    /* yield thread for specified timeout */
    ABT_mutex_lock(sleep_cb_dat.mutex);
609 610
    while(sleep_cb_dat.is_asleep)
        ABT_cond_wait(sleep_cb_dat.cond, sleep_cb_dat.mutex);
611 612 613 614 615
    ABT_mutex_unlock(sleep_cb_dat.mutex);

    return;
}

616
margo_instance_id margo_hg_class_to_instance(hg_class_t *cl)
617 618 619 620 621
{
    int i;

    for(i=0; i<handler_mapping_table_size; i++)
    {
622
        if(handler_mapping_table[i].class == cl)
623 624 625 626
            return(handler_mapping_table[i].mid);
    }
    return(NULL);
}
627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644

/* 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);
}