margo.c 13.2 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"
Philip Carns's avatar
Philip Carns committed
17
#include "utlist.h"
18

19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37
/* TODO: including core.h for cancel definition, presumably this will be 
 * available in top level later?
 */
#include <mercury_core.h>

struct timed_element
{
    struct timespec expiration;
    hg_handle_t handle;
    struct timed_element *next;
    struct timed_element *prev;
};

#ifdef CLOCK_REALTIME_COARSE
clockid_t clk_id = CLOCK_REALTIME_COARSE;
#else
clockid_t clk_id = CLOCK_REALTIME;
#endif

38 39
struct margo_instance
{
40
    /* provided by caller */
41 42
    hg_context_t *hg_context;
    hg_class_t *hg_class;
43 44 45
    ABT_pool handler_pool;
    ABT_pool progress_pool;

46
    /* internal to margo for this particular instance */
47 48
    ABT_thread hg_progress_tid;
    int hg_progress_shutdown_flag;
49 50 51 52 53 54 55

    /* 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;

56 57 58 59
    /* pending operations with timeouts */
    ABT_mutex timer_mutex;
    struct timed_element *timer_head;

60 61 62 63 64 65 66 67 68 69 70 71
    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};
72

73
static void hg_progress_fn(void* foo);
74 75 76 77 78 79 80 81 82


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

83 84
margo_instance_id margo_init(ABT_pool progress_pool, ABT_pool handler_pool,
    hg_context_t *hg_context, hg_class_t *hg_class)
85 86
{
    int ret;
87 88 89
    struct margo_instance *mid;

    if(handler_mapping_table_size >= MAX_HANDLER_MAPPING)
90
        return(MARGO_INSTANCE_NULL);
91 92 93

    mid = malloc(sizeof(*mid));
    if(!mid)
94
        return(MARGO_INSTANCE_NULL);
95
    memset(mid, 0, sizeof(*mid));
96

97 98 99
    ABT_mutex_create(&mid->finalize_mutex);
    ABT_cond_create(&mid->finalize_cond);

100 101
    ABT_mutex_create(&mid->timer_mutex);

102 103
    mid->progress_pool = progress_pool;
    mid->handler_pool = handler_pool;
104 105
    mid->hg_class = hg_class;
    mid->hg_context = hg_context;
106

107
    ret = ABT_thread_create(mid->progress_pool, hg_progress_fn, mid, 
108
        ABT_THREAD_ATTR_NULL, &mid->hg_progress_tid);
109 110 111
    if(ret != 0)
    {
        fprintf(stderr, "Error: ABT_thread_create()\n");
112
        free(mid);
113
        return(MARGO_INSTANCE_NULL);
114 115
    }

116 117 118 119 120 121
    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;
122 123
}

124
void margo_finalize(margo_instance_id mid)
125
{
126 127
    int i;

128
    /* tell progress thread to wrap things up */
129
    mid->hg_progress_shutdown_flag = 1;
130 131

    /* wait for it to shutdown cleanly */
132 133
    ABT_thread_join(mid->hg_progress_tid);
    ABT_thread_free(&mid->hg_progress_tid);
134

135 136 137 138 139
    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--;
140

141 142 143 144 145 146 147 148 149 150 151 152 153 154 155
    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
    ABT_mutex_free(&mid->finalize_mutex);
    ABT_cond_free(&mid->finalize_cond);
156
    ABT_mutex_free(&mid->timer_mutex);
157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191
    free(mid);
#endif

    return;
}

void margo_wait_for_finalize(margo_instance_id mid)
{
    ABT_xstream xstream;
    ABT_pool pool;
    int ret;
    int in_pool = 0;

    ret = ABT_xstream_self(&xstream);
    if(ret != 0)
        return;
    ret = ABT_xstream_get_main_pools(xstream, 1, &pool);
    if(ret != 0)
        return;

    /* Is this waiter in the same pool as the pool running the progress
     * thread?
     */
    if(pool == mid->progress_pool)
        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);
    
192 193 194 195
    return;
}

/* dedicated thread function to drive Mercury progress */
196
static void hg_progress_fn(void* foo)
197 198 199
{
    int ret;
    unsigned int actual_count;
200
    struct margo_instance *mid = (struct margo_instance *)foo;
201
    size_t size;
202 203
    struct timespec now;
    struct timed_element *cur;
204

205
    while(!mid->hg_progress_shutdown_flag)
206 207
    {
        do {
208
            ret = HG_Trigger(mid->hg_context, 0, 1, &actual_count);
209
        } while((ret == HG_SUCCESS) && actual_count && !mid->hg_progress_shutdown_flag);
210

211
        if(!mid->hg_progress_shutdown_flag)
212
        {
213
            ABT_pool_get_total_size(mid->progress_pool, &size);
214 215 216 217 218 219
            /* 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)
220
            {
221
                HG_Progress(mid->hg_context, 0);
222 223 224 225
                ABT_thread_yield();
            }
            else
            {
226
                HG_Progress(mid->hg_context, 100);
227 228
            }
        }
229

230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246
        clock_gettime(clk_id, &now);

        ABT_mutex_lock(mid->timer_mutex);
        while(mid->timer_head && 
            (mid->timer_head->expiration.tv_sec < now.tv_sec ||
             (mid->timer_head->expiration.tv_sec == now.tv_sec &&
              mid->timer_head->expiration.tv_nsec < now.tv_nsec)))
        {
            cur = mid->timer_head;
            DL_DELETE(mid->timer_head, cur);
            cur->next = NULL;
            cur->prev = NULL;
            /* TODO: actually cancel here, of course */
            printf("FOO: I would like to cancel operation with tv_sec %ld, tv_nsec %ld\n", (long)cur->expiration.tv_sec, cur->expiration.tv_nsec);
        }
        ABT_mutex_unlock(mid->timer_mutex);

247 248 249 250
        /* TODO: check for timeouts here.  If timer_head not null, then check
         * current time and compare against first element.  Keep walking list
         * cancelling operations until we find non-expired element.
         */
251 252
    }

253
    return;
254 255
}

256
ABT_pool* margo_get_handler_pool(margo_instance_id mid)
257
{
258
    return(&mid->handler_pool);
259 260
}

Jonathan Jenkins's avatar
Jonathan Jenkins committed
261
static hg_return_t margo_cb(const struct hg_cb_info *info)
262 263 264 265 266 267 268 269 270 271
{
    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);
}

272 273 274 275 276 277
hg_return_t margo_forward_timed(
    margo_instance_id mid,
    hg_handle_t handle,
    void *in_struct,
    double timeout_ms)
{
278 279 280 281 282
    hg_return_t hret = HG_TIMEOUT;
    ABT_eventual eventual;
    int ret;
    hg_return_t* waited_hret;
    struct timed_element el;
Philip Carns's avatar
Philip Carns committed
283
    struct timed_element *cur;
284 285 286

    /* calculate expiration time */
    el.handle = handle;
Philip Carns's avatar
Philip Carns committed
287 288
    el.prev = NULL;
    el.next = NULL;
289 290
    clock_gettime(clk_id, &el.expiration);
    el.expiration.tv_sec += timeout_ms/1000;
Philip Carns's avatar
bug fix  
Philip Carns committed
291
    el.expiration.tv_nsec += fmod(timeout_ms, 1000)*1000.0*1000.0;
292 293 294 295 296 297 298 299 300 301 302 303
    if(el.expiration.tv_nsec > 1000000000)
    {
        el.expiration.tv_nsec -= 1000000000;
        el.expiration.tv_sec++;
    }

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

Philip Carns's avatar
Philip Carns committed
304
    /* TODO: split this out into a subroutine */
305 306
    /* track timer */
    ABT_mutex_lock(mid->timer_mutex);
Philip Carns's avatar
Philip Carns committed
307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338

    /* if queue of expiring ops is empty, put ourselves on it */
    if(!mid->timer_head)
        DL_APPEND(mid->timer_head, &el);
    else
    {
        /* something else already in queue, keep it sorted in ascending order
         * of expiration time
         */
        cur = mid->timer_head;
        do
        {
            /* walk backwards through queue */
            cur = cur->prev;
            /* as soon as we find an element that expires before this one, 
             * then we add ours after it
             */
            if(cur->expiration.tv_sec < el.expiration.tv_sec ||
                (cur->expiration.tv_sec == el.expiration.tv_sec &&
                 cur->expiration.tv_nsec < el.expiration.tv_nsec))
            {
                DL_APPEND_ELEM(mid->timer_head, cur, &el);
                break;
            }
        }while(cur != mid->timer_head);

        /* if we never found one with an expiration before this one, then
         * this one is the new head
         */
        if(el.prev == NULL && el.next == NULL)
            DL_PREPEND(mid->timer_head, &el);
    }
339 340 341 342 343 344 345 346 347
    ABT_mutex_unlock(mid->timer_mutex);

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

348
    /* remove timer if it is still in place */
349
    ABT_mutex_lock(mid->timer_mutex);
350 351
    if(el.prev || el.next)
        DL_DELETE(mid->timer_head, &el);
352 353 354 355 356
    ABT_mutex_unlock(mid->timer_mutex);

    ABT_eventual_free(&eventual);

    return(hret);
357 358 359 360

}


361
hg_return_t margo_forward(
362
    margo_instance_id mid,
363 364 365 366 367 368 369 370 371 372 373 374 375 376
    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
377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405
    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);
406 407 408 409 410 411 412 413 414 415 416
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

417

418 419 420 421 422 423 424 425 426 427 428
static hg_return_t margo_bulk_transfer_cb(const struct hg_bulk_cb_info *hg_bulk_cb_info)
{
    hg_return_t hret = hg_bulk_cb_info->ret;
    ABT_eventual *eventual = hg_bulk_cb_info->arg;

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

Philip Carns's avatar
Philip Carns committed
429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449
struct lookup_cb_evt
{
    na_return_t nret;
    na_addr_t addr;
};

static na_return_t margo_na_addr_lookup_cb(const struct na_cb_info *callback_info)
{
    struct lookup_cb_evt evt;
    evt.nret = callback_info->ret;
    evt.addr = callback_info->info.lookup.addr;

    ABT_eventual *eventual = callback_info->arg;

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


450
na_return_t margo_na_addr_lookup(
451
    margo_instance_id mid,
452 453
    na_class_t   *na_class,
    na_context_t *context,
Philip Carns's avatar
Philip Carns committed
454 455
    const char   *name,
    na_addr_t    *addr)
456 457
{
    na_return_t nret;
Philip Carns's avatar
Philip Carns committed
458
    struct lookup_cb_evt *evt;
459 460 461
    ABT_eventual eventual;
    int ret;

Philip Carns's avatar
Philip Carns committed
462
    ret = ABT_eventual_create(sizeof(*evt), &eventual);
463 464 465 466 467 468 469 470 471
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);        
    }

    nret = NA_Addr_lookup(na_class, context, margo_na_addr_lookup_cb,
        &eventual, name, NA_OP_ID_IGNORE);
    if(nret == 0)
    {
Philip Carns's avatar
Philip Carns committed
472 473 474
        ABT_eventual_wait(eventual, (void**)&evt);
        *addr = evt->addr;
        nret = evt->nret;
475 476 477 478 479 480 481
    }

    ABT_eventual_free(&eventual);

    return(nret);
}

482
hg_return_t margo_bulk_transfer(
483
    margo_instance_id mid,
484
    hg_context_t *context,
485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517
    hg_bulk_op_t op,
    na_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)
{
    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);
}

518
margo_instance_id margo_hg_class_to_instance(hg_class_t *cl)
519 520 521 522 523
{
    int i;

    for(i=0; i<handler_mapping_table_size; i++)
    {
524
        if(handler_mapping_table[i].class == cl)
525 526 527 528
            return(handler_mapping_table[i].mid);
    }
    return(NULL);
}