margo.c 11 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>
14 15

#include "margo.h"
Philip Carns's avatar
Philip Carns committed
16
#include "utlist.h"
17

18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
/* 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

37 38
struct margo_instance
{
39
    /* provided by caller */
40 41
    hg_context_t *hg_context;
    hg_class_t *hg_class;
42 43 44
    ABT_pool handler_pool;
    ABT_pool progress_pool;

45
    /* internal to margo for this particular instance */
46 47
    ABT_thread hg_progress_tid;
    int hg_progress_shutdown_flag;
48 49 50 51 52 53 54

    /* 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;

55 56 57 58
    /* pending operations with timeouts */
    ABT_mutex timer_mutex;
    struct timed_element *timer_head;

59 60 61 62 63 64 65 66 67 68 69 70
    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};
71

72
static void hg_progress_fn(void* foo);
73 74 75 76 77 78 79 80 81


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

82 83
margo_instance_id margo_init(ABT_pool progress_pool, ABT_pool handler_pool,
    hg_context_t *hg_context, hg_class_t *hg_class)
84 85
{
    int ret;
86 87 88
    struct margo_instance *mid;

    if(handler_mapping_table_size >= MAX_HANDLER_MAPPING)
89
        return(MARGO_INSTANCE_NULL);
90 91 92

    mid = malloc(sizeof(*mid));
    if(!mid)
93
        return(MARGO_INSTANCE_NULL);
94
    memset(mid, 0, sizeof(*mid));
95

96 97 98
    ABT_mutex_create(&mid->finalize_mutex);
    ABT_cond_create(&mid->finalize_cond);

99 100
    ABT_mutex_create(&mid->timer_mutex);

101 102
    mid->progress_pool = progress_pool;
    mid->handler_pool = handler_pool;
103 104
    mid->hg_class = hg_class;
    mid->hg_context = hg_context;
105

106
    ret = ABT_thread_create(mid->progress_pool, hg_progress_fn, mid, 
107
        ABT_THREAD_ATTR_NULL, &mid->hg_progress_tid);
108 109 110
    if(ret != 0)
    {
        fprintf(stderr, "Error: ABT_thread_create()\n");
111
        free(mid);
112
        return(MARGO_INSTANCE_NULL);
113 114
    }

115 116 117 118 119 120
    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;
121 122
}

123
void margo_finalize(margo_instance_id mid)
124
{
125 126
    int i;

127
    /* tell progress thread to wrap things up */
128
    mid->hg_progress_shutdown_flag = 1;
129 130

    /* wait for it to shutdown cleanly */
131 132
    ABT_thread_join(mid->hg_progress_tid);
    ABT_thread_free(&mid->hg_progress_tid);
133

134 135 136 137 138
    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--;
139

140 141 142 143 144 145 146 147 148 149 150 151 152 153 154
    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);
155
    ABT_mutex_free(&mid->timer_mutex);
156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190
    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);
    
191 192 193 194
    return;
}

/* dedicated thread function to drive Mercury progress */
195
static void hg_progress_fn(void* foo)
196 197 198
{
    int ret;
    unsigned int actual_count;
199
    struct margo_instance *mid = (struct margo_instance *)foo;
200
    size_t size;
201

202
    while(!mid->hg_progress_shutdown_flag)
203 204
    {
        do {
205
            ret = HG_Trigger(mid->hg_context, 0, 1, &actual_count);
206
        } while((ret == HG_SUCCESS) && actual_count && !mid->hg_progress_shutdown_flag);
207

208
        if(!mid->hg_progress_shutdown_flag)
209
        {
210
            ABT_pool_get_total_size(mid->progress_pool, &size);
211 212 213 214 215 216
            /* 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)
217
            {
218
                HG_Progress(mid->hg_context, 0);
219 220 221 222
                ABT_thread_yield();
            }
            else
            {
223
                HG_Progress(mid->hg_context, 100);
224 225
            }
        }
226 227
    }

228
    return;
229 230
}

231
ABT_pool* margo_get_handler_pool(margo_instance_id mid)
232
{
233
    return(&mid->handler_pool);
234 235
}

Jonathan Jenkins's avatar
Jonathan Jenkins committed
236
static hg_return_t margo_cb(const struct hg_cb_info *info)
237 238 239 240 241 242 243 244 245 246
{
    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);
}

247 248 249 250 251 252
hg_return_t margo_forward_timed(
    margo_instance_id mid,
    hg_handle_t handle,
    void *in_struct,
    double timeout_ms)
{
253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296
    hg_return_t hret = HG_TIMEOUT;
    ABT_eventual eventual;
    int ret;
    hg_return_t* waited_hret;
    struct timed_element el;

    /* calculate expiration time */
    el.handle = handle;
    clock_gettime(clk_id, &el.expiration);
    el.expiration.tv_sec += timeout_ms/1000;
    el.expiration.tv_nsec += (timeout_ms - (timeout_ms/1000))*1000.0*1000.0;
    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);        
    }

    /* track timer */
    /* TODO: sort properly */
    ABT_mutex_lock(mid->timer_mutex);
    DL_APPEND(mid->timer_head, &el);
    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;
    }

    /* remove timer */
    ABT_mutex_lock(mid->timer_mutex);
    DL_DELETE(mid->timer_head, &el);
    ABT_mutex_unlock(mid->timer_mutex);

    ABT_eventual_free(&eventual);

    return(hret);
297 298 299 300

}


301
hg_return_t margo_forward(
302
    margo_instance_id mid,
303 304 305 306 307 308 309 310 311 312 313 314 315 316
    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
317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345
    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);
346 347 348 349 350 351 352 353 354 355 356
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

357

358 359 360 361 362 363 364 365 366 367 368
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
369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389
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);
}


390
na_return_t margo_na_addr_lookup(
391
    margo_instance_id mid,
392 393
    na_class_t   *na_class,
    na_context_t *context,
Philip Carns's avatar
Philip Carns committed
394 395
    const char   *name,
    na_addr_t    *addr)
396 397
{
    na_return_t nret;
Philip Carns's avatar
Philip Carns committed
398
    struct lookup_cb_evt *evt;
399 400 401
    ABT_eventual eventual;
    int ret;

Philip Carns's avatar
Philip Carns committed
402
    ret = ABT_eventual_create(sizeof(*evt), &eventual);
403 404 405 406 407 408 409 410 411
    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
412 413 414
        ABT_eventual_wait(eventual, (void**)&evt);
        *addr = evt->addr;
        nret = evt->nret;
415 416 417 418 419 420 421
    }

    ABT_eventual_free(&eventual);

    return(nret);
}

422
hg_return_t margo_bulk_transfer(
423
    margo_instance_id mid,
424
    hg_context_t *context,
425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457
    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);
}

458
margo_instance_id margo_hg_class_to_instance(hg_class_t *cl)
459 460 461 462 463
{
    int i;

    for(i=0; i<handler_mapping_table_size; i++)
    {
464
        if(handler_mapping_table[i].class == cl)
465 466 467 468
            return(handler_mapping_table[i].mid);
    }
    return(NULL);
}