margo.c 10.9 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
/* TODO: including core.h for cancel definition, presumably this will be 
 * available in top level later?
 */
#include <mercury_core.h>

24 25
struct margo_instance
{
26
    /* provided by caller */
27 28
    hg_context_t *hg_context;
    hg_class_t *hg_class;
29 30 31
    ABT_pool handler_pool;
    ABT_pool progress_pool;

32
    /* internal to margo for this particular instance */
33 34
    ABT_thread hg_progress_tid;
    int hg_progress_shutdown_flag;
35 36 37 38 39 40 41

    /* 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;

42 43 44 45 46 47 48 49 50 51 52 53
    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};
54

55
static void hg_progress_fn(void* foo);
56 57 58 59 60 61 62 63 64


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

65 66
margo_instance_id margo_init(ABT_pool progress_pool, ABT_pool handler_pool,
    hg_context_t *hg_context, hg_class_t *hg_class)
67 68
{
    int ret;
69 70 71
    struct margo_instance *mid;

    if(handler_mapping_table_size >= MAX_HANDLER_MAPPING)
72
        return(MARGO_INSTANCE_NULL);
73 74 75

    mid = malloc(sizeof(*mid));
    if(!mid)
76
        return(MARGO_INSTANCE_NULL);
77
    memset(mid, 0, sizeof(*mid));
78

79 80 81
    ABT_mutex_create(&mid->finalize_mutex);
    ABT_cond_create(&mid->finalize_cond);

82
    margo_timer_init();
83

84 85
    mid->progress_pool = progress_pool;
    mid->handler_pool = handler_pool;
86 87
    mid->hg_class = hg_class;
    mid->hg_context = hg_context;
88

89
    ret = ABT_thread_create(mid->progress_pool, hg_progress_fn, mid, 
90
        ABT_THREAD_ATTR_NULL, &mid->hg_progress_tid);
91 92 93
    if(ret != 0)
    {
        fprintf(stderr, "Error: ABT_thread_create()\n");
94
        free(mid);
95
        return(MARGO_INSTANCE_NULL);
96 97
    }

98 99 100 101 102 103
    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;
104 105
}

106
void margo_finalize(margo_instance_id mid)
107
{
108 109
    int i;

110
    /* tell progress thread to wrap things up */
111
    mid->hg_progress_shutdown_flag = 1;
112 113

    /* wait for it to shutdown cleanly */
114 115
    ABT_thread_join(mid->hg_progress_tid);
    ABT_thread_free(&mid->hg_progress_tid);
116

117 118 119 120 121
    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--;
122

123 124 125 126 127 128 129 130 131 132 133 134 135
    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
136 137
    margo_timer_cleanup();

138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174
    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)
{
    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);
    
175 176 177 178
    return;
}

/* dedicated thread function to drive Mercury progress */
179
static void hg_progress_fn(void* foo)
180 181 182
{
    int ret;
    unsigned int actual_count;
183
    struct margo_instance *mid = (struct margo_instance *)foo;
184
    size_t size;
185

186
    while(!mid->hg_progress_shutdown_flag)
187 188
    {
        do {
189
            ret = HG_Trigger(mid->hg_context, 0, 1, &actual_count);
190
        } while((ret == HG_SUCCESS) && actual_count && !mid->hg_progress_shutdown_flag);
191

192
        if(!mid->hg_progress_shutdown_flag)
193
        {
194
            ABT_pool_get_total_size(mid->progress_pool, &size);
195 196 197 198 199 200
            /* 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)
201
            {
202
                HG_Progress(mid->hg_context, 0);
203 204 205 206
                ABT_thread_yield();
            }
            else
            {
207
                HG_Progress(mid->hg_context, 100);
208 209
            }
        }
210

211 212
        /* check for any expired timers */
        margo_check_timers();
213 214
    }

215
    return;
216 217
}

218
ABT_pool* margo_get_handler_pool(margo_instance_id mid)
219
{
220
    return(&mid->handler_pool);
221 222
}

223 224 225 226 227 228 229 230 231 232 233
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
234
static hg_return_t margo_cb(const struct hg_cb_info *info)
235 236 237 238 239 240 241 242 243 244
{
    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);
}

245 246 247 248 249 250 251 252 253 254 255 256 257 258 259
typedef struct
{
    hg_handle_t handle;
} margo_hg_cancel_cb_dat;

static void margo_forward_timeout_cb(void *arg)
{
    margo_hg_cancel_cb_dat *cb_dat =
        (margo_hg_cancel_cb_dat *)arg;

    /* cancel the Mercury op if the forward timed out */
    HG_Core_cancel(cb_dat->handle);
    return;
}

260 261 262 263 264 265
hg_return_t margo_forward_timed(
    margo_instance_id mid,
    hg_handle_t handle,
    void *in_struct,
    double timeout_ms)
{
266 267 268 269
    hg_return_t hret = HG_TIMEOUT;
    ABT_eventual eventual;
    int ret;
    hg_return_t* waited_hret;
270 271
    margo_hg_cancel_cb_dat *cb_dat;
    margo_timer_handle timer_handle;
272 273 274 275 276 277 278

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

279 280 281 282 283 284 285
    /* create a timer that expires when this request has timed out */
    cb_dat = malloc(sizeof(margo_hg_cancel_cb_dat));
    assert(cb_dat);
    memset(cb_dat, 0, sizeof(*cb_dat));
    cb_dat->handle = handle;
    margo_timer_create(margo_forward_timeout_cb, cb_dat,
        timeout_ms, &timer_handle);
286 287 288 289 290 291 292 293

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

294 295 296
    /* remove timer if it is still in place (i.e., not timed out) */
    if(hret != HG_TIMEOUT)
        margo_timer_free(timer_handle);
297 298 299 300

    ABT_eventual_free(&eventual);

    return(hret);
301 302 303 304

}


305
hg_return_t margo_forward(
306
    margo_instance_id mid,
307 308 309 310 311 312 313 314 315 316 317 318 319 320
    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
321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349
    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);
350 351 352 353 354 355 356 357 358 359 360
    if(hret == 0)
    {
        ABT_eventual_wait(eventual, (void**)&waited_hret);
        hret = *waited_hret;
    }

    ABT_eventual_free(&eventual);

    return(hret);
}

361

362
static hg_return_t margo_bulk_transfer_cb(const struct hg_cb_info *info)
363
{
364 365
    hg_return_t hret = info->ret;
    ABT_eventual *eventual = info->arg;
366 367 368 369 370 371 372

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

Philip Carns's avatar
Philip Carns committed
373 374
struct lookup_cb_evt
{
375 376
    hg_return_t nret;
    hg_addr_t addr;
Philip Carns's avatar
Philip Carns committed
377 378
};

379
static hg_return_t margo_addr_lookup_cb(const struct hg_cb_info *info)
Philip Carns's avatar
Philip Carns committed
380 381
{
    struct lookup_cb_evt evt;
382 383
    evt.nret = info->ret;
    evt.addr = info->info.lookup.addr;
Philip Carns's avatar
Philip Carns committed
384

385
    ABT_eventual *eventual = info->arg;
Philip Carns's avatar
Philip Carns committed
386 387 388 389

    /* propagate return code out through eventual */
    ABT_eventual_set(*eventual, &evt, sizeof(evt));
    
390
    return(HG_SUCCESS);
Philip Carns's avatar
Philip Carns committed
391 392 393
}


394
hg_return_t margo_addr_lookup(
395
    margo_instance_id mid,
396
    hg_context_t *context,
Philip Carns's avatar
Philip Carns committed
397
    const char   *name,
398
    hg_addr_t    *addr)
399
{
400
    hg_return_t nret;
Philip Carns's avatar
Philip Carns committed
401
    struct lookup_cb_evt *evt;
402 403 404
    ABT_eventual eventual;
    int ret;

Philip Carns's avatar
Philip Carns committed
405
    ret = ABT_eventual_create(sizeof(*evt), &eventual);
406 407 408 409 410
    if(ret != 0)
    {
        return(HG_NOMEM_ERROR);        
    }

411 412
    nret = HG_Addr_lookup(context, margo_addr_lookup_cb,
        &eventual, name, HG_OP_ID_IGNORE);
413 414
    if(nret == 0)
    {
Philip Carns's avatar
Philip Carns committed
415 416 417
        ABT_eventual_wait(eventual, (void**)&evt);
        *addr = evt->addr;
        nret = evt->nret;
418 419 420 421 422 423 424
    }

    ABT_eventual_free(&eventual);

    return(nret);
}

425
hg_return_t margo_bulk_transfer(
426
    margo_instance_id mid,
427
    hg_context_t *context,
428
    hg_bulk_op_t op,
429
    hg_addr_t origin_addr,
430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460
    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);
}

461
margo_instance_id margo_hg_class_to_instance(hg_class_t *cl)
462 463 464 465 466
{
    int i;

    for(i=0; i<handler_mapping_table_size; i++)
    {
467
        if(handler_mapping_table[i].class == cl)
468 469 470 471
            return(handler_mapping_table[i].mid);
    }
    return(NULL);
}