ssg.c 22.7 KB
Newer Older
1 2 3 4 5 6
/*
 * Copyright (c) 2016 UChicago Argonne, LLC
 *
 * See COPYRIGHT in top-level directory.
 */

Shane Snyder's avatar
Shane Snyder committed
7 8
#include <ssg-config.h>

Jonathan Jenkins's avatar
Jonathan Jenkins committed
9 10 11 12
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <unistd.h>
13
#include <errno.h>
Jonathan Jenkins's avatar
Jonathan Jenkins committed
14 15
#include <stdlib.h>
#include <string.h>
Jonathan Jenkins's avatar
Jonathan Jenkins committed
16
#include <assert.h>
Jonathan Jenkins's avatar
Jonathan Jenkins committed
17

Shane Snyder's avatar
Shane Snyder committed
18
#include <mercury.h>
Jonathan Jenkins's avatar
Jonathan Jenkins committed
19
#include <mercury_proc.h>
Jonathan Jenkins's avatar
Jonathan Jenkins committed
20
#include <mercury_macros.h>
Shane Snyder's avatar
Shane Snyder committed
21 22
#include <abt.h>
#include <margo.h>
23
#include <ssg.h>
Jonathan Jenkins's avatar
Jonathan Jenkins committed
24 25
#include "def.h"

26
#define DO_DEBUG 0
Jonathan Jenkins's avatar
Jonathan Jenkins committed
27 28 29 30 31 32 33 34
#define DEBUG(...) \
    do { \
        if(DO_DEBUG) { \
            printf(__VA_ARGS__); \
            fflush(stdout); \
        } \
    } while(0)

Shane Snyder's avatar
Shane Snyder committed
35
// internal initialization of ssg data structures
36 37 38 39 40
static ssg_t ssg_init_internal(margo_instance_id mid, hg_addr_t self_addr,
    int rank, int num_addrs, char *addr_buf, int addr_buf_size);

// lookup peer addresses
static hg_return_t ssg_lookup(ssg_t s);
Shane Snyder's avatar
Shane Snyder committed
41

Jonathan Jenkins's avatar
Jonathan Jenkins committed
42 43
static char** setup_addr_str_list(int num_addrs, char * buf);

44
#if 0
Jonathan Jenkins's avatar
Jonathan Jenkins committed
45 46 47 48 49 50 51 52 53 54 55
// helper for hashing (don't want to pull in jenkins hash)
// see http://www.isthe.com/chongo/tech/comp/fnv/index.html
static uint64_t fnv1a_64(void *data, size_t size);

MERCURY_GEN_PROC(barrier_in_t,
        ((int32_t)(barrier_id)) \
        ((int32_t)(rank)))

// barrier RPC decls
static void proc_barrier(void *arg);
DEFINE_MARGO_RPC_HANDLER(proc_barrier)
56
#endif
Jonathan Jenkins's avatar
Jonathan Jenkins committed
57

58
ssg_t ssg_init_config(margo_instance_id mid, const char * fname, int is_member)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76
{
    // file to read
    int fd = -1;
    struct stat st;

    // file content to parse
    char *rdbuf = NULL;
    ssize_t rdsz;

    // parse metadata (strtok)
    char *tok;

    // vars to build up the addr string list
    int addr_cap = 128;
    int addr_len = 0;
    int num_addrs = 0;
    void *buf = NULL;

Shane Snyder's avatar
Shane Snyder committed
77
    // self rank/addr resolution helpers
78
    hg_class_t *hgcl = NULL;
Shane Snyder's avatar
Shane Snyder committed
79 80 81 82 83 84 85 86
    hg_addr_t self_addr = HG_ADDR_NULL;
    char * self_addr_str = NULL;
    const char * self_addr_substr = NULL;
    hg_size_t self_addr_size = 0;
    const char * addr_substr = NULL;
    // TODO: what is the purpose of an "external" rank (is_member flag)?
    int rank = is_member ? SSG_RANK_UNKNOWN : SSG_EXTERNAL_RANK;

Jonathan Jenkins's avatar
Jonathan Jenkins committed
87 88
    // misc return codes
    int ret;
89 90 91 92
    hg_return_t hret;

    // return data
    ssg_t s = NULL;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
93 94 95 96 97 98 99 100 101 102 103 104 105 106 107

    // open file for reading
    fd = open(fname, O_RDONLY);
    if (fd == -1) goto fini;

    // get file size
    ret = fstat(fd, &st);
    if (ret == -1) goto fini;

    // slurp file in all at once
    rdbuf = malloc(st.st_size+1);
    if (rdbuf == NULL) goto fini;

    // load it all in one fell swoop
    rdsz = read(fd, rdbuf, st.st_size);
Shane Snyder's avatar
Shane Snyder committed
108
    if (rdsz != st.st_size) goto fini;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
109 110
    rdbuf[rdsz]='\0';

111 112 113
    hgcl = margo_get_class(mid);
    if(!hgcl) goto fini;

Shane Snyder's avatar
Shane Snyder committed
114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132
    if(is_member) {
        // get my address
        hret = HG_Addr_self(hgcl, &self_addr);
        if (hret != HG_SUCCESS) goto fini;
        hret = HG_Addr_to_string(hgcl, NULL, &self_addr_size, self_addr);
        if (hret != HG_SUCCESS) goto fini;
        self_addr_str = malloc(self_addr_size);
        if (self_addr_str == NULL) goto fini;
        hret = HG_Addr_to_string(hgcl, self_addr_str, &self_addr_size, self_addr);
        if (hret != HG_SUCCESS) goto fini;

        // strstr is used here b/c there may be inconsistencies in whether the class
        // is included in the address or not (it's not in HG_Addr_to_string, it
        // should be in ssg_init_config)
        self_addr_substr = strstr(self_addr_str, "://");
        if (self_addr_substr == NULL) goto fini;
        self_addr_substr += 3;
    }

Jonathan Jenkins's avatar
Jonathan Jenkins committed
133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149
    // strtok the result - each space-delimited address is assumed to be
    // a unique mercury address
    tok = strtok(rdbuf, "\r\n\t ");
    if (tok == NULL) goto fini;

    // build up the address buffer
    buf = malloc(addr_cap);
    if (buf == NULL) goto fini;
    do {
        int tok_sz = strlen(tok);
        if (tok_sz + addr_len + 1 > addr_cap) {
            void * tmp;
            addr_cap *= 2;
            tmp = realloc(buf, addr_cap);
            if (tmp == NULL) goto fini;
            buf = tmp;
        }
Shane Snyder's avatar
Shane Snyder committed
150 151 152 153 154 155 156 157
        if(is_member) {
            // check if this is my addr to resolve rank
            addr_substr = strstr(tok, "://");
            if (addr_substr == NULL) goto fini;
            addr_substr+= 3;
            if (strcmp(self_addr_substr, addr_substr) == 0)
                rank = num_addrs;
        }
Jonathan Jenkins's avatar
Jonathan Jenkins committed
158 159 160 161 162 163
        memcpy((char*)buf + addr_len, tok, tok_sz+1);
        addr_len += tok_sz+1;
        num_addrs++;
        tok = strtok(NULL, "\r\n\t ");
    } while (tok != NULL);

Shane Snyder's avatar
Shane Snyder committed
164
    // init ssg internal structures
165
    s = ssg_init_internal(mid, self_addr, rank, num_addrs, buf, addr_len);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
166 167
    if (s == NULL) goto fini;

Shane Snyder's avatar
Shane Snyder committed
168 169 170
    // don't free this on success
    buf = NULL;
    self_addr = HG_ADDR_NULL;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
171 172 173 174
fini:
    if (fd != -1) close(fd);
    free(rdbuf);
    free(buf);
175
    if (hgcl && self_addr != HG_ADDR_NULL) HG_Addr_free(hgcl, self_addr);
Shane Snyder's avatar
Shane Snyder committed
176
    free(self_addr_str);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
177 178 179 180
    return s;
}

#ifdef HAVE_MPI
181
ssg_t ssg_init_mpi(margo_instance_id mid, MPI_Comm comm)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
182 183
{
    // my addr
184
    hg_class_t *hgcl = NULL;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
185
    hg_addr_t self_addr = HG_ADDR_NULL;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
186
    char * self_addr_str = NULL;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
187
    hg_size_t self_addr_size = 0;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
188 189 190 191 192 193 194 195 196
    int self_addr_size_int = 0; // for mpi-friendly conversion

    // collective helpers
    char * buf = NULL;
    int * sizes = NULL;
    int * sizes_psum = NULL;
    int comm_size = 0;
    int comm_rank = 0;

197 198 199
    // misc return codes
    hg_return_t hret;

Jonathan Jenkins's avatar
Jonathan Jenkins committed
200 201 202
    // return data
    ssg_t s = NULL;

203 204
    hgcl = margo_get_class(mid);
    if(!hgcl) goto fini;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
205 206

    // get my address
Jonathan Jenkins's avatar
Jonathan Jenkins committed
207 208 209
    hret = HG_Addr_self(hgcl, &self_addr);
    if (hret != HG_SUCCESS) goto fini;
    hret = HG_Addr_to_string(hgcl, NULL, &self_addr_size, self_addr);
Shane Snyder's avatar
Shane Snyder committed
210
    if (hret != HG_SUCCESS) goto fini;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
211 212
    self_addr_str = malloc(self_addr_size);
    if (self_addr_str == NULL) goto fini;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
213 214
    hret = HG_Addr_to_string(hgcl, self_addr_str, &self_addr_size, self_addr);
    if (hret != HG_SUCCESS) goto fini;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238
    self_addr_size_int = (int)self_addr_size; // null char included in call

    // gather the buffer sizes
    MPI_Comm_size(comm, &comm_size);
    MPI_Comm_rank(comm, &comm_rank);
    sizes = malloc(comm_size * sizeof(*sizes));
    if (sizes == NULL) goto fini;
    sizes[comm_rank] = self_addr_size_int;
    MPI_Allgather(MPI_IN_PLACE, 0, MPI_BYTE, sizes, 1, MPI_INT, comm);

    // compute a exclusive prefix sum of the data sizes,
    // including the total at the end
    sizes_psum = malloc((comm_size+1) * sizeof(*sizes_psum));
    if (sizes_psum == NULL) goto fini;
    sizes_psum[0] = 0;
    for (int i = 1; i < comm_size+1; i++)
        sizes_psum[i] = sizes_psum[i-1] + sizes[i-1];

    // allgather the addresses
    buf = malloc(sizes_psum[comm_size]);
    if (buf == NULL) goto fini;
    MPI_Allgatherv(self_addr_str, self_addr_size_int, MPI_BYTE,
            buf, sizes, sizes_psum, MPI_BYTE, comm);

Shane Snyder's avatar
Shane Snyder committed
239
    // init ssg internal structures
240
    s = ssg_init_internal(mid, self_addr, comm_rank, comm_size,
Shane Snyder's avatar
Shane Snyder committed
241 242 243 244 245
        buf, sizes_psum[comm_size]);
    if (s == NULL) goto fini;

    // don't free these on success
    buf = NULL;
246
    self_addr = HG_ADDR_NULL;
Shane Snyder's avatar
Shane Snyder committed
247 248 249 250
fini:
    free(sizes);
    free(sizes_psum);
    free(buf);
251 252
    if (hgcl && self_addr != HG_ADDR_NULL) HG_Addr_free(hgcl, self_addr);
    free(self_addr_str);
Shane Snyder's avatar
Shane Snyder committed
253 254 255 256
    return s;
}
#endif

257 258
static ssg_t ssg_init_internal(margo_instance_id mid, hg_addr_t self_addr,
    int rank, int num_addrs, char *addr_buf, int addr_buf_size)
Shane Snyder's avatar
Shane Snyder committed
259 260 261 262 263
{
    // arrays of peer address strings and addresses
    char **addr_strs = NULL;
    hg_addr_t *addrs = NULL;

264 265 266
    // misc return codes
    hg_return_t hret;

Shane Snyder's avatar
Shane Snyder committed
267 268 269
    // return data
    ssg_t s = NULL;

270 271
    if (rank == SSG_RANK_UNKNOWN) goto fini;
    if (self_addr == HG_ADDR_NULL && rank != SSG_EXTERNAL_RANK) goto fini;
Shane Snyder's avatar
Shane Snyder committed
272 273 274

    // set peer address strings
    addr_strs = setup_addr_str_list(num_addrs, addr_buf);
275
    if (addr_strs == NULL) goto fini;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
276 277

    // init peer addresses
Shane Snyder's avatar
Shane Snyder committed
278
    addrs = malloc(num_addrs*sizeof(*addrs));
Jonathan Jenkins's avatar
Jonathan Jenkins committed
279
    if (addrs == NULL) goto fini;
Shane Snyder's avatar
Shane Snyder committed
280
    for (int i = 0; i < num_addrs; i++) addrs[i] = HG_ADDR_NULL;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
281 282 283 284

    // set up the output
    s = malloc(sizeof(*s));
    if (s == NULL) goto fini;
285 286 287
    s->mid = mid;
    s->rank = rank;
    s->num_addrs = num_addrs;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
288 289
    s->addr_strs = addr_strs; addr_strs = NULL;
    s->addrs = addrs; addrs = NULL;
Shane Snyder's avatar
Shane Snyder committed
290 291 292
    s->backing_buf = addr_buf;
    s->buf_size = addr_buf_size;
    s->addrs[rank] = self_addr; // NOTE: remaining addrs are set in ssg_lookup
293 294 295 296 297 298
    s->barrier_rpc_id = 0;
    s->barrier_id = 0;
    s->barrier_count = 0;
    s->barrier_mutex = ABT_MUTEX_NULL;
    s->barrier_cond  = ABT_COND_NULL;
    s->barrier_eventual = ABT_EVENTUAL_NULL;
299
#if USE_SWIM_FD
300
    s->swim_ctx = NULL;
301
#endif
302 303 304 305 306 307 308 309

    // lookup hg addr information for all group members
    hret = ssg_lookup(s);
    if (hret != HG_SUCCESS)
    {
        ssg_finalize(s); s = NULL;
        goto fini;
    }
Jonathan Jenkins's avatar
Jonathan Jenkins committed
310

311
#if USE_SWIM_FD
312 313 314 315 316 317 318 319 320 321 322
    // initialize swim failure detector
    if (s->rank != SSG_EXTERNAL_RANK)
    {
        s->swim_ctx = swim_init(s->mid, s, 1);
        if (s->swim_ctx == NULL)
        {
            ssg_finalize(s); s = NULL;
        }
    }
#endif
    
Jonathan Jenkins's avatar
Jonathan Jenkins committed
323 324 325 326 327 328
fini:
    free(addr_strs);
    free(addrs);
    return s;
}

329 330 331 332 333 334 335 336 337 338 339
struct lookup_ult_args
{
    ssg_t ssg;
    int rank;
    hg_return_t out;
};

static void lookup_ult(void *arg)
{
    struct lookup_ult_args *l = arg;

Jonathan Jenkins's avatar
Jonathan Jenkins committed
340
    DEBUG("%d (ult): looking up rank %d\n", l->ssg->rank, l->rank);
341
    l->out = margo_addr_lookup(l->ssg->mid, l->ssg->addr_strs[l->rank],
342
            &l->ssg->addrs[l->rank]);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
343
    DEBUG("%d (ult): looked up rank %d\n", l->ssg->rank, l->rank);
344 345
}

346
static hg_return_t ssg_lookup(ssg_t s)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
347 348
{
    hg_context_t *hgctx;
349 350 351
    ABT_thread *ults;
    struct lookup_ult_args *args;
    hg_return_t hret = HG_SUCCESS;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
352 353 354 355

    // "effective" rank for the lookup loop
    int eff_rank = 0;

Jonathan Jenkins's avatar
Jonathan Jenkins committed
356 357
    DEBUG("%d: entered lookup\n", s->rank);

Jonathan Jenkins's avatar
Jonathan Jenkins committed
358
    // set the hg class up front - need for destructing addrs
Jonathan Jenkins's avatar
Jonathan Jenkins committed
359
    hgctx = margo_get_context(s->mid);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
360 361 362 363 364 365 366 367 368 369
    if (hgctx == NULL) return HG_INVALID_PARAM;

    if (s->rank == SSG_EXTERNAL_RANK) {
        // do a completely arbitrary effective rank determination to try and
        // prevent everyone talking to the same member at once
        eff_rank = (((intptr_t)hgctx)/sizeof(hgctx)) % s->num_addrs;
    }
    else
        eff_rank = s->rank;

370 371 372 373 374 375 376 377 378 379 380
    // initialize ULTs
    ults = malloc(s->num_addrs * sizeof(*ults));
    if (ults == NULL) return HG_NOMEM_ERROR;
    args = malloc(s->num_addrs * sizeof(*args));
    if (args == NULL) {
        free(ults);
        return HG_NOMEM_ERROR;
    }
    for (int i = 0; i < s->num_addrs; i++)
        ults[i] = ABT_THREAD_NULL;

Jonathan Jenkins's avatar
Jonathan Jenkins committed
381 382 383 384 385 386 387 388 389 390
    DEBUG("%d: beginning thread create loop (%d...%d)\n", s->rank,
            (s->rank!=SSG_EXTERNAL_RANK), s->num_addrs);
    for (int i = (s->rank!=SSG_EXTERNAL_RANK); i < s->num_addrs; i++) {
        int r = (eff_rank+i) % s->num_addrs;
        args[r].ssg = s;
        args[r].rank = r;

        DEBUG("%d: lookup: create thread for rank %d\n", s->rank, r);
        int aret = ABT_thread_create(*margo_get_handler_pool(s->mid), &lookup_ult,
                &args[r], ABT_THREAD_ATTR_NULL, &ults[r]);
391 392
        if (aret != ABT_SUCCESS) {
            hret = HG_OTHER_ERROR;
393
            goto fini;
394
        }
Jonathan Jenkins's avatar
Jonathan Jenkins committed
395
    }
396

Jonathan Jenkins's avatar
Jonathan Jenkins committed
397 398 399 400 401
    // wait on all
    for (int i = (s->rank!=SSG_EXTERNAL_RANK); i < s->num_addrs; i++) {
        int r = (eff_rank+i) % s->num_addrs;
        DEBUG("%d: lookup: join thread for rank %d\n", s->rank, r);
        int aret = ABT_thread_join(ults[r]);
402
        DEBUG("%d: lookup: join thread for rank %d fini\n", s->rank, r);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
403 404
        ABT_thread_free(&ults[r]);
        ults[r] = ABT_THREAD_NULL; // in case of cascading failure from join
405 406
        if (aret != ABT_SUCCESS) {
            hret = HG_OTHER_ERROR;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
407
            break;
408
        }
Jonathan Jenkins's avatar
Jonathan Jenkins committed
409 410 411
        else if (args[r].out != HG_SUCCESS) {
            hret = args[r].out;
            break;
412 413 414
        }
    }

415
fini:
416 417 418 419
    // cleanup
    if (ults != NULL) {
        for (int i = 0; i < s->num_addrs; i++) {
            if (ults[i] != ABT_THREAD_NULL) {
Jonathan Jenkins's avatar
Jonathan Jenkins committed
420
                DEBUG("%d: lookup failed, cancelling ULT %d\n", s->rank, i);
421 422 423 424 425 426 427 428 429
                ABT_thread_cancel(ults[i]);
                ABT_thread_free(ults[i]);
            }
        }
        free(ults);
    }
    if (args != NULL) free(args);

    return hret;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
430
}
Jonathan Jenkins's avatar
Jonathan Jenkins committed
431

432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452
#if 0
// TODO: handle hash collision, misc errors
void ssg_register_barrier(ssg_t s, hg_class_t *hgcl)
{
    if (s->num_addrs == 1) return;

    s->barrier_rpc_id = fnv1a_64(s->backing_buf, s->buf_size);
    hg_return_t hret = HG_Register(hgcl, s->barrier_rpc_id,
            hg_proc_barrier_in_t, NULL, &proc_barrier_handler);
    assert(hret == HG_SUCCESS);
    hret = HG_Register_data(hgcl, s->barrier_rpc_id, s, NULL);
    assert(hret == HG_SUCCESS);

    int aret = ABT_mutex_create(&s->barrier_mutex);
    assert(aret == ABT_SUCCESS);
    aret = ABT_cond_create(&s->barrier_cond);
    assert(aret == ABT_SUCCESS);
    aret = ABT_eventual_create(0, &s->barrier_eventual);
    assert(aret == ABT_SUCCESS);
}

Jonathan Jenkins's avatar
Jonathan Jenkins committed
453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564
// TODO: process errors in a sane manner
static void proc_barrier(void *arg)
{
    barrier_in_t in;
    hg_return_t hret;
    int aret;
    hg_handle_t h = arg;
    struct hg_info *info = HG_Get_info(h);
    ssg_t s = HG_Registered_data(info->hg_class, info->id);

    assert(s->rank == 0);

    hret = HG_Get_input(h, &in);
    assert(hret == HG_SUCCESS);

    DEBUG("%d: barrier ult: rx round %d from %d\n", s->rank, in.barrier_id,
            in.rank);
    // first wait until the nth barrier has been processed
    aret = ABT_mutex_lock(s->barrier_mutex);
    assert(aret == ABT_SUCCESS);
    while (s->barrier_id < in.barrier_id) {
        DEBUG("%d: barrier ult: waiting to enter round %d\n", s->rank,
                in.barrier_id);
        aret = ABT_cond_wait(s->barrier_cond, s->barrier_mutex);
        assert(aret == ABT_SUCCESS);
    }

    // inform all other ULTs waiting on this
    aret = ABT_cond_signal(s->barrier_cond);
    assert(aret == ABT_SUCCESS);
    // now wait until all barriers have been rx'd
    DEBUG("%d: barrier ult: out, incr count to %d\n", s->rank,
            s->barrier_count+1);
    s->barrier_count++;
    while (s->barrier_count < s->num_addrs-1) {
        DEBUG("%d: barrier ult: waiting (count at %d)\n", s->rank,
                s->barrier_count);
        aret = ABT_cond_wait(s->barrier_cond, s->barrier_mutex);
        assert(aret == ABT_SUCCESS);
    }
    DEBUG("%d: barrier ult: count compl, signal and respond\n", s->rank);
    // all barriers rx'd, inform other ULTs
    ABT_cond_signal(s->barrier_cond);
    ABT_mutex_unlock(s->barrier_mutex);

    hret = margo_respond(s->mid, h, NULL);
    assert(hret == HG_SUCCESS);
    HG_Destroy(h);

    DEBUG("%d: barrier ult: respond compl, count at %d\n", s->rank,
            s->barrier_count);

    aret = ABT_mutex_lock(s->barrier_mutex);
    assert(aret == ABT_SUCCESS);
    // done -> I'm the last ULT to enter, I do the eventual set
    int is_done = (++s->barrier_count) == 2*(s->num_addrs-1);
    if (is_done) s->barrier_count = 0;
    aret = ABT_mutex_unlock(s->barrier_mutex);
    assert(aret == ABT_SUCCESS);
    if (is_done) {
        aret = ABT_eventual_set(s->barrier_eventual, NULL, 0);
        assert(aret == ABT_SUCCESS);
    }
}

hg_return_t ssg_barrier_margo(ssg_t s)
{
    // non-members can't barrier
    if (s->rank < 0) return HG_INVALID_PARAM;

    // return immediately if a singleton group
    if (s->num_addrs == 1) return HG_SUCCESS;

    int aret = ABT_eventual_reset(s->barrier_eventual);
    if (aret != ABT_SUCCESS) return HG_OTHER_ERROR;

    DEBUG("%d: barrier: lock and incr id to %d\n", s->rank, s->barrier_id+1);
    int bid;
    // init the barrier state
    aret = ABT_mutex_lock(s->barrier_mutex);
    if (aret != ABT_SUCCESS) return HG_OTHER_ERROR;
    bid = ++s->barrier_id;
    aret = ABT_cond_broadcast(s->barrier_cond);
    if (aret != ABT_SUCCESS) {
        ABT_mutex_unlock(s->barrier_mutex); return HG_OTHER_ERROR;
    }
    aret = ABT_mutex_unlock(s->barrier_mutex);
    if (aret != ABT_SUCCESS) return HG_OTHER_ERROR;

    if (s->rank > 0) {
        DEBUG("%d: barrier: create and forward to 0\n", s->rank);
        barrier_in_t in;
        hg_handle_t h;
        hg_return_t hret = HG_Create(margo_get_context(s->mid),
                ssg_get_addr(s, 0), s->barrier_rpc_id, &h);
        if (hret != HG_SUCCESS) return hret;

        in.rank = s->rank;
        in.barrier_id = bid;
        hret = margo_forward(s->mid, h, &in);
        DEBUG("%d: barrier: finish\n", s->rank);
        HG_Destroy(h);
        if (hret != HG_SUCCESS) return hret;
    }
    else {
        DEBUG("%d: barrier: wait on eventual\n", s->rank);
        aret = ABT_eventual_wait(s->barrier_eventual, NULL);
        if (aret != ABT_SUCCESS) return HG_OTHER_ERROR;
    }

    return HG_SUCCESS;
}
565
#endif
Jonathan Jenkins's avatar
Jonathan Jenkins committed
566

Jonathan Jenkins's avatar
Jonathan Jenkins committed
567 568
void ssg_finalize(ssg_t s)
{
569 570
    if (s == SSG_NULL) return;

571
#if USE_SWIM_FD
572 573 574 575
    if(s->swim_ctx)
        swim_finalize(s->swim_ctx);
#endif

Jonathan Jenkins's avatar
Jonathan Jenkins committed
576 577 578 579 580 581
    if (s->barrier_mutex != ABT_MUTEX_NULL)
        ABT_mutex_free(&s->barrier_mutex);
    if (s->barrier_cond != ABT_COND_NULL)
        ABT_cond_free(&s->barrier_cond);
    if (s->barrier_eventual != ABT_EVENTUAL_NULL)
        ABT_eventual_free(&s->barrier_eventual);
582

Jonathan Jenkins's avatar
Jonathan Jenkins committed
583
    for (int i = 0; i < s->num_addrs; i++) {
584 585
        if (s->addrs[i] != HG_ADDR_NULL)
            HG_Addr_free(margo_get_class(s->mid), s->addrs[i]);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602
    }
    free(s->backing_buf);
    free(s->addr_strs);
    free(s->addrs);
    free(s);
}

int ssg_get_rank(const ssg_t s)
{
    return s->rank;
}

int ssg_get_count(const ssg_t s)
{
    return s->num_addrs;
}

Jonathan Jenkins's avatar
Jonathan Jenkins committed
603
hg_addr_t ssg_get_addr(const ssg_t s, int rank)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
604
{
605
    if (rank >= 0 && rank < s->num_addrs)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
606 607
        return s->addrs[rank];
    else
Jonathan Jenkins's avatar
Jonathan Jenkins committed
608
        return HG_ADDR_NULL;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
609 610 611 612
}

const char * ssg_get_addr_str(const ssg_t s, int rank)
{
613
    if (rank >= 0 && rank < s->num_addrs)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
614 615 616 617 618
        return s->addr_strs[rank];
    else
        return NULL;
}

619
#if 0
Jonathan Jenkins's avatar
Jonathan Jenkins committed
620 621
// serialization format looks like:
// < num members, buffer size, buffer... >
Jonathan Jenkins's avatar
Jonathan Jenkins committed
622
// doesn't attempt to grab hg_addr's, string buffers, etc. - client will be
Jonathan Jenkins's avatar
Jonathan Jenkins committed
623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693
// responsible for doing a separate address lookup routine
hg_return_t hg_proc_ssg_t(hg_proc_t proc, ssg_t *s)
{
    // error and return handling
    hg_return_t hret = HG_SUCCESS;
    char * err_str = NULL;

    // input/output vars + helpers for ssg decode setup
    ssg_t ss = NULL;

    switch(hg_proc_get_op(proc)) {
        case HG_ENCODE:
            ss = *s;
            // encode address count
            hret = hg_proc_int32_t(proc, &ss->num_addrs);
            if (hret != HG_SUCCESS) { err_str = "ssg num addrs"; goto end; }
            // encode addr
            hret = hg_proc_int32_t(proc, &ss->buf_size);
            if (hret != HG_SUCCESS) { err_str = "ssg buf size"; goto end; }
            // encode addr string, simple as blitting the backing buffer
            hret = hg_proc_memcpy(proc, ss->backing_buf, ss->buf_size);
            if (hret != HG_SUCCESS) { err_str = "ssg addr buf"; goto end; }
            break;

        case HG_DECODE:
            // create the output
            *s = NULL;
            ss = malloc(sizeof(*ss));
            if (ss == NULL) {
                err_str = "ssg alloc";
                hret = HG_NOMEM_ERROR;
                goto end;
            }
            ss->addr_strs = NULL;
            ss->addrs = NULL;
            ss->backing_buf = NULL;
            // get address count
            hret = hg_proc_int32_t(proc, &ss->num_addrs);
            if (hret != HG_SUCCESS) { err_str = "ssg num addrs"; goto end; }
            // get number of bytes for the address
            hret = hg_proc_int32_t(proc, &ss->buf_size);
            if (hret != HG_SUCCESS) { err_str = "ssg buf size"; goto end; }
            // allocate output buffer
            ss->backing_buf = malloc(ss->buf_size);
            if (hret != HG_SUCCESS) {
                err_str = "ssg buf alloc";
                hret = HG_NOMEM_ERROR;
                goto end;
            }
            hret = hg_proc_memcpy(proc, ss->backing_buf, ss->buf_size);
            if (hret != HG_SUCCESS) { err_str = "ssg addr buf"; goto end; }

            // set the remaining ssg vars

            ss->addr_strs = NULL; ss->addrs = NULL;
            ss->rank = -1; // receivers aren't part of the group

            ss->addr_strs = setup_addr_str_list(ss->num_addrs, ss->backing_buf);
            if (ss->addr_strs == NULL) {
                err_str = "ssg addr strs alloc";
                hret = HG_NOMEM_ERROR;
                goto end;
            }

            ss->addrs = malloc(ss->num_addrs * sizeof(*ss->addrs));
            if (ss->addrs == NULL) {
                err_str = "ssg addrs alloc";
                hret = HG_NOMEM_ERROR;
                goto end;
            }
            for (int i = 0; i < ss->num_addrs; i++) {
Jonathan Jenkins's avatar
Jonathan Jenkins committed
694
                ss->addrs[i] = HG_ADDR_NULL;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724
            }

            // success: set the output
            *s = ss;
            break;

        case HG_FREE:
            if (s != NULL && *s != NULL) {
                err_str = "ssg shouldn't be freed via HG_Free_*";
                hret = HG_INVALID_PARAM;
            }
            goto end;

        default:
            err_str = "bad proc mode";
            hret = HG_INVALID_PARAM;
    }
end:
    if (err_str) {
        HG_LOG_ERROR("Proc error: %s", err_str);
        if (hg_proc_get_op(proc) == HG_DECODE) {
            free(ss->addr_strs);
            free(ss->addrs);
            free(ss->backing_buf);
            free(ss);
        }
    }
    return hret;
}

725 726 727 728 729 730 731 732 733 734 735 736 737 738
int ssg_dump(const ssg_t s, const char *fname)
{
    // file to write to
    int fd = -1;
    ssize_t written;

    // string to xform and dump
    char * addrs_dup = NULL;
    char * tok = NULL;
    char * addrs_dup_end = NULL;

    // return code
    int ret = 0;

739 740
    // copy the backing buffer, replacing all null chars with
    // newlines
741 742 743 744 745
    addrs_dup = malloc(s->buf_size);
    if (addrs_dup == NULL) { errno = ENOMEM; ret = -1; goto end; }
    memcpy(addrs_dup, s->backing_buf, s->buf_size);
    tok = addrs_dup;
    addrs_dup_end = addrs_dup + s->buf_size;
746
    for (int i = 0; i < s->num_addrs; i++) {
747 748 749 750 751 752
        tok = memchr(tok, '\0', addrs_dup_end - tok);
        if (tok == NULL) { errno = EINVAL; ret = -1; goto end; }
        *tok = '\n';
    }

    // open the file and dump in a single call
753
    fd = open(fname, O_WRONLY | O_CREAT | O_EXCL, 0644);
754 755
    if (fd == -1) { ret = -1; goto end; }
    // don't include the null char at the end
756 757
    written = write(fd, addrs_dup, s->buf_size);
    if (written != s->buf_size) ret = -1;
758 759 760 761 762 763 764

end:
    free(addrs_dup);
    if (fd != -1) close(fd);

    return ret;
}
765
#endif
766

Jonathan Jenkins's avatar
Jonathan Jenkins committed
767 768 769 770 771 772 773 774 775 776 777 778
static char** setup_addr_str_list(int num_addrs, char * buf)
{
    char ** ret = malloc(num_addrs * sizeof(*ret));
    if (ret == NULL) return NULL;

    ret[0] = buf;
    for (int i = 1; i < num_addrs; i++) {
        char * a = ret[i-1];
        ret[i] = a + strlen(a) + 1;
    }
    return ret;
}
Jonathan Jenkins's avatar
Jonathan Jenkins committed
779

780
#if 0
Jonathan Jenkins's avatar
Jonathan Jenkins committed
781 782 783 784 785 786 787 788 789 790 791
static uint64_t fnv1a_64(void *data, size_t size)
{
    uint64_t hash = 14695981039346656037ul;
    unsigned char *d = data;

    for (size_t i = 0; i < size; i++) {
        hash ^= (uint64_t)*d++;
        hash *= 1099511628211;
    }
    return hash;
}
792 793
#endif