ssg.c 23 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 19
#include <mercury.h>
#include <margo.h>
20

21
#include <ssg.h>
22
#include "ssg-internal.h"
Jonathan Jenkins's avatar
Jonathan Jenkins committed
23

24 25 26 27
#if USE_SWIM_FD
#include "swim-fd/swim-fd.h"
#endif

28
#define DO_DEBUG 0
Jonathan Jenkins's avatar
Jonathan Jenkins committed
29 30 31 32 33 34 35 36
#define DEBUG(...) \
    do { \
        if(DO_DEBUG) { \
            printf(__VA_ARGS__); \
            fflush(stdout); \
        } \
    } while(0)

Shane Snyder's avatar
Shane Snyder committed
37
// internal initialization of ssg data structures
38 39
static ssg_t ssg_init_internal(margo_instance_id mid, int self_rank,
    int group_size, hg_addr_t self_addr, char *addr_str_buf, int addr_str_buf_size);
40 41 42

// lookup peer addresses
static hg_return_t ssg_lookup(ssg_t s);
Shane Snyder's avatar
Shane Snyder committed
43

Jonathan Jenkins's avatar
Jonathan Jenkins committed
44 45
static char** setup_addr_str_list(int num_addrs, char * buf);

46
#if 0
Jonathan Jenkins's avatar
Jonathan Jenkins committed
47 48 49 50 51 52 53 54 55 56 57
// 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)
58
#endif
Jonathan Jenkins's avatar
Jonathan Jenkins committed
59

60

61
ssg_t ssg_init_config(margo_instance_id mid, const char * fname, int is_member)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77
{
    // 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;
78
    void *addr_buf = NULL;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
79

Shane Snyder's avatar
Shane Snyder committed
80
    // self rank/addr resolution helpers
81
    hg_class_t *hgcl = NULL;
Shane Snyder's avatar
Shane Snyder committed
82 83 84 85 86 87 88 89
    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
90 91
    // misc return codes
    int ret;
92 93 94 95
    hg_return_t hret;

    // return data
    ssg_t s = NULL;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
96 97 98 99 100 101 102 103 104 105 106 107 108 109 110

    // 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
111
    if (rdsz != st.st_size) goto fini;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
112 113
    rdbuf[rdsz]='\0';

114 115 116
    hgcl = margo_get_class(mid);
    if(!hgcl) goto fini;

Shane Snyder's avatar
Shane Snyder committed
117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135
    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
136 137 138 139 140 141
    // 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
142 143
    addr_buf = malloc(addr_cap);
    if (addr_buf == NULL) goto fini;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
144 145 146 147 148
    do {
        int tok_sz = strlen(tok);
        if (tok_sz + addr_len + 1 > addr_cap) {
            void * tmp;
            addr_cap *= 2;
149
            tmp = realloc(addr_buf, addr_cap);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
150
            if (tmp == NULL) goto fini;
151
            addr_buf = tmp;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
152
        }
Shane Snyder's avatar
Shane Snyder committed
153 154 155 156 157 158 159 160
        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;
        }
161
        memcpy((char*)addr_buf + addr_len, tok, tok_sz+1);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
162 163 164 165 166
        addr_len += tok_sz+1;
        num_addrs++;
        tok = strtok(NULL, "\r\n\t ");
    } while (tok != NULL);

Shane Snyder's avatar
Shane Snyder committed
167
    // init ssg internal structures
168
    s = ssg_init_internal(mid, rank, num_addrs, self_addr, addr_buf, addr_len);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
169 170
    if (s == NULL) goto fini;

Shane Snyder's avatar
Shane Snyder committed
171
    // don't free this on success
172
    addr_buf = NULL;
Shane Snyder's avatar
Shane Snyder committed
173
    self_addr = HG_ADDR_NULL;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
174 175 176
fini:
    if (fd != -1) close(fd);
    free(rdbuf);
177
    free(addr_buf);
178
    if (hgcl && self_addr != HG_ADDR_NULL) HG_Addr_free(hgcl, self_addr);
Shane Snyder's avatar
Shane Snyder committed
179
    free(self_addr_str);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
180 181 182 183
    return s;
}

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

    // collective helpers
194
    char * addr_buf = NULL;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
195 196 197 198 199
    int * sizes = NULL;
    int * sizes_psum = NULL;
    int comm_size = 0;
    int comm_rank = 0;

200 201 202
    // misc return codes
    hg_return_t hret;

Jonathan Jenkins's avatar
Jonathan Jenkins committed
203 204 205
    // return data
    ssg_t s = NULL;

206 207
    hgcl = margo_get_class(mid);
    if(!hgcl) goto fini;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
208 209

    // get my address
Jonathan Jenkins's avatar
Jonathan Jenkins committed
210 211 212
    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
213
    if (hret != HG_SUCCESS) goto fini;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
214 215
    self_addr_str = malloc(self_addr_size);
    if (self_addr_str == NULL) goto fini;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
216 217
    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
218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236
    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
237 238
    addr_buf = malloc(sizes_psum[comm_size]);
    if (addr_buf == NULL) goto fini;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
239
    MPI_Allgatherv(self_addr_str, self_addr_size_int, MPI_BYTE,
240
            addr_buf, sizes, sizes_psum, MPI_BYTE, comm);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
241

Shane Snyder's avatar
Shane Snyder committed
242
    // init ssg internal structures
243 244
    s = ssg_init_internal(mid, comm_rank, comm_size, self_addr,
        addr_buf, sizes_psum[comm_size]);
Shane Snyder's avatar
Shane Snyder committed
245 246 247
    if (s == NULL) goto fini;

    // don't free these on success
248
    addr_buf = NULL;
249
    self_addr = HG_ADDR_NULL;
Shane Snyder's avatar
Shane Snyder committed
250 251 252
fini:
    free(sizes);
    free(sizes_psum);
253
    free(addr_buf);
254 255
    if (hgcl && self_addr != HG_ADDR_NULL) HG_Addr_free(hgcl, self_addr);
    free(self_addr_str);
Shane Snyder's avatar
Shane Snyder committed
256 257 258 259
    return s;
}
#endif

260 261
static ssg_t ssg_init_internal(margo_instance_id mid, int self_rank,
    int group_size, hg_addr_t self_addr, char *addr_str_buf, int addr_str_buf_size)
Shane Snyder's avatar
Shane Snyder committed
262
{
263 264 265
    int i;

    // arrays of peer address strings
Shane Snyder's avatar
Shane Snyder committed
266 267
    char **addr_strs = NULL;

268 269 270
    // misc return codes
    hg_return_t hret;

Shane Snyder's avatar
Shane Snyder committed
271 272 273
    // return data
    ssg_t s = NULL;

274 275
    if (self_rank == SSG_RANK_UNKNOWN) goto fini;
    if (self_addr == HG_ADDR_NULL && self_rank != SSG_EXTERNAL_RANK) goto fini;
Shane Snyder's avatar
Shane Snyder committed
276 277

    // set peer address strings
278
    addr_strs = setup_addr_str_list(group_size, addr_str_buf);
279
    if (addr_strs == NULL) goto fini;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
280 281 282 283

    // set up the output
    s = malloc(sizeof(*s));
    if (s == NULL) goto fini;
284
    memset(s, 0, sizeof(*s));
285
    s->mid = mid;
286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304
    s->addr_str_buf = addr_str_buf;
    s->addr_str_buf_size = addr_str_buf_size;

    // initialize the group "view"
    s->view.self_rank = self_rank;
    s->view.group_size = group_size;
    for (i = self_rank + 1; i < group_size; i++)
    {
        s->view.member_states[i].status = SSG_MEMBER_UNKNOWN;
        // NOTE: remote addrs are set in ssg_lookup
        s->view.member_states[i].addr = HG_ADDR_NULL;
        s->view.member_states[i].addr_str = addr_strs[i];
    }
    // set view info for self
    s->view.member_states[i].status = SSG_MEMBER_ALIVE;
    s->view.member_states[i].addr = self_addr;
    s->view.member_states[i].addr_str = addr_strs[i];

#if 0
305 306 307
    s->barrier_mutex = ABT_MUTEX_NULL;
    s->barrier_cond  = ABT_COND_NULL;
    s->barrier_eventual = ABT_EVENTUAL_NULL;
308
#endif
309

310
#if 0
311 312 313 314
    // lookup hg addr information for all group members
    hret = ssg_lookup(s);
    if (hret != HG_SUCCESS)
    {
315
        ssg_finalize(s); s = NULL; /* TODO: is finalize needed? or just free? */
316 317
        goto fini;
    }
318
#endif
Jonathan Jenkins's avatar
Jonathan Jenkins committed
319

320
#if USE_SWIM_FD
321
    // initialize swim failure detector
322
    if (self_rank != SSG_EXTERNAL_RANK)
323 324 325 326 327 328 329 330 331
    {
        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
332 333 334 335 336
fini:
    free(addr_strs);
    return s;
}

337
#if 0
338 339 340 341 342 343 344 345 346 347 348
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
349
    DEBUG("%d (ult): looking up rank %d\n", l->ssg->rank, l->rank);
350
    l->out = margo_addr_lookup(l->ssg->mid, l->ssg->addr_strs[l->rank],
351
            &l->ssg->addrs[l->rank]);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
352
    DEBUG("%d (ult): looked up rank %d\n", l->ssg->rank, l->rank);
353 354
}

355
static hg_return_t ssg_lookup(ssg_t s)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
356 357
{
    hg_context_t *hgctx;
358 359 360
    ABT_thread *ults;
    struct lookup_ult_args *args;
    hg_return_t hret = HG_SUCCESS;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
361 362 363 364

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

Jonathan Jenkins's avatar
Jonathan Jenkins committed
365 366
    DEBUG("%d: entered lookup\n", s->rank);

Jonathan Jenkins's avatar
Jonathan Jenkins committed
367
    // set the hg class up front - need for destructing addrs
Jonathan Jenkins's avatar
Jonathan Jenkins committed
368
    hgctx = margo_get_context(s->mid);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
369 370 371 372 373 374 375 376 377 378
    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;

379 380 381 382 383 384 385 386 387 388 389
    // 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
390 391 392 393 394 395 396 397 398 399
    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]);
400 401
        if (aret != ABT_SUCCESS) {
            hret = HG_OTHER_ERROR;
402
            goto fini;
403
        }
Jonathan Jenkins's avatar
Jonathan Jenkins committed
404
    }
405

Jonathan Jenkins's avatar
Jonathan Jenkins committed
406 407 408 409 410
    // 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]);
411
        DEBUG("%d: lookup: join thread for rank %d fini\n", s->rank, r);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
412 413
        ABT_thread_free(&ults[r]);
        ults[r] = ABT_THREAD_NULL; // in case of cascading failure from join
414 415
        if (aret != ABT_SUCCESS) {
            hret = HG_OTHER_ERROR;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
416
            break;
417
        }
Jonathan Jenkins's avatar
Jonathan Jenkins committed
418 419 420
        else if (args[r].out != HG_SUCCESS) {
            hret = args[r].out;
            break;
421 422 423
        }
    }

424
fini:
425 426 427 428
    // 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
429
                DEBUG("%d: lookup failed, cancelling ULT %d\n", s->rank, i);
430 431 432 433 434 435 436 437 438
                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
439
}
440
#endif
Jonathan Jenkins's avatar
Jonathan Jenkins committed
441

442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462
#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
463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574
// 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;
}
575
#endif
Jonathan Jenkins's avatar
Jonathan Jenkins committed
576

Jonathan Jenkins's avatar
Jonathan Jenkins committed
577 578
void ssg_finalize(ssg_t s)
{
579 580
    if (s == SSG_NULL) return;

581
#if USE_SWIM_FD
582 583 584 585
    if(s->swim_ctx)
        swim_finalize(s->swim_ctx);
#endif

586
#if 0
Jonathan Jenkins's avatar
Jonathan Jenkins committed
587 588 589 590 591 592
    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);
593
#endif
594

595 596 597
    for (int i = 0; i < s->view.group_size; i++) {
        if (s->view.member_states[i].addr != HG_ADDR_NULL)
            HG_Addr_free(margo_get_class(s->mid), s->view.member_states[i].addr);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
598
    }
599
    free(s->addr_str_buf);
Jonathan Jenkins's avatar
Jonathan Jenkins committed
600 601 602
    free(s);
}

603
int ssg_get_group_rank(const ssg_t s)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
604
{
605
    return s->view.self_rank;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
606 607
}

608
int ssg_get_group_size(const ssg_t s)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
609
{
610
    return s->view.group_size;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
611 612
}

Jonathan Jenkins's avatar
Jonathan Jenkins committed
613
hg_addr_t ssg_get_addr(const ssg_t s, int rank)
Jonathan Jenkins's avatar
Jonathan Jenkins committed
614
{
615 616
    if (rank >= 0 && rank < s->view.group_size)
        return s->view.member_states[rank].addr;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
617
    else
Jonathan Jenkins's avatar
Jonathan Jenkins committed
618
        return HG_ADDR_NULL;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
619 620 621 622
}

const char * ssg_get_addr_str(const ssg_t s, int rank)
{
623 624
    if (rank >= 0 && rank < s->view.group_size)
        return s->view.member_states[rank].addr_str;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
625 626 627 628
    else
        return NULL;
}

629
#if 0
Jonathan Jenkins's avatar
Jonathan Jenkins committed
630 631
// serialization format looks like:
// < num members, buffer size, buffer... >
Jonathan Jenkins's avatar
Jonathan Jenkins committed
632
// doesn't attempt to grab hg_addr's, string buffers, etc. - client will be
Jonathan Jenkins's avatar
Jonathan Jenkins committed
633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703
// 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
704
                ss->addrs[i] = HG_ADDR_NULL;
Jonathan Jenkins's avatar
Jonathan Jenkins committed
705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734
            }

            // 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;
}

735 736 737 738 739 740 741 742 743 744 745 746 747 748
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;

749 750
    // copy the backing buffer, replacing all null chars with
    // newlines
751 752 753 754 755
    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;
756
    for (int i = 0; i < s->num_addrs; i++) {
757 758 759 760 761 762
        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
763
    fd = open(fname, O_WRONLY | O_CREAT | O_EXCL, 0644);
764 765
    if (fd == -1) { ret = -1; goto end; }
    // don't include the null char at the end
766 767
    written = write(fd, addrs_dup, s->buf_size);
    if (written != s->buf_size) ret = -1;
768 769 770 771 772 773 774

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

    return ret;
}
775
#endif
776

Jonathan Jenkins's avatar
Jonathan Jenkins committed
777 778 779 780 781 782 783 784 785 786 787 788
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
789

790
#if 0
Jonathan Jenkins's avatar
Jonathan Jenkins committed
791 792 793 794 795 796 797 798 799 800 801
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;
}
802 803
#endif