resource-lp.c 19.1 KB
Newer Older
1 2 3 4 5 6 7 8 9
/*
 * Copyright (C) 2014 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
*/

#include "codes/resource-lp.h"
#include "codes/resource.h"
#include "codes/codes_mapping.h"
10
#include "codes/configuration.h"
11
#include "codes/jenkins-hash.h"
12
#include "codes/quicklist.h"
13
#include "codes/lp-io.h"
14 15 16
#include "ross.h"
#include <assert.h>
#include <stdio.h>
17
#include <string.h>
18 19 20 21 22


/**** BEGIN SIMULATION DATA STRUCTURES ****/

static int resource_magic; /* use this as sanity check on events */
23 24 25 26 27

/* configuration globals (will be consumed by LP when they init) */
static uint64_t avail_unanno;
static uint64_t *avail_per_anno;
static const config_anno_map_t *anno_map;
28 29 30

typedef struct resource_state resource_state;
typedef struct resource_msg resource_msg;
31
typedef struct pending_op pending_op;
32 33 34 35 36 37

#define TOKEN_DUMMY ((resource_token_t)-1)

/* event types */
enum resource_event
{
38
    RESOURCE_GET = 100,
39
    RESOURCE_FREE,
40
    RESOURCE_DEQ,
41 42 43 44 45
    RESOURCE_RESERVE,
};

struct resource_state {
    resource r;
46 47 48 49 50 51
    /* pending operations - if OOM and we are using the 'blocking' method, 
     * then need to stash parameters.
     * Index 0 is the general pool, index 1.. are the reservation-specific
     * pools. We take advantage of resource_token_t's status as a simple 
     * array index to do the proper indexing */
    struct qlist_head pending[MAX_RESERVE+1];
52 53
};

54 55 56
/* following struct exists because we want to basically cache a message within
 * a message for rc (ewww) */
struct resource_msg_internal{
57 58 59 60
    msg_header h;
    /* request data */
    uint64_t req;
    resource_token_t tok; /* only for reserved calls */
61 62 63 64 65
    /* behavior when sending response to caller
     * 0 - send the callback immediately if resource unavailable. 
     * 1 - send the callback when memory is available (danger - deadlock
     * possible) */
    int block_on_unavail; 
66 67 68 69 70
    /* callback data */
    msg_header h_callback;
    int msg_size;
    int msg_header_offset;
    int msg_callback_offset;
71 72 73 74
    /* user-provided data */
    int msg_callback_misc_size;
    int msg_callback_misc_offset;
    char msg_callback_misc[RESOURCE_MAX_CALLBACK_PAYLOAD];
75 76 77 78
}; 

struct resource_msg {
    struct resource_msg_internal i, i_rc;
79 80 81
    // for RC (asides from the message itself): the previous minimum resource
    // value
    uint64_t min_avail_rc;
82 83 84 85 86
};

struct pending_op {
    struct resource_msg_internal m;
    struct qlist_head ql;
87 88 89 90 91 92 93 94
};

/**** END SIMULATION DATA STRUCTURES ****/

/**** BEGIN LP, EVENT PROCESSING FUNCTION DECLS ****/

/* ROSS LP processing functions */  
static void resource_lp_ind_init(
95 96
        resource_state * ns,
        tw_lp * lp);
97
static void resource_event_handler(
98 99 100 101
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp);
102
static void resource_rev_handler(
103 104 105 106
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp);
107
static void resource_finalize(
108 109
        resource_state * ns,
        tw_lp * lp);
110 111 112

/* ROSS function pointer table for this LP */
static tw_lptype resource_lp = {
113 114 115 116 117 118 119
    (init_f) resource_lp_ind_init,
    (pre_run_f) NULL,
    (event_f) resource_event_handler,
    (revent_f) resource_rev_handler,
    (final_f)  resource_finalize, 
    (map_f) codes_mapping,
    sizeof(resource_state),
120 121 122 123 124 125 126 127 128
};

/**** END LP, EVENT PROCESSING FUNCTION DECLS ****/

/**** BEGIN IMPLEMENTATIONS ****/

void resource_lp_ind_init(
        resource_state * ns,
        tw_lp * lp){
129 130 131 132 133 134 135 136 137 138 139 140 141 142 143
    // get my annotation
    const char * anno = codes_mapping_get_annotation_by_lpid(lp->gid);
    if (anno == NULL){
        resource_init(avail_unanno, &ns->r);
    }
    else{
        int idx = configuration_get_annotation_index(anno, anno_map);
        if (idx < 0){
            tw_error("resource LP %lu: unable to find annotation "
                    "%s in configuration\n", lp->gid, anno);
        }
        else{
            resource_init(avail_per_anno[idx], &ns->r);
        }
    }
144 145 146 147 148 149 150
    int i;
    for (i = 0; i < MAX_RESERVE+1; i++){
        INIT_QLIST_HEAD(&ns->pending[i]);
    }
}

static void resource_response(
Jonathan Jenkins's avatar
oops!  
Jonathan Jenkins committed
151
        struct resource_msg_internal *m,
152 153 154 155 156
        tw_lp *lp,
        int ret,
        resource_token_t tok){
    /* send return message */
    msg_header h;
157
    msg_set_header(m->h_callback.magic, m->h_callback.event_type, 
158 159 160 161 162 163 164
            lp->gid, &h);

    resource_callback c;
    c.ret = ret;
    c.tok = tok;

    /* before we send the message, sanity check the sizes */
165 166 167 168
    if (m->msg_size >= m->msg_header_offset+sizeof(h) &&
            m->msg_size >= m->msg_callback_offset+sizeof(c) &&
            m->msg_size >= m->msg_callback_offset+m->msg_callback_misc_size){
        tw_event *e = codes_event_new(m->h_callback.src, 
169 170
                codes_local_latency(lp), lp);
        void *msg = tw_event_data(e);
171 172 173 174 175
        memcpy(((char*)msg)+m->msg_header_offset, &h, sizeof(h));
        memcpy(((char*)msg)+m->msg_callback_offset, &c, sizeof(c));
        if (m->msg_callback_misc_size > 0){
            memcpy(((char*)msg)+m->msg_callback_misc_offset, 
                        m->msg_callback_misc, m->msg_callback_misc_size);
176
        }
177 178 179 180
        tw_event_send(e);
    }
    else{
        tw_error(TW_LOC, 
181
                "message size not large enough to hold header/callback/misc"
182 183 184 185
                " structures\n"
                "msg size: %3d, header   off/size:  %d, %d\n"
                "               callback off/size:  %d, %d\n"
                "               callback misc size: %d",
186 187 188
                m->msg_size, m->msg_header_offset, (int)sizeof(h),
                m->msg_callback_offset, (int)sizeof(c),
                m->msg_callback_misc_size);
189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205
    }
}
static void resource_response_rc(tw_lp *lp){
    codes_local_latency_reverse(lp);
}

/* bitfield usage:
 * c0 - enqueued a message 
 * c1 - sent an ack 
 * c2 - successfully got the resource */
static void handle_resource_get(
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp){
    int ret = 1;
    int send_ack = 1;
206 207
    // save the previous minimum for RC
    assert(!resource_get_min_avail(m->i.tok, &m->min_avail_rc, &ns->r));
208 209 210 211 212 213 214 215 216 217
    if (!qlist_empty(&ns->pending[m->i.tok]) || 
            (ret = resource_get(m->i.req, m->i.tok, &ns->r))){
        /* failed to receive data */
        assert(ret != 2);
        if (m->i.block_on_unavail){
            /* queue up operation, save til later */
            b->c0 = 1;
            pending_op *op = malloc(sizeof(pending_op));
            op->m = m->i; /* no need to set rc msg here */
            qlist_add_tail(&op->ql, &ns->pending[m->i.tok]);
218
            send_ack = 0;
219 220 221 222
        }
    }
    if (send_ack){
        b->c1 = 1;
223
        resource_response(&m->i, lp, ret, TOKEN_DUMMY);
224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247
    }

    b->c2 = !ret;
}

/* bitfield usage:
 * c0 - enqueued a message 
 * c1 - sent an ack 
 * c2 - successfully got the resource */
static void handle_resource_get_rc(
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp){
    if (b->c0){
        assert(!qlist_empty(&ns->pending[m->i.tok]));
        struct qlist_head *ql = qlist_pop_back(&ns->pending[m->i.tok]);
        free(qlist_entry(ql, pending_op, ql));
    }
    else if (b->c1){
        resource_response_rc(lp);
    }

    if (b->c2){
248 249
        assert(!resource_restore_min_avail(m->i.tok, m->min_avail_rc, &ns->r));
        assert(!resource_free(m->i.req, m->i.tok, &ns->r));
250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275
    }
}

static void handle_resource_free(
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp){
    assert(!resource_free(m->i.req, m->i.tok, &ns->r));
    /* create an event to pop the next queue item */
    tw_event *e = codes_event_new(lp->gid, codes_local_latency(lp), lp);
    resource_msg *m_deq = tw_event_data(e);
    msg_set_header(resource_magic, RESOURCE_DEQ, lp->gid, &m_deq->i.h);
    m_deq->i.tok = m->i.tok; /* only tok is needed, all others grabbed from q */
    tw_event_send(e);
}
static void handle_resource_free_rc(
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp){
    assert(!resource_get(m->i.req, m->i.tok, &ns->r));
    codes_local_latency_reverse(lp);
}

/* bitfield usage:
276 277
 * c0 - queue was empty to begin with
 * c1 - assuming !c0, alloc succeeded */ 
278 279 280 281 282 283 284
static void handle_resource_deq(
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp){
    if (qlist_empty(&ns->pending[m->i.tok])){
        /* nothing to do */
285
        b->c0 = 1;
286 287 288 289 290
        return;
    }

    struct qlist_head *front = ns->pending[m->i.tok].next;
    pending_op *p = qlist_entry(front, pending_op, ql);
291
    assert(!resource_get_min_avail(m->i.tok, &m->min_avail_rc, &ns->r));
292 293 294
    int ret = resource_get(p->m.req, p->m.tok, &ns->r);
    assert(ret != 2);
    if (!ret){
295
        b->c1 = 1;
296 297 298
        /* success, dequeue (saving as rc) and send to client */
        qlist_del(front);
        m->i_rc = p->m;
299
        resource_response(&p->m, lp, ret, TOKEN_DUMMY);
300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318
        free(p);
        /* additionally attempt to dequeue next one down */
        tw_event *e = codes_event_new(lp->gid, codes_local_latency(lp), lp);
        resource_msg *m_deq = tw_event_data(e);
        msg_set_header(resource_magic, RESOURCE_DEQ, lp->gid, &m_deq->i.h);
        /* only tok is needed, all others grabbed from q */
        m_deq->i.tok = m->i.tok; 
        tw_event_send(e);
    }
    /* else do nothing */
}

/* bitfield usage:
 * c0 - dequeue+alloc success */ 
static void handle_resource_deq_rc(
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp){
319
    if (b->c0){
320 321 322
        return;
    }

323
    if (b->c1){
324 325 326 327 328
        /* add operation back to the front of the queue */
        pending_op *op = malloc(sizeof(pending_op));
        op->m = m->i_rc;
        qlist_add(&op->ql, &ns->pending[m->i.tok]);
        resource_response_rc(lp);
329
        assert(!resource_restore_min_avail(m->i.tok, m->min_avail_rc, &ns->r)); 
Jonathan Jenkins's avatar
Jonathan Jenkins committed
330
        assert(!resource_free(op->m.req, op->m.tok, &ns->r));
331 332 333 334 335 336 337 338 339 340 341 342 343
        /* reverse "deq next" op */
        codes_local_latency_reverse(lp);
    }
}

static void handle_resource_reserve(
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp){
    resource_token_t tok;
    int ret = resource_reserve(m->i.req, &tok, &ns->r);
    assert(!ret);
344
    resource_response(&m->i, lp, ret, tok);
345 346 347 348 349 350 351 352 353 354 355 356
}
static void handle_resource_reserve_rc(
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp){
    /* this reversal method is essentially a hack that relies on each
     * sequential reserve appending to the end of the list 
     * - we expect reserves to happen strictly at the beginning of the
     *   simulation */
    ns->r.num_tokens--;
    resource_response_rc(lp);
357 358 359 360 361 362 363
}

void resource_event_handler(
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp){
364 365
    assert(m->i.h.magic == resource_magic);
    switch(m->i.h.event_type){
366
        case RESOURCE_GET:
367
            handle_resource_get(ns,b,m,lp);
368 369
            break;
        case RESOURCE_FREE:
370
            handle_resource_free(ns,b,m,lp);
371
            break;
372 373
        case RESOURCE_DEQ:
            handle_resource_deq(ns,b,m,lp);
374
            break;
375 376
        case RESOURCE_RESERVE:
            handle_resource_reserve(ns,b,m,lp);
377 378
            break;
        default:
379
            assert(0);
380 381 382 383 384 385 386
    }
}
void resource_rev_handler(
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp){
387 388
    assert(m->i.h.magic == resource_magic);
    switch(m->i.h.event_type){
389
        case RESOURCE_GET:
390
            handle_resource_get_rc(ns,b,m,lp);
391 392
            break;
        case RESOURCE_FREE:
393
            handle_resource_free_rc(ns,b,m,lp);
394
            break;
395 396
        case RESOURCE_DEQ:
            handle_resource_deq_rc(ns,b,m,lp);
397
            break;
398 399
        case RESOURCE_RESERVE:
            handle_resource_reserve_rc(ns,b,m,lp);
400 401
            break;
        default:
402
            assert(0);
403 404 405 406 407 408
    }
}

void resource_finalize(
        resource_state * ns,
        tw_lp * lp){
409
    struct qlist_head *ent;
410 411 412 413 414 415 416 417 418
    for (int i = 0; i < MAX_RESERVE+1; i++){
        qlist_for_each(ent, &ns->pending[i]){
            fprintf(stderr, "WARNING: resource LP %lu has a pending allocation\n",
                    lp->gid);
        }
    }

    char *out_buf = malloc(1<<12);
    int written;
419 420
    // see if I'm the "first" resource (currently doing it globally)
    if (codes_mapping_get_lp_relative_id(lp->gid, 0, 0) == 0){
421 422
        written = sprintf(out_buf, 
                "# format: <LP> <max used general> <max used token...>\n");
423
        lp_io_write(lp->gid, RESOURCE_LP_NM, written, out_buf);
424 425 426
    }
    written = sprintf(out_buf, "%lu", lp->gid);

427
    // compute peak resource usage
428 429 430
    // TODO: wrap this up in the resource interface
    for (int i = 0; i < ns->r.num_tokens+1; i++){
        written += sprintf(out_buf+written, " %lu", ns->r.max[i]-ns->r.min_avail[i]);
431
    }
432
    written += sprintf(out_buf+written, "\n");
433
    lp_io_write(lp->gid, RESOURCE_LP_NM, written, out_buf);
434 435 436 437 438 439 440 441
}

/**** END IMPLEMENTATIONS ****/

/**** BEGIN USER-FACING FUNCTIONS ****/
void resource_lp_init(){
    uint32_t h1=0, h2=0;

442
    bj_hashlittle2(RESOURCE_LP_NM, strlen(RESOURCE_LP_NM), &h1, &h2);
443 444
    resource_magic = h1+h2;

445
    lp_type_register(RESOURCE_LP_NM, &resource_lp);
446 447 448
}

void resource_lp_configure(){
449 450 451 452 453 454

    anno_map = codes_mapping_get_lp_anno_map(RESOURCE_LP_NM);
    avail_per_anno = (anno_map->num_annos > 0) ?
            malloc(anno_map->num_annos * sizeof(*avail_per_anno)) :
            NULL;
    // get the unannotated version
455
    long int avail;
456
    int ret;
457
    if (anno_map->has_unanno_lp > 0){
458
        ret = configuration_get_value_longint(&config, RESOURCE_LP_NM,
459
            "available", NULL, &avail);
460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479
        if (ret){
            fprintf(stderr,
                    "Could not find section:resource value:available for "
                    "resource LP\n");
            exit(1);
        }
        assert(avail > 0);
        avail_unanno = (uint64_t)avail;
    }
    for (uint64_t i = 0; i < anno_map->num_annos; i++){
        ret = configuration_get_value_longint(&config, RESOURCE_LP_NM,
            "available", anno_map->annotations[i], &avail);
        if (ret){
            fprintf(stderr,
                    "Could not find section:resource value:available@%s for "
                    "resource LP\n", anno_map->annotations[i]);
            exit(1);
        }
        assert(avail > 0);
        avail_per_anno[i] = (uint64_t)avail;
480 481 482 483 484 485 486
    }
}

static void resource_lp_issue_event(
        msg_header *header,
        uint64_t req,
        resource_token_t tok, /* only used in reserve_get/free */
487
        int block_on_unavail,
488 489 490
        int msg_size,
        int msg_header_offset,
        int msg_callback_offset,
491 492 493
        int msg_callback_misc_size,
        int msg_callback_misc_offset,
        void *msg_callback_misc_data,
494 495 496 497 498 499
        enum resource_event type,
        tw_lp *sender){

    tw_lpid resource_lpid;

    /* map out the lpid of the resource */
500 501 502 503 504 505
    int mapping_rep_id, mapping_offset, dummy;
    char lp_group_name[MAX_NAME_LENGTH];
    // TODO: currently ignoring annotations... perhaps give annotation as a
    // parameter?
    codes_mapping_get_lp_info(sender->gid, lp_group_name, &dummy,
            NULL, &dummy, NULL,
506
            &mapping_rep_id, &mapping_offset);
507 508
    codes_mapping_get_lp_id(lp_group_name, RESOURCE_LP_NM, NULL, 1,
            mapping_rep_id, mapping_offset, &resource_lpid); 
509 510 511 512 513 514

    tw_event *e = codes_event_new(resource_lpid, codes_local_latency(sender),
            sender);

    /* set message info */
    resource_msg *m = tw_event_data(e);
515 516 517 518
    msg_set_header(resource_magic, type, sender->gid, &m->i.h);
    m->i.req = req;
    m->i.tok = tok;
    m->i.block_on_unavail = block_on_unavail;
519 520

    /* set callback info */
521
    if (header != NULL){
522
        m->i.h_callback = *header;
523
    }
524 525 526
    m->i.msg_size = msg_size;
    m->i.msg_header_offset = msg_header_offset;
    m->i.msg_callback_offset = msg_callback_offset;
527

528 529 530 531 532 533 534 535 536 537 538 539
    if (msg_callback_misc_size > 0){
        assert(msg_callback_misc_size <= RESOURCE_MAX_CALLBACK_PAYLOAD);
        m->i.msg_callback_misc_size = msg_callback_misc_size;
        m->i.msg_callback_misc_offset = msg_callback_misc_offset;
        memcpy(m->i.msg_callback_misc, msg_callback_misc_data,
                msg_callback_misc_size);
    }
    else{
        m->i.msg_callback_misc_size = 0;
        m->i.msg_callback_misc_offset = 0;
    }

540 541 542 543 544 545
    tw_event_send(e);
}

void resource_lp_get(
        msg_header *header,
        uint64_t req, 
546
        int block_on_unavail,
547 548 549
        int msg_size, 
        int msg_header_offset,
        int msg_callback_offset,
550 551 552
        int msg_callback_misc_size,
        int msg_callback_misc_offset,
        void *msg_callback_misc_data,
553
        tw_lp *sender){
554
    resource_lp_issue_event(header, req, 0, block_on_unavail,
555 556 557
            msg_size, msg_header_offset, msg_callback_offset,
            msg_callback_misc_size, msg_callback_misc_offset,
            msg_callback_misc_data, RESOURCE_GET, sender);
558 559 560
}

/* no callback for frees thus far */
561
void resource_lp_free(uint64_t req, tw_lp *sender){
562
    resource_lp_issue_event(NULL, req, 0, -1, -1,-1,-1, 0, 0, NULL,
563 564 565 566 567
            RESOURCE_FREE, sender);
}
void resource_lp_reserve(
        msg_header *header, 
        uint64_t req,
568
        int block_on_unavail,
569 570 571
        int msg_size,
        int msg_header_offset,
        int msg_callback_offset,
572 573 574
        int msg_callback_misc_size,
        int msg_callback_misc_offset,
        void *msg_callback_misc_data,
575
        tw_lp *sender){
576 577 578 579
    resource_lp_issue_event(header, req, 0, block_on_unavail, msg_size,
            msg_header_offset, msg_callback_offset, msg_callback_misc_size,
            msg_callback_misc_offset, msg_callback_misc_data, RESOURCE_RESERVE,
            sender);
580 581 582 583 584
}
void resource_lp_get_reserved(
        msg_header *header,
        uint64_t req,
        resource_token_t tok,
585
        int block_on_unavail,
586 587 588
        int msg_size, 
        int msg_header_offset,
        int msg_callback_offset,
589 590 591
        int msg_callback_misc_size,
        int msg_callback_misc_offset,
        void *msg_callback_misc_data,
592
        tw_lp *sender){
593 594 595
    resource_lp_issue_event(header, req, tok, block_on_unavail, msg_size,
            msg_header_offset, msg_callback_offset, msg_callback_misc_size,
            msg_callback_misc_offset, msg_callback_misc_data, RESOURCE_GET,
596 597 598 599 600 601
            sender);
}
void resource_lp_free_reserved(
        uint64_t req, 
        resource_token_t tok,
        tw_lp *sender){
602 603
    resource_lp_issue_event(NULL, req, tok, -1,-1,-1,-1, 0,0,NULL,
            RESOURCE_FREE, sender);
604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636
}

/* rc functions - thankfully, they only use codes-local-latency, so no need 
 * to pass in any arguments */

static void resource_lp_issue_event_rc(tw_lp *sender){
    codes_local_latency_reverse(sender);
}

void resource_lp_get_rc(tw_lp *sender){
    resource_lp_issue_event_rc(sender);
}
void resource_lp_free_rc(tw_lp *sender){
    resource_lp_issue_event_rc(sender);
}
void resource_lp_reserve_rc(tw_lp *sender){
    resource_lp_issue_event_rc(sender);
}
void resource_lp_get_reserved_rc(tw_lp *sender){
    resource_lp_issue_event_rc(sender);
}
void resource_lp_free_reserved_rc(tw_lp *sender){
    resource_lp_issue_event_rc(sender);
}

/*
 * Local variables:
 *  c-indent-level: 4
 *  c-basic-offset: 4
 * End:
 *
 * vim: ts=8 sts=4 sw=4 expandtab
 */