resource-lp.c 17.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10
/*
 * 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"
#include "codes/jenkins-hash.h"
11
#include "codes/quicklist.h"
12
#include "codes/lp-io.h"
13 14 15
#include "ross.h"
#include <assert.h>
#include <stdio.h>
16
#include <string.h>
17 18 19 20 21 22 23 24 25 26 27 28


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

static int resource_magic; /* use this as sanity check on events */
/* TODO: we currently use a single config value to initialize the resource unit
 * count for all resources in the system. Later on, we'll want to do this on a
 * per-group basis */
static uint64_t avail_global;

typedef struct resource_state resource_state;
typedef struct resource_msg resource_msg;
29
typedef struct pending_op pending_op;
30 31 32 33 34 35

#define TOKEN_DUMMY ((resource_token_t)-1)

/* event types */
enum resource_event
{
36
    RESOURCE_GET = 100,
37
    RESOURCE_FREE,
38
    RESOURCE_DEQ,
39 40 41 42 43
    RESOURCE_RESERVE,
};

struct resource_state {
    resource r;
44 45 46 47 48 49
    /* 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];
50 51
};

52 53 54
/* following struct exists because we want to basically cache a message within
 * a message for rc (ewww) */
struct resource_msg_internal{
55 56 57 58
    msg_header h;
    /* request data */
    uint64_t req;
    resource_token_t tok; /* only for reserved calls */
59 60 61 62 63
    /* 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; 
64 65 66 67 68
    /* callback data */
    msg_header h_callback;
    int msg_size;
    int msg_header_offset;
    int msg_callback_offset;
69 70 71 72
    /* user-provided data */
    int msg_callback_misc_size;
    int msg_callback_misc_offset;
    char msg_callback_misc[RESOURCE_MAX_CALLBACK_PAYLOAD];
73 74 75 76
}; 

struct resource_msg {
    struct resource_msg_internal i, i_rc;
77 78 79
    // for RC (asides from the message itself): the previous minimum resource
    // value
    uint64_t min_avail_rc;
80 81 82 83 84
};

struct pending_op {
    struct resource_msg_internal m;
    struct qlist_head ql;
85 86 87 88 89 90 91 92
};

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

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

/* ROSS LP processing functions */  
static void resource_lp_ind_init(
93 94
        resource_state * ns,
        tw_lp * lp);
95
static void resource_event_handler(
96 97 98 99
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp);
100
static void resource_rev_handler(
101 102 103 104
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp);
105
static void resource_finalize(
106 107
        resource_state * ns,
        tw_lp * lp);
108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127

/* ROSS function pointer table for this LP */
static tw_lptype resource_lp = {
     (init_f) resource_lp_ind_init,
     (event_f) resource_event_handler,
     (revent_f) resource_rev_handler,
     (final_f)  resource_finalize, 
     (map_f) codes_mapping,
     sizeof(resource_state),
};

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

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

void resource_lp_ind_init(
        resource_state * ns,
        tw_lp * lp){
    /* currently use global to initialize, may need to have other LPs init */
    resource_init(avail_global, &ns->r);
128 129 130 131 132 133 134
    int i;
    for (i = 0; i < MAX_RESERVE+1; i++){
        INIT_QLIST_HEAD(&ns->pending[i]);
    }
}

static void resource_response(
Jonathan Jenkins's avatar
Jonathan Jenkins committed
135
        struct resource_msg_internal *m,
136 137 138 139 140
        tw_lp *lp,
        int ret,
        resource_token_t tok){
    /* send return message */
    msg_header h;
141
    msg_set_header(m->h_callback.magic, m->h_callback.event_type, 
142 143 144 145 146 147 148
            lp->gid, &h);

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

    /* before we send the message, sanity check the sizes */
149 150 151 152
    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, 
153 154
                codes_local_latency(lp), lp);
        void *msg = tw_event_data(e);
155 156 157 158 159
        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);
160
        }
161 162 163 164
        tw_event_send(e);
    }
    else{
        tw_error(TW_LOC, 
165
                "message size not large enough to hold header/callback/misc"
166 167 168 169
                " structures\n"
                "msg size: %3d, header   off/size:  %d, %d\n"
                "               callback off/size:  %d, %d\n"
                "               callback misc size: %d",
170 171 172
                m->msg_size, m->msg_header_offset, (int)sizeof(h),
                m->msg_callback_offset, (int)sizeof(c),
                m->msg_callback_misc_size);
173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189
    }
}
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;
190 191
    // save the previous minimum for RC
    assert(!resource_get_min_avail(m->i.tok, &m->min_avail_rc, &ns->r));
192 193 194 195 196 197 198 199 200 201
    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]);
202
            send_ack = 0;
203 204 205 206
        }
    }
    if (send_ack){
        b->c1 = 1;
207
        resource_response(&m->i, lp, ret, TOKEN_DUMMY);
208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231
    }

    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){
232 233
        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));
234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259
    }
}

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:
260 261
 * c0 - queue was empty to begin with
 * c1 - assuming !c0, alloc succeeded */ 
262 263 264 265 266 267 268
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 */
269
        b->c0 = 1;
270 271 272 273 274
        return;
    }

    struct qlist_head *front = ns->pending[m->i.tok].next;
    pending_op *p = qlist_entry(front, pending_op, ql);
275
    assert(!resource_get_min_avail(m->i.tok, &m->min_avail_rc, &ns->r));
276 277 278
    int ret = resource_get(p->m.req, p->m.tok, &ns->r);
    assert(ret != 2);
    if (!ret){
279
        b->c1 = 1;
280 281 282
        /* success, dequeue (saving as rc) and send to client */
        qlist_del(front);
        m->i_rc = p->m;
283
        resource_response(&p->m, lp, ret, TOKEN_DUMMY);
284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302
        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){
303
    if (b->c0){
304 305 306
        return;
    }

307
    if (b->c1){
308 309 310 311 312
        /* 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);
313
        assert(!resource_restore_min_avail(m->i.tok, m->min_avail_rc, &ns->r)); 
Jonathan Jenkins's avatar
Jonathan Jenkins committed
314
        assert(!resource_free(op->m.req, op->m.tok, &ns->r));
315 316 317 318 319 320 321 322 323 324 325 326 327
        /* 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);
328
    resource_response(&m->i, lp, ret, tok);
329 330 331 332 333 334 335 336 337 338 339 340
}
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);
341 342 343 344 345 346 347
}

void resource_event_handler(
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp){
348 349
    assert(m->i.h.magic == resource_magic);
    switch(m->i.h.event_type){
350
        case RESOURCE_GET:
351
            handle_resource_get(ns,b,m,lp);
352 353
            break;
        case RESOURCE_FREE:
354
            handle_resource_free(ns,b,m,lp);
355
            break;
356 357
        case RESOURCE_DEQ:
            handle_resource_deq(ns,b,m,lp);
358
            break;
359 360
        case RESOURCE_RESERVE:
            handle_resource_reserve(ns,b,m,lp);
361 362
            break;
        default:
363
            assert(0);
364 365 366 367 368 369 370
    }
}
void resource_rev_handler(
        resource_state * ns,
        tw_bf * b,
        resource_msg * m,
        tw_lp * lp){
371 372
    assert(m->i.h.magic == resource_magic);
    switch(m->i.h.event_type){
373
        case RESOURCE_GET:
374
            handle_resource_get_rc(ns,b,m,lp);
375 376
            break;
        case RESOURCE_FREE:
377
            handle_resource_free_rc(ns,b,m,lp);
378
            break;
379 380
        case RESOURCE_DEQ:
            handle_resource_deq_rc(ns,b,m,lp);
381
            break;
382 383
        case RESOURCE_RESERVE:
            handle_resource_reserve_rc(ns,b,m,lp);
384 385
            break;
        default:
386
            assert(0);
387 388 389 390 391 392
    }
}

void resource_finalize(
        resource_state * ns,
        tw_lp * lp){
393
    struct qlist_head *ent;
394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410
    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;
    // see if I'm the "first" resource
    if (codes_mapping_get_lp_global_rel_id(lp->gid) == 0){
        written = sprintf(out_buf, 
                "# format: <LP> <max used general> <max used token...>\n");
        lp_io_write(lp->gid, "resource", written, out_buf);
    }
    written = sprintf(out_buf, "%lu", lp->gid);

411
    // compute peak resource usage
412 413 414
    // 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]);
415
    }
416 417
    written += sprintf(out_buf+written, "\n");
    lp_io_write(lp->gid, "resource", written, out_buf);
418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434
}

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

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

    bj_hashlittle2("resource", strlen("resource"), &h1, &h2);
    resource_magic = h1+h2;

    lp_type_register("resource", &resource_lp);
}

void resource_lp_configure(){
    long int avail;
    int ret = configuration_get_value_longint(&config, "resource", 
435
            "available", NULL, &avail);
436 437 438 439 440 441 442 443 444 445 446 447 448
    if (ret){
        fprintf(stderr, "Could not find section:resource value:available for "
                        "resource LP\n");
        exit(1);
    }
    assert(avail > 0);
    avail_global = (uint64_t) avail;
}

static void resource_lp_issue_event(
        msg_header *header,
        uint64_t req,
        resource_token_t tok, /* only used in reserve_get/free */
449
        int block_on_unavail,
450 451 452
        int msg_size,
        int msg_header_offset,
        int msg_callback_offset,
453 454 455
        int msg_callback_misc_size,
        int msg_callback_misc_offset,
        void *msg_callback_misc_data,
456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474
        enum resource_event type,
        tw_lp *sender){

    tw_lpid resource_lpid;

    /* map out the lpid of the resource */
    int mapping_grp_id, mapping_type_id, mapping_rep_id, mapping_offset;
    char lp_type_name[MAX_NAME_LENGTH], lp_group_name[MAX_NAME_LENGTH];
    codes_mapping_get_lp_info(sender->gid, lp_group_name, 
            &mapping_grp_id, &mapping_type_id, lp_type_name, 
            &mapping_rep_id, &mapping_offset);
    codes_mapping_get_lp_id(lp_group_name, "resource", mapping_rep_id, 
            mapping_offset, &resource_lpid); 

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

    /* set message info */
    resource_msg *m = tw_event_data(e);
475 476 477 478
    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;
479 480

    /* set callback info */
481
    if (header != NULL){
482
        m->i.h_callback = *header;
483
    }
484 485 486
    m->i.msg_size = msg_size;
    m->i.msg_header_offset = msg_header_offset;
    m->i.msg_callback_offset = msg_callback_offset;
487

488 489 490 491 492 493 494 495 496 497 498 499
    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;
    }

500 501 502 503 504 505
    tw_event_send(e);
}

void resource_lp_get(
        msg_header *header,
        uint64_t req, 
506
        int block_on_unavail,
507 508 509
        int msg_size, 
        int msg_header_offset,
        int msg_callback_offset,
510 511 512
        int msg_callback_misc_size,
        int msg_callback_misc_offset,
        void *msg_callback_misc_data,
513
        tw_lp *sender){
514
    resource_lp_issue_event(header, req, 0, block_on_unavail,
515 516 517
            msg_size, msg_header_offset, msg_callback_offset,
            msg_callback_misc_size, msg_callback_misc_offset,
            msg_callback_misc_data, RESOURCE_GET, sender);
518 519 520
}

/* no callback for frees thus far */
521
void resource_lp_free(uint64_t req, tw_lp *sender){
522
    resource_lp_issue_event(NULL, req, 0, -1, -1,-1,-1, 0, 0, NULL,
523 524 525 526 527
            RESOURCE_FREE, sender);
}
void resource_lp_reserve(
        msg_header *header, 
        uint64_t req,
528
        int block_on_unavail,
529 530 531
        int msg_size,
        int msg_header_offset,
        int msg_callback_offset,
532 533 534
        int msg_callback_misc_size,
        int msg_callback_misc_offset,
        void *msg_callback_misc_data,
535
        tw_lp *sender){
536 537 538 539
    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);
540 541 542 543 544
}
void resource_lp_get_reserved(
        msg_header *header,
        uint64_t req,
        resource_token_t tok,
545
        int block_on_unavail,
546 547 548
        int msg_size, 
        int msg_header_offset,
        int msg_callback_offset,
549 550 551
        int msg_callback_misc_size,
        int msg_callback_misc_offset,
        void *msg_callback_misc_data,
552
        tw_lp *sender){
553 554 555
    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,
556 557 558 559 560 561
            sender);
}
void resource_lp_free_reserved(
        uint64_t req, 
        resource_token_t tok,
        tw_lp *sender){
562 563
    resource_lp_issue_event(NULL, req, tok, -1,-1,-1,-1, 0,0,NULL,
            RESOURCE_FREE, sender);
564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596
}

/* 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
 */