lib/librte_distributor/rte_distributor.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2017 Intel Corporation
   3  */
   4
   5 #include <stdio.h>
   6 #include <sys/queue.h>
   7 #include <string.h>
   8 #include <rte_mbuf.h>
   9 #include <rte_memory.h>
  10 #include <rte_cycles.h>
  11 #include <rte_function_versioning.h>
  12 #include <rte_memzone.h>
  13 #include <rte_errno.h>
  14 #include <rte_string_fns.h>
  15 #include <rte_eal_memconfig.h>
  16 #include <rte_pause.h>
  17 #include <rte_tailq.h>
  18
  19 #include "rte_distributor.h"
  20 #include "rte_distributor_v20.h"
  21 #include "distributor_private.h"
  22
  23 TAILQ_HEAD(rte_dist_burst_list, rte_distributor);
  24
  25 static struct rte_tailq_elem rte_dist_burst_tailq = {
  26         .name = "RTE_DIST_BURST",
  27 };
  28 EAL_REGISTER_TAILQ(rte_dist_burst_tailq)
  29
  30 /**** APIs called by workers ****/
  31
  32 /**** Burst Packet APIs called by workers ****/
  33
  34 void
  35 rte_distributor_request_pkt(struct rte_distributor *d,
  36                 unsigned int worker_id, struct rte_mbuf **oldpkt,
  37                 unsigned int count)
  38 {
  39         struct rte_distributor_buffer *buf = &(d->bufs[worker_id]);
  40         unsigned int i;
  41
  42         volatile int64_t *retptr64;
  43
  44         if (unlikely(d->alg_type == RTE_DIST_ALG_SINGLE)) {
  45                 rte_distributor_request_pkt_v20(d->d_v20,
  46                         worker_id, oldpkt[0]);
  47                 return;
  48         }
  49
  50         retptr64 = &(buf->retptr64[0]);
  51         /* Spin while handshake bits are set (scheduler clears it).
  52          * Sync with worker on GET_BUF flag.
  53          */
  54         while (unlikely(__atomic_load_n(retptr64, __ATOMIC_ACQUIRE)
  55                         & RTE_DISTRIB_GET_BUF)) {
  56                 rte_pause();
  57                 uint64_t t = rte_rdtsc()+100;
  58
  59                 while (rte_rdtsc() < t)
  60                         rte_pause();
  61         }
  62
  63         /*
  64          * OK, if we've got here, then the scheduler has just cleared the
  65          * handshake bits. Populate the retptrs with returning packets.
  66          */
  67
  68         for (i = count; i < RTE_DIST_BURST_SIZE; i++)
  69                 buf->retptr64[i] = 0;
  70
  71         /* Set Return bit for each packet returned */
  72         for (i = count; i-- > 0; )
  73                 buf->retptr64[i] =
  74                         (((int64_t)(uintptr_t)(oldpkt[i])) <<
  75                         RTE_DISTRIB_FLAG_BITS) | RTE_DISTRIB_RETURN_BUF;
  76
  77         /*
  78          * Finally, set the GET_BUF  to signal to distributor that cache
  79          * line is ready for processing
  80          * Sync with distributor to release retptrs
  81          */
  82         __atomic_store_n(retptr64, *retptr64 | RTE_DISTRIB_GET_BUF,
  83                         __ATOMIC_RELEASE);
  84 }
  85
  86 int
  87 rte_distributor_poll_pkt(struct rte_distributor *d,
  88                 unsigned int worker_id, struct rte_mbuf **pkts)
  89 {
  90         struct rte_distributor_buffer *buf = &d->bufs[worker_id];
  91         uint64_t ret;
  92         int count = 0;
  93         unsigned int i;
  94
  95         if (unlikely(d->alg_type == RTE_DIST_ALG_SINGLE)) {
  96                 pkts[0] = rte_distributor_poll_pkt_v20(d->d_v20, worker_id);
  97                 return (pkts[0]) ? 1 : 0;
  98         }
  99
 100         /* If bit is set, return
 101          * Sync with distributor to acquire bufptrs
 102          */
 103         if (__atomic_load_n(&(buf->bufptr64[0]), __ATOMIC_ACQUIRE)
 104                 & RTE_DISTRIB_GET_BUF)
 105                 return -1;
 106
 107         /* since bufptr64 is signed, this should be an arithmetic shift */
 108         for (i = 0; i < RTE_DIST_BURST_SIZE; i++) {
 109                 if (likely(buf->bufptr64[i] & RTE_DISTRIB_VALID_BUF)) {
 110                         ret = buf->bufptr64[i] >> RTE_DISTRIB_FLAG_BITS;
 111                         pkts[count++] = (struct rte_mbuf *)((uintptr_t)(ret));
 112                 }
 113         }
 114
 115         /*
 116          * so now we've got the contents of the cacheline into an  array of
 117          * mbuf pointers, so toggle the bit so scheduler can start working
 118          * on the next cacheline while we're working.
 119          * Sync with distributor on GET_BUF flag. Release bufptrs.
 120          */
 121         __atomic_store_n(&(buf->bufptr64[0]),
 122                 buf->bufptr64[0] | RTE_DISTRIB_GET_BUF, __ATOMIC_RELEASE);
 123
 124         return count;
 125 }
 126
 127 int
 128 rte_distributor_get_pkt(struct rte_distributor *d,
 129                 unsigned int worker_id, struct rte_mbuf **pkts,
 130                 struct rte_mbuf **oldpkt, unsigned int return_count)
 131 {
 132         int count;
 133
 134         if (unlikely(d->alg_type == RTE_DIST_ALG_SINGLE)) {
 135                 if (return_count <= 1) {
 136                         pkts[0] = rte_distributor_get_pkt_v20(d->d_v20,
 137                                 worker_id, oldpkt[0]);
 138                         return (pkts[0]) ? 1 : 0;
 139                 } else
 140                         return -EINVAL;
 141         }
 142
 143         rte_distributor_request_pkt(d, worker_id, oldpkt, return_count);
 144
 145         count = rte_distributor_poll_pkt(d, worker_id, pkts);
 146         while (count == -1) {
 147                 uint64_t t = rte_rdtsc() + 100;
 148
 149                 while (rte_rdtsc() < t)
 150                         rte_pause();
 151
 152                 count = rte_distributor_poll_pkt(d, worker_id, pkts);
 153         }
 154         return count;
 155 }
 156
 157 int
 158 rte_distributor_return_pkt(struct rte_distributor *d,
 159                 unsigned int worker_id, struct rte_mbuf **oldpkt, int num)
 160 {
 161         struct rte_distributor_buffer *buf = &d->bufs[worker_id];
 162         unsigned int i;
 163
 164         if (unlikely(d->alg_type == RTE_DIST_ALG_SINGLE)) {
 165                 if (num == 1)
 166                         return rte_distributor_return_pkt_v20(d->d_v20,
 167                                 worker_id, oldpkt[0]);
 168                 else
 169                         return -EINVAL;
 170         }
 171
 172         /* Sync with distributor to acquire retptrs */
 173         __atomic_thread_fence(__ATOMIC_ACQUIRE);
 174         for (i = 0; i < RTE_DIST_BURST_SIZE; i++)
 175                 /* Switch off the return bit first */
 176                 buf->retptr64[i] &= ~RTE_DISTRIB_RETURN_BUF;
 177
 178         for (i = num; i-- > 0; )
 179                 buf->retptr64[i] = (((int64_t)(uintptr_t)oldpkt[i]) <<
 180                         RTE_DISTRIB_FLAG_BITS) | RTE_DISTRIB_RETURN_BUF;
 181
 182         /* set the GET_BUF but even if we got no returns.
 183          * Sync with distributor on GET_BUF flag. Release retptrs.
 184          */
 185         __atomic_store_n(&(buf->retptr64[0]),
 186                 buf->retptr64[0] | RTE_DISTRIB_GET_BUF, __ATOMIC_RELEASE);
 187
 188         return 0;
 189 }
 190
 191 /**** APIs called on distributor core ***/
 192
 193 /* stores a packet returned from a worker inside the returns array */
 194 static inline void
 195 store_return(uintptr_t oldbuf, struct rte_distributor *d,
 196                 unsigned int *ret_start, unsigned int *ret_count)
 197 {
 198         if (!oldbuf)
 199                 return;
 200         /* store returns in a circular buffer */
 201         d->returns.mbufs[(*ret_start + *ret_count) & RTE_DISTRIB_RETURNS_MASK]
 202                         = (void *)oldbuf;
 203         *ret_start += (*ret_count == RTE_DISTRIB_RETURNS_MASK);
 204         *ret_count += (*ret_count != RTE_DISTRIB_RETURNS_MASK);
 205 }
 206
 207 /*
 208  * Match then flow_ids (tags) of the incoming packets to the flow_ids
 209  * of the inflight packets (both inflight on the workers and in each worker
 210  * backlog). This will then allow us to pin those packets to the relevant
 211  * workers to give us our atomic flow pinning.
 212  */
 213 void
 214 find_match_scalar(struct rte_distributor *d,
 215                         uint16_t *data_ptr,
 216                         uint16_t *output_ptr)
 217 {
 218         struct rte_distributor_backlog *bl;
 219         uint16_t i, j, w;
 220
 221         /*
 222          * Function overview:
 223          * 1. Loop through all worker ID's
 224          * 2. Compare the current inflights to the incoming tags
 225          * 3. Compare the current backlog to the incoming tags
 226          * 4. Add any matches to the output
 227          */
 228
 229         for (j = 0 ; j < RTE_DIST_BURST_SIZE; j++)
 230                 output_ptr[j] = 0;
 231
 232         for (i = 0; i < d->num_workers; i++) {
 233                 bl = &d->backlog[i];
 234
 235                 for (j = 0; j < RTE_DIST_BURST_SIZE ; j++)
 236                         for (w = 0; w < RTE_DIST_BURST_SIZE; w++)
 237                                 if (d->in_flight_tags[i][j] == data_ptr[w]) {
 238                                         output_ptr[j] = i+1;
 239                                         break;
 240                                 }
 241                 for (j = 0; j < RTE_DIST_BURST_SIZE; j++)
 242                         for (w = 0; w < RTE_DIST_BURST_SIZE; w++)
 243                                 if (bl->tags[j] == data_ptr[w]) {
 244                                         output_ptr[j] = i+1;
 245                                         break;
 246                                 }
 247         }
 248
 249         /*
 250          * At this stage, the output contains 8 16-bit values, with
 251          * each non-zero value containing the worker ID on which the
 252          * corresponding flow is pinned to.
 253          */
 254 }
 255
 256
 257 /*
 258  * When the handshake bits indicate that there are packets coming
 259  * back from the worker, this function is called to copy and store
 260  * the valid returned pointers (store_return).
 261  */
 262 static unsigned int
 263 handle_returns(struct rte_distributor *d, unsigned int wkr)
 264 {
 265         struct rte_distributor_buffer *buf = &(d->bufs[wkr]);
 266         uintptr_t oldbuf;
 267         unsigned int ret_start = d->returns.start,
 268                         ret_count = d->returns.count;
 269         unsigned int count = 0;
 270         unsigned int i;
 271
 272         /* Sync on GET_BUF flag. Acquire retptrs. */
 273         if (__atomic_load_n(&(buf->retptr64[0]), __ATOMIC_ACQUIRE)
 274                 & RTE_DISTRIB_GET_BUF) {
 275                 for (i = 0; i < RTE_DIST_BURST_SIZE; i++) {
 276                         if (buf->retptr64[i] & RTE_DISTRIB_RETURN_BUF) {
 277                                 oldbuf = ((uintptr_t)(buf->retptr64[i] >>
 278                                         RTE_DISTRIB_FLAG_BITS));
 279                                 /* store returns in a circular buffer */
 280                                 store_return(oldbuf, d, &ret_start, &ret_count);
 281                                 count++;
 282                                 buf->retptr64[i] &= ~RTE_DISTRIB_RETURN_BUF;
 283                         }
 284                 }
 285                 d->returns.start = ret_start;
 286                 d->returns.count = ret_count;
 287                 /* Clear for the worker to populate with more returns.
 288                  * Sync with distributor on GET_BUF flag. Release retptrs.
 289                  */
 290                 __atomic_store_n(&(buf->retptr64[0]), 0, __ATOMIC_RELEASE);
 291         }
 292         return count;
 293 }
 294
 295 /*
 296  * This function releases a burst (cache line) to a worker.
 297  * It is called from the process function when a cacheline is
 298  * full to make room for more packets for that worker, or when
 299  * all packets have been assigned to bursts and need to be flushed
 300  * to the workers.
 301  * It also needs to wait for any outstanding packets from the worker
 302  * before sending out new packets.
 303  */
 304 static unsigned int
 305 release(struct rte_distributor *d, unsigned int wkr)
 306 {
 307         struct rte_distributor_buffer *buf = &(d->bufs[wkr]);
 308         unsigned int i;
 309
 310         /* Sync with worker on GET_BUF flag */
 311         while (!(__atomic_load_n(&(d->bufs[wkr].bufptr64[0]), __ATOMIC_ACQUIRE)
 312                 & RTE_DISTRIB_GET_BUF))
 313                 rte_pause();
 314
 315         handle_returns(d, wkr);
 316
 317         buf->count = 0;
 318
 319         for (i = 0; i < d->backlog[wkr].count; i++) {
 320                 d->bufs[wkr].bufptr64[i] = d->backlog[wkr].pkts[i] |
 321                                 RTE_DISTRIB_GET_BUF | RTE_DISTRIB_VALID_BUF;
 322                 d->in_flight_tags[wkr][i] = d->backlog[wkr].tags[i];
 323         }
 324         buf->count = i;
 325         for ( ; i < RTE_DIST_BURST_SIZE ; i++) {
 326                 buf->bufptr64[i] = RTE_DISTRIB_GET_BUF;
 327                 d->in_flight_tags[wkr][i] = 0;
 328         }
 329
 330         d->backlog[wkr].count = 0;
 331
 332         /* Clear the GET bit.
 333          * Sync with worker on GET_BUF flag. Release bufptrs.
 334          */
 335         __atomic_store_n(&(buf->bufptr64[0]),
 336                 buf->bufptr64[0] & ~RTE_DISTRIB_GET_BUF, __ATOMIC_RELEASE);
 337         return  buf->count;
 338
 339 }
 340
 341
 342 /* process a set of packets to distribute them to workers */
 343 int
 344 rte_distributor_process(struct rte_distributor *d,
 345                 struct rte_mbuf **mbufs, unsigned int num_mbufs)
 346 {
 347         unsigned int next_idx = 0;
 348         static unsigned int wkr;
 349         struct rte_mbuf *next_mb = NULL;
 350         int64_t next_value = 0;
 351         uint16_t new_tag = 0;
 352         uint16_t flows[RTE_DIST_BURST_SIZE] __rte_cache_aligned;
 353         unsigned int i, j, w, wid;
 354
 355         if (d->alg_type == RTE_DIST_ALG_SINGLE) {
 356                 /* Call the old API */
 357                 return rte_distributor_process_v20(d->d_v20, mbufs, num_mbufs);
 358         }
 359
 360         if (unlikely(num_mbufs == 0)) {
 361                 /* Flush out all non-full cache-lines to workers. */
 362                 for (wid = 0 ; wid < d->num_workers; wid++) {
 363                         /* Sync with worker on GET_BUF flag. */
 364                         if (__atomic_load_n(&(d->bufs[wid].bufptr64[0]),
 365                                 __ATOMIC_ACQUIRE) & RTE_DISTRIB_GET_BUF) {
 366                                 release(d, wid);
 367                                 handle_returns(d, wid);
 368                         }
 369                 }
 370                 return 0;
 371         }
 372
 373         while (next_idx < num_mbufs) {
 374                 uint16_t matches[RTE_DIST_BURST_SIZE];
 375                 unsigned int pkts;
 376
 377                 /* Sync with worker on GET_BUF flag. */
 378                 if (__atomic_load_n(&(d->bufs[wkr].bufptr64[0]),
 379                         __ATOMIC_ACQUIRE) & RTE_DISTRIB_GET_BUF)
 380                         d->bufs[wkr].count = 0;
 381
 382                 if ((num_mbufs - next_idx) < RTE_DIST_BURST_SIZE)
 383                         pkts = num_mbufs - next_idx;
 384                 else
 385                         pkts = RTE_DIST_BURST_SIZE;
 386
 387                 for (i = 0; i < pkts; i++) {
 388                         if (mbufs[next_idx + i]) {
 389                                 /* flows have to be non-zero */
 390                                 flows[i] = mbufs[next_idx + i]->hash.usr | 1;
 391                         } else
 392                                 flows[i] = 0;
 393                 }
 394                 for (; i < RTE_DIST_BURST_SIZE; i++)
 395                         flows[i] = 0;
 396
 397                 switch (d->dist_match_fn) {
 398                 case RTE_DIST_MATCH_VECTOR:
 399                         find_match_vec(d, &flows[0], &matches[0]);
 400                         break;
 401                 default:
 402                         find_match_scalar(d, &flows[0], &matches[0]);
 403                 }
 404
 405                 /*
 406                  * Matches array now contain the intended worker ID (+1) of
 407                  * the incoming packets. Any zeroes need to be assigned
 408                  * workers.
 409                  */
 410
 411                 for (j = 0; j < pkts; j++) {
 412
 413                         next_mb = mbufs[next_idx++];
 414                         next_value = (((int64_t)(uintptr_t)next_mb) <<
 415                                         RTE_DISTRIB_FLAG_BITS);
 416                         /*
 417                          * User is advocated to set tag value for each
 418                          * mbuf before calling rte_distributor_process.
 419                          * User defined tags are used to identify flows,
 420                          * or sessions.
 421                          */
 422                         /* flows MUST be non-zero */
 423                         new_tag = (uint16_t)(next_mb->hash.usr) | 1;
 424
 425                         /*
 426                          * Uncommenting the next line will cause the find_match
 427                          * function to be optimized out, making this function
 428                          * do parallel (non-atomic) distribution
 429                          */
 430                         /* matches[j] = 0; */
 431
 432                         if (matches[j]) {
 433                                 struct rte_distributor_backlog *bl =
 434                                                 &d->backlog[matches[j]-1];
 435                                 if (unlikely(bl->count ==
 436                                                 RTE_DIST_BURST_SIZE)) {
 437                                         release(d, matches[j]-1);
 438                                 }
 439
 440                                 /* Add to worker that already has flow */
 441                                 unsigned int idx = bl->count++;
 442
 443                                 bl->tags[idx] = new_tag;
 444                                 bl->pkts[idx] = next_value;
 445
 446                         } else {
 447                                 struct rte_distributor_backlog *bl =
 448                                                 &d->backlog[wkr];
 449                                 if (unlikely(bl->count ==
 450                                                 RTE_DIST_BURST_SIZE)) {
 451                                         release(d, wkr);
 452                                 }
 453
 454                                 /* Add to current worker worker */
 455                                 unsigned int idx = bl->count++;
 456
 457                                 bl->tags[idx] = new_tag;
 458                                 bl->pkts[idx] = next_value;
 459                                 /*
 460                                  * Now that we've just added an unpinned flow
 461                                  * to a worker, we need to ensure that all
 462                                  * other packets with that same flow will go
 463                                  * to the same worker in this burst.
 464                                  */
 465                                 for (w = j; w < pkts; w++)
 466                                         if (flows[w] == new_tag)
 467                                                 matches[w] = wkr+1;
 468                         }
 469                 }
 470                 wkr++;
 471                 if (wkr >= d->num_workers)
 472                         wkr = 0;
 473         }
 474
 475         /* Flush out all non-full cache-lines to workers. */
 476         for (wid = 0 ; wid < d->num_workers; wid++)
 477                 /* Sync with worker on GET_BUF flag. */
 478                 if ((__atomic_load_n(&(d->bufs[wid].bufptr64[0]),
 479                         __ATOMIC_ACQUIRE) & RTE_DISTRIB_GET_BUF))
 480                         release(d, wid);
 481
 482         return num_mbufs;
 483 }
 484
 485 /* return to the caller, packets returned from workers */
 486 int
 487 rte_distributor_returned_pkts(struct rte_distributor *d,
 488                 struct rte_mbuf **mbufs, unsigned int max_mbufs)
 489 {
 490         struct rte_distributor_returned_pkts *returns = &d->returns;
 491         unsigned int retval = (max_mbufs < returns->count) ?
 492                         max_mbufs : returns->count;
 493         unsigned int i;
 494
 495         if (d->alg_type == RTE_DIST_ALG_SINGLE) {
 496                 /* Call the old API */
 497                 return rte_distributor_returned_pkts_v20(d->d_v20,
 498                                 mbufs, max_mbufs);
 499         }
 500
 501         for (i = 0; i < retval; i++) {
 502                 unsigned int idx = (returns->start + i) &
 503                                 RTE_DISTRIB_RETURNS_MASK;
 504
 505                 mbufs[i] = returns->mbufs[idx];
 506         }
 507         returns->start += i;
 508         returns->count -= i;
 509
 510         return retval;
 511 }
 512
 513 /*
 514  * Return the number of packets in-flight in a distributor, i.e. packets
 515  * being worked on or queued up in a backlog.
 516  */
 517 static inline unsigned int
 518 total_outstanding(const struct rte_distributor *d)
 519 {
 520         unsigned int wkr, total_outstanding = 0;
 521
 522         for (wkr = 0; wkr < d->num_workers; wkr++)
 523                 total_outstanding += d->backlog[wkr].count;
 524
 525         return total_outstanding;
 526 }
 527
 528 /*
 529  * Flush the distributor, so that there are no outstanding packets in flight or
 530  * queued up.
 531  */
 532 int
 533 rte_distributor_flush(struct rte_distributor *d)
 534 {
 535         unsigned int flushed;
 536         unsigned int wkr;
 537
 538         if (d->alg_type == RTE_DIST_ALG_SINGLE) {
 539                 /* Call the old API */
 540                 return rte_distributor_flush_v20(d->d_v20);
 541         }
 542
 543         flushed = total_outstanding(d);
 544
 545         while (total_outstanding(d) > 0)
 546                 rte_distributor_process(d, NULL, 0);
 547
 548         /* wait 10ms to allow all worker drain the pkts */
 549         rte_delay_us(10000);
 550
 551         /*
 552          * Send empty burst to all workers to allow them to exit
 553          * gracefully, should they need to.
 554          */
 555         rte_distributor_process(d, NULL, 0);
 556
 557         for (wkr = 0; wkr < d->num_workers; wkr++)
 558                 handle_returns(d, wkr);
 559
 560         return flushed;
 561 }
 562
 563 /* clears the internal returns array in the distributor */
 564 void
 565 rte_distributor_clear_returns(struct rte_distributor *d)
 566 {
 567         unsigned int wkr;
 568
 569         if (d->alg_type == RTE_DIST_ALG_SINGLE) {
 570                 /* Call the old API */
 571                 rte_distributor_clear_returns_v20(d->d_v20);
 572                 return;
 573         }
 574
 575         /* throw away returns, so workers can exit */
 576         for (wkr = 0; wkr < d->num_workers; wkr++)
 577                 /* Sync with worker. Release retptrs. */
 578                 __atomic_store_n(&(d->bufs[wkr].retptr64[0]), 0,
 579                                 __ATOMIC_RELEASE);
 580 }
 581
 582 /* creates a distributor instance */
 583 struct rte_distributor *
 584 rte_distributor_create(const char *name,
 585                 unsigned int socket_id,
 586                 unsigned int num_workers,
 587                 unsigned int alg_type)
 588 {
 589         struct rte_distributor *d;
 590         struct rte_dist_burst_list *dist_burst_list;
 591         char mz_name[RTE_MEMZONE_NAMESIZE];
 592         const struct rte_memzone *mz;
 593         unsigned int i;
 594
 595         /* TODO Reorganise function properly around RTE_DIST_ALG_SINGLE/BURST */
 596
 597         /* compilation-time checks */
 598         RTE_BUILD_BUG_ON((sizeof(*d) & RTE_CACHE_LINE_MASK) != 0);
 599         RTE_BUILD_BUG_ON((RTE_DISTRIB_MAX_WORKERS & 7) != 0);
 600
 601         if (name == NULL || num_workers >=
 602                 (unsigned int)RTE_MIN(RTE_DISTRIB_MAX_WORKERS, RTE_MAX_LCORE)) {
 603                 rte_errno = EINVAL;
 604                 return NULL;
 605         }
 606
 607         if (alg_type == RTE_DIST_ALG_SINGLE) {
 608                 d = malloc(sizeof(struct rte_distributor));
 609                 if (d == NULL) {
 610                         rte_errno = ENOMEM;
 611                         return NULL;
 612                 }
 613                 d->d_v20 = rte_distributor_create_v20(name,
 614                                 socket_id, num_workers);
 615                 if (d->d_v20 == NULL) {
 616                         free(d);
 617                         /* rte_errno will have been set */
 618                         return NULL;
 619                 }
 620                 d->alg_type = alg_type;
 621                 return d;
 622         }
 623
 624         snprintf(mz_name, sizeof(mz_name), RTE_DISTRIB_PREFIX"%s", name);
 625         mz = rte_memzone_reserve(mz_name, sizeof(*d), socket_id, NO_FLAGS);
 626         if (mz == NULL) {
 627                 rte_errno = ENOMEM;
 628                 return NULL;
 629         }
 630
 631         d = mz->addr;
 632         strlcpy(d->name, name, sizeof(d->name));
 633         d->num_workers = num_workers;
 634         d->alg_type = alg_type;
 635
 636         d->dist_match_fn = RTE_DIST_MATCH_SCALAR;
 637 #if defined(RTE_ARCH_X86)
 638         d->dist_match_fn = RTE_DIST_MATCH_VECTOR;
 639 #endif
 640
 641         /*
 642          * Set up the backlog tags so they're pointing at the second cache
 643          * line for performance during flow matching
 644          */
 645         for (i = 0 ; i < num_workers ; i++)
 646                 d->backlog[i].tags = &d->in_flight_tags[i][RTE_DIST_BURST_SIZE];
 647
 648         dist_burst_list = RTE_TAILQ_CAST(rte_dist_burst_tailq.head,
 649                                           rte_dist_burst_list);
 650
 651
 652         rte_mcfg_tailq_write_lock();
 653         TAILQ_INSERT_TAIL(dist_burst_list, d, next);
 654         rte_mcfg_tailq_write_unlock();
 655
 656         return d;
 657 }