lib/librte_distributor/rte_distributor_v20.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stdio.h>
  35 #include <sys/queue.h>
  36 #include <string.h>
  37 #include <rte_mbuf.h>
  38 #include <rte_memory.h>
  39 #include <rte_memzone.h>
  40 #include <rte_errno.h>
  41 #include <rte_string_fns.h>
  42 #include <rte_eal_memconfig.h>
  43 #include "rte_distributor_v20.h"
  44 #include "rte_distributor_private.h"
  45
  46 TAILQ_HEAD(rte_distributor_list, rte_distributor_v20);
  47
  48 static struct rte_tailq_elem rte_distributor_tailq = {
  49         .name = "RTE_DISTRIBUTOR",
  50 };
  51 EAL_REGISTER_TAILQ(rte_distributor_tailq)
  52
  53 /**** APIs called by workers ****/
  54
  55 void
  56 rte_distributor_request_pkt_v20(struct rte_distributor_v20 *d,
  57                 unsigned worker_id, struct rte_mbuf *oldpkt)
  58 {
  59         union rte_distributor_buffer_v20 *buf = &d->bufs[worker_id];
  60         int64_t req = (((int64_t)(uintptr_t)oldpkt) << RTE_DISTRIB_FLAG_BITS)
  61                         | RTE_DISTRIB_GET_BUF;
  62         while (unlikely(buf->bufptr64 & RTE_DISTRIB_FLAGS_MASK))
  63                 rte_pause();
  64         buf->bufptr64 = req;
  65 }
  66
  67 struct rte_mbuf *
  68 rte_distributor_poll_pkt_v20(struct rte_distributor_v20 *d,
  69                 unsigned worker_id)
  70 {
  71         union rte_distributor_buffer_v20 *buf = &d->bufs[worker_id];
  72         if (buf->bufptr64 & RTE_DISTRIB_GET_BUF)
  73                 return NULL;
  74
  75         /* since bufptr64 is signed, this should be an arithmetic shift */
  76         int64_t ret = buf->bufptr64 >> RTE_DISTRIB_FLAG_BITS;
  77         return (struct rte_mbuf *)((uintptr_t)ret);
  78 }
  79
  80 struct rte_mbuf *
  81 rte_distributor_get_pkt_v20(struct rte_distributor_v20 *d,
  82                 unsigned worker_id, struct rte_mbuf *oldpkt)
  83 {
  84         struct rte_mbuf *ret;
  85         rte_distributor_request_pkt_v20(d, worker_id, oldpkt);
  86         while ((ret = rte_distributor_poll_pkt_v20(d, worker_id)) == NULL)
  87                 rte_pause();
  88         return ret;
  89 }
  90
  91 int
  92 rte_distributor_return_pkt_v20(struct rte_distributor_v20 *d,
  93                 unsigned worker_id, struct rte_mbuf *oldpkt)
  94 {
  95         union rte_distributor_buffer_v20 *buf = &d->bufs[worker_id];
  96         uint64_t req = (((int64_t)(uintptr_t)oldpkt) << RTE_DISTRIB_FLAG_BITS)
  97                         | RTE_DISTRIB_RETURN_BUF;
  98         buf->bufptr64 = req;
  99         return 0;
 100 }
 101
 102 /**** APIs called on distributor core ***/
 103
 104 /* as name suggests, adds a packet to the backlog for a particular worker */
 105 static int
 106 add_to_backlog(struct rte_distributor_backlog *bl, int64_t item)
 107 {
 108         if (bl->count == RTE_DISTRIB_BACKLOG_SIZE)
 109                 return -1;
 110
 111         bl->pkts[(bl->start + bl->count++) & (RTE_DISTRIB_BACKLOG_MASK)]
 112                         = item;
 113         return 0;
 114 }
 115
 116 /* takes the next packet for a worker off the backlog */
 117 static int64_t
 118 backlog_pop(struct rte_distributor_backlog *bl)
 119 {
 120         bl->count--;
 121         return bl->pkts[bl->start++ & RTE_DISTRIB_BACKLOG_MASK];
 122 }
 123
 124 /* stores a packet returned from a worker inside the returns array */
 125 static inline void
 126 store_return(uintptr_t oldbuf, struct rte_distributor_v20 *d,
 127                 unsigned *ret_start, unsigned *ret_count)
 128 {
 129         /* store returns in a circular buffer - code is branch-free */
 130         d->returns.mbufs[(*ret_start + *ret_count) & RTE_DISTRIB_RETURNS_MASK]
 131                         = (void *)oldbuf;
 132         *ret_start += (*ret_count == RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
 133         *ret_count += (*ret_count != RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
 134 }
 135
 136 static inline void
 137 handle_worker_shutdown(struct rte_distributor_v20 *d, unsigned int wkr)
 138 {
 139         d->in_flight_tags[wkr] = 0;
 140         d->in_flight_bitmask &= ~(1UL << wkr);
 141         d->bufs[wkr].bufptr64 = 0;
 142         if (unlikely(d->backlog[wkr].count != 0)) {
 143                 /* On return of a packet, we need to move the
 144                  * queued packets for this core elsewhere.
 145                  * Easiest solution is to set things up for
 146                  * a recursive call. That will cause those
 147                  * packets to be queued up for the next free
 148                  * core, i.e. it will return as soon as a
 149                  * core becomes free to accept the first
 150                  * packet, as subsequent ones will be added to
 151                  * the backlog for that core.
 152                  */
 153                 struct rte_mbuf *pkts[RTE_DISTRIB_BACKLOG_SIZE];
 154                 unsigned i;
 155                 struct rte_distributor_backlog *bl = &d->backlog[wkr];
 156
 157                 for (i = 0; i < bl->count; i++) {
 158                         unsigned idx = (bl->start + i) &
 159                                         RTE_DISTRIB_BACKLOG_MASK;
 160                         pkts[i] = (void *)((uintptr_t)(bl->pkts[idx] >>
 161                                         RTE_DISTRIB_FLAG_BITS));
 162                 }
 163                 /* recursive call.
 164                  * Note that the tags were set before first level call
 165                  * to rte_distributor_process.
 166                  */
 167                 rte_distributor_process_v20(d, pkts, i);
 168                 bl->count = bl->start = 0;
 169         }
 170 }
 171
 172 /* this function is called when process() fn is called without any new
 173  * packets. It goes through all the workers and clears any returned packets
 174  * to do a partial flush.
 175  */
 176 static int
 177 process_returns(struct rte_distributor_v20 *d)
 178 {
 179         unsigned wkr;
 180         unsigned flushed = 0;
 181         unsigned ret_start = d->returns.start,
 182                         ret_count = d->returns.count;
 183
 184         for (wkr = 0; wkr < d->num_workers; wkr++) {
 185
 186                 const int64_t data = d->bufs[wkr].bufptr64;
 187                 uintptr_t oldbuf = 0;
 188
 189                 if (data & RTE_DISTRIB_GET_BUF) {
 190                         flushed++;
 191                         if (d->backlog[wkr].count)
 192                                 d->bufs[wkr].bufptr64 =
 193                                                 backlog_pop(&d->backlog[wkr]);
 194                         else {
 195                                 d->bufs[wkr].bufptr64 = RTE_DISTRIB_GET_BUF;
 196                                 d->in_flight_tags[wkr] = 0;
 197                                 d->in_flight_bitmask &= ~(1UL << wkr);
 198                         }
 199                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 200                 } else if (data & RTE_DISTRIB_RETURN_BUF) {
 201                         handle_worker_shutdown(d, wkr);
 202                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 203                 }
 204
 205                 store_return(oldbuf, d, &ret_start, &ret_count);
 206         }
 207
 208         d->returns.start = ret_start;
 209         d->returns.count = ret_count;
 210
 211         return flushed;
 212 }
 213
 214 /* process a set of packets to distribute them to workers */
 215 int
 216 rte_distributor_process_v20(struct rte_distributor_v20 *d,
 217                 struct rte_mbuf **mbufs, unsigned num_mbufs)
 218 {
 219         unsigned next_idx = 0;
 220         unsigned wkr = 0;
 221         struct rte_mbuf *next_mb = NULL;
 222         int64_t next_value = 0;
 223         uint32_t new_tag = 0;
 224         unsigned ret_start = d->returns.start,
 225                         ret_count = d->returns.count;
 226
 227         if (unlikely(num_mbufs == 0))
 228                 return process_returns(d);
 229
 230         while (next_idx < num_mbufs || next_mb != NULL) {
 231
 232                 int64_t data = d->bufs[wkr].bufptr64;
 233                 uintptr_t oldbuf = 0;
 234
 235                 if (!next_mb) {
 236                         next_mb = mbufs[next_idx++];
 237                         next_value = (((int64_t)(uintptr_t)next_mb)
 238                                         << RTE_DISTRIB_FLAG_BITS);
 239                         /*
 240                          * User is advocated to set tag vaue for each
 241                          * mbuf before calling rte_distributor_process.
 242                          * User defined tags are used to identify flows,
 243                          * or sessions.
 244                          */
 245                         new_tag = next_mb->hash.usr;
 246
 247                         /*
 248                          * Note that if RTE_DISTRIB_MAX_WORKERS is larger than 64
 249                          * then the size of match has to be expanded.
 250                          */
 251                         uint64_t match = 0;
 252                         unsigned i;
 253                         /*
 254                          * to scan for a match use "xor" and "not" to get a 0/1
 255                          * value, then use shifting to merge to single "match"
 256                          * variable, where a one-bit indicates a match for the
 257                          * worker given by the bit-position
 258                          */
 259                         for (i = 0; i < d->num_workers; i++)
 260                                 match |= (!(d->in_flight_tags[i] ^ new_tag)
 261                                         << i);
 262
 263                         /* Only turned-on bits are considered as match */
 264                         match &= d->in_flight_bitmask;
 265
 266                         if (match) {
 267                                 next_mb = NULL;
 268                                 unsigned worker = __builtin_ctzl(match);
 269                                 if (add_to_backlog(&d->backlog[worker],
 270                                                 next_value) < 0)
 271                                         next_idx--;
 272                         }
 273                 }
 274
 275                 if ((data & RTE_DISTRIB_GET_BUF) &&
 276                                 (d->backlog[wkr].count || next_mb)) {
 277
 278                         if (d->backlog[wkr].count)
 279                                 d->bufs[wkr].bufptr64 =
 280                                                 backlog_pop(&d->backlog[wkr]);
 281
 282                         else {
 283                                 d->bufs[wkr].bufptr64 = next_value;
 284                                 d->in_flight_tags[wkr] = new_tag;
 285                                 d->in_flight_bitmask |= (1UL << wkr);
 286                                 next_mb = NULL;
 287                         }
 288                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 289                 } else if (data & RTE_DISTRIB_RETURN_BUF) {
 290                         handle_worker_shutdown(d, wkr);
 291                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 292                 }
 293
 294                 /* store returns in a circular buffer */
 295                 store_return(oldbuf, d, &ret_start, &ret_count);
 296
 297                 if (++wkr == d->num_workers)
 298                         wkr = 0;
 299         }
 300         /* to finish, check all workers for backlog and schedule work for them
 301          * if they are ready */
 302         for (wkr = 0; wkr < d->num_workers; wkr++)
 303                 if (d->backlog[wkr].count &&
 304                                 (d->bufs[wkr].bufptr64 & RTE_DISTRIB_GET_BUF)) {
 305
 306                         int64_t oldbuf = d->bufs[wkr].bufptr64 >>
 307                                         RTE_DISTRIB_FLAG_BITS;
 308                         store_return(oldbuf, d, &ret_start, &ret_count);
 309
 310                         d->bufs[wkr].bufptr64 = backlog_pop(&d->backlog[wkr]);
 311                 }
 312
 313         d->returns.start = ret_start;
 314         d->returns.count = ret_count;
 315         return num_mbufs;
 316 }
 317
 318 /* return to the caller, packets returned from workers */
 319 int
 320 rte_distributor_returned_pkts_v20(struct rte_distributor_v20 *d,
 321                 struct rte_mbuf **mbufs, unsigned max_mbufs)
 322 {
 323         struct rte_distributor_returned_pkts *returns = &d->returns;
 324         unsigned retval = (max_mbufs < returns->count) ?
 325                         max_mbufs : returns->count;
 326         unsigned i;
 327
 328         for (i = 0; i < retval; i++) {
 329                 unsigned idx = (returns->start + i) & RTE_DISTRIB_RETURNS_MASK;
 330                 mbufs[i] = returns->mbufs[idx];
 331         }
 332         returns->start += i;
 333         returns->count -= i;
 334
 335         return retval;
 336 }
 337
 338 /* return the number of packets in-flight in a distributor, i.e. packets
 339  * being workered on or queued up in a backlog. */
 340 static inline unsigned
 341 total_outstanding(const struct rte_distributor_v20 *d)
 342 {
 343         unsigned wkr, total_outstanding;
 344
 345         total_outstanding = __builtin_popcountl(d->in_flight_bitmask);
 346
 347         for (wkr = 0; wkr < d->num_workers; wkr++)
 348                 total_outstanding += d->backlog[wkr].count;
 349
 350         return total_outstanding;
 351 }
 352
 353 /* flush the distributor, so that there are no outstanding packets in flight or
 354  * queued up. */
 355 int
 356 rte_distributor_flush_v20(struct rte_distributor_v20 *d)
 357 {
 358         const unsigned flushed = total_outstanding(d);
 359
 360         while (total_outstanding(d) > 0)
 361                 rte_distributor_process_v20(d, NULL, 0);
 362
 363         return flushed;
 364 }
 365
 366 /* clears the internal returns array in the distributor */
 367 void
 368 rte_distributor_clear_returns_v20(struct rte_distributor_v20 *d)
 369 {
 370         d->returns.start = d->returns.count = 0;
 371 #ifndef __OPTIMIZE__
 372         memset(d->returns.mbufs, 0, sizeof(d->returns.mbufs));
 373 #endif
 374 }
 375
 376 /* creates a distributor instance */
 377 struct rte_distributor_v20 *
 378 rte_distributor_create_v20(const char *name,
 379                 unsigned socket_id,
 380                 unsigned num_workers)
 381 {
 382         struct rte_distributor_v20 *d;
 383         struct rte_distributor_list *distributor_list;
 384         char mz_name[RTE_MEMZONE_NAMESIZE];
 385         const struct rte_memzone *mz;
 386
 387         /* compilation-time checks */
 388         RTE_BUILD_BUG_ON((sizeof(*d) & RTE_CACHE_LINE_MASK) != 0);
 389         RTE_BUILD_BUG_ON((RTE_DISTRIB_MAX_WORKERS & 7) != 0);
 390         RTE_BUILD_BUG_ON(RTE_DISTRIB_MAX_WORKERS >
 391                                 sizeof(d->in_flight_bitmask) * CHAR_BIT);
 392
 393         if (name == NULL || num_workers >= RTE_DISTRIB_MAX_WORKERS) {
 394                 rte_errno = EINVAL;
 395                 return NULL;
 396         }
 397
 398         snprintf(mz_name, sizeof(mz_name), RTE_DISTRIB_PREFIX"%s", name);
 399         mz = rte_memzone_reserve(mz_name, sizeof(*d), socket_id, NO_FLAGS);
 400         if (mz == NULL) {
 401                 rte_errno = ENOMEM;
 402                 return NULL;
 403         }
 404
 405         d = mz->addr;
 406         snprintf(d->name, sizeof(d->name), "%s", name);
 407         d->num_workers = num_workers;
 408
 409         distributor_list = RTE_TAILQ_CAST(rte_distributor_tailq.head,
 410                                           rte_distributor_list);
 411
 412         rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
 413         TAILQ_INSERT_TAIL(distributor_list, d, next);
 414         rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
 415
 416         return d;
 417 }