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_compat.h>
  42 #include <rte_string_fns.h>
  43 #include <rte_eal_memconfig.h>
  44 #include "rte_distributor_v20.h"
  45 #include "rte_distributor_private.h"
  46
  47 TAILQ_HEAD(rte_distributor_list, rte_distributor_v20);
  48
  49 static struct rte_tailq_elem rte_distributor_tailq = {
  50         .name = "RTE_DISTRIBUTOR",
  51 };
  52 EAL_REGISTER_TAILQ(rte_distributor_tailq)
  53
  54 /**** APIs called by workers ****/
  55
  56 void
  57 rte_distributor_request_pkt_v20(struct rte_distributor_v20 *d,
  58                 unsigned worker_id, struct rte_mbuf *oldpkt)
  59 {
  60         union rte_distributor_buffer_v20 *buf = &d->bufs[worker_id];
  61         int64_t req = (((int64_t)(uintptr_t)oldpkt) << RTE_DISTRIB_FLAG_BITS)
  62                         | RTE_DISTRIB_GET_BUF;
  63         while (unlikely(buf->bufptr64 & RTE_DISTRIB_FLAGS_MASK))
  64                 rte_pause();
  65         buf->bufptr64 = req;
  66 }
  67 VERSION_SYMBOL(rte_distributor_request_pkt, _v20, 2.0);
  68
  69 struct rte_mbuf *
  70 rte_distributor_poll_pkt_v20(struct rte_distributor_v20 *d,
  71                 unsigned worker_id)
  72 {
  73         union rte_distributor_buffer_v20 *buf = &d->bufs[worker_id];
  74         if (buf->bufptr64 & RTE_DISTRIB_GET_BUF)
  75                 return NULL;
  76
  77         /* since bufptr64 is signed, this should be an arithmetic shift */
  78         int64_t ret = buf->bufptr64 >> RTE_DISTRIB_FLAG_BITS;
  79         return (struct rte_mbuf *)((uintptr_t)ret);
  80 }
  81 VERSION_SYMBOL(rte_distributor_poll_pkt, _v20, 2.0);
  82
  83 struct rte_mbuf *
  84 rte_distributor_get_pkt_v20(struct rte_distributor_v20 *d,
  85                 unsigned worker_id, struct rte_mbuf *oldpkt)
  86 {
  87         struct rte_mbuf *ret;
  88         rte_distributor_request_pkt_v20(d, worker_id, oldpkt);
  89         while ((ret = rte_distributor_poll_pkt_v20(d, worker_id)) == NULL)
  90                 rte_pause();
  91         return ret;
  92 }
  93 VERSION_SYMBOL(rte_distributor_get_pkt, _v20, 2.0);
  94
  95 int
  96 rte_distributor_return_pkt_v20(struct rte_distributor_v20 *d,
  97                 unsigned worker_id, struct rte_mbuf *oldpkt)
  98 {
  99         union rte_distributor_buffer_v20 *buf = &d->bufs[worker_id];
 100         uint64_t req = (((int64_t)(uintptr_t)oldpkt) << RTE_DISTRIB_FLAG_BITS)
 101                         | RTE_DISTRIB_RETURN_BUF;
 102         buf->bufptr64 = req;
 103         return 0;
 104 }
 105 VERSION_SYMBOL(rte_distributor_return_pkt, _v20, 2.0);
 106
 107 /**** APIs called on distributor core ***/
 108
 109 /* as name suggests, adds a packet to the backlog for a particular worker */
 110 static int
 111 add_to_backlog(struct rte_distributor_backlog *bl, int64_t item)
 112 {
 113         if (bl->count == RTE_DISTRIB_BACKLOG_SIZE)
 114                 return -1;
 115
 116         bl->pkts[(bl->start + bl->count++) & (RTE_DISTRIB_BACKLOG_MASK)]
 117                         = item;
 118         return 0;
 119 }
 120
 121 /* takes the next packet for a worker off the backlog */
 122 static int64_t
 123 backlog_pop(struct rte_distributor_backlog *bl)
 124 {
 125         bl->count--;
 126         return bl->pkts[bl->start++ & RTE_DISTRIB_BACKLOG_MASK];
 127 }
 128
 129 /* stores a packet returned from a worker inside the returns array */
 130 static inline void
 131 store_return(uintptr_t oldbuf, struct rte_distributor_v20 *d,
 132                 unsigned *ret_start, unsigned *ret_count)
 133 {
 134         /* store returns in a circular buffer - code is branch-free */
 135         d->returns.mbufs[(*ret_start + *ret_count) & RTE_DISTRIB_RETURNS_MASK]
 136                         = (void *)oldbuf;
 137         *ret_start += (*ret_count == RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
 138         *ret_count += (*ret_count != RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
 139 }
 140
 141 static inline void
 142 handle_worker_shutdown(struct rte_distributor_v20 *d, unsigned int wkr)
 143 {
 144         d->in_flight_tags[wkr] = 0;
 145         d->in_flight_bitmask &= ~(1UL << wkr);
 146         d->bufs[wkr].bufptr64 = 0;
 147         if (unlikely(d->backlog[wkr].count != 0)) {
 148                 /* On return of a packet, we need to move the
 149                  * queued packets for this core elsewhere.
 150                  * Easiest solution is to set things up for
 151                  * a recursive call. That will cause those
 152                  * packets to be queued up for the next free
 153                  * core, i.e. it will return as soon as a
 154                  * core becomes free to accept the first
 155                  * packet, as subsequent ones will be added to
 156                  * the backlog for that core.
 157                  */
 158                 struct rte_mbuf *pkts[RTE_DISTRIB_BACKLOG_SIZE];
 159                 unsigned i;
 160                 struct rte_distributor_backlog *bl = &d->backlog[wkr];
 161
 162                 for (i = 0; i < bl->count; i++) {
 163                         unsigned idx = (bl->start + i) &
 164                                         RTE_DISTRIB_BACKLOG_MASK;
 165                         pkts[i] = (void *)((uintptr_t)(bl->pkts[idx] >>
 166                                         RTE_DISTRIB_FLAG_BITS));
 167                 }
 168                 /* recursive call.
 169                  * Note that the tags were set before first level call
 170                  * to rte_distributor_process.
 171                  */
 172                 rte_distributor_process_v20(d, pkts, i);
 173                 bl->count = bl->start = 0;
 174         }
 175 }
 176
 177 /* this function is called when process() fn is called without any new
 178  * packets. It goes through all the workers and clears any returned packets
 179  * to do a partial flush.
 180  */
 181 static int
 182 process_returns(struct rte_distributor_v20 *d)
 183 {
 184         unsigned wkr;
 185         unsigned flushed = 0;
 186         unsigned ret_start = d->returns.start,
 187                         ret_count = d->returns.count;
 188
 189         for (wkr = 0; wkr < d->num_workers; wkr++) {
 190
 191                 const int64_t data = d->bufs[wkr].bufptr64;
 192                 uintptr_t oldbuf = 0;
 193
 194                 if (data & RTE_DISTRIB_GET_BUF) {
 195                         flushed++;
 196                         if (d->backlog[wkr].count)
 197                                 d->bufs[wkr].bufptr64 =
 198                                                 backlog_pop(&d->backlog[wkr]);
 199                         else {
 200                                 d->bufs[wkr].bufptr64 = RTE_DISTRIB_GET_BUF;
 201                                 d->in_flight_tags[wkr] = 0;
 202                                 d->in_flight_bitmask &= ~(1UL << wkr);
 203                         }
 204                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 205                 } else if (data & RTE_DISTRIB_RETURN_BUF) {
 206                         handle_worker_shutdown(d, wkr);
 207                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 208                 }
 209
 210                 store_return(oldbuf, d, &ret_start, &ret_count);
 211         }
 212
 213         d->returns.start = ret_start;
 214         d->returns.count = ret_count;
 215
 216         return flushed;
 217 }
 218
 219 /* process a set of packets to distribute them to workers */
 220 int
 221 rte_distributor_process_v20(struct rte_distributor_v20 *d,
 222                 struct rte_mbuf **mbufs, unsigned num_mbufs)
 223 {
 224         unsigned next_idx = 0;
 225         unsigned wkr = 0;
 226         struct rte_mbuf *next_mb = NULL;
 227         int64_t next_value = 0;
 228         uint32_t new_tag = 0;
 229         unsigned ret_start = d->returns.start,
 230                         ret_count = d->returns.count;
 231
 232         if (unlikely(num_mbufs == 0))
 233                 return process_returns(d);
 234
 235         while (next_idx < num_mbufs || next_mb != NULL) {
 236
 237                 int64_t data = d->bufs[wkr].bufptr64;
 238                 uintptr_t oldbuf = 0;
 239
 240                 if (!next_mb) {
 241                         next_mb = mbufs[next_idx++];
 242                         next_value = (((int64_t)(uintptr_t)next_mb)
 243                                         << RTE_DISTRIB_FLAG_BITS);
 244                         /*
 245                          * User is advocated to set tag vaue for each
 246                          * mbuf before calling rte_distributor_process.
 247                          * User defined tags are used to identify flows,
 248                          * or sessions.
 249                          */
 250                         new_tag = next_mb->hash.usr;
 251
 252                         /*
 253                          * Note that if RTE_DISTRIB_MAX_WORKERS is larger than 64
 254                          * then the size of match has to be expanded.
 255                          */
 256                         uint64_t match = 0;
 257                         unsigned i;
 258                         /*
 259                          * to scan for a match use "xor" and "not" to get a 0/1
 260                          * value, then use shifting to merge to single "match"
 261                          * variable, where a one-bit indicates a match for the
 262                          * worker given by the bit-position
 263                          */
 264                         for (i = 0; i < d->num_workers; i++)
 265                                 match |= (!(d->in_flight_tags[i] ^ new_tag)
 266                                         << i);
 267
 268                         /* Only turned-on bits are considered as match */
 269                         match &= d->in_flight_bitmask;
 270
 271                         if (match) {
 272                                 next_mb = NULL;
 273                                 unsigned worker = __builtin_ctzl(match);
 274                                 if (add_to_backlog(&d->backlog[worker],
 275                                                 next_value) < 0)
 276                                         next_idx--;
 277                         }
 278                 }
 279
 280                 if ((data & RTE_DISTRIB_GET_BUF) &&
 281                                 (d->backlog[wkr].count || next_mb)) {
 282
 283                         if (d->backlog[wkr].count)
 284                                 d->bufs[wkr].bufptr64 =
 285                                                 backlog_pop(&d->backlog[wkr]);
 286
 287                         else {
 288                                 d->bufs[wkr].bufptr64 = next_value;
 289                                 d->in_flight_tags[wkr] = new_tag;
 290                                 d->in_flight_bitmask |= (1UL << wkr);
 291                                 next_mb = NULL;
 292                         }
 293                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 294                 } else if (data & RTE_DISTRIB_RETURN_BUF) {
 295                         handle_worker_shutdown(d, wkr);
 296                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 297                 }
 298
 299                 /* store returns in a circular buffer */
 300                 store_return(oldbuf, d, &ret_start, &ret_count);
 301
 302                 if (++wkr == d->num_workers)
 303                         wkr = 0;
 304         }
 305         /* to finish, check all workers for backlog and schedule work for them
 306          * if they are ready */
 307         for (wkr = 0; wkr < d->num_workers; wkr++)
 308                 if (d->backlog[wkr].count &&
 309                                 (d->bufs[wkr].bufptr64 & RTE_DISTRIB_GET_BUF)) {
 310
 311                         int64_t oldbuf = d->bufs[wkr].bufptr64 >>
 312                                         RTE_DISTRIB_FLAG_BITS;
 313                         store_return(oldbuf, d, &ret_start, &ret_count);
 314
 315                         d->bufs[wkr].bufptr64 = backlog_pop(&d->backlog[wkr]);
 316                 }
 317
 318         d->returns.start = ret_start;
 319         d->returns.count = ret_count;
 320         return num_mbufs;
 321 }
 322 VERSION_SYMBOL(rte_distributor_process, _v20, 2.0);
 323
 324 /* return to the caller, packets returned from workers */
 325 int
 326 rte_distributor_returned_pkts_v20(struct rte_distributor_v20 *d,
 327                 struct rte_mbuf **mbufs, unsigned max_mbufs)
 328 {
 329         struct rte_distributor_returned_pkts *returns = &d->returns;
 330         unsigned retval = (max_mbufs < returns->count) ?
 331                         max_mbufs : returns->count;
 332         unsigned i;
 333
 334         for (i = 0; i < retval; i++) {
 335                 unsigned idx = (returns->start + i) & RTE_DISTRIB_RETURNS_MASK;
 336                 mbufs[i] = returns->mbufs[idx];
 337         }
 338         returns->start += i;
 339         returns->count -= i;
 340
 341         return retval;
 342 }
 343 VERSION_SYMBOL(rte_distributor_returned_pkts, _v20, 2.0);
 344
 345 /* return the number of packets in-flight in a distributor, i.e. packets
 346  * being workered on or queued up in a backlog. */
 347 static inline unsigned
 348 total_outstanding(const struct rte_distributor_v20 *d)
 349 {
 350         unsigned wkr, total_outstanding;
 351
 352         total_outstanding = __builtin_popcountl(d->in_flight_bitmask);
 353
 354         for (wkr = 0; wkr < d->num_workers; wkr++)
 355                 total_outstanding += d->backlog[wkr].count;
 356
 357         return total_outstanding;
 358 }
 359
 360 /* flush the distributor, so that there are no outstanding packets in flight or
 361  * queued up. */
 362 int
 363 rte_distributor_flush_v20(struct rte_distributor_v20 *d)
 364 {
 365         const unsigned flushed = total_outstanding(d);
 366
 367         while (total_outstanding(d) > 0)
 368                 rte_distributor_process_v20(d, NULL, 0);
 369
 370         return flushed;
 371 }
 372 VERSION_SYMBOL(rte_distributor_flush, _v20, 2.0);
 373
 374 /* clears the internal returns array in the distributor */
 375 void
 376 rte_distributor_clear_returns_v20(struct rte_distributor_v20 *d)
 377 {
 378         d->returns.start = d->returns.count = 0;
 379 #ifndef __OPTIMIZE__
 380         memset(d->returns.mbufs, 0, sizeof(d->returns.mbufs));
 381 #endif
 382 }
 383 VERSION_SYMBOL(rte_distributor_clear_returns, _v20, 2.0);
 384
 385 /* creates a distributor instance */
 386 struct rte_distributor_v20 *
 387 rte_distributor_create_v20(const char *name,
 388                 unsigned socket_id,
 389                 unsigned num_workers)
 390 {
 391         struct rte_distributor_v20 *d;
 392         struct rte_distributor_list *distributor_list;
 393         char mz_name[RTE_MEMZONE_NAMESIZE];
 394         const struct rte_memzone *mz;
 395
 396         /* compilation-time checks */
 397         RTE_BUILD_BUG_ON((sizeof(*d) & RTE_CACHE_LINE_MASK) != 0);
 398         RTE_BUILD_BUG_ON((RTE_DISTRIB_MAX_WORKERS & 7) != 0);
 399         RTE_BUILD_BUG_ON(RTE_DISTRIB_MAX_WORKERS >
 400                                 sizeof(d->in_flight_bitmask) * CHAR_BIT);
 401
 402         if (name == NULL || num_workers >= RTE_DISTRIB_MAX_WORKERS) {
 403                 rte_errno = EINVAL;
 404                 return NULL;
 405         }
 406
 407         snprintf(mz_name, sizeof(mz_name), RTE_DISTRIB_PREFIX"%s", name);
 408         mz = rte_memzone_reserve(mz_name, sizeof(*d), socket_id, NO_FLAGS);
 409         if (mz == NULL) {
 410                 rte_errno = ENOMEM;
 411                 return NULL;
 412         }
 413
 414         d = mz->addr;
 415         snprintf(d->name, sizeof(d->name), "%s", name);
 416         d->num_workers = num_workers;
 417
 418         distributor_list = RTE_TAILQ_CAST(rte_distributor_tailq.head,
 419                                           rte_distributor_list);
 420
 421         rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
 422         TAILQ_INSERT_TAIL(distributor_list, d, next);
 423         rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
 424
 425         return d;
 426 }
 427 VERSION_SYMBOL(rte_distributor_create, _v20, 2.0);