lib/librte_distributor/rte_distributor_v20.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2014 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_memzone.h>
  11 #include <rte_errno.h>
  12 #include <rte_compat.h>
  13 #include <rte_string_fns.h>
  14 #include <rte_eal_memconfig.h>
  15 #include <rte_pause.h>
  16 #include <rte_tailq.h>
  17
  18 #include "rte_distributor_v20.h"
  19 #include "rte_distributor_private.h"
  20
  21 TAILQ_HEAD(rte_distributor_list, rte_distributor_v20);
  22
  23 static struct rte_tailq_elem rte_distributor_tailq = {
  24         .name = "RTE_DISTRIBUTOR",
  25 };
  26 EAL_REGISTER_TAILQ(rte_distributor_tailq)
  27
  28 /**** APIs called by workers ****/
  29
  30 void
  31 rte_distributor_request_pkt_v20(struct rte_distributor_v20 *d,
  32                 unsigned worker_id, struct rte_mbuf *oldpkt)
  33 {
  34         union rte_distributor_buffer_v20 *buf = &d->bufs[worker_id];
  35         int64_t req = (((int64_t)(uintptr_t)oldpkt) << RTE_DISTRIB_FLAG_BITS)
  36                         | RTE_DISTRIB_GET_BUF;
  37         while (unlikely(__atomic_load_n(&buf->bufptr64, __ATOMIC_RELAXED)
  38                         & RTE_DISTRIB_FLAGS_MASK))
  39                 rte_pause();
  40
  41         /* Sync with distributor on GET_BUF flag. */
  42         __atomic_store_n(&(buf->bufptr64), req, __ATOMIC_RELEASE);
  43 }
  44 VERSION_SYMBOL(rte_distributor_request_pkt, _v20, 2.0);
  45
  46 struct rte_mbuf *
  47 rte_distributor_poll_pkt_v20(struct rte_distributor_v20 *d,
  48                 unsigned worker_id)
  49 {
  50         union rte_distributor_buffer_v20 *buf = &d->bufs[worker_id];
  51         /* Sync with distributor. Acquire bufptr64. */
  52         if (__atomic_load_n(&buf->bufptr64, __ATOMIC_ACQUIRE)
  53                 & RTE_DISTRIB_GET_BUF)
  54                 return NULL;
  55
  56         /* since bufptr64 is signed, this should be an arithmetic shift */
  57         int64_t ret = buf->bufptr64 >> RTE_DISTRIB_FLAG_BITS;
  58         return (struct rte_mbuf *)((uintptr_t)ret);
  59 }
  60 VERSION_SYMBOL(rte_distributor_poll_pkt, _v20, 2.0);
  61
  62 struct rte_mbuf *
  63 rte_distributor_get_pkt_v20(struct rte_distributor_v20 *d,
  64                 unsigned worker_id, struct rte_mbuf *oldpkt)
  65 {
  66         struct rte_mbuf *ret;
  67         rte_distributor_request_pkt_v20(d, worker_id, oldpkt);
  68         while ((ret = rte_distributor_poll_pkt_v20(d, worker_id)) == NULL)
  69                 rte_pause();
  70         return ret;
  71 }
  72 VERSION_SYMBOL(rte_distributor_get_pkt, _v20, 2.0);
  73
  74 int
  75 rte_distributor_return_pkt_v20(struct rte_distributor_v20 *d,
  76                 unsigned worker_id, struct rte_mbuf *oldpkt)
  77 {
  78         union rte_distributor_buffer_v20 *buf = &d->bufs[worker_id];
  79         uint64_t req = (((int64_t)(uintptr_t)oldpkt) << RTE_DISTRIB_FLAG_BITS)
  80                         | RTE_DISTRIB_RETURN_BUF;
  81         /* Sync with distributor on RETURN_BUF flag. */
  82         __atomic_store_n(&(buf->bufptr64), req, __ATOMIC_RELEASE);
  83         return 0;
  84 }
  85 VERSION_SYMBOL(rte_distributor_return_pkt, _v20, 2.0);
  86
  87 /**** APIs called on distributor core ***/
  88
  89 /* as name suggests, adds a packet to the backlog for a particular worker */
  90 static int
  91 add_to_backlog(struct rte_distributor_backlog *bl, int64_t item)
  92 {
  93         if (bl->count == RTE_DISTRIB_BACKLOG_SIZE)
  94                 return -1;
  95
  96         bl->pkts[(bl->start + bl->count++) & (RTE_DISTRIB_BACKLOG_MASK)]
  97                         = item;
  98         return 0;
  99 }
 100
 101 /* takes the next packet for a worker off the backlog */
 102 static int64_t
 103 backlog_pop(struct rte_distributor_backlog *bl)
 104 {
 105         bl->count--;
 106         return bl->pkts[bl->start++ & RTE_DISTRIB_BACKLOG_MASK];
 107 }
 108
 109 /* stores a packet returned from a worker inside the returns array */
 110 static inline void
 111 store_return(uintptr_t oldbuf, struct rte_distributor_v20 *d,
 112                 unsigned *ret_start, unsigned *ret_count)
 113 {
 114         /* store returns in a circular buffer - code is branch-free */
 115         d->returns.mbufs[(*ret_start + *ret_count) & RTE_DISTRIB_RETURNS_MASK]
 116                         = (void *)oldbuf;
 117         *ret_start += (*ret_count == RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
 118         *ret_count += (*ret_count != RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
 119 }
 120
 121 static inline void
 122 handle_worker_shutdown(struct rte_distributor_v20 *d, unsigned int wkr)
 123 {
 124         d->in_flight_tags[wkr] = 0;
 125         d->in_flight_bitmask &= ~(1UL << wkr);
 126         /* Sync with worker. Release bufptr64. */
 127         __atomic_store_n(&(d->bufs[wkr].bufptr64), 0, __ATOMIC_RELEASE);
 128         if (unlikely(d->backlog[wkr].count != 0)) {
 129                 /* On return of a packet, we need to move the
 130                  * queued packets for this core elsewhere.
 131                  * Easiest solution is to set things up for
 132                  * a recursive call. That will cause those
 133                  * packets to be queued up for the next free
 134                  * core, i.e. it will return as soon as a
 135                  * core becomes free to accept the first
 136                  * packet, as subsequent ones will be added to
 137                  * the backlog for that core.
 138                  */
 139                 struct rte_mbuf *pkts[RTE_DISTRIB_BACKLOG_SIZE];
 140                 unsigned i;
 141                 struct rte_distributor_backlog *bl = &d->backlog[wkr];
 142
 143                 for (i = 0; i < bl->count; i++) {
 144                         unsigned idx = (bl->start + i) &
 145                                         RTE_DISTRIB_BACKLOG_MASK;
 146                         pkts[i] = (void *)((uintptr_t)(bl->pkts[idx] >>
 147                                         RTE_DISTRIB_FLAG_BITS));
 148                 }
 149                 /* recursive call.
 150                  * Note that the tags were set before first level call
 151                  * to rte_distributor_process.
 152                  */
 153                 rte_distributor_process_v20(d, pkts, i);
 154                 bl->count = bl->start = 0;
 155         }
 156 }
 157
 158 /* this function is called when process() fn is called without any new
 159  * packets. It goes through all the workers and clears any returned packets
 160  * to do a partial flush.
 161  */
 162 static int
 163 process_returns(struct rte_distributor_v20 *d)
 164 {
 165         unsigned wkr;
 166         unsigned flushed = 0;
 167         unsigned ret_start = d->returns.start,
 168                         ret_count = d->returns.count;
 169
 170         for (wkr = 0; wkr < d->num_workers; wkr++) {
 171                 uintptr_t oldbuf = 0;
 172                 /* Sync with worker. Acquire bufptr64. */
 173                 const int64_t data = __atomic_load_n(&(d->bufs[wkr].bufptr64),
 174                                                         __ATOMIC_ACQUIRE);
 175
 176                 if (data & RTE_DISTRIB_GET_BUF) {
 177                         flushed++;
 178                         if (d->backlog[wkr].count)
 179                                 /* Sync with worker. Release bufptr64. */
 180                                 __atomic_store_n(&(d->bufs[wkr].bufptr64),
 181                                         backlog_pop(&d->backlog[wkr]),
 182                                         __ATOMIC_RELEASE);
 183                         else {
 184                                 /* Sync with worker on GET_BUF flag. */
 185                                 __atomic_store_n(&(d->bufs[wkr].bufptr64),
 186                                         RTE_DISTRIB_GET_BUF,
 187                                         __ATOMIC_RELEASE);
 188                                 d->in_flight_tags[wkr] = 0;
 189                                 d->in_flight_bitmask &= ~(1UL << wkr);
 190                         }
 191                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 192                 } else if (data & RTE_DISTRIB_RETURN_BUF) {
 193                         handle_worker_shutdown(d, wkr);
 194                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 195                 }
 196
 197                 store_return(oldbuf, d, &ret_start, &ret_count);
 198         }
 199
 200         d->returns.start = ret_start;
 201         d->returns.count = ret_count;
 202
 203         return flushed;
 204 }
 205
 206 /* process a set of packets to distribute them to workers */
 207 int
 208 rte_distributor_process_v20(struct rte_distributor_v20 *d,
 209                 struct rte_mbuf **mbufs, unsigned num_mbufs)
 210 {
 211         unsigned next_idx = 0;
 212         unsigned wkr = 0;
 213         struct rte_mbuf *next_mb = NULL;
 214         int64_t next_value = 0;
 215         uint32_t new_tag = 0;
 216         unsigned ret_start = d->returns.start,
 217                         ret_count = d->returns.count;
 218
 219         if (unlikely(num_mbufs == 0))
 220                 return process_returns(d);
 221
 222         while (next_idx < num_mbufs || next_mb != NULL) {
 223                 uintptr_t oldbuf = 0;
 224                 /* Sync with worker. Acquire bufptr64. */
 225                 int64_t data = __atomic_load_n(&(d->bufs[wkr].bufptr64),
 226                                                 __ATOMIC_ACQUIRE);
 227
 228                 if (!next_mb) {
 229                         next_mb = mbufs[next_idx++];
 230                         next_value = (((int64_t)(uintptr_t)next_mb)
 231                                         << RTE_DISTRIB_FLAG_BITS);
 232                         /*
 233                          * User is advocated to set tag value for each
 234                          * mbuf before calling rte_distributor_process.
 235                          * User defined tags are used to identify flows,
 236                          * or sessions.
 237                          */
 238                         new_tag = next_mb->hash.usr;
 239
 240                         /*
 241                          * Note that if RTE_DISTRIB_MAX_WORKERS is larger than 64
 242                          * then the size of match has to be expanded.
 243                          */
 244                         uint64_t match = 0;
 245                         unsigned i;
 246                         /*
 247                          * to scan for a match use "xor" and "not" to get a 0/1
 248                          * value, then use shifting to merge to single "match"
 249                          * variable, where a one-bit indicates a match for the
 250                          * worker given by the bit-position
 251                          */
 252                         for (i = 0; i < d->num_workers; i++)
 253                                 match |= (!(d->in_flight_tags[i] ^ new_tag)
 254                                         << i);
 255
 256                         /* Only turned-on bits are considered as match */
 257                         match &= d->in_flight_bitmask;
 258
 259                         if (match) {
 260                                 next_mb = NULL;
 261                                 unsigned worker = __builtin_ctzl(match);
 262                                 if (add_to_backlog(&d->backlog[worker],
 263                                                 next_value) < 0)
 264                                         next_idx--;
 265                         }
 266                 }
 267
 268                 if ((data & RTE_DISTRIB_GET_BUF) &&
 269                                 (d->backlog[wkr].count || next_mb)) {
 270
 271                         if (d->backlog[wkr].count)
 272                                 /* Sync with worker. Release bufptr64. */
 273                                 __atomic_store_n(&(d->bufs[wkr].bufptr64),
 274                                                 backlog_pop(&d->backlog[wkr]),
 275                                                 __ATOMIC_RELEASE);
 276
 277                         else {
 278                                 /* Sync with worker. Release bufptr64.  */
 279                                 __atomic_store_n(&(d->bufs[wkr].bufptr64),
 280                                                 next_value,
 281                                                 __ATOMIC_RELEASE);
 282                                 d->in_flight_tags[wkr] = new_tag;
 283                                 d->in_flight_bitmask |= (1UL << wkr);
 284                                 next_mb = NULL;
 285                         }
 286                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 287                 } else if (data & RTE_DISTRIB_RETURN_BUF) {
 288                         handle_worker_shutdown(d, wkr);
 289                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 290                 }
 291
 292                 /* store returns in a circular buffer */
 293                 store_return(oldbuf, d, &ret_start, &ret_count);
 294
 295                 if (++wkr == d->num_workers)
 296                         wkr = 0;
 297         }
 298         /* to finish, check all workers for backlog and schedule work for them
 299          * if they are ready */
 300         for (wkr = 0; wkr < d->num_workers; wkr++)
 301                 if (d->backlog[wkr].count &&
 302                                 /* Sync with worker. Acquire bufptr64. */
 303                                 (__atomic_load_n(&(d->bufs[wkr].bufptr64),
 304                                 __ATOMIC_ACQUIRE) & RTE_DISTRIB_GET_BUF)) {
 305
 306                         int64_t oldbuf = d->bufs[wkr].bufptr64 >>
 307                                         RTE_DISTRIB_FLAG_BITS;
 308
 309                         store_return(oldbuf, d, &ret_start, &ret_count);
 310
 311                         /* Sync with worker. Release bufptr64. */
 312                         __atomic_store_n(&(d->bufs[wkr].bufptr64),
 313                                 backlog_pop(&d->backlog[wkr]),
 314                                 __ATOMIC_RELEASE);
 315                 }
 316
 317         d->returns.start = ret_start;
 318         d->returns.count = ret_count;
 319         return num_mbufs;
 320 }
 321 VERSION_SYMBOL(rte_distributor_process, _v20, 2.0);
 322
 323 /* return to the caller, packets returned from workers */
 324 int
 325 rte_distributor_returned_pkts_v20(struct rte_distributor_v20 *d,
 326                 struct rte_mbuf **mbufs, unsigned max_mbufs)
 327 {
 328         struct rte_distributor_returned_pkts *returns = &d->returns;
 329         unsigned retval = (max_mbufs < returns->count) ?
 330                         max_mbufs : returns->count;
 331         unsigned i;
 332
 333         for (i = 0; i < retval; i++) {
 334                 unsigned idx = (returns->start + i) & RTE_DISTRIB_RETURNS_MASK;
 335                 mbufs[i] = returns->mbufs[idx];
 336         }
 337         returns->start += i;
 338         returns->count -= i;
 339
 340         return retval;
 341 }
 342 VERSION_SYMBOL(rte_distributor_returned_pkts, _v20, 2.0);
 343
 344 /* return the number of packets in-flight in a distributor, i.e. packets
 345  * being worked on or queued up in a backlog.
 346  */
 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         strlcpy(d->name, name, sizeof(d->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_mcfg_tailq_write_lock();
 422         TAILQ_INSERT_TAIL(distributor_list, d, next);
 423         rte_mcfg_tailq_write_unlock();
 424
 425         return d;
 426 }
 427 VERSION_SYMBOL(rte_distributor_create, _v20, 2.0);