lib/librte_distributor/rte_distributor_single.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_string_fns.h>
  13 #include <rte_eal_memconfig.h>
  14 #include <rte_pause.h>
  15 #include <rte_tailq.h>
  16
  17 #include "rte_distributor_single.h"
  18 #include "distributor_private.h"
  19
  20 TAILQ_HEAD(rte_distributor_list, rte_distributor_single);
  21
  22 static struct rte_tailq_elem rte_distributor_tailq = {
  23         .name = "RTE_DISTRIBUTOR",
  24 };
  25 EAL_REGISTER_TAILQ(rte_distributor_tailq)
  26
  27 /**** APIs called by workers ****/
  28
  29 void
  30 rte_distributor_request_pkt_single(struct rte_distributor_single *d,
  31                 unsigned worker_id, struct rte_mbuf *oldpkt)
  32 {
  33         union rte_distributor_buffer_single *buf = &d->bufs[worker_id];
  34         int64_t req = (((int64_t)(uintptr_t)oldpkt) << RTE_DISTRIB_FLAG_BITS)
  35                         | RTE_DISTRIB_GET_BUF;
  36         while (unlikely(__atomic_load_n(&buf->bufptr64, __ATOMIC_RELAXED)
  37                         & RTE_DISTRIB_FLAGS_MASK))
  38                 rte_pause();
  39
  40         /* Sync with distributor on GET_BUF flag. */
  41         __atomic_store_n(&(buf->bufptr64), req, __ATOMIC_RELEASE);
  42 }
  43
  44 struct rte_mbuf *
  45 rte_distributor_poll_pkt_single(struct rte_distributor_single *d,
  46                 unsigned worker_id)
  47 {
  48         union rte_distributor_buffer_single *buf = &d->bufs[worker_id];
  49         /* Sync with distributor. Acquire bufptr64. */
  50         if (__atomic_load_n(&buf->bufptr64, __ATOMIC_ACQUIRE)
  51                 & RTE_DISTRIB_GET_BUF)
  52                 return NULL;
  53
  54         /* since bufptr64 is signed, this should be an arithmetic shift */
  55         int64_t ret = buf->bufptr64 >> RTE_DISTRIB_FLAG_BITS;
  56         return (struct rte_mbuf *)((uintptr_t)ret);
  57 }
  58
  59 struct rte_mbuf *
  60 rte_distributor_get_pkt_single(struct rte_distributor_single *d,
  61                 unsigned worker_id, struct rte_mbuf *oldpkt)
  62 {
  63         struct rte_mbuf *ret;
  64         rte_distributor_request_pkt_single(d, worker_id, oldpkt);
  65         while ((ret = rte_distributor_poll_pkt_single(d, worker_id)) == NULL)
  66                 rte_pause();
  67         return ret;
  68 }
  69
  70 int
  71 rte_distributor_return_pkt_single(struct rte_distributor_single *d,
  72                 unsigned worker_id, struct rte_mbuf *oldpkt)
  73 {
  74         union rte_distributor_buffer_single *buf = &d->bufs[worker_id];
  75         uint64_t req = (((int64_t)(uintptr_t)oldpkt) << RTE_DISTRIB_FLAG_BITS)
  76                         | RTE_DISTRIB_RETURN_BUF;
  77         /* Sync with distributor on RETURN_BUF flag. */
  78         __atomic_store_n(&(buf->bufptr64), req, __ATOMIC_RELEASE);
  79         return 0;
  80 }
  81
  82 /**** APIs called on distributor core ***/
  83
  84 /* as name suggests, adds a packet to the backlog for a particular worker */
  85 static int
  86 add_to_backlog(struct rte_distributor_backlog *bl, int64_t item)
  87 {
  88         if (bl->count == RTE_DISTRIB_BACKLOG_SIZE)
  89                 return -1;
  90
  91         bl->pkts[(bl->start + bl->count++) & (RTE_DISTRIB_BACKLOG_MASK)]
  92                         = item;
  93         return 0;
  94 }
  95
  96 /* takes the next packet for a worker off the backlog */
  97 static int64_t
  98 backlog_pop(struct rte_distributor_backlog *bl)
  99 {
 100         bl->count--;
 101         return bl->pkts[bl->start++ & RTE_DISTRIB_BACKLOG_MASK];
 102 }
 103
 104 /* stores a packet returned from a worker inside the returns array */
 105 static inline void
 106 store_return(uintptr_t oldbuf, struct rte_distributor_single *d,
 107                 unsigned *ret_start, unsigned *ret_count)
 108 {
 109         /* store returns in a circular buffer - code is branch-free */
 110         d->returns.mbufs[(*ret_start + *ret_count) & RTE_DISTRIB_RETURNS_MASK]
 111                         = (void *)oldbuf;
 112         *ret_start += (*ret_count == RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
 113         *ret_count += (*ret_count != RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
 114 }
 115
 116 static inline void
 117 handle_worker_shutdown(struct rte_distributor_single *d, unsigned int wkr)
 118 {
 119         d->in_flight_tags[wkr] = 0;
 120         d->in_flight_bitmask &= ~(1UL << wkr);
 121         /* Sync with worker. Release bufptr64. */
 122         __atomic_store_n(&(d->bufs[wkr].bufptr64), 0, __ATOMIC_RELEASE);
 123         if (unlikely(d->backlog[wkr].count != 0)) {
 124                 /* On return of a packet, we need to move the
 125                  * queued packets for this core elsewhere.
 126                  * Easiest solution is to set things up for
 127                  * a recursive call. That will cause those
 128                  * packets to be queued up for the next free
 129                  * core, i.e. it will return as soon as a
 130                  * core becomes free to accept the first
 131                  * packet, as subsequent ones will be added to
 132                  * the backlog for that core.
 133                  */
 134                 struct rte_mbuf *pkts[RTE_DISTRIB_BACKLOG_SIZE];
 135                 unsigned i;
 136                 struct rte_distributor_backlog *bl = &d->backlog[wkr];
 137
 138                 for (i = 0; i < bl->count; i++) {
 139                         unsigned idx = (bl->start + i) &
 140                                         RTE_DISTRIB_BACKLOG_MASK;
 141                         pkts[i] = (void *)((uintptr_t)(bl->pkts[idx] >>
 142                                         RTE_DISTRIB_FLAG_BITS));
 143                 }
 144                 /* recursive call.
 145                  * Note that the tags were set before first level call
 146                  * to rte_distributor_process.
 147                  */
 148                 rte_distributor_process_single(d, pkts, i);
 149                 bl->count = bl->start = 0;
 150         }
 151 }
 152
 153 /* this function is called when process() fn is called without any new
 154  * packets. It goes through all the workers and clears any returned packets
 155  * to do a partial flush.
 156  */
 157 static int
 158 process_returns(struct rte_distributor_single *d)
 159 {
 160         unsigned wkr;
 161         unsigned flushed = 0;
 162         unsigned ret_start = d->returns.start,
 163                         ret_count = d->returns.count;
 164
 165         for (wkr = 0; wkr < d->num_workers; wkr++) {
 166                 uintptr_t oldbuf = 0;
 167                 /* Sync with worker. Acquire bufptr64. */
 168                 const int64_t data = __atomic_load_n(&(d->bufs[wkr].bufptr64),
 169                                                         __ATOMIC_ACQUIRE);
 170
 171                 if (data & RTE_DISTRIB_GET_BUF) {
 172                         flushed++;
 173                         if (d->backlog[wkr].count)
 174                                 /* Sync with worker. Release bufptr64. */
 175                                 __atomic_store_n(&(d->bufs[wkr].bufptr64),
 176                                         backlog_pop(&d->backlog[wkr]),
 177                                         __ATOMIC_RELEASE);
 178                         else {
 179                                 /* Sync with worker on GET_BUF flag. */
 180                                 __atomic_store_n(&(d->bufs[wkr].bufptr64),
 181                                         RTE_DISTRIB_GET_BUF,
 182                                         __ATOMIC_RELEASE);
 183                                 d->in_flight_tags[wkr] = 0;
 184                                 d->in_flight_bitmask &= ~(1UL << wkr);
 185                         }
 186                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 187                 } else if (data & RTE_DISTRIB_RETURN_BUF) {
 188                         handle_worker_shutdown(d, wkr);
 189                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 190                 }
 191
 192                 store_return(oldbuf, d, &ret_start, &ret_count);
 193         }
 194
 195         d->returns.start = ret_start;
 196         d->returns.count = ret_count;
 197
 198         return flushed;
 199 }
 200
 201 /* process a set of packets to distribute them to workers */
 202 int
 203 rte_distributor_process_single(struct rte_distributor_single *d,
 204                 struct rte_mbuf **mbufs, unsigned num_mbufs)
 205 {
 206         unsigned next_idx = 0;
 207         unsigned wkr = 0;
 208         struct rte_mbuf *next_mb = NULL;
 209         int64_t next_value = 0;
 210         uint32_t new_tag = 0;
 211         unsigned ret_start = d->returns.start,
 212                         ret_count = d->returns.count;
 213
 214         if (unlikely(num_mbufs == 0))
 215                 return process_returns(d);
 216
 217         while (next_idx < num_mbufs || next_mb != NULL) {
 218                 uintptr_t oldbuf = 0;
 219                 /* Sync with worker. Acquire bufptr64. */
 220                 int64_t data = __atomic_load_n(&(d->bufs[wkr].bufptr64),
 221                                                 __ATOMIC_ACQUIRE);
 222
 223                 if (!next_mb) {
 224                         next_mb = mbufs[next_idx++];
 225                         next_value = (((int64_t)(uintptr_t)next_mb)
 226                                         << RTE_DISTRIB_FLAG_BITS);
 227                         /*
 228                          * User is advocated to set tag value for each
 229                          * mbuf before calling rte_distributor_process.
 230                          * User defined tags are used to identify flows,
 231                          * or sessions.
 232                          */
 233                         new_tag = next_mb->hash.usr;
 234
 235                         /*
 236                          * Note that if RTE_DISTRIB_MAX_WORKERS is larger than 64
 237                          * then the size of match has to be expanded.
 238                          */
 239                         uint64_t match = 0;
 240                         unsigned i;
 241                         /*
 242                          * to scan for a match use "xor" and "not" to get a 0/1
 243                          * value, then use shifting to merge to single "match"
 244                          * variable, where a one-bit indicates a match for the
 245                          * worker given by the bit-position
 246                          */
 247                         for (i = 0; i < d->num_workers; i++)
 248                                 match |= (!(d->in_flight_tags[i] ^ new_tag)
 249                                         << i);
 250
 251                         /* Only turned-on bits are considered as match */
 252                         match &= d->in_flight_bitmask;
 253
 254                         if (match) {
 255                                 next_mb = NULL;
 256                                 unsigned worker = __builtin_ctzl(match);
 257                                 if (add_to_backlog(&d->backlog[worker],
 258                                                 next_value) < 0)
 259                                         next_idx--;
 260                         }
 261                 }
 262
 263                 if ((data & RTE_DISTRIB_GET_BUF) &&
 264                                 (d->backlog[wkr].count || next_mb)) {
 265
 266                         if (d->backlog[wkr].count)
 267                                 /* Sync with worker. Release bufptr64. */
 268                                 __atomic_store_n(&(d->bufs[wkr].bufptr64),
 269                                                 backlog_pop(&d->backlog[wkr]),
 270                                                 __ATOMIC_RELEASE);
 271
 272                         else {
 273                                 /* Sync with worker. Release bufptr64.  */
 274                                 __atomic_store_n(&(d->bufs[wkr].bufptr64),
 275                                                 next_value,
 276                                                 __ATOMIC_RELEASE);
 277                                 d->in_flight_tags[wkr] = new_tag;
 278                                 d->in_flight_bitmask |= (1UL << wkr);
 279                                 next_mb = NULL;
 280                         }
 281                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 282                 } else if (data & RTE_DISTRIB_RETURN_BUF) {
 283                         handle_worker_shutdown(d, wkr);
 284                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 285                 }
 286
 287                 /* store returns in a circular buffer */
 288                 store_return(oldbuf, d, &ret_start, &ret_count);
 289
 290                 if (++wkr == d->num_workers)
 291                         wkr = 0;
 292         }
 293         /* to finish, check all workers for backlog and schedule work for them
 294          * if they are ready */
 295         for (wkr = 0; wkr < d->num_workers; wkr++)
 296                 if (d->backlog[wkr].count &&
 297                                 /* Sync with worker. Acquire bufptr64. */
 298                                 (__atomic_load_n(&(d->bufs[wkr].bufptr64),
 299                                 __ATOMIC_ACQUIRE) & RTE_DISTRIB_GET_BUF)) {
 300
 301                         int64_t oldbuf = d->bufs[wkr].bufptr64 >>
 302                                         RTE_DISTRIB_FLAG_BITS;
 303
 304                         store_return(oldbuf, d, &ret_start, &ret_count);
 305
 306                         /* Sync with worker. Release bufptr64. */
 307                         __atomic_store_n(&(d->bufs[wkr].bufptr64),
 308                                 backlog_pop(&d->backlog[wkr]),
 309                                 __ATOMIC_RELEASE);
 310                 }
 311
 312         d->returns.start = ret_start;
 313         d->returns.count = ret_count;
 314         return num_mbufs;
 315 }
 316
 317 /* return to the caller, packets returned from workers */
 318 int
 319 rte_distributor_returned_pkts_single(struct rte_distributor_single *d,
 320                 struct rte_mbuf **mbufs, unsigned max_mbufs)
 321 {
 322         struct rte_distributor_returned_pkts *returns = &d->returns;
 323         unsigned retval = (max_mbufs < returns->count) ?
 324                         max_mbufs : returns->count;
 325         unsigned i;
 326
 327         for (i = 0; i < retval; i++) {
 328                 unsigned idx = (returns->start + i) & RTE_DISTRIB_RETURNS_MASK;
 329                 mbufs[i] = returns->mbufs[idx];
 330         }
 331         returns->start += i;
 332         returns->count -= i;
 333
 334         return retval;
 335 }
 336
 337 /* return the number of packets in-flight in a distributor, i.e. packets
 338  * being worked on or queued up in a backlog.
 339  */
 340 static inline unsigned
 341 total_outstanding(const struct rte_distributor_single *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_single(struct rte_distributor_single *d)
 357 {
 358         const unsigned flushed = total_outstanding(d);
 359
 360         while (total_outstanding(d) > 0)
 361                 rte_distributor_process_single(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_single(struct rte_distributor_single *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_single *
 378 rte_distributor_create_single(const char *name,
 379                 unsigned socket_id,
 380                 unsigned num_workers)
 381 {
 382         struct rte_distributor_single *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         strlcpy(d->name, name, sizeof(d->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_mcfg_tailq_write_lock();
 413         TAILQ_INSERT_TAIL(distributor_list, d, next);
 414         rte_mcfg_tailq_write_unlock();
 415
 416         return d;
 417 }