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