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_function_versioning.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_single.h"
  19 #include "distributor_private.h"
  20
  21 TAILQ_HEAD(rte_distributor_list, rte_distributor_single);
  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_single(struct rte_distributor_single *d,
  32                 unsigned worker_id, struct rte_mbuf *oldpkt)
  33 {
  34         union rte_distributor_buffer_single *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
  45 struct rte_mbuf *
  46 rte_distributor_poll_pkt_single(struct rte_distributor_single *d,
  47                 unsigned worker_id)
  48 {
  49         union rte_distributor_buffer_single *buf = &d->bufs[worker_id];
  50         /* Sync with distributor. Acquire bufptr64. */
  51         if (__atomic_load_n(&buf->bufptr64, __ATOMIC_ACQUIRE)
  52                 & RTE_DISTRIB_GET_BUF)
  53                 return NULL;
  54
  55         /* since bufptr64 is signed, this should be an arithmetic shift */
  56         int64_t ret = buf->bufptr64 >> RTE_DISTRIB_FLAG_BITS;
  57         return (struct rte_mbuf *)((uintptr_t)ret);
  58 }
  59
  60 struct rte_mbuf *
  61 rte_distributor_get_pkt_single(struct rte_distributor_single *d,
  62                 unsigned worker_id, struct rte_mbuf *oldpkt)
  63 {
  64         struct rte_mbuf *ret;
  65         rte_distributor_request_pkt_single(d, worker_id, oldpkt);
  66         while ((ret = rte_distributor_poll_pkt_single(d, worker_id)) == NULL)
  67                 rte_pause();
  68         return ret;
  69 }
  70
  71 int
  72 rte_distributor_return_pkt_single(struct rte_distributor_single *d,
  73                 unsigned worker_id, struct rte_mbuf *oldpkt)
  74 {
  75         union rte_distributor_buffer_single *buf = &d->bufs[worker_id];
  76         uint64_t req = (((int64_t)(uintptr_t)oldpkt) << RTE_DISTRIB_FLAG_BITS)
  77                         | RTE_DISTRIB_RETURN_BUF;
  78         /* Sync with distributor on RETURN_BUF flag. */
  79         __atomic_store_n(&(buf->bufptr64), req, __ATOMIC_RELEASE);
  80         return 0;
  81 }
  82
  83 /**** APIs called on distributor core ***/
  84
  85 /* as name suggests, adds a packet to the backlog for a particular worker */
  86 static int
  87 add_to_backlog(struct rte_distributor_backlog *bl, int64_t item)
  88 {
  89         if (bl->count == RTE_DISTRIB_BACKLOG_SIZE)
  90                 return -1;
  91
  92         bl->pkts[(bl->start + bl->count++) & (RTE_DISTRIB_BACKLOG_MASK)]
  93                         = item;
  94         return 0;
  95 }
  96
  97 /* takes the next packet for a worker off the backlog */
  98 static int64_t
  99 backlog_pop(struct rte_distributor_backlog *bl)
 100 {
 101         bl->count--;
 102         return bl->pkts[bl->start++ & RTE_DISTRIB_BACKLOG_MASK];
 103 }
 104
 105 /* stores a packet returned from a worker inside the returns array */
 106 static inline void
 107 store_return(uintptr_t oldbuf, struct rte_distributor_single *d,
 108                 unsigned *ret_start, unsigned *ret_count)
 109 {
 110         /* store returns in a circular buffer - code is branch-free */
 111         d->returns.mbufs[(*ret_start + *ret_count) & RTE_DISTRIB_RETURNS_MASK]
 112                         = (void *)oldbuf;
 113         *ret_start += (*ret_count == RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
 114         *ret_count += (*ret_count != RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
 115 }
 116
 117 static inline void
 118 handle_worker_shutdown(struct rte_distributor_single *d, unsigned int wkr)
 119 {
 120         d->in_flight_tags[wkr] = 0;
 121         d->in_flight_bitmask &= ~(1UL << wkr);
 122         /* Sync with worker. Release bufptr64. */
 123         __atomic_store_n(&(d->bufs[wkr].bufptr64), 0, __ATOMIC_RELEASE);
 124         if (unlikely(d->backlog[wkr].count != 0)) {
 125                 /* On return of a packet, we need to move the
 126                  * queued packets for this core elsewhere.
 127                  * Easiest solution is to set things up for
 128                  * a recursive call. That will cause those
 129                  * packets to be queued up for the next free
 130                  * core, i.e. it will return as soon as a
 131                  * core becomes free to accept the first
 132                  * packet, as subsequent ones will be added to
 133                  * the backlog for that core.
 134                  */
 135                 struct rte_mbuf *pkts[RTE_DISTRIB_BACKLOG_SIZE];
 136                 unsigned i;
 137                 struct rte_distributor_backlog *bl = &d->backlog[wkr];
 138
 139                 for (i = 0; i < bl->count; i++) {
 140                         unsigned idx = (bl->start + i) &
 141                                         RTE_DISTRIB_BACKLOG_MASK;
 142                         pkts[i] = (void *)((uintptr_t)(bl->pkts[idx] >>
 143                                         RTE_DISTRIB_FLAG_BITS));
 144                 }
 145                 /* recursive call.
 146                  * Note that the tags were set before first level call
 147                  * to rte_distributor_process.
 148                  */
 149                 rte_distributor_process_single(d, pkts, i);
 150                 bl->count = bl->start = 0;
 151         }
 152 }
 153
 154 /* this function is called when process() fn is called without any new
 155  * packets. It goes through all the workers and clears any returned packets
 156  * to do a partial flush.
 157  */
 158 static int
 159 process_returns(struct rte_distributor_single *d)
 160 {
 161         unsigned wkr;
 162         unsigned flushed = 0;
 163         unsigned ret_start = d->returns.start,
 164                         ret_count = d->returns.count;
 165
 166         for (wkr = 0; wkr < d->num_workers; wkr++) {
 167                 uintptr_t oldbuf = 0;
 168                 /* Sync with worker. Acquire bufptr64. */
 169                 const int64_t data = __atomic_load_n(&(d->bufs[wkr].bufptr64),
 170                                                         __ATOMIC_ACQUIRE);
 171
 172                 if (data & RTE_DISTRIB_GET_BUF) {
 173                         flushed++;
 174                         if (d->backlog[wkr].count)
 175                                 /* Sync with worker. Release bufptr64. */
 176                                 __atomic_store_n(&(d->bufs[wkr].bufptr64),
 177                                         backlog_pop(&d->backlog[wkr]),
 178                                         __ATOMIC_RELEASE);
 179                         else {
 180                                 /* Sync with worker on GET_BUF flag. */
 181                                 __atomic_store_n(&(d->bufs[wkr].bufptr64),
 182                                         RTE_DISTRIB_GET_BUF,
 183                                         __ATOMIC_RELEASE);
 184                                 d->in_flight_tags[wkr] = 0;
 185                                 d->in_flight_bitmask &= ~(1UL << wkr);
 186                         }
 187                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 188                 } else if (data & RTE_DISTRIB_RETURN_BUF) {
 189                         handle_worker_shutdown(d, wkr);
 190                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 191                 }
 192
 193                 store_return(oldbuf, d, &ret_start, &ret_count);
 194         }
 195
 196         d->returns.start = ret_start;
 197         d->returns.count = ret_count;
 198
 199         return flushed;
 200 }
 201
 202 /* process a set of packets to distribute them to workers */
 203 int
 204 rte_distributor_process_single(struct rte_distributor_single *d,
 205                 struct rte_mbuf **mbufs, unsigned num_mbufs)
 206 {
 207         unsigned next_idx = 0;
 208         unsigned wkr = 0;
 209         struct rte_mbuf *next_mb = NULL;
 210         int64_t next_value = 0;
 211         uint32_t new_tag = 0;
 212         unsigned ret_start = d->returns.start,
 213                         ret_count = d->returns.count;
 214
 215         if (unlikely(num_mbufs == 0))
 216                 return process_returns(d);
 217
 218         while (next_idx < num_mbufs || next_mb != NULL) {
 219                 uintptr_t oldbuf = 0;
 220                 /* Sync with worker. Acquire bufptr64. */
 221                 int64_t data = __atomic_load_n(&(d->bufs[wkr].bufptr64),
 222                                                 __ATOMIC_ACQUIRE);
 223
 224                 if (!next_mb) {
 225                         next_mb = mbufs[next_idx++];
 226                         next_value = (((int64_t)(uintptr_t)next_mb)
 227                                         << RTE_DISTRIB_FLAG_BITS);
 228                         /*
 229                          * User is advocated to set tag value for each
 230                          * mbuf before calling rte_distributor_process.
 231                          * User defined tags are used to identify flows,
 232                          * or sessions.
 233                          */
 234                         new_tag = next_mb->hash.usr;
 235
 236                         /*
 237                          * Note that if RTE_DISTRIB_MAX_WORKERS is larger than 64
 238                          * then the size of match has to be expanded.
 239                          */
 240                         uint64_t match = 0;
 241                         unsigned i;
 242                         /*
 243                          * to scan for a match use "xor" and "not" to get a 0/1
 244                          * value, then use shifting to merge to single "match"
 245                          * variable, where a one-bit indicates a match for the
 246                          * worker given by the bit-position
 247                          */
 248                         for (i = 0; i < d->num_workers; i++)
 249                                 match |= (!(d->in_flight_tags[i] ^ new_tag)
 250                                         << i);
 251
 252                         /* Only turned-on bits are considered as match */
 253                         match &= d->in_flight_bitmask;
 254
 255                         if (match) {
 256                                 next_mb = NULL;
 257                                 unsigned worker = __builtin_ctzl(match);
 258                                 if (add_to_backlog(&d->backlog[worker],
 259                                                 next_value) < 0)
 260                                         next_idx--;
 261                         }
 262                 }
 263
 264                 if ((data & RTE_DISTRIB_GET_BUF) &&
 265                                 (d->backlog[wkr].count || next_mb)) {
 266
 267                         if (d->backlog[wkr].count)
 268                                 /* Sync with worker. Release bufptr64. */
 269                                 __atomic_store_n(&(d->bufs[wkr].bufptr64),
 270                                                 backlog_pop(&d->backlog[wkr]),
 271                                                 __ATOMIC_RELEASE);
 272
 273                         else {
 274                                 /* Sync with worker. Release bufptr64.  */
 275                                 __atomic_store_n(&(d->bufs[wkr].bufptr64),
 276                                                 next_value,
 277                                                 __ATOMIC_RELEASE);
 278                                 d->in_flight_tags[wkr] = new_tag;
 279                                 d->in_flight_bitmask |= (1UL << wkr);
 280                                 next_mb = NULL;
 281                         }
 282                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 283                 } else if (data & RTE_DISTRIB_RETURN_BUF) {
 284                         handle_worker_shutdown(d, wkr);
 285                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 286                 }
 287
 288                 /* store returns in a circular buffer */
 289                 store_return(oldbuf, d, &ret_start, &ret_count);
 290
 291                 if (++wkr == d->num_workers)
 292                         wkr = 0;
 293         }
 294         /* to finish, check all workers for backlog and schedule work for them
 295          * if they are ready */
 296         for (wkr = 0; wkr < d->num_workers; wkr++)
 297                 if (d->backlog[wkr].count &&
 298                                 /* Sync with worker. Acquire bufptr64. */
 299                                 (__atomic_load_n(&(d->bufs[wkr].bufptr64),
 300                                 __ATOMIC_ACQUIRE) & RTE_DISTRIB_GET_BUF)) {
 301
 302                         int64_t oldbuf = d->bufs[wkr].bufptr64 >>
 303                                         RTE_DISTRIB_FLAG_BITS;
 304
 305                         store_return(oldbuf, d, &ret_start, &ret_count);
 306
 307                         /* Sync with worker. Release bufptr64. */
 308                         __atomic_store_n(&(d->bufs[wkr].bufptr64),
 309                                 backlog_pop(&d->backlog[wkr]),
 310                                 __ATOMIC_RELEASE);
 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_single(struct rte_distributor_single *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 worked on or queued up in a backlog.
 340  */
 341 static inline unsigned
 342 total_outstanding(const struct rte_distributor_single *d)
 343 {
 344         unsigned wkr, total_outstanding;
 345
 346         total_outstanding = __builtin_popcountl(d->in_flight_bitmask);
 347
 348         for (wkr = 0; wkr < d->num_workers; wkr++)
 349                 total_outstanding += d->backlog[wkr].count;
 350
 351         return total_outstanding;
 352 }
 353
 354 /* flush the distributor, so that there are no outstanding packets in flight or
 355  * queued up. */
 356 int
 357 rte_distributor_flush_single(struct rte_distributor_single *d)
 358 {
 359         const unsigned flushed = total_outstanding(d);
 360
 361         while (total_outstanding(d) > 0)
 362                 rte_distributor_process_single(d, NULL, 0);
 363
 364         return flushed;
 365 }
 366
 367 /* clears the internal returns array in the distributor */
 368 void
 369 rte_distributor_clear_returns_single(struct rte_distributor_single *d)
 370 {
 371         d->returns.start = d->returns.count = 0;
 372 #ifndef __OPTIMIZE__
 373         memset(d->returns.mbufs, 0, sizeof(d->returns.mbufs));
 374 #endif
 375 }
 376
 377 /* creates a distributor instance */
 378 struct rte_distributor_single *
 379 rte_distributor_create_single(const char *name,
 380                 unsigned socket_id,
 381                 unsigned num_workers)
 382 {
 383         struct rte_distributor_single *d;
 384         struct rte_distributor_list *distributor_list;
 385         char mz_name[RTE_MEMZONE_NAMESIZE];
 386         const struct rte_memzone *mz;
 387
 388         /* compilation-time checks */
 389         RTE_BUILD_BUG_ON((sizeof(*d) & RTE_CACHE_LINE_MASK) != 0);
 390         RTE_BUILD_BUG_ON((RTE_DISTRIB_MAX_WORKERS & 7) != 0);
 391         RTE_BUILD_BUG_ON(RTE_DISTRIB_MAX_WORKERS >
 392                                 sizeof(d->in_flight_bitmask) * CHAR_BIT);
 393
 394         if (name == NULL || num_workers >= RTE_DISTRIB_MAX_WORKERS) {
 395                 rte_errno = EINVAL;
 396                 return NULL;
 397         }
 398
 399         snprintf(mz_name, sizeof(mz_name), RTE_DISTRIB_PREFIX"%s", name);
 400         mz = rte_memzone_reserve(mz_name, sizeof(*d), socket_id, NO_FLAGS);
 401         if (mz == NULL) {
 402                 rte_errno = ENOMEM;
 403                 return NULL;
 404         }
 405
 406         d = mz->addr;
 407         strlcpy(d->name, name, sizeof(d->name));
 408         d->num_workers = num_workers;
 409
 410         distributor_list = RTE_TAILQ_CAST(rte_distributor_tailq.head,
 411                                           rte_distributor_list);
 412
 413         rte_mcfg_tailq_write_lock();
 414         TAILQ_INSERT_TAIL(distributor_list, d, next);
 415         rte_mcfg_tailq_write_unlock();
 416
 417         return d;
 418 }