ceph/src/seastar/dpdk/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
  17 #include "rte_distributor_v20.h"
  18 #include "rte_distributor_private.h"
  19
  20 TAILQ_HEAD(rte_distributor_list, rte_distributor_v20);
  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_v20(struct rte_distributor_v20 *d,
  31                 unsigned worker_id, struct rte_mbuf *oldpkt)
  32 {
  33         union rte_distributor_buffer_v20 *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(buf->bufptr64 & RTE_DISTRIB_FLAGS_MASK))
  37                 rte_pause();
  38         buf->bufptr64 = req;
  39 }
  40 VERSION_SYMBOL(rte_distributor_request_pkt, _v20, 2.0);
  41
  42 struct rte_mbuf *
  43 rte_distributor_poll_pkt_v20(struct rte_distributor_v20 *d,
  44                 unsigned worker_id)
  45 {
  46         union rte_distributor_buffer_v20 *buf = &d->bufs[worker_id];
  47         if (buf->bufptr64 & RTE_DISTRIB_GET_BUF)
  48                 return NULL;
  49
  50         /* since bufptr64 is signed, this should be an arithmetic shift */
  51         int64_t ret = buf->bufptr64 >> RTE_DISTRIB_FLAG_BITS;
  52         return (struct rte_mbuf *)((uintptr_t)ret);
  53 }
  54 VERSION_SYMBOL(rte_distributor_poll_pkt, _v20, 2.0);
  55
  56 struct rte_mbuf *
  57 rte_distributor_get_pkt_v20(struct rte_distributor_v20 *d,
  58                 unsigned worker_id, struct rte_mbuf *oldpkt)
  59 {
  60         struct rte_mbuf *ret;
  61         rte_distributor_request_pkt_v20(d, worker_id, oldpkt);
  62         while ((ret = rte_distributor_poll_pkt_v20(d, worker_id)) == NULL)
  63                 rte_pause();
  64         return ret;
  65 }
  66 VERSION_SYMBOL(rte_distributor_get_pkt, _v20, 2.0);
  67
  68 int
  69 rte_distributor_return_pkt_v20(struct rte_distributor_v20 *d,
  70                 unsigned worker_id, struct rte_mbuf *oldpkt)
  71 {
  72         union rte_distributor_buffer_v20 *buf = &d->bufs[worker_id];
  73         uint64_t req = (((int64_t)(uintptr_t)oldpkt) << RTE_DISTRIB_FLAG_BITS)
  74                         | RTE_DISTRIB_RETURN_BUF;
  75         buf->bufptr64 = req;
  76         return 0;
  77 }
  78 VERSION_SYMBOL(rte_distributor_return_pkt, _v20, 2.0);
  79
  80 /**** APIs called on distributor core ***/
  81
  82 /* as name suggests, adds a packet to the backlog for a particular worker */
  83 static int
  84 add_to_backlog(struct rte_distributor_backlog *bl, int64_t item)
  85 {
  86         if (bl->count == RTE_DISTRIB_BACKLOG_SIZE)
  87                 return -1;
  88
  89         bl->pkts[(bl->start + bl->count++) & (RTE_DISTRIB_BACKLOG_MASK)]
  90                         = item;
  91         return 0;
  92 }
  93
  94 /* takes the next packet for a worker off the backlog */
  95 static int64_t
  96 backlog_pop(struct rte_distributor_backlog *bl)
  97 {
  98         bl->count--;
  99         return bl->pkts[bl->start++ & RTE_DISTRIB_BACKLOG_MASK];
 100 }
 101
 102 /* stores a packet returned from a worker inside the returns array */
 103 static inline void
 104 store_return(uintptr_t oldbuf, struct rte_distributor_v20 *d,
 105                 unsigned *ret_start, unsigned *ret_count)
 106 {
 107         /* store returns in a circular buffer - code is branch-free */
 108         d->returns.mbufs[(*ret_start + *ret_count) & RTE_DISTRIB_RETURNS_MASK]
 109                         = (void *)oldbuf;
 110         *ret_start += (*ret_count == RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
 111         *ret_count += (*ret_count != RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
 112 }
 113
 114 static inline void
 115 handle_worker_shutdown(struct rte_distributor_v20 *d, unsigned int wkr)
 116 {
 117         d->in_flight_tags[wkr] = 0;
 118         d->in_flight_bitmask &= ~(1UL << wkr);
 119         d->bufs[wkr].bufptr64 = 0;
 120         if (unlikely(d->backlog[wkr].count != 0)) {
 121                 /* On return of a packet, we need to move the
 122                  * queued packets for this core elsewhere.
 123                  * Easiest solution is to set things up for
 124                  * a recursive call. That will cause those
 125                  * packets to be queued up for the next free
 126                  * core, i.e. it will return as soon as a
 127                  * core becomes free to accept the first
 128                  * packet, as subsequent ones will be added to
 129                  * the backlog for that core.
 130                  */
 131                 struct rte_mbuf *pkts[RTE_DISTRIB_BACKLOG_SIZE];
 132                 unsigned i;
 133                 struct rte_distributor_backlog *bl = &d->backlog[wkr];
 134
 135                 for (i = 0; i < bl->count; i++) {
 136                         unsigned idx = (bl->start + i) &
 137                                         RTE_DISTRIB_BACKLOG_MASK;
 138                         pkts[i] = (void *)((uintptr_t)(bl->pkts[idx] >>
 139                                         RTE_DISTRIB_FLAG_BITS));
 140                 }
 141                 /* recursive call.
 142                  * Note that the tags were set before first level call
 143                  * to rte_distributor_process.
 144                  */
 145                 rte_distributor_process_v20(d, pkts, i);
 146                 bl->count = bl->start = 0;
 147         }
 148 }
 149
 150 /* this function is called when process() fn is called without any new
 151  * packets. It goes through all the workers and clears any returned packets
 152  * to do a partial flush.
 153  */
 154 static int
 155 process_returns(struct rte_distributor_v20 *d)
 156 {
 157         unsigned wkr;
 158         unsigned flushed = 0;
 159         unsigned ret_start = d->returns.start,
 160                         ret_count = d->returns.count;
 161
 162         for (wkr = 0; wkr < d->num_workers; wkr++) {
 163
 164                 const int64_t data = d->bufs[wkr].bufptr64;
 165                 uintptr_t oldbuf = 0;
 166
 167                 if (data & RTE_DISTRIB_GET_BUF) {
 168                         flushed++;
 169                         if (d->backlog[wkr].count)
 170                                 d->bufs[wkr].bufptr64 =
 171                                                 backlog_pop(&d->backlog[wkr]);
 172                         else {
 173                                 d->bufs[wkr].bufptr64 = RTE_DISTRIB_GET_BUF;
 174                                 d->in_flight_tags[wkr] = 0;
 175                                 d->in_flight_bitmask &= ~(1UL << wkr);
 176                         }
 177                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 178                 } else if (data & RTE_DISTRIB_RETURN_BUF) {
 179                         handle_worker_shutdown(d, wkr);
 180                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 181                 }
 182
 183                 store_return(oldbuf, d, &ret_start, &ret_count);
 184         }
 185
 186         d->returns.start = ret_start;
 187         d->returns.count = ret_count;
 188
 189         return flushed;
 190 }
 191
 192 /* process a set of packets to distribute them to workers */
 193 int
 194 rte_distributor_process_v20(struct rte_distributor_v20 *d,
 195                 struct rte_mbuf **mbufs, unsigned num_mbufs)
 196 {
 197         unsigned next_idx = 0;
 198         unsigned wkr = 0;
 199         struct rte_mbuf *next_mb = NULL;
 200         int64_t next_value = 0;
 201         uint32_t new_tag = 0;
 202         unsigned ret_start = d->returns.start,
 203                         ret_count = d->returns.count;
 204
 205         if (unlikely(num_mbufs == 0))
 206                 return process_returns(d);
 207
 208         while (next_idx < num_mbufs || next_mb != NULL) {
 209
 210                 int64_t data = d->bufs[wkr].bufptr64;
 211                 uintptr_t oldbuf = 0;
 212
 213                 if (!next_mb) {
 214                         next_mb = mbufs[next_idx++];
 215                         next_value = (((int64_t)(uintptr_t)next_mb)
 216                                         << RTE_DISTRIB_FLAG_BITS);
 217                         /*
 218                          * User is advocated to set tag value for each
 219                          * mbuf before calling rte_distributor_process.
 220                          * User defined tags are used to identify flows,
 221                          * or sessions.
 222                          */
 223                         new_tag = next_mb->hash.usr;
 224
 225                         /*
 226                          * Note that if RTE_DISTRIB_MAX_WORKERS is larger than 64
 227                          * then the size of match has to be expanded.
 228                          */
 229                         uint64_t match = 0;
 230                         unsigned i;
 231                         /*
 232                          * to scan for a match use "xor" and "not" to get a 0/1
 233                          * value, then use shifting to merge to single "match"
 234                          * variable, where a one-bit indicates a match for the
 235                          * worker given by the bit-position
 236                          */
 237                         for (i = 0; i < d->num_workers; i++)
 238                                 match |= (!(d->in_flight_tags[i] ^ new_tag)
 239                                         << i);
 240
 241                         /* Only turned-on bits are considered as match */
 242                         match &= d->in_flight_bitmask;
 243
 244                         if (match) {
 245                                 next_mb = NULL;
 246                                 unsigned worker = __builtin_ctzl(match);
 247                                 if (add_to_backlog(&d->backlog[worker],
 248                                                 next_value) < 0)
 249                                         next_idx--;
 250                         }
 251                 }
 252
 253                 if ((data & RTE_DISTRIB_GET_BUF) &&
 254                                 (d->backlog[wkr].count || next_mb)) {
 255
 256                         if (d->backlog[wkr].count)
 257                                 d->bufs[wkr].bufptr64 =
 258                                                 backlog_pop(&d->backlog[wkr]);
 259
 260                         else {
 261                                 d->bufs[wkr].bufptr64 = next_value;
 262                                 d->in_flight_tags[wkr] = new_tag;
 263                                 d->in_flight_bitmask |= (1UL << wkr);
 264                                 next_mb = NULL;
 265                         }
 266                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 267                 } else if (data & RTE_DISTRIB_RETURN_BUF) {
 268                         handle_worker_shutdown(d, wkr);
 269                         oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
 270                 }
 271
 272                 /* store returns in a circular buffer */
 273                 store_return(oldbuf, d, &ret_start, &ret_count);
 274
 275                 if (++wkr == d->num_workers)
 276                         wkr = 0;
 277         }
 278         /* to finish, check all workers for backlog and schedule work for them
 279          * if they are ready */
 280         for (wkr = 0; wkr < d->num_workers; wkr++)
 281                 if (d->backlog[wkr].count &&
 282                                 (d->bufs[wkr].bufptr64 & RTE_DISTRIB_GET_BUF)) {
 283
 284                         int64_t oldbuf = d->bufs[wkr].bufptr64 >>
 285                                         RTE_DISTRIB_FLAG_BITS;
 286                         store_return(oldbuf, d, &ret_start, &ret_count);
 287
 288                         d->bufs[wkr].bufptr64 = backlog_pop(&d->backlog[wkr]);
 289                 }
 290
 291         d->returns.start = ret_start;
 292         d->returns.count = ret_count;
 293         return num_mbufs;
 294 }
 295 VERSION_SYMBOL(rte_distributor_process, _v20, 2.0);
 296
 297 /* return to the caller, packets returned from workers */
 298 int
 299 rte_distributor_returned_pkts_v20(struct rte_distributor_v20 *d,
 300                 struct rte_mbuf **mbufs, unsigned max_mbufs)
 301 {
 302         struct rte_distributor_returned_pkts *returns = &d->returns;
 303         unsigned retval = (max_mbufs < returns->count) ?
 304                         max_mbufs : returns->count;
 305         unsigned i;
 306
 307         for (i = 0; i < retval; i++) {
 308                 unsigned idx = (returns->start + i) & RTE_DISTRIB_RETURNS_MASK;
 309                 mbufs[i] = returns->mbufs[idx];
 310         }
 311         returns->start += i;
 312         returns->count -= i;
 313
 314         return retval;
 315 }
 316 VERSION_SYMBOL(rte_distributor_returned_pkts, _v20, 2.0);
 317
 318 /* return the number of packets in-flight in a distributor, i.e. packets
 319  * being worked on or queued up in a backlog.
 320  */
 321 static inline unsigned
 322 total_outstanding(const struct rte_distributor_v20 *d)
 323 {
 324         unsigned wkr, total_outstanding;
 325
 326         total_outstanding = __builtin_popcountl(d->in_flight_bitmask);
 327
 328         for (wkr = 0; wkr < d->num_workers; wkr++)
 329                 total_outstanding += d->backlog[wkr].count;
 330
 331         return total_outstanding;
 332 }
 333
 334 /* flush the distributor, so that there are no outstanding packets in flight or
 335  * queued up. */
 336 int
 337 rte_distributor_flush_v20(struct rte_distributor_v20 *d)
 338 {
 339         const unsigned flushed = total_outstanding(d);
 340
 341         while (total_outstanding(d) > 0)
 342                 rte_distributor_process_v20(d, NULL, 0);
 343
 344         return flushed;
 345 }
 346 VERSION_SYMBOL(rte_distributor_flush, _v20, 2.0);
 347
 348 /* clears the internal returns array in the distributor */
 349 void
 350 rte_distributor_clear_returns_v20(struct rte_distributor_v20 *d)
 351 {
 352         d->returns.start = d->returns.count = 0;
 353 #ifndef __OPTIMIZE__
 354         memset(d->returns.mbufs, 0, sizeof(d->returns.mbufs));
 355 #endif
 356 }
 357 VERSION_SYMBOL(rte_distributor_clear_returns, _v20, 2.0);
 358
 359 /* creates a distributor instance */
 360 struct rte_distributor_v20 *
 361 rte_distributor_create_v20(const char *name,
 362                 unsigned socket_id,
 363                 unsigned num_workers)
 364 {
 365         struct rte_distributor_v20 *d;
 366         struct rte_distributor_list *distributor_list;
 367         char mz_name[RTE_MEMZONE_NAMESIZE];
 368         const struct rte_memzone *mz;
 369
 370         /* compilation-time checks */
 371         RTE_BUILD_BUG_ON((sizeof(*d) & RTE_CACHE_LINE_MASK) != 0);
 372         RTE_BUILD_BUG_ON((RTE_DISTRIB_MAX_WORKERS & 7) != 0);
 373         RTE_BUILD_BUG_ON(RTE_DISTRIB_MAX_WORKERS >
 374                                 sizeof(d->in_flight_bitmask) * CHAR_BIT);
 375
 376         if (name == NULL || num_workers >= RTE_DISTRIB_MAX_WORKERS) {
 377                 rte_errno = EINVAL;
 378                 return NULL;
 379         }
 380
 381         snprintf(mz_name, sizeof(mz_name), RTE_DISTRIB_PREFIX"%s", name);
 382         mz = rte_memzone_reserve(mz_name, sizeof(*d), socket_id, NO_FLAGS);
 383         if (mz == NULL) {
 384                 rte_errno = ENOMEM;
 385                 return NULL;
 386         }
 387
 388         d = mz->addr;
 389         strlcpy(d->name, name, sizeof(d->name));
 390         d->num_workers = num_workers;
 391
 392         distributor_list = RTE_TAILQ_CAST(rte_distributor_tailq.head,
 393                                           rte_distributor_list);
 394
 395         rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
 396         TAILQ_INSERT_TAIL(distributor_list, d, next);
 397         rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
 398
 399         return d;
 400 }
 401 VERSION_SYMBOL(rte_distributor_create, _v20, 2.0);