1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation
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>
17 #include "rte_distributor_single.h"
18 #include "distributor_private.h"
20 TAILQ_HEAD(rte_distributor_list, rte_distributor_single);
22 static struct rte_tailq_elem rte_distributor_tailq = {
23 .name = "RTE_DISTRIBUTOR",
25 EAL_REGISTER_TAILQ(rte_distributor_tailq)
27 /**** APIs called by workers ****/
30 rte_distributor_request_pkt_single(struct rte_distributor_single *d,
31 unsigned worker_id, struct rte_mbuf *oldpkt)
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))
40 /* Sync with distributor on GET_BUF flag. */
41 __atomic_store_n(&(buf->bufptr64), req, __ATOMIC_RELEASE);
45 rte_distributor_poll_pkt_single(struct rte_distributor_single *d,
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)
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);
60 rte_distributor_get_pkt_single(struct rte_distributor_single *d,
61 unsigned worker_id, struct rte_mbuf *oldpkt)
64 rte_distributor_request_pkt_single(d, worker_id, oldpkt);
65 while ((ret = rte_distributor_poll_pkt_single(d, worker_id)) == NULL)
71 rte_distributor_return_pkt_single(struct rte_distributor_single *d,
72 unsigned worker_id, struct rte_mbuf *oldpkt)
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);
82 /**** APIs called on distributor core ***/
84 /* as name suggests, adds a packet to the backlog for a particular worker */
86 add_to_backlog(struct rte_distributor_backlog *bl, int64_t item)
88 if (bl->count == RTE_DISTRIB_BACKLOG_SIZE)
91 bl->pkts[(bl->start + bl->count++) & (RTE_DISTRIB_BACKLOG_MASK)]
96 /* takes the next packet for a worker off the backlog */
98 backlog_pop(struct rte_distributor_backlog *bl)
101 return bl->pkts[bl->start++ & RTE_DISTRIB_BACKLOG_MASK];
104 /* stores a packet returned from a worker inside the returns array */
106 store_return(uintptr_t oldbuf, struct rte_distributor_single *d,
107 unsigned *ret_start, unsigned *ret_count)
109 /* store returns in a circular buffer - code is branch-free */
110 d->returns.mbufs[(*ret_start + *ret_count) & RTE_DISTRIB_RETURNS_MASK]
112 *ret_start += (*ret_count == RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
113 *ret_count += (*ret_count != RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
117 handle_worker_shutdown(struct rte_distributor_single *d, unsigned int wkr)
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.
134 struct rte_mbuf *pkts[RTE_DISTRIB_BACKLOG_SIZE];
136 struct rte_distributor_backlog *bl = &d->backlog[wkr];
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));
145 * Note that the tags were set before first level call
146 * to rte_distributor_process.
148 rte_distributor_process_single(d, pkts, i);
149 bl->count = bl->start = 0;
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.
158 process_returns(struct rte_distributor_single *d)
161 unsigned flushed = 0;
162 unsigned ret_start = d->returns.start,
163 ret_count = d->returns.count;
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),
171 if (data & RTE_DISTRIB_GET_BUF) {
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]),
179 /* Sync with worker on GET_BUF flag. */
180 __atomic_store_n(&(d->bufs[wkr].bufptr64),
183 d->in_flight_tags[wkr] = 0;
184 d->in_flight_bitmask &= ~(1UL << wkr);
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;
192 store_return(oldbuf, d, &ret_start, &ret_count);
195 d->returns.start = ret_start;
196 d->returns.count = ret_count;
201 /* process a set of packets to distribute them to workers */
203 rte_distributor_process_single(struct rte_distributor_single *d,
204 struct rte_mbuf **mbufs, unsigned num_mbufs)
206 unsigned next_idx = 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;
214 if (unlikely(num_mbufs == 0))
215 return process_returns(d);
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),
224 next_mb = mbufs[next_idx++];
225 next_value = (((int64_t)(uintptr_t)next_mb)
226 << RTE_DISTRIB_FLAG_BITS);
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,
233 new_tag = next_mb->hash.usr;
236 * Note that if RTE_DISTRIB_MAX_WORKERS is larger than 64
237 * then the size of match has to be expanded.
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
247 for (i = 0; i < d->num_workers; i++)
248 match |= (!(d->in_flight_tags[i] ^ new_tag)
251 /* Only turned-on bits are considered as match */
252 match &= d->in_flight_bitmask;
256 unsigned worker = __builtin_ctzl(match);
257 if (add_to_backlog(&d->backlog[worker],
263 if ((data & RTE_DISTRIB_GET_BUF) &&
264 (d->backlog[wkr].count || next_mb)) {
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]),
273 /* Sync with worker. Release bufptr64. */
274 __atomic_store_n(&(d->bufs[wkr].bufptr64),
277 d->in_flight_tags[wkr] = new_tag;
278 d->in_flight_bitmask |= (1UL << wkr);
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;
287 /* store returns in a circular buffer */
288 store_return(oldbuf, d, &ret_start, &ret_count);
290 if (++wkr == d->num_workers)
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)) {
301 int64_t oldbuf = d->bufs[wkr].bufptr64 >>
302 RTE_DISTRIB_FLAG_BITS;
304 store_return(oldbuf, d, &ret_start, &ret_count);
306 /* Sync with worker. Release bufptr64. */
307 __atomic_store_n(&(d->bufs[wkr].bufptr64),
308 backlog_pop(&d->backlog[wkr]),
312 d->returns.start = ret_start;
313 d->returns.count = ret_count;
317 /* return to the caller, packets returned from workers */
319 rte_distributor_returned_pkts_single(struct rte_distributor_single *d,
320 struct rte_mbuf **mbufs, unsigned max_mbufs)
322 struct rte_distributor_returned_pkts *returns = &d->returns;
323 unsigned retval = (max_mbufs < returns->count) ?
324 max_mbufs : returns->count;
327 for (i = 0; i < retval; i++) {
328 unsigned idx = (returns->start + i) & RTE_DISTRIB_RETURNS_MASK;
329 mbufs[i] = returns->mbufs[idx];
337 /* return the number of packets in-flight in a distributor, i.e. packets
338 * being worked on or queued up in a backlog.
340 static inline unsigned
341 total_outstanding(const struct rte_distributor_single *d)
343 unsigned wkr, total_outstanding;
345 total_outstanding = __builtin_popcountl(d->in_flight_bitmask);
347 for (wkr = 0; wkr < d->num_workers; wkr++)
348 total_outstanding += d->backlog[wkr].count;
350 return total_outstanding;
353 /* flush the distributor, so that there are no outstanding packets in flight or
356 rte_distributor_flush_single(struct rte_distributor_single *d)
358 const unsigned flushed = total_outstanding(d);
360 while (total_outstanding(d) > 0)
361 rte_distributor_process_single(d, NULL, 0);
366 /* clears the internal returns array in the distributor */
368 rte_distributor_clear_returns_single(struct rte_distributor_single *d)
370 d->returns.start = d->returns.count = 0;
372 memset(d->returns.mbufs, 0, sizeof(d->returns.mbufs));
376 /* creates a distributor instance */
377 struct rte_distributor_single *
378 rte_distributor_create_single(const char *name,
380 unsigned num_workers)
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;
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);
393 if (name == NULL || num_workers >= RTE_DISTRIB_MAX_WORKERS) {
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);
406 strlcpy(d->name, name, sizeof(d->name));
407 d->num_workers = num_workers;
409 distributor_list = RTE_TAILQ_CAST(rte_distributor_tailq.head,
410 rte_distributor_list);
412 rte_mcfg_tailq_write_lock();
413 TAILQ_INSERT_TAIL(distributor_list, d, next);
414 rte_mcfg_tailq_write_unlock();