4 * Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of Intel Corporation nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
35 #include <sys/queue.h>
38 #include <rte_memzone.h>
39 #include <rte_errno.h>
40 #include <rte_string_fns.h>
41 #include <rte_tailq.h>
42 #include <rte_eal_memconfig.h>
43 #include "rte_distributor.h"
46 #define RTE_DISTRIB_PREFIX "DT_"
48 /* we will use the bottom four bits of pointer for flags, shifting out
49 * the top four bits to make room (since a 64-bit pointer actually only uses
50 * 48 bits). An arithmetic-right-shift will then appropriately restore the
51 * original pointer value with proper sign extension into the top bits. */
52 #define RTE_DISTRIB_FLAG_BITS 4
53 #define RTE_DISTRIB_FLAGS_MASK (0x0F)
54 #define RTE_DISTRIB_NO_BUF 0 /**< empty flags: no buffer requested */
55 #define RTE_DISTRIB_GET_BUF (1) /**< worker requests a buffer, returns old */
56 #define RTE_DISTRIB_RETURN_BUF (2) /**< worker returns a buffer, no request */
58 #define RTE_DISTRIB_BACKLOG_SIZE 8
59 #define RTE_DISTRIB_BACKLOG_MASK (RTE_DISTRIB_BACKLOG_SIZE - 1)
61 #define RTE_DISTRIB_MAX_RETURNS 128
62 #define RTE_DISTRIB_RETURNS_MASK (RTE_DISTRIB_MAX_RETURNS - 1)
65 * Buffer structure used to pass the pointer data between cores. This is cache
66 * line aligned, but to improve performance and prevent adjacent cache-line
67 * prefetches of buffers for other workers, e.g. when worker 1's buffer is on
68 * the next cache line to worker 0, we pad this out to three cache lines.
69 * Only 64-bits of the memory is actually used though.
71 union rte_distributor_buffer {
72 volatile int64_t bufptr64;
73 char pad[CACHE_LINE_SIZE*3];
74 } __rte_cache_aligned;
76 struct rte_distributor_backlog {
79 int64_t pkts[RTE_DISTRIB_BACKLOG_SIZE];
82 struct rte_distributor_returned_pkts {
85 struct rte_mbuf *mbufs[RTE_DISTRIB_MAX_RETURNS];
88 struct rte_distributor {
89 TAILQ_ENTRY(rte_distributor) next; /**< Next in list. */
91 char name[RTE_DISTRIBUTOR_NAMESIZE]; /**< Name of the ring. */
92 unsigned num_workers; /**< Number of workers polling */
94 uint32_t in_flight_tags[RTE_MAX_LCORE];
95 /**< Tracks the tag being processed per core, 0 == no pkt */
96 struct rte_distributor_backlog backlog[RTE_MAX_LCORE];
98 union rte_distributor_buffer bufs[RTE_MAX_LCORE];
100 struct rte_distributor_returned_pkts returns;
103 TAILQ_HEAD(rte_distributor_list, rte_distributor);
105 /**** APIs called by workers ****/
108 rte_distributor_request_pkt(struct rte_distributor *d,
109 unsigned worker_id, struct rte_mbuf *oldpkt)
111 union rte_distributor_buffer *buf = &d->bufs[worker_id];
112 int64_t req = (((int64_t)(uintptr_t)oldpkt) << RTE_DISTRIB_FLAG_BITS)
113 | RTE_DISTRIB_GET_BUF;
114 while (unlikely(buf->bufptr64 & RTE_DISTRIB_FLAGS_MASK))
120 rte_distributor_poll_pkt(struct rte_distributor *d,
123 union rte_distributor_buffer *buf = &d->bufs[worker_id];
124 if (buf->bufptr64 & RTE_DISTRIB_GET_BUF)
127 /* since bufptr64 is signed, this should be an arithmetic shift */
128 int64_t ret = buf->bufptr64 >> RTE_DISTRIB_FLAG_BITS;
129 return (struct rte_mbuf *)((uintptr_t)ret);
133 rte_distributor_get_pkt(struct rte_distributor *d,
134 unsigned worker_id, struct rte_mbuf *oldpkt)
136 struct rte_mbuf *ret;
137 rte_distributor_request_pkt(d, worker_id, oldpkt);
138 while ((ret = rte_distributor_poll_pkt(d, worker_id)) == NULL)
144 rte_distributor_return_pkt(struct rte_distributor *d,
145 unsigned worker_id, struct rte_mbuf *oldpkt)
147 union rte_distributor_buffer *buf = &d->bufs[worker_id];
148 uint64_t req = (((int64_t)(uintptr_t)oldpkt) << RTE_DISTRIB_FLAG_BITS)
149 | RTE_DISTRIB_RETURN_BUF;
154 /**** APIs called on distributor core ***/
156 /* as name suggests, adds a packet to the backlog for a particular worker */
158 add_to_backlog(struct rte_distributor_backlog *bl, int64_t item)
160 if (bl->count == RTE_DISTRIB_BACKLOG_SIZE)
163 bl->pkts[(bl->start + bl->count++) & (RTE_DISTRIB_BACKLOG_MASK)]
168 /* takes the next packet for a worker off the backlog */
170 backlog_pop(struct rte_distributor_backlog *bl)
173 return bl->pkts[bl->start++ & RTE_DISTRIB_BACKLOG_MASK];
176 /* stores a packet returned from a worker inside the returns array */
178 store_return(uintptr_t oldbuf, struct rte_distributor *d,
179 unsigned *ret_start, unsigned *ret_count)
181 /* store returns in a circular buffer - code is branch-free */
182 d->returns.mbufs[(*ret_start + *ret_count) & RTE_DISTRIB_RETURNS_MASK]
184 *ret_start += (*ret_count == RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
185 *ret_count += (*ret_count != RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
189 handle_worker_shutdown(struct rte_distributor *d, unsigned wkr)
191 d->in_flight_tags[wkr] = 0;
192 d->bufs[wkr].bufptr64 = 0;
193 if (unlikely(d->backlog[wkr].count != 0)) {
194 /* On return of a packet, we need to move the
195 * queued packets for this core elsewhere.
196 * Easiest solution is to set things up for
197 * a recursive call. That will cause those
198 * packets to be queued up for the next free
199 * core, i.e. it will return as soon as a
200 * core becomes free to accept the first
201 * packet, as subsequent ones will be added to
202 * the backlog for that core.
204 struct rte_mbuf *pkts[RTE_DISTRIB_BACKLOG_SIZE];
206 struct rte_distributor_backlog *bl = &d->backlog[wkr];
208 for (i = 0; i < bl->count; i++) {
209 unsigned idx = (bl->start + i) &
210 RTE_DISTRIB_BACKLOG_MASK;
211 pkts[i] = (void *)((uintptr_t)(bl->pkts[idx] >>
212 RTE_DISTRIB_FLAG_BITS));
215 rte_distributor_process(d, pkts, i);
216 bl->count = bl->start = 0;
220 /* this function is called when process() fn is called without any new
221 * packets. It goes through all the workers and clears any returned packets
222 * to do a partial flush.
225 process_returns(struct rte_distributor *d)
228 unsigned flushed = 0;
229 unsigned ret_start = d->returns.start,
230 ret_count = d->returns.count;
232 for (wkr = 0; wkr < d->num_workers; wkr++) {
234 const int64_t data = d->bufs[wkr].bufptr64;
235 uintptr_t oldbuf = 0;
237 if (data & RTE_DISTRIB_GET_BUF) {
239 if (d->backlog[wkr].count)
240 d->bufs[wkr].bufptr64 =
241 backlog_pop(&d->backlog[wkr]);
243 d->bufs[wkr].bufptr64 = RTE_DISTRIB_GET_BUF;
244 d->in_flight_tags[wkr] = 0;
246 oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
247 } else if (data & RTE_DISTRIB_RETURN_BUF) {
248 handle_worker_shutdown(d, wkr);
249 oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
252 store_return(oldbuf, d, &ret_start, &ret_count);
255 d->returns.start = ret_start;
256 d->returns.count = ret_count;
261 /* process a set of packets to distribute them to workers */
263 rte_distributor_process(struct rte_distributor *d,
264 struct rte_mbuf **mbufs, unsigned num_mbufs)
266 unsigned next_idx = 0;
268 struct rte_mbuf *next_mb = NULL;
269 int64_t next_value = 0;
270 uint32_t new_tag = 0;
271 unsigned ret_start = d->returns.start,
272 ret_count = d->returns.count;
274 if (unlikely(num_mbufs == 0))
275 return process_returns(d);
277 while (next_idx < num_mbufs || next_mb != NULL) {
279 int64_t data = d->bufs[wkr].bufptr64;
280 uintptr_t oldbuf = 0;
283 next_mb = mbufs[next_idx++];
284 next_value = (((int64_t)(uintptr_t)next_mb)
285 << RTE_DISTRIB_FLAG_BITS);
287 * Set the low bit on the tag, so we can guarantee that
288 * we never store a tag value of zero. That means we can
289 * use the zero-value to indicate that no packet is
290 * being processed by a worker.
292 new_tag = (next_mb->hash.usr | 1);
297 * to scan for a match use "xor" and "not" to get a 0/1
298 * value, then use shifting to merge to single "match"
299 * variable, where a one-bit indicates a match for the
300 * worker given by the bit-position
302 for (i = 0; i < d->num_workers; i++)
303 match |= (!(d->in_flight_tags[i] ^ new_tag)
308 unsigned worker = __builtin_ctz(match);
309 if (add_to_backlog(&d->backlog[worker],
315 if ((data & RTE_DISTRIB_GET_BUF) &&
316 (d->backlog[wkr].count || next_mb)) {
318 if (d->backlog[wkr].count)
319 d->bufs[wkr].bufptr64 =
320 backlog_pop(&d->backlog[wkr]);
323 d->bufs[wkr].bufptr64 = next_value;
324 d->in_flight_tags[wkr] = new_tag;
327 oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
328 } else if (data & RTE_DISTRIB_RETURN_BUF) {
329 handle_worker_shutdown(d, wkr);
330 oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
333 /* store returns in a circular buffer */
334 store_return(oldbuf, d, &ret_start, &ret_count);
336 if (++wkr == d->num_workers)
339 /* to finish, check all workers for backlog and schedule work for them
340 * if they are ready */
341 for (wkr = 0; wkr < d->num_workers; wkr++)
342 if (d->backlog[wkr].count &&
343 (d->bufs[wkr].bufptr64 & RTE_DISTRIB_GET_BUF)) {
345 int64_t oldbuf = d->bufs[wkr].bufptr64 >>
346 RTE_DISTRIB_FLAG_BITS;
347 store_return(oldbuf, d, &ret_start, &ret_count);
349 d->bufs[wkr].bufptr64 = backlog_pop(&d->backlog[wkr]);
352 d->returns.start = ret_start;
353 d->returns.count = ret_count;
357 /* return to the caller, packets returned from workers */
359 rte_distributor_returned_pkts(struct rte_distributor *d,
360 struct rte_mbuf **mbufs, unsigned max_mbufs)
362 struct rte_distributor_returned_pkts *returns = &d->returns;
363 unsigned retval = (max_mbufs < returns->count) ?
364 max_mbufs : returns->count;
367 for (i = 0; i < retval; i++) {
368 unsigned idx = (returns->start + i) & RTE_DISTRIB_RETURNS_MASK;
369 mbufs[i] = returns->mbufs[idx];
377 /* return the number of packets in-flight in a distributor, i.e. packets
378 * being workered on or queued up in a backlog. */
379 static inline unsigned
380 total_outstanding(const struct rte_distributor *d)
382 unsigned wkr, total_outstanding = 0;
384 for (wkr = 0; wkr < d->num_workers; wkr++)
385 total_outstanding += d->backlog[wkr].count +
386 !!(d->in_flight_tags[wkr]);
387 return total_outstanding;
390 /* flush the distributor, so that there are no outstanding packets in flight or
393 rte_distributor_flush(struct rte_distributor *d)
395 const unsigned flushed = total_outstanding(d);
397 while (total_outstanding(d) > 0)
398 rte_distributor_process(d, NULL, 0);
403 /* clears the internal returns array in the distributor */
405 rte_distributor_clear_returns(struct rte_distributor *d)
407 d->returns.start = d->returns.count = 0;
409 memset(d->returns.mbufs, 0, sizeof(d->returns.mbufs));
413 /* creates a distributor instance */
414 struct rte_distributor *
415 rte_distributor_create(const char *name,
417 unsigned num_workers)
419 struct rte_distributor *d;
420 struct rte_distributor_list *distributor_list;
421 char mz_name[RTE_MEMZONE_NAMESIZE];
422 const struct rte_memzone *mz;
424 /* compilation-time checks */
425 RTE_BUILD_BUG_ON((sizeof(*d) & CACHE_LINE_MASK) != 0);
426 RTE_BUILD_BUG_ON((RTE_MAX_LCORE & 7) != 0);
428 if (name == NULL || num_workers >= RTE_MAX_LCORE) {
433 /* check that we have an initialised tail queue */
434 distributor_list = RTE_TAILQ_LOOKUP_BY_IDX(RTE_TAILQ_DISTRIBUTOR,
435 rte_distributor_list);
436 if (distributor_list == NULL) {
437 rte_errno = E_RTE_NO_TAILQ;
441 snprintf(mz_name, sizeof(mz_name), RTE_DISTRIB_PREFIX"%s", name);
442 mz = rte_memzone_reserve(mz_name, sizeof(*d), socket_id, NO_FLAGS);
449 snprintf(d->name, sizeof(d->name), "%s", name);
450 d->num_workers = num_workers;
452 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
453 TAILQ_INSERT_TAIL(distributor_list, d, next);
454 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);