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_memory.h>
39 #include <rte_memzone.h>
40 #include <rte_errno.h>
41 #include <rte_string_fns.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 * Maximum number of workers allowed.
66 * Be aware of increasing the limit, becaus it is limited by how we track
67 * in-flight tags. See @in_flight_bitmask and @rte_distributor_process
69 #define RTE_DISTRIB_MAX_WORKERS 64
72 * Buffer structure used to pass the pointer data between cores. This is cache
73 * line aligned, but to improve performance and prevent adjacent cache-line
74 * prefetches of buffers for other workers, e.g. when worker 1's buffer is on
75 * the next cache line to worker 0, we pad this out to three cache lines.
76 * Only 64-bits of the memory is actually used though.
78 union rte_distributor_buffer {
79 volatile int64_t bufptr64;
80 char pad[RTE_CACHE_LINE_SIZE*3];
81 } __rte_cache_aligned;
83 struct rte_distributor_backlog {
86 int64_t pkts[RTE_DISTRIB_BACKLOG_SIZE];
89 struct rte_distributor_returned_pkts {
92 struct rte_mbuf *mbufs[RTE_DISTRIB_MAX_RETURNS];
95 struct rte_distributor {
96 TAILQ_ENTRY(rte_distributor) next; /**< Next in list. */
98 char name[RTE_DISTRIBUTOR_NAMESIZE]; /**< Name of the ring. */
99 unsigned num_workers; /**< Number of workers polling */
101 uint32_t in_flight_tags[RTE_DISTRIB_MAX_WORKERS];
102 /**< Tracks the tag being processed per core */
103 uint64_t in_flight_bitmask;
104 /**< on/off bits for in-flight tags.
105 * Note that if RTE_DISTRIB_MAX_WORKERS is larger than 64 then
106 * the bitmask has to expand.
109 struct rte_distributor_backlog backlog[RTE_DISTRIB_MAX_WORKERS];
111 union rte_distributor_buffer bufs[RTE_DISTRIB_MAX_WORKERS];
113 struct rte_distributor_returned_pkts returns;
116 TAILQ_HEAD(rte_distributor_list, rte_distributor);
118 static struct rte_tailq_elem rte_distributor_tailq = {
119 .name = "RTE_DISTRIBUTOR",
121 EAL_REGISTER_TAILQ(rte_distributor_tailq)
123 /**** APIs called by workers ****/
126 rte_distributor_request_pkt(struct rte_distributor *d,
127 unsigned worker_id, struct rte_mbuf *oldpkt)
129 union rte_distributor_buffer *buf = &d->bufs[worker_id];
130 int64_t req = (((int64_t)(uintptr_t)oldpkt) << RTE_DISTRIB_FLAG_BITS)
131 | RTE_DISTRIB_GET_BUF;
132 while (unlikely(buf->bufptr64 & RTE_DISTRIB_FLAGS_MASK))
138 rte_distributor_poll_pkt(struct rte_distributor *d,
141 union rte_distributor_buffer *buf = &d->bufs[worker_id];
142 if (buf->bufptr64 & RTE_DISTRIB_GET_BUF)
145 /* since bufptr64 is signed, this should be an arithmetic shift */
146 int64_t ret = buf->bufptr64 >> RTE_DISTRIB_FLAG_BITS;
147 return (struct rte_mbuf *)((uintptr_t)ret);
151 rte_distributor_get_pkt(struct rte_distributor *d,
152 unsigned worker_id, struct rte_mbuf *oldpkt)
154 struct rte_mbuf *ret;
155 rte_distributor_request_pkt(d, worker_id, oldpkt);
156 while ((ret = rte_distributor_poll_pkt(d, worker_id)) == NULL)
162 rte_distributor_return_pkt(struct rte_distributor *d,
163 unsigned worker_id, struct rte_mbuf *oldpkt)
165 union rte_distributor_buffer *buf = &d->bufs[worker_id];
166 uint64_t req = (((int64_t)(uintptr_t)oldpkt) << RTE_DISTRIB_FLAG_BITS)
167 | RTE_DISTRIB_RETURN_BUF;
172 /**** APIs called on distributor core ***/
174 /* as name suggests, adds a packet to the backlog for a particular worker */
176 add_to_backlog(struct rte_distributor_backlog *bl, int64_t item)
178 if (bl->count == RTE_DISTRIB_BACKLOG_SIZE)
181 bl->pkts[(bl->start + bl->count++) & (RTE_DISTRIB_BACKLOG_MASK)]
186 /* takes the next packet for a worker off the backlog */
188 backlog_pop(struct rte_distributor_backlog *bl)
191 return bl->pkts[bl->start++ & RTE_DISTRIB_BACKLOG_MASK];
194 /* stores a packet returned from a worker inside the returns array */
196 store_return(uintptr_t oldbuf, struct rte_distributor *d,
197 unsigned *ret_start, unsigned *ret_count)
199 /* store returns in a circular buffer - code is branch-free */
200 d->returns.mbufs[(*ret_start + *ret_count) & RTE_DISTRIB_RETURNS_MASK]
202 *ret_start += (*ret_count == RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
203 *ret_count += (*ret_count != RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
207 handle_worker_shutdown(struct rte_distributor *d, unsigned wkr)
209 d->in_flight_tags[wkr] = 0;
210 d->in_flight_bitmask &= ~(1UL << wkr);
211 d->bufs[wkr].bufptr64 = 0;
212 if (unlikely(d->backlog[wkr].count != 0)) {
213 /* On return of a packet, we need to move the
214 * queued packets for this core elsewhere.
215 * Easiest solution is to set things up for
216 * a recursive call. That will cause those
217 * packets to be queued up for the next free
218 * core, i.e. it will return as soon as a
219 * core becomes free to accept the first
220 * packet, as subsequent ones will be added to
221 * the backlog for that core.
223 struct rte_mbuf *pkts[RTE_DISTRIB_BACKLOG_SIZE];
225 struct rte_distributor_backlog *bl = &d->backlog[wkr];
227 for (i = 0; i < bl->count; i++) {
228 unsigned idx = (bl->start + i) &
229 RTE_DISTRIB_BACKLOG_MASK;
230 pkts[i] = (void *)((uintptr_t)(bl->pkts[idx] >>
231 RTE_DISTRIB_FLAG_BITS));
234 * Note that the tags were set before first level call
235 * to rte_distributor_process.
237 rte_distributor_process(d, pkts, i);
238 bl->count = bl->start = 0;
242 /* this function is called when process() fn is called without any new
243 * packets. It goes through all the workers and clears any returned packets
244 * to do a partial flush.
247 process_returns(struct rte_distributor *d)
250 unsigned flushed = 0;
251 unsigned ret_start = d->returns.start,
252 ret_count = d->returns.count;
254 for (wkr = 0; wkr < d->num_workers; wkr++) {
256 const int64_t data = d->bufs[wkr].bufptr64;
257 uintptr_t oldbuf = 0;
259 if (data & RTE_DISTRIB_GET_BUF) {
261 if (d->backlog[wkr].count)
262 d->bufs[wkr].bufptr64 =
263 backlog_pop(&d->backlog[wkr]);
265 d->bufs[wkr].bufptr64 = RTE_DISTRIB_GET_BUF;
266 d->in_flight_tags[wkr] = 0;
267 d->in_flight_bitmask &= ~(1UL << wkr);
269 oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
270 } else if (data & RTE_DISTRIB_RETURN_BUF) {
271 handle_worker_shutdown(d, wkr);
272 oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
275 store_return(oldbuf, d, &ret_start, &ret_count);
278 d->returns.start = ret_start;
279 d->returns.count = ret_count;
284 /* process a set of packets to distribute them to workers */
286 rte_distributor_process(struct rte_distributor *d,
287 struct rte_mbuf **mbufs, unsigned num_mbufs)
289 unsigned next_idx = 0;
291 struct rte_mbuf *next_mb = NULL;
292 int64_t next_value = 0;
293 uint32_t new_tag = 0;
294 unsigned ret_start = d->returns.start,
295 ret_count = d->returns.count;
297 if (unlikely(num_mbufs == 0))
298 return process_returns(d);
300 while (next_idx < num_mbufs || next_mb != NULL) {
302 int64_t data = d->bufs[wkr].bufptr64;
303 uintptr_t oldbuf = 0;
306 next_mb = mbufs[next_idx++];
307 next_value = (((int64_t)(uintptr_t)next_mb)
308 << RTE_DISTRIB_FLAG_BITS);
310 * User is advocated to set tag vaue for each
311 * mbuf before calling rte_distributor_process.
312 * User defined tags are used to identify flows,
315 new_tag = next_mb->hash.usr;
318 * Note that if RTE_DISTRIB_MAX_WORKERS is larger than 64
319 * then the size of match has to be expanded.
324 * to scan for a match use "xor" and "not" to get a 0/1
325 * value, then use shifting to merge to single "match"
326 * variable, where a one-bit indicates a match for the
327 * worker given by the bit-position
329 for (i = 0; i < d->num_workers; i++)
330 match |= (!(d->in_flight_tags[i] ^ new_tag)
333 /* Only turned-on bits are considered as match */
334 match &= d->in_flight_bitmask;
338 unsigned worker = __builtin_ctzl(match);
339 if (add_to_backlog(&d->backlog[worker],
345 if ((data & RTE_DISTRIB_GET_BUF) &&
346 (d->backlog[wkr].count || next_mb)) {
348 if (d->backlog[wkr].count)
349 d->bufs[wkr].bufptr64 =
350 backlog_pop(&d->backlog[wkr]);
353 d->bufs[wkr].bufptr64 = next_value;
354 d->in_flight_tags[wkr] = new_tag;
355 d->in_flight_bitmask |= (1UL << wkr);
358 oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
359 } else if (data & RTE_DISTRIB_RETURN_BUF) {
360 handle_worker_shutdown(d, wkr);
361 oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
364 /* store returns in a circular buffer */
365 store_return(oldbuf, d, &ret_start, &ret_count);
367 if (++wkr == d->num_workers)
370 /* to finish, check all workers for backlog and schedule work for them
371 * if they are ready */
372 for (wkr = 0; wkr < d->num_workers; wkr++)
373 if (d->backlog[wkr].count &&
374 (d->bufs[wkr].bufptr64 & RTE_DISTRIB_GET_BUF)) {
376 int64_t oldbuf = d->bufs[wkr].bufptr64 >>
377 RTE_DISTRIB_FLAG_BITS;
378 store_return(oldbuf, d, &ret_start, &ret_count);
380 d->bufs[wkr].bufptr64 = backlog_pop(&d->backlog[wkr]);
383 d->returns.start = ret_start;
384 d->returns.count = ret_count;
388 /* return to the caller, packets returned from workers */
390 rte_distributor_returned_pkts(struct rte_distributor *d,
391 struct rte_mbuf **mbufs, unsigned max_mbufs)
393 struct rte_distributor_returned_pkts *returns = &d->returns;
394 unsigned retval = (max_mbufs < returns->count) ?
395 max_mbufs : returns->count;
398 for (i = 0; i < retval; i++) {
399 unsigned idx = (returns->start + i) & RTE_DISTRIB_RETURNS_MASK;
400 mbufs[i] = returns->mbufs[idx];
408 /* return the number of packets in-flight in a distributor, i.e. packets
409 * being workered on or queued up in a backlog. */
410 static inline unsigned
411 total_outstanding(const struct rte_distributor *d)
413 unsigned wkr, total_outstanding;
415 total_outstanding = __builtin_popcountl(d->in_flight_bitmask);
417 for (wkr = 0; wkr < d->num_workers; wkr++)
418 total_outstanding += d->backlog[wkr].count;
420 return total_outstanding;
423 /* flush the distributor, so that there are no outstanding packets in flight or
426 rte_distributor_flush(struct rte_distributor *d)
428 const unsigned flushed = total_outstanding(d);
430 while (total_outstanding(d) > 0)
431 rte_distributor_process(d, NULL, 0);
436 /* clears the internal returns array in the distributor */
438 rte_distributor_clear_returns(struct rte_distributor *d)
440 d->returns.start = d->returns.count = 0;
442 memset(d->returns.mbufs, 0, sizeof(d->returns.mbufs));
446 /* creates a distributor instance */
447 struct rte_distributor *
448 rte_distributor_create(const char *name,
450 unsigned num_workers)
452 struct rte_distributor *d;
453 struct rte_distributor_list *distributor_list;
454 char mz_name[RTE_MEMZONE_NAMESIZE];
455 const struct rte_memzone *mz;
457 /* compilation-time checks */
458 RTE_BUILD_BUG_ON((sizeof(*d) & RTE_CACHE_LINE_MASK) != 0);
459 RTE_BUILD_BUG_ON((RTE_DISTRIB_MAX_WORKERS & 7) != 0);
460 RTE_BUILD_BUG_ON(RTE_DISTRIB_MAX_WORKERS >
461 sizeof(d->in_flight_bitmask) * CHAR_BIT);
463 if (name == NULL || num_workers >= RTE_DISTRIB_MAX_WORKERS) {
468 snprintf(mz_name, sizeof(mz_name), RTE_DISTRIB_PREFIX"%s", name);
469 mz = rte_memzone_reserve(mz_name, sizeof(*d), socket_id, NO_FLAGS);
476 snprintf(d->name, sizeof(d->name), "%s", name);
477 d->num_workers = num_workers;
479 distributor_list = RTE_TAILQ_CAST(rte_distributor_tailq.head,
480 rte_distributor_list);
482 rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
483 TAILQ_INSERT_TAIL(distributor_list, d, next);
484 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);