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_string_fns.h>
13 #include <rte_eal_memconfig.h>
14 #include <rte_pause.h>
15 #include <rte_tailq.h>
16 
17 #include "rte_distributor_single.h"
18 #include "distributor_private.h"
19 
20 TAILQ_HEAD(rte_distributor_list, rte_distributor_single);
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_single(struct rte_distributor_single *d,
31 		unsigned worker_id, struct rte_mbuf *oldpkt)
32 {
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 	RTE_WAIT_UNTIL_MASKED(&buf->bufptr64, RTE_DISTRIB_FLAGS_MASK,
37 		==, 0, __ATOMIC_RELAXED);
38 
39 	/* Sync with distributor on GET_BUF flag. */
40 	__atomic_store_n(&(buf->bufptr64), req, __ATOMIC_RELEASE);
41 }
42 
43 struct rte_mbuf *
44 rte_distributor_poll_pkt_single(struct rte_distributor_single *d,
45 		unsigned worker_id)
46 {
47 	union rte_distributor_buffer_single *buf = &d->bufs[worker_id];
48 	/* Sync with distributor. Acquire bufptr64. */
49 	if (__atomic_load_n(&buf->bufptr64, __ATOMIC_ACQUIRE)
50 		& RTE_DISTRIB_GET_BUF)
51 		return NULL;
52 
53 	/* since bufptr64 is signed, this should be an arithmetic shift */
54 	int64_t ret = buf->bufptr64 >> RTE_DISTRIB_FLAG_BITS;
55 	return (struct rte_mbuf *)((uintptr_t)ret);
56 }
57 
58 struct rte_mbuf *
59 rte_distributor_get_pkt_single(struct rte_distributor_single *d,
60 		unsigned worker_id, struct rte_mbuf *oldpkt)
61 {
62 	struct rte_mbuf *ret;
63 	rte_distributor_request_pkt_single(d, worker_id, oldpkt);
64 	while ((ret = rte_distributor_poll_pkt_single(d, worker_id)) == NULL)
65 		rte_pause();
66 	return ret;
67 }
68 
69 int
70 rte_distributor_return_pkt_single(struct rte_distributor_single *d,
71 		unsigned worker_id, struct rte_mbuf *oldpkt)
72 {
73 	union rte_distributor_buffer_single *buf = &d->bufs[worker_id];
74 	uint64_t req = (((int64_t)(uintptr_t)oldpkt) << RTE_DISTRIB_FLAG_BITS)
75 			| RTE_DISTRIB_RETURN_BUF;
76 	RTE_WAIT_UNTIL_MASKED(&buf->bufptr64, RTE_DISTRIB_FLAGS_MASK,
77 		==, 0, __ATOMIC_RELAXED);
78 
79 	/* Sync with distributor on RETURN_BUF flag. */
80 	__atomic_store_n(&(buf->bufptr64), req, __ATOMIC_RELEASE);
81 	return 0;
82 }
83 
84 /**** APIs called on distributor core ***/
85 
86 /* as name suggests, adds a packet to the backlog for a particular worker */
87 static int
88 add_to_backlog(struct rte_distributor_backlog *bl, int64_t item)
89 {
90 	if (bl->count == RTE_DISTRIB_BACKLOG_SIZE)
91 		return -1;
92 
93 	bl->pkts[(bl->start + bl->count++) & (RTE_DISTRIB_BACKLOG_MASK)]
94 			= item;
95 	return 0;
96 }
97 
98 /* takes the next packet for a worker off the backlog */
99 static int64_t
100 backlog_pop(struct rte_distributor_backlog *bl)
101 {
102 	bl->count--;
103 	return bl->pkts[bl->start++ & RTE_DISTRIB_BACKLOG_MASK];
104 }
105 
106 /* stores a packet returned from a worker inside the returns array */
107 static inline void
108 store_return(uintptr_t oldbuf, struct rte_distributor_single *d,
109 		unsigned *ret_start, unsigned *ret_count)
110 {
111 	/* store returns in a circular buffer - code is branch-free */
112 	d->returns.mbufs[(*ret_start + *ret_count) & RTE_DISTRIB_RETURNS_MASK]
113 			= (void *)oldbuf;
114 	*ret_start += (*ret_count == RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
115 	*ret_count += (*ret_count != RTE_DISTRIB_RETURNS_MASK) & !!(oldbuf);
116 }
117 
118 static inline void
119 handle_worker_shutdown(struct rte_distributor_single *d, unsigned int wkr)
120 {
121 	d->in_flight_tags[wkr] = 0;
122 	d->in_flight_bitmask &= ~(1UL << wkr);
123 	/* Sync with worker. Release bufptr64. */
124 	__atomic_store_n(&(d->bufs[wkr].bufptr64), 0, __ATOMIC_RELEASE);
125 	if (unlikely(d->backlog[wkr].count != 0)) {
126 		/* On return of a packet, we need to move the
127 		 * queued packets for this core elsewhere.
128 		 * Easiest solution is to set things up for
129 		 * a recursive call. That will cause those
130 		 * packets to be queued up for the next free
131 		 * core, i.e. it will return as soon as a
132 		 * core becomes free to accept the first
133 		 * packet, as subsequent ones will be added to
134 		 * the backlog for that core.
135 		 */
136 		struct rte_mbuf *pkts[RTE_DISTRIB_BACKLOG_SIZE];
137 		unsigned i;
138 		struct rte_distributor_backlog *bl = &d->backlog[wkr];
139 
140 		for (i = 0; i < bl->count; i++) {
141 			unsigned idx = (bl->start + i) &
142 					RTE_DISTRIB_BACKLOG_MASK;
143 			pkts[i] = (void *)((uintptr_t)(bl->pkts[idx] >>
144 					RTE_DISTRIB_FLAG_BITS));
145 		}
146 		/* recursive call.
147 		 * Note that the tags were set before first level call
148 		 * to rte_distributor_process.
149 		 */
150 		rte_distributor_process_single(d, pkts, i);
151 		bl->count = bl->start = 0;
152 	}
153 }
154 
155 /* this function is called when process() fn is called without any new
156  * packets. It goes through all the workers and clears any returned packets
157  * to do a partial flush.
158  */
159 static int
160 process_returns(struct rte_distributor_single *d)
161 {
162 	unsigned wkr;
163 	unsigned flushed = 0;
164 	unsigned ret_start = d->returns.start,
165 			ret_count = d->returns.count;
166 
167 	for (wkr = 0; wkr < d->num_workers; wkr++) {
168 		uintptr_t oldbuf = 0;
169 		/* Sync with worker. Acquire bufptr64. */
170 		const int64_t data = __atomic_load_n(&(d->bufs[wkr].bufptr64),
171 							__ATOMIC_ACQUIRE);
172 
173 		if (data & RTE_DISTRIB_GET_BUF) {
174 			flushed++;
175 			if (d->backlog[wkr].count)
176 				/* Sync with worker. Release bufptr64. */
177 				__atomic_store_n(&(d->bufs[wkr].bufptr64),
178 					backlog_pop(&d->backlog[wkr]),
179 					__ATOMIC_RELEASE);
180 			else {
181 				/* Sync with worker on GET_BUF flag. */
182 				__atomic_store_n(&(d->bufs[wkr].bufptr64),
183 					RTE_DISTRIB_GET_BUF,
184 					__ATOMIC_RELEASE);
185 				d->in_flight_tags[wkr] = 0;
186 				d->in_flight_bitmask &= ~(1UL << wkr);
187 			}
188 			oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
189 		} else if (data & RTE_DISTRIB_RETURN_BUF) {
190 			handle_worker_shutdown(d, wkr);
191 			oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
192 		}
193 
194 		store_return(oldbuf, d, &ret_start, &ret_count);
195 	}
196 
197 	d->returns.start = ret_start;
198 	d->returns.count = ret_count;
199 
200 	return flushed;
201 }
202 
203 /* process a set of packets to distribute them to workers */
204 int
205 rte_distributor_process_single(struct rte_distributor_single *d,
206 		struct rte_mbuf **mbufs, unsigned num_mbufs)
207 {
208 	unsigned next_idx = 0;
209 	unsigned wkr = 0;
210 	struct rte_mbuf *next_mb = NULL;
211 	int64_t next_value = 0;
212 	uint32_t new_tag = 0;
213 	unsigned ret_start = d->returns.start,
214 			ret_count = d->returns.count;
215 
216 	if (unlikely(num_mbufs == 0))
217 		return process_returns(d);
218 
219 	while (next_idx < num_mbufs || next_mb != NULL) {
220 		uintptr_t oldbuf = 0;
221 		/* Sync with worker. Acquire bufptr64. */
222 		int64_t data = __atomic_load_n(&(d->bufs[wkr].bufptr64),
223 						__ATOMIC_ACQUIRE);
224 
225 		if (!next_mb) {
226 			next_mb = mbufs[next_idx++];
227 			next_value = (((int64_t)(uintptr_t)next_mb)
228 					<< RTE_DISTRIB_FLAG_BITS);
229 			/*
230 			 * User is advocated to set tag value for each
231 			 * mbuf before calling rte_distributor_process.
232 			 * User defined tags are used to identify flows,
233 			 * or sessions.
234 			 */
235 			new_tag = next_mb->hash.usr;
236 
237 			/*
238 			 * Note that if RTE_DISTRIB_MAX_WORKERS is larger than 64
239 			 * then the size of match has to be expanded.
240 			 */
241 			uint64_t match = 0;
242 			unsigned i;
243 			/*
244 			 * to scan for a match use "xor" and "not" to get a 0/1
245 			 * value, then use shifting to merge to single "match"
246 			 * variable, where a one-bit indicates a match for the
247 			 * worker given by the bit-position
248 			 */
249 			for (i = 0; i < d->num_workers; i++)
250 				match |= (!(d->in_flight_tags[i] ^ new_tag)
251 					<< i);
252 
253 			/* Only turned-on bits are considered as match */
254 			match &= d->in_flight_bitmask;
255 
256 			if (match) {
257 				next_mb = NULL;
258 				unsigned worker = __builtin_ctzl(match);
259 				if (add_to_backlog(&d->backlog[worker],
260 						next_value) < 0)
261 					next_idx--;
262 			}
263 		}
264 
265 		if ((data & RTE_DISTRIB_GET_BUF) &&
266 				(d->backlog[wkr].count || next_mb)) {
267 
268 			if (d->backlog[wkr].count)
269 				/* Sync with worker. Release bufptr64. */
270 				__atomic_store_n(&(d->bufs[wkr].bufptr64),
271 						backlog_pop(&d->backlog[wkr]),
272 						__ATOMIC_RELEASE);
273 
274 			else {
275 				/* Sync with worker. Release bufptr64.  */
276 				__atomic_store_n(&(d->bufs[wkr].bufptr64),
277 						next_value,
278 						__ATOMIC_RELEASE);
279 				d->in_flight_tags[wkr] = new_tag;
280 				d->in_flight_bitmask |= (1UL << wkr);
281 				next_mb = NULL;
282 			}
283 			oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
284 		} else if (data & RTE_DISTRIB_RETURN_BUF) {
285 			handle_worker_shutdown(d, wkr);
286 			oldbuf = data >> RTE_DISTRIB_FLAG_BITS;
287 		}
288 
289 		/* store returns in a circular buffer */
290 		store_return(oldbuf, d, &ret_start, &ret_count);
291 
292 		if (++wkr == d->num_workers)
293 			wkr = 0;
294 	}
295 	/* to finish, check all workers for backlog and schedule work for them
296 	 * if they are ready */
297 	for (wkr = 0; wkr < d->num_workers; wkr++)
298 		if (d->backlog[wkr].count &&
299 				/* Sync with worker. Acquire bufptr64. */
300 				(__atomic_load_n(&(d->bufs[wkr].bufptr64),
301 				__ATOMIC_ACQUIRE) & RTE_DISTRIB_GET_BUF)) {
302 
303 			int64_t oldbuf = d->bufs[wkr].bufptr64 >>
304 					RTE_DISTRIB_FLAG_BITS;
305 
306 			store_return(oldbuf, d, &ret_start, &ret_count);
307 
308 			/* Sync with worker. Release bufptr64. */
309 			__atomic_store_n(&(d->bufs[wkr].bufptr64),
310 				backlog_pop(&d->backlog[wkr]),
311 				__ATOMIC_RELEASE);
312 		}
313 
314 	d->returns.start = ret_start;
315 	d->returns.count = ret_count;
316 	return num_mbufs;
317 }
318 
319 /* return to the caller, packets returned from workers */
320 int
321 rte_distributor_returned_pkts_single(struct rte_distributor_single *d,
322 		struct rte_mbuf **mbufs, unsigned max_mbufs)
323 {
324 	struct rte_distributor_returned_pkts *returns = &d->returns;
325 	unsigned retval = (max_mbufs < returns->count) ?
326 			max_mbufs : returns->count;
327 	unsigned i;
328 
329 	for (i = 0; i < retval; i++) {
330 		unsigned idx = (returns->start + i) & RTE_DISTRIB_RETURNS_MASK;
331 		mbufs[i] = returns->mbufs[idx];
332 	}
333 	returns->start += i;
334 	returns->count -= i;
335 
336 	return retval;
337 }
338 
339 /* return the number of packets in-flight in a distributor, i.e. packets
340  * being worked on or queued up in a backlog.
341  */
342 static inline unsigned
343 total_outstanding(const struct rte_distributor_single *d)
344 {
345 	unsigned wkr, total_outstanding;
346 
347 	total_outstanding = __builtin_popcountl(d->in_flight_bitmask);
348 
349 	for (wkr = 0; wkr < d->num_workers; wkr++)
350 		total_outstanding += d->backlog[wkr].count;
351 
352 	return total_outstanding;
353 }
354 
355 /* flush the distributor, so that there are no outstanding packets in flight or
356  * queued up. */
357 int
358 rte_distributor_flush_single(struct rte_distributor_single *d)
359 {
360 	const unsigned flushed = total_outstanding(d);
361 
362 	while (total_outstanding(d) > 0)
363 		rte_distributor_process_single(d, NULL, 0);
364 
365 	return flushed;
366 }
367 
368 /* clears the internal returns array in the distributor */
369 void
370 rte_distributor_clear_returns_single(struct rte_distributor_single *d)
371 {
372 	d->returns.start = d->returns.count = 0;
373 #ifndef __OPTIMIZE__
374 	memset(d->returns.mbufs, 0, sizeof(d->returns.mbufs));
375 #endif
376 }
377 
378 /* creates a distributor instance */
379 struct rte_distributor_single *
380 rte_distributor_create_single(const char *name,
381 		unsigned socket_id,
382 		unsigned num_workers)
383 {
384 	struct rte_distributor_single *d;
385 	struct rte_distributor_list *distributor_list;
386 	char mz_name[RTE_MEMZONE_NAMESIZE];
387 	const struct rte_memzone *mz;
388 
389 	/* compilation-time checks */
390 	RTE_BUILD_BUG_ON((sizeof(*d) & RTE_CACHE_LINE_MASK) != 0);
391 	RTE_BUILD_BUG_ON((RTE_DISTRIB_MAX_WORKERS & 7) != 0);
392 	RTE_BUILD_BUG_ON(RTE_DISTRIB_MAX_WORKERS >
393 				sizeof(d->in_flight_bitmask) * CHAR_BIT);
394 
395 	if (name == NULL || num_workers >= RTE_DISTRIB_MAX_WORKERS) {
396 		rte_errno = EINVAL;
397 		return NULL;
398 	}
399 
400 	snprintf(mz_name, sizeof(mz_name), RTE_DISTRIB_PREFIX"%s", name);
401 	mz = rte_memzone_reserve(mz_name, sizeof(*d), socket_id, NO_FLAGS);
402 	if (mz == NULL) {
403 		rte_errno = ENOMEM;
404 		return NULL;
405 	}
406 
407 	d = mz->addr;
408 	strlcpy(d->name, name, sizeof(d->name));
409 	d->num_workers = num_workers;
410 
411 	distributor_list = RTE_TAILQ_CAST(rte_distributor_tailq.head,
412 					  rte_distributor_list);
413 
414 	rte_mcfg_tailq_write_lock();
415 	TAILQ_INSERT_TAIL(distributor_list, d, next);
416 	rte_mcfg_tailq_write_unlock();
417 
418 	return d;
419 }
420