2 * Copyright (c) 2007-2009 Kip Macy <kmacy@freebsd.org>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 #ifndef _SYS_BUF_RING_H_
31 #define _SYS_BUF_RING_H_
33 #include <machine/cpu.h>
35 #if defined(INVARIANTS) && !defined(DEBUG_BUFRING)
36 #define DEBUG_BUFRING 1
41 #include <sys/mutex.h>
45 volatile uint32_t br_prod_head;
46 volatile uint32_t br_prod_tail;
50 volatile uint32_t br_cons_head __aligned(CACHE_LINE_SIZE);
51 volatile uint32_t br_cons_tail;
57 void *br_ring[0] __aligned(CACHE_LINE_SIZE);
61 * multi-producer safe lock-free ring buffer enqueue
65 buf_ring_enqueue(struct buf_ring *br, void *buf)
67 uint32_t prod_head, prod_next, cons_tail;
70 for (i = br->br_cons_head; i != br->br_prod_head;
71 i = ((i + 1) & br->br_cons_mask))
72 if(br->br_ring[i] == buf)
73 panic("buf=%p already enqueue at %d prod=%d cons=%d",
74 buf, i, br->br_prod_tail, br->br_cons_tail);
78 prod_head = br->br_prod_head;
79 prod_next = (prod_head + 1) & br->br_prod_mask;
80 cons_tail = br->br_cons_tail;
82 if (prod_next == cons_tail) {
84 if (prod_head == br->br_prod_head &&
85 cons_tail == br->br_cons_tail) {
92 } while (!atomic_cmpset_acq_int(&br->br_prod_head, prod_head, prod_next));
94 if (br->br_ring[prod_head] != NULL)
95 panic("dangling value in enqueue");
97 br->br_ring[prod_head] = buf;
100 * If there are other enqueues in progress
101 * that preceeded us, we need to wait for them
104 while (br->br_prod_tail != prod_head)
106 atomic_store_rel_int(&br->br_prod_tail, prod_next);
112 * multi-consumer safe dequeue
115 static __inline void *
116 buf_ring_dequeue_mc(struct buf_ring *br)
118 uint32_t cons_head, cons_next;
123 cons_head = br->br_cons_head;
124 cons_next = (cons_head + 1) & br->br_cons_mask;
126 if (cons_head == br->br_prod_tail) {
130 } while (!atomic_cmpset_acq_int(&br->br_cons_head, cons_head, cons_next));
132 buf = br->br_ring[cons_head];
134 br->br_ring[cons_head] = NULL;
137 * If there are other dequeues in progress
138 * that preceeded us, we need to wait for them
141 while (br->br_cons_tail != cons_head)
144 atomic_store_rel_int(&br->br_cons_tail, cons_next);
151 * single-consumer dequeue
152 * use where dequeue is protected by a lock
153 * e.g. a network driver's tx queue lock
155 static __inline void *
156 buf_ring_dequeue_sc(struct buf_ring *br)
158 uint32_t cons_head, cons_next;
159 #ifdef PREFETCH_DEFINED
160 uint32_t cons_next_next;
165 cons_head = br->br_cons_head;
166 prod_tail = br->br_prod_tail;
168 cons_next = (cons_head + 1) & br->br_cons_mask;
169 #ifdef PREFETCH_DEFINED
170 cons_next_next = (cons_head + 2) & br->br_cons_mask;
173 if (cons_head == prod_tail)
176 #ifdef PREFETCH_DEFINED
177 if (cons_next != prod_tail) {
178 prefetch(br->br_ring[cons_next]);
179 if (cons_next_next != prod_tail)
180 prefetch(br->br_ring[cons_next_next]);
183 br->br_cons_head = cons_next;
184 buf = br->br_ring[cons_head];
187 br->br_ring[cons_head] = NULL;
188 if (!mtx_owned(br->br_lock))
189 panic("lock not held on single consumer dequeue");
190 if (br->br_cons_tail != cons_head)
191 panic("inconsistent list cons_tail=%d cons_head=%d",
192 br->br_cons_tail, cons_head);
194 br->br_cons_tail = cons_next;
199 * single-consumer advance after a peek
200 * use where it is protected by a lock
201 * e.g. a network driver's tx queue lock
204 buf_ring_advance_sc(struct buf_ring *br)
206 uint32_t cons_head, cons_next;
209 cons_head = br->br_cons_head;
210 prod_tail = br->br_prod_tail;
212 cons_next = (cons_head + 1) & br->br_cons_mask;
213 if (cons_head == prod_tail)
215 br->br_cons_head = cons_next;
217 br->br_ring[cons_head] = NULL;
219 br->br_cons_tail = cons_next;
223 * Used to return a buffer (most likely already there)
224 * to the top od the ring. The caller should *not*
225 * have used any dequeue to pull it out of the ring
226 * but instead should have used the peek() function.
227 * This is normally used where the transmit queue
228 * of a driver is full, and an mubf must be returned.
229 * Most likely whats in the ring-buffer is what
230 * is being put back (since it was not removed), but
231 * sometimes the lower transmit function may have
232 * done a pullup or other function that will have
233 * changed it. As an optimzation we always put it
234 * back (since jhb says the store is probably cheaper),
235 * if we have to do a multi-queue version we will need
236 * the compare and an atomic.
239 buf_ring_putback_sc(struct buf_ring *br, void *new)
241 KASSERT(br->br_cons_head != br->br_prod_tail,
242 ("Buf-Ring has none in putback")) ;
243 br->br_ring[br->br_cons_head] = new;
247 * return a pointer to the first entry in the ring
248 * without modifying it, or NULL if the ring is empty
249 * race-prone if not protected by a lock
251 static __inline void *
252 buf_ring_peek(struct buf_ring *br)
256 if ((br->br_lock != NULL) && !mtx_owned(br->br_lock))
257 panic("lock not held on single consumer dequeue");
260 * I believe it is safe to not have a memory barrier
261 * here because we control cons and tail is worst case
262 * a lagging indicator so we worst case we might
263 * return NULL immediately after a buffer has been enqueued
265 if (br->br_cons_head == br->br_prod_tail)
268 return (br->br_ring[br->br_cons_head]);
271 static __inline void *
272 buf_ring_peek_clear_sc(struct buf_ring *br)
277 if (!mtx_owned(br->br_lock))
278 panic("lock not held on single consumer dequeue");
281 * I believe it is safe to not have a memory barrier
282 * here because we control cons and tail is worst case
283 * a lagging indicator so we worst case we might
284 * return NULL immediately after a buffer has been enqueued
286 if (br->br_cons_head == br->br_prod_tail)
291 * Single consumer, i.e. cons_head will not move while we are
292 * running, so atomic_swap_ptr() is not necessary here.
294 ret = br->br_ring[br->br_cons_head];
295 br->br_ring[br->br_cons_head] = NULL;
298 return (br->br_ring[br->br_cons_head]);
303 buf_ring_full(struct buf_ring *br)
306 return (((br->br_prod_head + 1) & br->br_prod_mask) == br->br_cons_tail);
310 buf_ring_empty(struct buf_ring *br)
313 return (br->br_cons_head == br->br_prod_tail);
317 buf_ring_count(struct buf_ring *br)
320 return ((br->br_prod_size + br->br_prod_tail - br->br_cons_tail)
324 struct buf_ring *buf_ring_alloc(int count, struct malloc_type *type, int flags,
326 void buf_ring_free(struct buf_ring *br, struct malloc_type *type);