2 * Definitions for the 'struct ptr_ring' datastructure.
5 * Michael S. Tsirkin <mst@redhat.com>
7 * Copyright (C) 2016 Red Hat, Inc.
9 * This program is free software; you can redistribute it and/or modify it
10 * under the terms of the GNU General Public License as published by the
11 * Free Software Foundation; either version 2 of the License, or (at your
12 * option) any later version.
14 * This is a limited-size FIFO maintaining pointers in FIFO order, with
15 * one CPU producing entries and another consuming entries from a FIFO.
17 * This implementation tries to minimize cache-contention when there is a
18 * single producer and a single consumer CPU.
21 #ifndef _LINUX_PTR_RING_H
22 #define _LINUX_PTR_RING_H 1
25 #include <linux/spinlock.h>
26 #include <linux/cache.h>
27 #include <linux/types.h>
28 #include <linux/compiler.h>
29 #include <linux/cache.h>
30 #include <linux/slab.h>
31 #include <asm/errno.h>
35 int producer ____cacheline_aligned_in_smp;
36 spinlock_t producer_lock;
37 int consumer_head ____cacheline_aligned_in_smp; /* next valid entry */
38 int consumer_tail; /* next entry to invalidate */
39 spinlock_t consumer_lock;
40 /* Shared consumer/producer data */
41 /* Read-only by both the producer and the consumer */
42 int size ____cacheline_aligned_in_smp; /* max entries in queue */
43 int batch; /* number of entries to consume in a batch */
47 /* Note: callers invoking this in a loop must use a compiler barrier,
48 * for example cpu_relax().
50 * NB: this is unlike __ptr_ring_empty in that callers must hold producer_lock:
51 * see e.g. ptr_ring_full.
53 static inline bool __ptr_ring_full(struct ptr_ring *r)
55 return r->queue[r->producer];
58 static inline bool ptr_ring_full(struct ptr_ring *r)
62 spin_lock(&r->producer_lock);
63 ret = __ptr_ring_full(r);
64 spin_unlock(&r->producer_lock);
69 static inline bool ptr_ring_full_irq(struct ptr_ring *r)
73 spin_lock_irq(&r->producer_lock);
74 ret = __ptr_ring_full(r);
75 spin_unlock_irq(&r->producer_lock);
80 static inline bool ptr_ring_full_any(struct ptr_ring *r)
85 spin_lock_irqsave(&r->producer_lock, flags);
86 ret = __ptr_ring_full(r);
87 spin_unlock_irqrestore(&r->producer_lock, flags);
92 static inline bool ptr_ring_full_bh(struct ptr_ring *r)
96 spin_lock_bh(&r->producer_lock);
97 ret = __ptr_ring_full(r);
98 spin_unlock_bh(&r->producer_lock);
103 /* Note: callers invoking this in a loop must use a compiler barrier,
104 * for example cpu_relax(). Callers must hold producer_lock.
105 * Callers are responsible for making sure pointer that is being queued
106 * points to a valid data.
108 static inline int __ptr_ring_produce(struct ptr_ring *r, void *ptr)
110 if (unlikely(!r->size) || r->queue[r->producer])
113 /* Make sure the pointer we are storing points to a valid data. */
114 /* Pairs with smp_read_barrier_depends in __ptr_ring_consume. */
117 WRITE_ONCE(r->queue[r->producer++], ptr);
118 if (unlikely(r->producer >= r->size))
124 * Note: resize (below) nests producer lock within consumer lock, so if you
125 * consume in interrupt or BH context, you must disable interrupts/BH when
128 static inline int ptr_ring_produce(struct ptr_ring *r, void *ptr)
132 spin_lock(&r->producer_lock);
133 ret = __ptr_ring_produce(r, ptr);
134 spin_unlock(&r->producer_lock);
139 static inline int ptr_ring_produce_irq(struct ptr_ring *r, void *ptr)
143 spin_lock_irq(&r->producer_lock);
144 ret = __ptr_ring_produce(r, ptr);
145 spin_unlock_irq(&r->producer_lock);
150 static inline int ptr_ring_produce_any(struct ptr_ring *r, void *ptr)
155 spin_lock_irqsave(&r->producer_lock, flags);
156 ret = __ptr_ring_produce(r, ptr);
157 spin_unlock_irqrestore(&r->producer_lock, flags);
162 static inline int ptr_ring_produce_bh(struct ptr_ring *r, void *ptr)
166 spin_lock_bh(&r->producer_lock);
167 ret = __ptr_ring_produce(r, ptr);
168 spin_unlock_bh(&r->producer_lock);
173 static inline void *__ptr_ring_peek(struct ptr_ring *r)
176 return READ_ONCE(r->queue[r->consumer_head]);
181 * Test ring empty status without taking any locks.
183 * NB: This is only safe to call if ring is never resized.
185 * However, if some other CPU consumes ring entries at the same time, the value
186 * returned is not guaranteed to be correct.
188 * In this case - to avoid incorrectly detecting the ring
189 * as empty - the CPU consuming the ring entries is responsible
190 * for either consuming all ring entries until the ring is empty,
191 * or synchronizing with some other CPU and causing it to
192 * re-test __ptr_ring_empty and/or consume the ring enteries
193 * after the synchronization point.
195 * Note: callers invoking this in a loop must use a compiler barrier,
196 * for example cpu_relax().
198 static inline bool __ptr_ring_empty(struct ptr_ring *r)
201 return !r->queue[READ_ONCE(r->consumer_head)];
205 static inline bool ptr_ring_empty(struct ptr_ring *r)
209 spin_lock(&r->consumer_lock);
210 ret = __ptr_ring_empty(r);
211 spin_unlock(&r->consumer_lock);
216 static inline bool ptr_ring_empty_irq(struct ptr_ring *r)
220 spin_lock_irq(&r->consumer_lock);
221 ret = __ptr_ring_empty(r);
222 spin_unlock_irq(&r->consumer_lock);
227 static inline bool ptr_ring_empty_any(struct ptr_ring *r)
232 spin_lock_irqsave(&r->consumer_lock, flags);
233 ret = __ptr_ring_empty(r);
234 spin_unlock_irqrestore(&r->consumer_lock, flags);
239 static inline bool ptr_ring_empty_bh(struct ptr_ring *r)
243 spin_lock_bh(&r->consumer_lock);
244 ret = __ptr_ring_empty(r);
245 spin_unlock_bh(&r->consumer_lock);
250 /* Must only be called after __ptr_ring_peek returned !NULL */
251 static inline void __ptr_ring_discard_one(struct ptr_ring *r)
253 /* Fundamentally, what we want to do is update consumer
254 * index and zero out the entry so producer can reuse it.
255 * Doing it naively at each consume would be as simple as:
256 * consumer = r->consumer;
257 * r->queue[consumer++] = NULL;
258 * if (unlikely(consumer >= r->size))
260 * r->consumer = consumer;
261 * but that is suboptimal when the ring is full as producer is writing
262 * out new entries in the same cache line. Defer these updates until a
263 * batch of entries has been consumed.
265 /* Note: we must keep consumer_head valid at all times for __ptr_ring_empty
268 int consumer_head = r->consumer_head;
269 int head = consumer_head++;
271 /* Once we have processed enough entries invalidate them in
272 * the ring all at once so producer can reuse their space in the ring.
273 * We also do this when we reach end of the ring - not mandatory
274 * but helps keep the implementation simple.
276 if (unlikely(consumer_head - r->consumer_tail >= r->batch ||
277 consumer_head >= r->size)) {
278 /* Zero out entries in the reverse order: this way we touch the
279 * cache line that producer might currently be reading the last;
280 * producer won't make progress and touch other cache lines
281 * besides the first one until we write out all entries.
283 while (likely(head >= r->consumer_tail))
284 r->queue[head--] = NULL;
285 r->consumer_tail = consumer_head;
287 if (unlikely(consumer_head >= r->size)) {
289 r->consumer_tail = 0;
291 /* matching READ_ONCE in __ptr_ring_empty for lockless tests */
292 WRITE_ONCE(r->consumer_head, consumer_head);
295 static inline void *__ptr_ring_consume(struct ptr_ring *r)
299 /* The READ_ONCE in __ptr_ring_peek guarantees that anyone
300 * accessing data through the pointer is up to date. Pairs
301 * with smp_wmb in __ptr_ring_produce.
303 ptr = __ptr_ring_peek(r);
305 __ptr_ring_discard_one(r);
310 static inline int __ptr_ring_consume_batched(struct ptr_ring *r,
316 for (i = 0; i < n; i++) {
317 ptr = __ptr_ring_consume(r);
327 * Note: resize (below) nests producer lock within consumer lock, so if you
328 * call this in interrupt or BH context, you must disable interrupts/BH when
331 static inline void *ptr_ring_consume(struct ptr_ring *r)
335 spin_lock(&r->consumer_lock);
336 ptr = __ptr_ring_consume(r);
337 spin_unlock(&r->consumer_lock);
342 static inline void *ptr_ring_consume_irq(struct ptr_ring *r)
346 spin_lock_irq(&r->consumer_lock);
347 ptr = __ptr_ring_consume(r);
348 spin_unlock_irq(&r->consumer_lock);
353 static inline void *ptr_ring_consume_any(struct ptr_ring *r)
358 spin_lock_irqsave(&r->consumer_lock, flags);
359 ptr = __ptr_ring_consume(r);
360 spin_unlock_irqrestore(&r->consumer_lock, flags);
365 static inline void *ptr_ring_consume_bh(struct ptr_ring *r)
369 spin_lock_bh(&r->consumer_lock);
370 ptr = __ptr_ring_consume(r);
371 spin_unlock_bh(&r->consumer_lock);
376 static inline int ptr_ring_consume_batched(struct ptr_ring *r,
381 spin_lock(&r->consumer_lock);
382 ret = __ptr_ring_consume_batched(r, array, n);
383 spin_unlock(&r->consumer_lock);
388 static inline int ptr_ring_consume_batched_irq(struct ptr_ring *r,
393 spin_lock_irq(&r->consumer_lock);
394 ret = __ptr_ring_consume_batched(r, array, n);
395 spin_unlock_irq(&r->consumer_lock);
400 static inline int ptr_ring_consume_batched_any(struct ptr_ring *r,
406 spin_lock_irqsave(&r->consumer_lock, flags);
407 ret = __ptr_ring_consume_batched(r, array, n);
408 spin_unlock_irqrestore(&r->consumer_lock, flags);
413 static inline int ptr_ring_consume_batched_bh(struct ptr_ring *r,
418 spin_lock_bh(&r->consumer_lock);
419 ret = __ptr_ring_consume_batched(r, array, n);
420 spin_unlock_bh(&r->consumer_lock);
425 /* Cast to structure type and call a function without discarding from FIFO.
426 * Function must return a value.
427 * Callers must take consumer_lock.
429 #define __PTR_RING_PEEK_CALL(r, f) ((f)(__ptr_ring_peek(r)))
431 #define PTR_RING_PEEK_CALL(r, f) ({ \
432 typeof((f)(NULL)) __PTR_RING_PEEK_CALL_v; \
434 spin_lock(&(r)->consumer_lock); \
435 __PTR_RING_PEEK_CALL_v = __PTR_RING_PEEK_CALL(r, f); \
436 spin_unlock(&(r)->consumer_lock); \
437 __PTR_RING_PEEK_CALL_v; \
440 #define PTR_RING_PEEK_CALL_IRQ(r, f) ({ \
441 typeof((f)(NULL)) __PTR_RING_PEEK_CALL_v; \
443 spin_lock_irq(&(r)->consumer_lock); \
444 __PTR_RING_PEEK_CALL_v = __PTR_RING_PEEK_CALL(r, f); \
445 spin_unlock_irq(&(r)->consumer_lock); \
446 __PTR_RING_PEEK_CALL_v; \
449 #define PTR_RING_PEEK_CALL_BH(r, f) ({ \
450 typeof((f)(NULL)) __PTR_RING_PEEK_CALL_v; \
452 spin_lock_bh(&(r)->consumer_lock); \
453 __PTR_RING_PEEK_CALL_v = __PTR_RING_PEEK_CALL(r, f); \
454 spin_unlock_bh(&(r)->consumer_lock); \
455 __PTR_RING_PEEK_CALL_v; \
458 #define PTR_RING_PEEK_CALL_ANY(r, f) ({ \
459 typeof((f)(NULL)) __PTR_RING_PEEK_CALL_v; \
460 unsigned long __PTR_RING_PEEK_CALL_f;\
462 spin_lock_irqsave(&(r)->consumer_lock, __PTR_RING_PEEK_CALL_f); \
463 __PTR_RING_PEEK_CALL_v = __PTR_RING_PEEK_CALL(r, f); \
464 spin_unlock_irqrestore(&(r)->consumer_lock, __PTR_RING_PEEK_CALL_f); \
465 __PTR_RING_PEEK_CALL_v; \
468 /* Not all gfp_t flags (besides GFP_KERNEL) are allowed. See
469 * documentation for vmalloc for which of them are legal.
471 static inline void **__ptr_ring_init_queue_alloc(unsigned int size, gfp_t gfp)
473 if (size > KMALLOC_MAX_SIZE / sizeof(void *))
475 return kvmalloc_array(size, sizeof(void *), gfp | __GFP_ZERO);
478 static inline void __ptr_ring_set_size(struct ptr_ring *r, int size)
481 r->batch = SMP_CACHE_BYTES * 2 / sizeof(*(r->queue));
482 /* We need to set batch at least to 1 to make logic
483 * in __ptr_ring_discard_one work correctly.
484 * Batching too much (because ring is small) would cause a lot of
485 * burstiness. Needs tuning, for now disable batching.
487 if (r->batch > r->size / 2 || !r->batch)
491 static inline int ptr_ring_init(struct ptr_ring *r, int size, gfp_t gfp)
493 r->queue = __ptr_ring_init_queue_alloc(size, gfp);
497 __ptr_ring_set_size(r, size);
498 r->producer = r->consumer_head = r->consumer_tail = 0;
499 spin_lock_init(&r->producer_lock);
500 spin_lock_init(&r->consumer_lock);
506 * Return entries into ring. Destroy entries that don't fit.
508 * Note: this is expected to be a rare slow path operation.
510 * Note: producer lock is nested within consumer lock, so if you
511 * resize you must make sure all uses nest correctly.
512 * In particular if you consume ring in interrupt or BH context, you must
513 * disable interrupts/BH when doing so.
515 static inline void ptr_ring_unconsume(struct ptr_ring *r, void **batch, int n,
516 void (*destroy)(void *))
521 spin_lock_irqsave(&r->consumer_lock, flags);
522 spin_lock(&r->producer_lock);
528 * Clean out buffered entries (for simplicity). This way following code
529 * can test entries for NULL and if not assume they are valid.
531 head = r->consumer_head - 1;
532 while (likely(head >= r->consumer_tail))
533 r->queue[head--] = NULL;
534 r->consumer_tail = r->consumer_head;
537 * Go over entries in batch, start moving head back and copy entries.
538 * Stop when we run into previously unconsumed entries.
541 head = r->consumer_head - 1;
544 if (r->queue[head]) {
545 /* This batch entry will have to be destroyed. */
548 r->queue[head] = batch[--n];
549 r->consumer_tail = head;
550 /* matching READ_ONCE in __ptr_ring_empty for lockless tests */
551 WRITE_ONCE(r->consumer_head, head);
555 /* Destroy all entries left in the batch. */
558 spin_unlock(&r->producer_lock);
559 spin_unlock_irqrestore(&r->consumer_lock, flags);
562 static inline void **__ptr_ring_swap_queue(struct ptr_ring *r, void **queue,
564 void (*destroy)(void *))
570 while ((ptr = __ptr_ring_consume(r)))
572 queue[producer++] = ptr;
576 __ptr_ring_set_size(r, size);
577 r->producer = producer;
578 r->consumer_head = 0;
579 r->consumer_tail = 0;
587 * Note: producer lock is nested within consumer lock, so if you
588 * resize you must make sure all uses nest correctly.
589 * In particular if you consume ring in interrupt or BH context, you must
590 * disable interrupts/BH when doing so.
592 static inline int ptr_ring_resize(struct ptr_ring *r, int size, gfp_t gfp,
593 void (*destroy)(void *))
596 void **queue = __ptr_ring_init_queue_alloc(size, gfp);
602 spin_lock_irqsave(&(r)->consumer_lock, flags);
603 spin_lock(&(r)->producer_lock);
605 old = __ptr_ring_swap_queue(r, queue, size, gfp, destroy);
607 spin_unlock(&(r)->producer_lock);
608 spin_unlock_irqrestore(&(r)->consumer_lock, flags);
616 * Note: producer lock is nested within consumer lock, so if you
617 * resize you must make sure all uses nest correctly.
618 * In particular if you consume ring in interrupt or BH context, you must
619 * disable interrupts/BH when doing so.
621 static inline int ptr_ring_resize_multiple(struct ptr_ring **rings,
624 gfp_t gfp, void (*destroy)(void *))
630 queues = kmalloc_array(nrings, sizeof(*queues), gfp);
634 for (i = 0; i < nrings; ++i) {
635 queues[i] = __ptr_ring_init_queue_alloc(size, gfp);
640 for (i = 0; i < nrings; ++i) {
641 spin_lock_irqsave(&(rings[i])->consumer_lock, flags);
642 spin_lock(&(rings[i])->producer_lock);
643 queues[i] = __ptr_ring_swap_queue(rings[i], queues[i],
645 spin_unlock(&(rings[i])->producer_lock);
646 spin_unlock_irqrestore(&(rings[i])->consumer_lock, flags);
649 for (i = 0; i < nrings; ++i)
666 static inline void ptr_ring_cleanup(struct ptr_ring *r, void (*destroy)(void *))
671 while ((ptr = ptr_ring_consume(r)))
676 #endif /* _LINUX_PTR_RING_H */