ptr_ring: array based FIFO for pointers
authorMichael S. Tsirkin <mst@redhat.com>
Mon, 13 Jun 2016 20:54:31 +0000 (23:54 +0300)
committerDavid S. Miller <davem@davemloft.net>
Wed, 15 Jun 2016 20:57:21 +0000 (13:57 -0700)
A simple array based FIFO of pointers.  Intended for net stack which
commonly has a single consumer/producer.

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/ptr_ring.h [new file with mode: 0644]

diff --git a/include/linux/ptr_ring.h b/include/linux/ptr_ring.h
new file mode 100644 (file)
index 0000000..633406f
--- /dev/null
@@ -0,0 +1,264 @@
+/*
+ *     Definitions for the 'struct ptr_ring' datastructure.
+ *
+ *     Author:
+ *             Michael S. Tsirkin <mst@redhat.com>
+ *
+ *     Copyright (C) 2016 Red Hat, Inc.
+ *
+ *     This program is free software; you can redistribute it and/or modify it
+ *     under the terms of the GNU General Public License as published by the
+ *     Free Software Foundation; either version 2 of the License, or (at your
+ *     option) any later version.
+ *
+ *     This is a limited-size FIFO maintaining pointers in FIFO order, with
+ *     one CPU producing entries and another consuming entries from a FIFO.
+ *
+ *     This implementation tries to minimize cache-contention when there is a
+ *     single producer and a single consumer CPU.
+ */
+
+#ifndef _LINUX_PTR_RING_H
+#define _LINUX_PTR_RING_H 1
+
+#ifdef __KERNEL__
+#include <linux/spinlock.h>
+#include <linux/cache.h>
+#include <linux/types.h>
+#include <linux/compiler.h>
+#include <linux/cache.h>
+#include <linux/slab.h>
+#include <asm/errno.h>
+#endif
+
+struct ptr_ring {
+       int producer ____cacheline_aligned_in_smp;
+       spinlock_t producer_lock;
+       int consumer ____cacheline_aligned_in_smp;
+       spinlock_t consumer_lock;
+       /* Shared consumer/producer data */
+       /* Read-only by both the producer and the consumer */
+       int size ____cacheline_aligned_in_smp; /* max entries in queue */
+       void **queue;
+};
+
+/* Note: callers invoking this in a loop must use a compiler barrier,
+ * for example cpu_relax().
+ * Callers don't need to take producer lock - if they don't
+ * the next call to __ptr_ring_produce may fail.
+ */
+static inline bool __ptr_ring_full(struct ptr_ring *r)
+{
+       return r->queue[r->producer];
+}
+
+static inline bool ptr_ring_full(struct ptr_ring *r)
+{
+       barrier();
+       return __ptr_ring_full(r);
+}
+
+/* Note: callers invoking this in a loop must use a compiler barrier,
+ * for example cpu_relax().
+ */
+static inline int __ptr_ring_produce(struct ptr_ring *r, void *ptr)
+{
+       if (__ptr_ring_full(r))
+               return -ENOSPC;
+
+       r->queue[r->producer++] = ptr;
+       if (unlikely(r->producer >= r->size))
+               r->producer = 0;
+       return 0;
+}
+
+static inline int ptr_ring_produce(struct ptr_ring *r, void *ptr)
+{
+       int ret;
+
+       spin_lock(&r->producer_lock);
+       ret = __ptr_ring_produce(r, ptr);
+       spin_unlock(&r->producer_lock);
+
+       return ret;
+}
+
+static inline int ptr_ring_produce_irq(struct ptr_ring *r, void *ptr)
+{
+       int ret;
+
+       spin_lock_irq(&r->producer_lock);
+       ret = __ptr_ring_produce(r, ptr);
+       spin_unlock_irq(&r->producer_lock);
+
+       return ret;
+}
+
+static inline int ptr_ring_produce_any(struct ptr_ring *r, void *ptr)
+{
+       unsigned long flags;
+       int ret;
+
+       spin_lock_irqsave(&r->producer_lock, flags);
+       ret = __ptr_ring_produce(r, ptr);
+       spin_unlock_irqrestore(&r->producer_lock, flags);
+
+       return ret;
+}
+
+static inline int ptr_ring_produce_bh(struct ptr_ring *r, void *ptr)
+{
+       int ret;
+
+       spin_lock_bh(&r->producer_lock);
+       ret = __ptr_ring_produce(r, ptr);
+       spin_unlock_bh(&r->producer_lock);
+
+       return ret;
+}
+
+/* Note: callers invoking this in a loop must use a compiler barrier,
+ * for example cpu_relax(). Callers must take consumer_lock
+ * if they dereference the pointer - see e.g. PTR_RING_PEEK_CALL.
+ * There's no need for a lock if pointer is merely tested - see e.g.
+ * ptr_ring_empty.
+ */
+static inline void *__ptr_ring_peek(struct ptr_ring *r)
+{
+       return r->queue[r->consumer];
+}
+
+static inline bool ptr_ring_empty(struct ptr_ring *r)
+{
+       barrier();
+       return !__ptr_ring_peek(r);
+}
+
+/* Must only be called after __ptr_ring_peek returned !NULL */
+static inline void __ptr_ring_discard_one(struct ptr_ring *r)
+{
+       r->queue[r->consumer++] = NULL;
+       if (unlikely(r->consumer >= r->size))
+               r->consumer = 0;
+}
+
+static inline void *__ptr_ring_consume(struct ptr_ring *r)
+{
+       void *ptr;
+
+       ptr = __ptr_ring_peek(r);
+       if (ptr)
+               __ptr_ring_discard_one(r);
+
+       return ptr;
+}
+
+static inline void *ptr_ring_consume(struct ptr_ring *r)
+{
+       void *ptr;
+
+       spin_lock(&r->consumer_lock);
+       ptr = __ptr_ring_consume(r);
+       spin_unlock(&r->consumer_lock);
+
+       return ptr;
+}
+
+static inline void *ptr_ring_consume_irq(struct ptr_ring *r)
+{
+       void *ptr;
+
+       spin_lock_irq(&r->consumer_lock);
+       ptr = __ptr_ring_consume(r);
+       spin_unlock_irq(&r->consumer_lock);
+
+       return ptr;
+}
+
+static inline void *ptr_ring_consume_any(struct ptr_ring *r)
+{
+       unsigned long flags;
+       void *ptr;
+
+       spin_lock_irqsave(&r->consumer_lock, flags);
+       ptr = __ptr_ring_consume(r);
+       spin_unlock_irqrestore(&r->consumer_lock, flags);
+
+       return ptr;
+}
+
+static inline void *ptr_ring_consume_bh(struct ptr_ring *r)
+{
+       void *ptr;
+
+       spin_lock_bh(&r->consumer_lock);
+       ptr = __ptr_ring_consume(r);
+       spin_unlock_bh(&r->consumer_lock);
+
+       return ptr;
+}
+
+/* Cast to structure type and call a function without discarding from FIFO.
+ * Function must return a value.
+ * Callers must take consumer_lock.
+ */
+#define __PTR_RING_PEEK_CALL(r, f) ((f)(__ptr_ring_peek(r)))
+
+#define PTR_RING_PEEK_CALL(r, f) ({ \
+       typeof((f)(NULL)) __PTR_RING_PEEK_CALL_v; \
+       \
+       spin_lock(&(r)->consumer_lock); \
+       __PTR_RING_PEEK_CALL_v = __PTR_RING_PEEK_CALL(r, f); \
+       spin_unlock(&(r)->consumer_lock); \
+       __PTR_RING_PEEK_CALL_v; \
+})
+
+#define PTR_RING_PEEK_CALL_IRQ(r, f) ({ \
+       typeof((f)(NULL)) __PTR_RING_PEEK_CALL_v; \
+       \
+       spin_lock_irq(&(r)->consumer_lock); \
+       __PTR_RING_PEEK_CALL_v = __PTR_RING_PEEK_CALL(r, f); \
+       spin_unlock_irq(&(r)->consumer_lock); \
+       __PTR_RING_PEEK_CALL_v; \
+})
+
+#define PTR_RING_PEEK_CALL_BH(r, f) ({ \
+       typeof((f)(NULL)) __PTR_RING_PEEK_CALL_v; \
+       \
+       spin_lock_bh(&(r)->consumer_lock); \
+       __PTR_RING_PEEK_CALL_v = __PTR_RING_PEEK_CALL(r, f); \
+       spin_unlock_bh(&(r)->consumer_lock); \
+       __PTR_RING_PEEK_CALL_v; \
+})
+
+#define PTR_RING_PEEK_CALL_ANY(r, f) ({ \
+       typeof((f)(NULL)) __PTR_RING_PEEK_CALL_v; \
+       unsigned long __PTR_RING_PEEK_CALL_f;\
+       \
+       spin_lock_irqsave(&(r)->consumer_lock, __PTR_RING_PEEK_CALL_f); \
+       __PTR_RING_PEEK_CALL_v = __PTR_RING_PEEK_CALL(r, f); \
+       spin_unlock_irqrestore(&(r)->consumer_lock, __PTR_RING_PEEK_CALL_f); \
+       __PTR_RING_PEEK_CALL_v; \
+})
+
+static inline int ptr_ring_init(struct ptr_ring *r, int size, gfp_t gfp)
+{
+       r->queue = kzalloc(ALIGN(size * sizeof *(r->queue), SMP_CACHE_BYTES),
+                          gfp);
+       if (!r->queue)
+               return -ENOMEM;
+
+       r->size = size;
+       r->producer = r->consumer = 0;
+       spin_lock_init(&r->producer_lock);
+       spin_lock_init(&r->consumer_lock);
+
+       return 0;
+}
+
+static inline void ptr_ring_cleanup(struct ptr_ring *r)
+{
+       kfree(r->queue);
+}
+
+#endif /* _LINUX_PTR_RING_H  */