[linux-2.6-block.git] / net / xdp / xsk_queue.h

/* SPDX-License-Identifier: GPL-2.0 */
/* XDP user-space ring structure
 * Copyright(c) 2018 Intel Corporation.
 */

#ifndef _LINUX_XSK_QUEUE_H
#define _LINUX_XSK_QUEUE_H

#include <linux/types.h>
#include <linux/if_xdp.h>
#include <net/xdp_sock.h>

#define RX_BATCH_SIZE 16
#define LAZY_UPDATE_THRESHOLD 128

struct xdp_ring {
	u32 producer ____cacheline_aligned_in_smp;
	u32 consumer ____cacheline_aligned_in_smp;
};

/* Used for the RX and TX queues for packets */
struct xdp_rxtx_ring {
	struct xdp_ring ptrs;
	struct xdp_desc desc[0] ____cacheline_aligned_in_smp;
};

/* Used for the fill and completion queues for buffers */
struct xdp_umem_ring {
	struct xdp_ring ptrs;
	u64 desc[0] ____cacheline_aligned_in_smp;
};

struct xsk_queue {
	u64 chunk_mask;
	u64 size;
	u32 ring_mask;
	u32 nentries;
	u32 prod_head;
	u32 prod_tail;
	u32 cons_head;
	u32 cons_tail;
	struct xdp_ring *ring;
	u64 invalid_descs;
};

/* The structure of the shared state of the rings are the same as the
 * ring buffer in kernel/events/ring_buffer.c. For the Rx and completion
 * ring, the kernel is the producer and user space is the consumer. For
 * the Tx and fill rings, the kernel is the consumer and user space is
 * the producer.
 *
 * producer                         consumer
 *
 * if (LOAD ->consumer) {           LOAD ->producer
 *                    (A)           smp_rmb()       (C)
 *    STORE $data                   LOAD $data
 *    smp_wmb()       (B)           smp_mb()        (D)
 *    STORE ->producer              STORE ->consumer
 * }
 *
 * (A) pairs with (D), and (B) pairs with (C).
 *
 * Starting with (B), it protects the data from being written after
 * the producer pointer. If this barrier was missing, the consumer
 * could observe the producer pointer being set and thus load the data
 * before the producer has written the new data. The consumer would in
 * this case load the old data.
 *
 * (C) protects the consumer from speculatively loading the data before
 * the producer pointer actually has been read. If we do not have this
 * barrier, some architectures could load old data as speculative loads
 * are not discarded as the CPU does not know there is a dependency
 * between ->producer and data.
 *
 * (A) is a control dependency that separates the load of ->consumer
 * from the stores of $data. In case ->consumer indicates there is no
 * room in the buffer to store $data we do not. So no barrier is needed.
 *
 * (D) protects the load of the data to be observed to happen after the
 * store of the consumer pointer. If we did not have this memory
 * barrier, the producer could observe the consumer pointer being set
 * and overwrite the data with a new value before the consumer got the
 * chance to read the old value. The consumer would thus miss reading
 * the old entry and very likely read the new entry twice, once right
 * now and again after circling through the ring.
 */

/* Common functions operating for both RXTX and umem queues */

static inline u64 xskq_nb_invalid_descs(struct xsk_queue *q)
{
	return q ? q->invalid_descs : 0;
}

static inline u32 xskq_nb_avail(struct xsk_queue *q, u32 dcnt)
{
	u32 entries = q->prod_tail - q->cons_tail;

	if (entries == 0) {
		/* Refresh the local pointer */
		q->prod_tail = READ_ONCE(q->ring->producer);
		entries = q->prod_tail - q->cons_tail;
	}

	return (entries > dcnt) ? dcnt : entries;
}

static inline u32 xskq_nb_free(struct xsk_queue *q, u32 producer, u32 dcnt)
{
	u32 free_entries = q->nentries - (producer - q->cons_tail);

	if (free_entries >= dcnt)
		return free_entries;

	/* Refresh the local tail pointer */
	q->cons_tail = READ_ONCE(q->ring->consumer);
	return q->nentries - (producer - q->cons_tail);
}

/* UMEM queue */

static inline bool xskq_is_valid_addr(struct xsk_queue *q, u64 addr)
{
	if (addr >= q->size) {
		q->invalid_descs++;
		return false;
	}

	return true;
}

static inline u64 *xskq_validate_addr(struct xsk_queue *q, u64 *addr)
{
	while (q->cons_tail != q->cons_head) {
		struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;
		unsigned int idx = q->cons_tail & q->ring_mask;

		*addr = READ_ONCE(ring->desc[idx]) & q->chunk_mask;
		if (xskq_is_valid_addr(q, *addr))
			return addr;

		q->cons_tail++;
	}

	return NULL;
}

static inline u64 *xskq_peek_addr(struct xsk_queue *q, u64 *addr)
{
	if (q->cons_tail == q->cons_head) {
		smp_mb(); /* D, matches A */
		WRITE_ONCE(q->ring->consumer, q->cons_tail);
		q->cons_head = q->cons_tail + xskq_nb_avail(q, RX_BATCH_SIZE);

		/* Order consumer and data */
		smp_rmb();
	}

	return xskq_validate_addr(q, addr);
}

static inline void xskq_discard_addr(struct xsk_queue *q)
{
	q->cons_tail++;
}

static inline int xskq_produce_addr(struct xsk_queue *q, u64 addr)
{
	struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;

	if (xskq_nb_free(q, q->prod_tail, 1) == 0)
		return -ENOSPC;

	/* A, matches D */
	ring->desc[q->prod_tail++ & q->ring_mask] = addr;

	/* Order producer and data */
	smp_wmb(); /* B, matches C */

	WRITE_ONCE(q->ring->producer, q->prod_tail);
	return 0;
}

static inline int xskq_produce_addr_lazy(struct xsk_queue *q, u64 addr)
{
	struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring;

	if (xskq_nb_free(q, q->prod_head, LAZY_UPDATE_THRESHOLD) == 0)
		return -ENOSPC;

	/* A, matches D */
	ring->desc[q->prod_head++ & q->ring_mask] = addr;
	return 0;
}

static inline void xskq_produce_flush_addr_n(struct xsk_queue *q,
					     u32 nb_entries)
{
	/* Order producer and data */
	smp_wmb(); /* B, matches C */

	q->prod_tail += nb_entries;
	WRITE_ONCE(q->ring->producer, q->prod_tail);
}

static inline int xskq_reserve_addr(struct xsk_queue *q)
{
	if (xskq_nb_free(q, q->prod_head, 1) == 0)
		return -ENOSPC;

	/* A, matches D */
	q->prod_head++;
	return 0;
}

/* Rx/Tx queue */

static inline bool xskq_is_valid_desc(struct xsk_queue *q, struct xdp_desc *d)
{
	if (!xskq_is_valid_addr(q, d->addr))
		return false;

	if (((d->addr + d->len) & q->chunk_mask) != (d->addr & q->chunk_mask) ||
	    d->options) {
		q->invalid_descs++;
		return false;
	}

	return true;
}

static inline struct xdp_desc *xskq_validate_desc(struct xsk_queue *q,
						  struct xdp_desc *desc)
{
	while (q->cons_tail != q->cons_head) {
		struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring;
		unsigned int idx = q->cons_tail & q->ring_mask;

		*desc = READ_ONCE(ring->desc[idx]);
		if (xskq_is_valid_desc(q, desc))
			return desc;

		q->cons_tail++;
	}

	return NULL;
}

static inline struct xdp_desc *xskq_peek_desc(struct xsk_queue *q,
					      struct xdp_desc *desc)
{
	if (q->cons_tail == q->cons_head) {
		smp_mb(); /* D, matches A */
		WRITE_ONCE(q->ring->consumer, q->cons_tail);
		q->cons_head = q->cons_tail + xskq_nb_avail(q, RX_BATCH_SIZE);

		/* Order consumer and data */
		smp_rmb(); /* C, matches B */
	}

	return xskq_validate_desc(q, desc);
}

static inline void xskq_discard_desc(struct xsk_queue *q)
{
	q->cons_tail++;
}

static inline int xskq_produce_batch_desc(struct xsk_queue *q,
					  u64 addr, u32 len)
{
	struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring;
	unsigned int idx;

	if (xskq_nb_free(q, q->prod_head, 1) == 0)
		return -ENOSPC;

	/* A, matches D */
	idx = (q->prod_head++) & q->ring_mask;
	ring->desc[idx].addr = addr;
	ring->desc[idx].len = len;

	return 0;
}

static inline void xskq_produce_flush_desc(struct xsk_queue *q)
{
	/* Order producer and data */
	smp_wmb(); /* B, matches C */

	q->prod_tail = q->prod_head,
	WRITE_ONCE(q->ring->producer, q->prod_tail);
}

static inline bool xskq_full_desc(struct xsk_queue *q)
{
	return xskq_nb_avail(q, q->nentries) == q->nentries;
}

static inline bool xskq_empty_desc(struct xsk_queue *q)
{
	return xskq_nb_free(q, q->prod_tail, q->nentries) == q->nentries;
}

void xskq_set_umem(struct xsk_queue *q, u64 size, u64 chunk_mask);
struct xsk_queue *xskq_create(u32 nentries, bool umem_queue);
void xskq_destroy(struct xsk_queue *q_ops);

/* Executed by the core when the entire UMEM gets freed */
void xsk_reuseq_destroy(struct xdp_umem *umem);

#endif /* _LINUX_XSK_QUEUE_H */
Commit	Line	Data
dac09149 BT	1	/* SPDX-License-Identifier: GPL-2.0 */
dac09149 BT	2	/* XDP user-space ring structure
423f3832	3	* Copyright(c) 2018 Intel Corporation.
423f3832 MK	4	*/
	5
	6	#ifndef _LINUX_XSK_QUEUE_H
	7	#define _LINUX_XSK_QUEUE_H
	8
	9	#include <linux/types.h>
	10	#include <linux/if_xdp.h>
e61e62b9	11	#include <net/xdp_sock.h>
423f3832	12
c497176c	13	#define RX_BATCH_SIZE 16
ac98d8aa	14	#define LAZY_UPDATE_THRESHOLD 128
c497176c	15
b3a9e0be BT	16	struct xdp_ring {
	17	u32 producer ____cacheline_aligned_in_smp;
	18	u32 consumer ____cacheline_aligned_in_smp;
	19	};
	20
	21	/* Used for the RX and TX queues for packets */
	22	struct xdp_rxtx_ring {
	23	struct xdp_ring ptrs;
	24	struct xdp_desc desc[0] ____cacheline_aligned_in_smp;
	25	};
	26
	27	/* Used for the fill and completion queues for buffers */
	28	struct xdp_umem_ring {
	29	struct xdp_ring ptrs;
bbff2f32	30	u64 desc[0] ____cacheline_aligned_in_smp;
b3a9e0be BT	31	};
b3a9e0be BT	32
423f3832	33	struct xsk_queue {
93ee30f3 MK	34	u64 chunk_mask;
93ee30f3 MK	35	u64 size;
423f3832 MK	36	u32 ring_mask;
	37	u32 nentries;
	38	u32 prod_head;
	39	u32 prod_tail;
	40	u32 cons_head;
	41	u32 cons_tail;
	42	struct xdp_ring *ring;
	43	u64 invalid_descs;
	44	};
	45
f63666de MK	46	/* The structure of the shared state of the rings are the same as the
	47	* ring buffer in kernel/events/ring_buffer.c. For the Rx and completion
	48	* ring, the kernel is the producer and user space is the consumer. For
	49	* the Tx and fill rings, the kernel is the consumer and user space is
	50	* the producer.
	51	*
	52	* producer consumer
	53	*
	54	* if (LOAD ->consumer) { LOAD ->producer
	55	* (A) smp_rmb() (C)
	56	* STORE $data LOAD $data
	57	* smp_wmb() (B) smp_mb() (D)
	58	* STORE ->producer STORE ->consumer
	59	* }
	60	*
	61	* (A) pairs with (D), and (B) pairs with (C).
	62	*
	63	* Starting with (B), it protects the data from being written after
	64	* the producer pointer. If this barrier was missing, the consumer
	65	* could observe the producer pointer being set and thus load the data
	66	* before the producer has written the new data. The consumer would in
	67	* this case load the old data.
	68	*
	69	* (C) protects the consumer from speculatively loading the data before
	70	* the producer pointer actually has been read. If we do not have this
	71	* barrier, some architectures could load old data as speculative loads
	72	* are not discarded as the CPU does not know there is a dependency
	73	* between ->producer and data.
	74	*
	75	* (A) is a control dependency that separates the load of ->consumer
	76	* from the stores of $data. In case ->consumer indicates there is no
	77	* room in the buffer to store $data we do not. So no barrier is needed.
	78	*
	79	* (D) protects the load of the data to be observed to happen after the
	80	* store of the consumer pointer. If we did not have this memory
	81	* barrier, the producer could observe the consumer pointer being set
	82	* and overwrite the data with a new value before the consumer got the
	83	* chance to read the old value. The consumer would thus miss reading
	84	* the old entry and very likely read the new entry twice, once right
	85	* now and again after circling through the ring.
	86	*/
	87
c497176c BT	88	/* Common functions operating for both RXTX and umem queues */
c497176c BT	89
af75d9e0 MK	90	static inline u64 xskq_nb_invalid_descs(struct xsk_queue *q)
	91	{
	92	return q ? q->invalid_descs : 0;
	93	}
	94
c497176c BT	95	static inline u32 xskq_nb_avail(struct xsk_queue *q, u32 dcnt)
	96	{
	97	u32 entries = q->prod_tail - q->cons_tail;
	98
	99	if (entries == 0) {
	100	/* Refresh the local pointer */
	101	q->prod_tail = READ_ONCE(q->ring->producer);
	102	entries = q->prod_tail - q->cons_tail;
	103	}
	104
	105	return (entries > dcnt) ? dcnt : entries;
	106	}
	107
	108	static inline u32 xskq_nb_free(struct xsk_queue *q, u32 producer, u32 dcnt)
	109	{
20b52a75	110	u32 free_entries = q->nentries - (producer - q->cons_tail);
c497176c BT	111
	112	if (free_entries >= dcnt)
	113	return free_entries;
	114
	115	/* Refresh the local tail pointer */
	116	q->cons_tail = READ_ONCE(q->ring->consumer);
	117	return q->nentries - (producer - q->cons_tail);
	118	}
	119
	120	/* UMEM queue */
	121
bbff2f32	122	static inline bool xskq_is_valid_addr(struct xsk_queue *q, u64 addr)
c497176c	123	{
93ee30f3	124	if (addr >= q->size) {
c497176c BT	125	q->invalid_descs++;
	126	return false;
	127	}
bbff2f32	128
c497176c BT	129	return true;
	130	}
	131
bbff2f32	132	static inline u64 xskq_validate_addr(struct xsk_queue q, u64 *addr)
c497176c BT	133	{
	134	while (q->cons_tail != q->cons_head) {
	135	struct xdp_umem_ring ring = (struct xdp_umem_ring )q->ring;
	136	unsigned int idx = q->cons_tail & q->ring_mask;
	137
93ee30f3	138	*addr = READ_ONCE(ring->desc[idx]) & q->chunk_mask;
bbff2f32 BT	139	if (xskq_is_valid_addr(q, *addr))
bbff2f32 BT	140	return addr;
c497176c BT	141
	142	q->cons_tail++;
	143	}
	144
	145	return NULL;
	146	}
	147
bbff2f32	148	static inline u64 xskq_peek_addr(struct xsk_queue q, u64 *addr)
c497176c	149	{
c497176c	150	if (q->cons_tail == q->cons_head) {
f63666de	151	smp_mb(); /* D, matches A */
c497176c BT	152	WRITE_ONCE(q->ring->consumer, q->cons_tail);
	153	q->cons_head = q->cons_tail + xskq_nb_avail(q, RX_BATCH_SIZE);
	154
	155	/* Order consumer and data */
	156	smp_rmb();
c497176c BT	157	}
c497176c BT	158
bbff2f32	159	return xskq_validate_addr(q, addr);
c497176c BT	160	}
c497176c BT	161
bbff2f32	162	static inline void xskq_discard_addr(struct xsk_queue *q)
c497176c BT	163	{
c497176c BT	164	q->cons_tail++;
c497176c BT	165	}
c497176c BT	166
bbff2f32	167	static inline int xskq_produce_addr(struct xsk_queue *q, u64 addr)
35fcde7f MK	168	{
	169	struct xdp_umem_ring ring = (struct xdp_umem_ring )q->ring;
	170
20b52a75	171	if (xskq_nb_free(q, q->prod_tail, 1) == 0)
ac98d8aa MK	172	return -ENOSPC;
ac98d8aa MK	173
f63666de	174	/* A, matches D */
bbff2f32	175	ring->desc[q->prod_tail++ & q->ring_mask] = addr;
35fcde7f MK	176
35fcde7f MK	177	/* Order producer and data */
f63666de	178	smp_wmb(); /* B, matches C */
35fcde7f MK	179
	180	WRITE_ONCE(q->ring->producer, q->prod_tail);
	181	return 0;
	182	}
	183
ac98d8aa MK	184	static inline int xskq_produce_addr_lazy(struct xsk_queue *q, u64 addr)
	185	{
	186	struct xdp_umem_ring ring = (struct xdp_umem_ring )q->ring;
	187
	188	if (xskq_nb_free(q, q->prod_head, LAZY_UPDATE_THRESHOLD) == 0)
	189	return -ENOSPC;
	190
f63666de	191	/* A, matches D */
ac98d8aa MK	192	ring->desc[q->prod_head++ & q->ring_mask] = addr;
	193	return 0;
	194	}
	195
	196	static inline void xskq_produce_flush_addr_n(struct xsk_queue *q,
	197	u32 nb_entries)
	198	{
	199	/* Order producer and data */
f63666de	200	smp_wmb(); /* B, matches C */
ac98d8aa MK	201
	202	q->prod_tail += nb_entries;
	203	WRITE_ONCE(q->ring->producer, q->prod_tail);
	204	}
	205
bbff2f32	206	static inline int xskq_reserve_addr(struct xsk_queue *q)
35fcde7f MK	207	{
	208	if (xskq_nb_free(q, q->prod_head, 1) == 0)
	209	return -ENOSPC;
	210
f63666de	211	/* A, matches D */
35fcde7f MK	212	q->prod_head++;
	213	return 0;
	214	}
	215
	216	/* Rx/Tx queue */
	217
	218	static inline bool xskq_is_valid_desc(struct xsk_queue q, struct xdp_desc d)
	219	{
bbff2f32	220	if (!xskq_is_valid_addr(q, d->addr))
35fcde7f	221	return false;
35fcde7f	222
c57b557b BT	223	if (((d->addr + d->len) & q->chunk_mask) != (d->addr & q->chunk_mask) \|\|
c57b557b BT	224	d->options) {
35fcde7f MK	225	q->invalid_descs++;
	226	return false;
	227	}
	228
	229	return true;
	230	}
	231
	232	static inline struct xdp_desc xskq_validate_desc(struct xsk_queue q,
	233	struct xdp_desc *desc)
	234	{
	235	while (q->cons_tail != q->cons_head) {
	236	struct xdp_rxtx_ring ring = (struct xdp_rxtx_ring )q->ring;
	237	unsigned int idx = q->cons_tail & q->ring_mask;
	238
4e64c835 BT	239	*desc = READ_ONCE(ring->desc[idx]);
4e64c835 BT	240	if (xskq_is_valid_desc(q, desc))
35fcde7f	241	return desc;
35fcde7f MK	242
	243	q->cons_tail++;
	244	}
	245
	246	return NULL;
	247	}
	248
	249	static inline struct xdp_desc xskq_peek_desc(struct xsk_queue q,
	250	struct xdp_desc *desc)
	251	{
35fcde7f	252	if (q->cons_tail == q->cons_head) {
f63666de	253	smp_mb(); /* D, matches A */
35fcde7f MK	254	WRITE_ONCE(q->ring->consumer, q->cons_tail);
	255	q->cons_head = q->cons_tail + xskq_nb_avail(q, RX_BATCH_SIZE);
	256
	257	/* Order consumer and data */
f63666de	258	smp_rmb(); /* C, matches B */
35fcde7f MK	259	}
35fcde7f MK	260
4e64c835	261	return xskq_validate_desc(q, desc);
35fcde7f MK	262	}
	263
	264	static inline void xskq_discard_desc(struct xsk_queue *q)
	265	{
	266	q->cons_tail++;
35fcde7f	267	}
c497176c BT	268
c497176c BT	269	static inline int xskq_produce_batch_desc(struct xsk_queue *q,
bbff2f32	270	u64 addr, u32 len)
c497176c BT	271	{
	272	struct xdp_rxtx_ring ring = (struct xdp_rxtx_ring )q->ring;
	273	unsigned int idx;
	274
	275	if (xskq_nb_free(q, q->prod_head, 1) == 0)
	276	return -ENOSPC;
	277
f63666de	278	/* A, matches D */
c497176c	279	idx = (q->prod_head++) & q->ring_mask;
bbff2f32	280	ring->desc[idx].addr = addr;
c497176c	281	ring->desc[idx].len = len;
c497176c BT	282
	283	return 0;
	284	}
	285
	286	static inline void xskq_produce_flush_desc(struct xsk_queue *q)
	287	{
	288	/* Order producer and data */
f63666de	289	smp_wmb(); /* B, matches C */
c497176c BT	290
	291	q->prod_tail = q->prod_head,
	292	WRITE_ONCE(q->ring->producer, q->prod_tail);
	293	}
	294
35fcde7f MK	295	static inline bool xskq_full_desc(struct xsk_queue *q)
35fcde7f MK	296	{
da60cf00	297	return xskq_nb_avail(q, q->nentries) == q->nentries;
35fcde7f MK	298	}
35fcde7f MK	299
c497176c BT	300	static inline bool xskq_empty_desc(struct xsk_queue *q)
c497176c BT	301	{
d24458e4	302	return xskq_nb_free(q, q->prod_tail, q->nentries) == q->nentries;
c497176c BT	303	}
c497176c BT	304
93ee30f3	305	void xskq_set_umem(struct xsk_queue *q, u64 size, u64 chunk_mask);
b9b6b68e	306	struct xsk_queue *xskq_create(u32 nentries, bool umem_queue);
c497176c	307	void xskq_destroy(struct xsk_queue *q_ops);
423f3832	308
f5bd9138 JK	309	/* Executed by the core when the entire UMEM gets freed */
	310	void xsk_reuseq_destroy(struct xdp_umem *umem);
	311
423f3832	312	#endif /* _LINUX_XSK_QUEUE_H */