[linux-2.6-block.git] / net / rds / ib.h

/* SPDX-License-Identifier: GPL-2.0 */
#ifndef _RDS_IB_H
#define _RDS_IB_H

#include <rdma/ib_verbs.h>
#include <rdma/rdma_cm.h>
#include <linux/interrupt.h>
#include <linux/pci.h>
#include <linux/slab.h>
#include "rds.h"
#include "rdma_transport.h"

#define RDS_IB_MAX_SGE			8
#define RDS_IB_RECV_SGE 		2

#define RDS_IB_DEFAULT_RECV_WR		1024
#define RDS_IB_DEFAULT_SEND_WR		256
#define RDS_IB_DEFAULT_FR_WR		512

#define RDS_IB_DEFAULT_RETRY_COUNT	1

#define RDS_IB_SUPPORTED_PROTOCOLS	0x00000003	/* minor versions supported */

#define RDS_IB_RECYCLE_BATCH_COUNT	32

#define RDS_IB_WC_MAX			32

extern struct rw_semaphore rds_ib_devices_lock;
extern struct list_head rds_ib_devices;

/*
 * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
 * try and minimize the amount of memory tied up both the device and
 * socket receive queues.
 */
struct rds_page_frag {
	struct list_head	f_item;
	struct list_head	f_cache_entry;
	struct scatterlist	f_sg;
};

struct rds_ib_incoming {
	struct list_head	ii_frags;
	struct list_head	ii_cache_entry;
	struct rds_incoming	ii_inc;
};

struct rds_ib_cache_head {
	struct list_head *first;
	unsigned long count;
};

struct rds_ib_refill_cache {
	struct rds_ib_cache_head __percpu *percpu;
	struct list_head	 *xfer;
	struct list_head	 *ready;
};

/* This is the common structure for the IB private data exchange in setting up
 * an RDS connection.  The exchange is different for IPv4 and IPv6 connections.
 * The reason is that the address size is different and the addresses
 * exchanged are in the beginning of the structure.  Hence it is not possible
 * for interoperability if same structure is used.
 */
struct rds_ib_conn_priv_cmn {
	u8			ricpc_protocol_major;
	u8			ricpc_protocol_minor;
	__be16			ricpc_protocol_minor_mask;	/* bitmask */
	u8			ricpc_dp_toss;
	u8			ripc_reserved1;
	__be16			ripc_reserved2;
	__be64			ricpc_ack_seq;
	__be32			ricpc_credit;	/* non-zero enables flow ctl */
};

struct rds_ib_connect_private {
	/* Add new fields at the end, and don't permute existing fields. */
	__be32				dp_saddr;
	__be32				dp_daddr;
	struct rds_ib_conn_priv_cmn	dp_cmn;
};

struct rds6_ib_connect_private {
	/* Add new fields at the end, and don't permute existing fields. */
	struct in6_addr			dp_saddr;
	struct in6_addr			dp_daddr;
	struct rds_ib_conn_priv_cmn	dp_cmn;
};

#define dp_protocol_major	dp_cmn.ricpc_protocol_major
#define dp_protocol_minor	dp_cmn.ricpc_protocol_minor
#define dp_protocol_minor_mask	dp_cmn.ricpc_protocol_minor_mask
#define dp_ack_seq		dp_cmn.ricpc_ack_seq
#define dp_credit		dp_cmn.ricpc_credit

union rds_ib_conn_priv {
	struct rds_ib_connect_private	ricp_v4;
	struct rds6_ib_connect_private	ricp_v6;
};

struct rds_ib_send_work {
	void			*s_op;
	union {
		struct ib_send_wr	s_wr;
		struct ib_rdma_wr	s_rdma_wr;
		struct ib_atomic_wr	s_atomic_wr;
	};
	struct ib_sge		s_sge[RDS_IB_MAX_SGE];
	unsigned long		s_queued;
};

struct rds_ib_recv_work {
	struct rds_ib_incoming 	*r_ibinc;
	struct rds_page_frag	*r_frag;
	struct ib_recv_wr	r_wr;
	struct ib_sge		r_sge[2];
};

struct rds_ib_work_ring {
	u32		w_nr;
	u32		w_alloc_ptr;
	u32		w_alloc_ctr;
	u32		w_free_ptr;
	atomic_t	w_free_ctr;
};

/* Rings are posted with all the allocations they'll need to queue the
 * incoming message to the receiving socket so this can't fail.
 * All fragments start with a header, so we can make sure we're not receiving
 * garbage, and we can tell a small 8 byte fragment from an ACK frame.
 */
struct rds_ib_ack_state {
	u64		ack_next;
	u64		ack_recv;
	unsigned int	ack_required:1;
	unsigned int	ack_next_valid:1;
	unsigned int	ack_recv_valid:1;
};


struct rds_ib_device;

struct rds_ib_connection {

	struct list_head	ib_node;
	struct rds_ib_device	*rds_ibdev;
	struct rds_connection	*conn;

	/* alphabet soup, IBTA style */
	struct rdma_cm_id	*i_cm_id;
	struct ib_pd		*i_pd;
	struct ib_cq		*i_send_cq;
	struct ib_cq		*i_recv_cq;
	struct ib_wc		i_send_wc[RDS_IB_WC_MAX];
	struct ib_wc		i_recv_wc[RDS_IB_WC_MAX];

	/* To control the number of wrs from fastreg */
	atomic_t		i_fastreg_wrs;
	atomic_t		i_fastreg_inuse_count;

	/* interrupt handling */
	struct tasklet_struct	i_send_tasklet;
	struct tasklet_struct	i_recv_tasklet;

	/* tx */
	struct rds_ib_work_ring	i_send_ring;
	struct rm_data_op	*i_data_op;
	struct rds_header	*i_send_hdrs;
	dma_addr_t		i_send_hdrs_dma;
	struct rds_ib_send_work *i_sends;
	atomic_t		i_signaled_sends;

	/* rx */
	struct mutex		i_recv_mutex;
	struct rds_ib_work_ring	i_recv_ring;
	struct rds_ib_incoming	*i_ibinc;
	u32			i_recv_data_rem;
	struct rds_header	*i_recv_hdrs;
	dma_addr_t		i_recv_hdrs_dma;
	struct rds_ib_recv_work *i_recvs;
	u64			i_ack_recv;	/* last ACK received */
	struct rds_ib_refill_cache i_cache_incs;
	struct rds_ib_refill_cache i_cache_frags;
	atomic_t		i_cache_allocs;

	/* sending acks */
	unsigned long		i_ack_flags;
#ifdef KERNEL_HAS_ATOMIC64
	atomic64_t		i_ack_next;	/* next ACK to send */
#else
	spinlock_t		i_ack_lock;	/* protect i_ack_next */
	u64			i_ack_next;	/* next ACK to send */
#endif
	struct rds_header	*i_ack;
	struct ib_send_wr	i_ack_wr;
	struct ib_sge		i_ack_sge;
	dma_addr_t		i_ack_dma;
	unsigned long		i_ack_queued;

	/* Flow control related information
	 *
	 * Our algorithm uses a pair variables that we need to access
	 * atomically - one for the send credits, and one posted
	 * recv credits we need to transfer to remote.
	 * Rather than protect them using a slow spinlock, we put both into
	 * a single atomic_t and update it using cmpxchg
	 */
	atomic_t		i_credits;

	/* Protocol version specific information */
	unsigned int		i_flowctl:1;	/* enable/disable flow ctl */

	/* Batched completions */
	unsigned int		i_unsignaled_wrs;

	/* Endpoint role in connection */
	bool			i_active_side;
	atomic_t		i_cq_quiesce;

	/* Send/Recv vectors */
	int			i_scq_vector;
	int			i_rcq_vector;
	u8			i_sl;
};

/* This assumes that atomic_t is at least 32 bits */
#define IB_GET_SEND_CREDITS(v)	((v) & 0xffff)
#define IB_GET_POST_CREDITS(v)	((v) >> 16)
#define IB_SET_SEND_CREDITS(v)	((v) & 0xffff)
#define IB_SET_POST_CREDITS(v)	((v) << 16)

struct rds_ib_ipaddr {
	struct list_head	list;
	__be32			ipaddr;
	struct rcu_head		rcu;
};

enum {
	RDS_IB_MR_8K_POOL,
	RDS_IB_MR_1M_POOL,
};

struct rds_ib_device {
	struct list_head	list;
	struct list_head	ipaddr_list;
	struct list_head	conn_list;
	struct ib_device	*dev;
	struct ib_pd		*pd;
	bool                    use_fastreg;

	unsigned int		max_mrs;
	struct rds_ib_mr_pool	*mr_1m_pool;
	struct rds_ib_mr_pool   *mr_8k_pool;
	unsigned int		fmr_max_remaps;
	unsigned int		max_8k_mrs;
	unsigned int		max_1m_mrs;
	int			max_sge;
	unsigned int		max_wrs;
	unsigned int		max_initiator_depth;
	unsigned int		max_responder_resources;
	spinlock_t		spinlock;	/* protect the above */
	refcount_t		refcount;
	struct work_struct	free_work;
	int			*vector_load;
};

#define ibdev_to_node(ibdev) dev_to_node((ibdev)->dev.parent)
#define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev)

/* bits for i_ack_flags */
#define IB_ACK_IN_FLIGHT	0
#define IB_ACK_REQUESTED	1

/* Magic WR_ID for ACKs */
#define RDS_IB_ACK_WR_ID	(~(u64) 0)

struct rds_ib_statistics {
	uint64_t	s_ib_connect_raced;
	uint64_t	s_ib_listen_closed_stale;
	uint64_t	s_ib_evt_handler_call;
	uint64_t	s_ib_tasklet_call;
	uint64_t	s_ib_tx_cq_event;
	uint64_t	s_ib_tx_ring_full;
	uint64_t	s_ib_tx_throttle;
	uint64_t	s_ib_tx_sg_mapping_failure;
	uint64_t	s_ib_tx_stalled;
	uint64_t	s_ib_tx_credit_updates;
	uint64_t	s_ib_rx_cq_event;
	uint64_t	s_ib_rx_ring_empty;
	uint64_t	s_ib_rx_refill_from_cq;
	uint64_t	s_ib_rx_refill_from_thread;
	uint64_t	s_ib_rx_alloc_limit;
	uint64_t	s_ib_rx_total_frags;
	uint64_t	s_ib_rx_total_incs;
	uint64_t	s_ib_rx_credit_updates;
	uint64_t	s_ib_ack_sent;
	uint64_t	s_ib_ack_send_failure;
	uint64_t	s_ib_ack_send_delayed;
	uint64_t	s_ib_ack_send_piggybacked;
	uint64_t	s_ib_ack_received;
	uint64_t	s_ib_rdma_mr_8k_alloc;
	uint64_t	s_ib_rdma_mr_8k_free;
	uint64_t	s_ib_rdma_mr_8k_used;
	uint64_t	s_ib_rdma_mr_8k_pool_flush;
	uint64_t	s_ib_rdma_mr_8k_pool_wait;
	uint64_t	s_ib_rdma_mr_8k_pool_depleted;
	uint64_t	s_ib_rdma_mr_1m_alloc;
	uint64_t	s_ib_rdma_mr_1m_free;
	uint64_t	s_ib_rdma_mr_1m_used;
	uint64_t	s_ib_rdma_mr_1m_pool_flush;
	uint64_t	s_ib_rdma_mr_1m_pool_wait;
	uint64_t	s_ib_rdma_mr_1m_pool_depleted;
	uint64_t	s_ib_rdma_mr_8k_reused;
	uint64_t	s_ib_rdma_mr_1m_reused;
	uint64_t	s_ib_atomic_cswp;
	uint64_t	s_ib_atomic_fadd;
	uint64_t	s_ib_recv_added_to_cache;
	uint64_t	s_ib_recv_removed_from_cache;
};

extern struct workqueue_struct *rds_ib_wq;

/*
 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
 * doesn't define it.
 */
static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev,
					      struct scatterlist *sglist,
					      unsigned int sg_dma_len,
					      int direction)
{
	struct scatterlist *sg;
	unsigned int i;

	for_each_sg(sglist, sg, sg_dma_len, i) {
		ib_dma_sync_single_for_cpu(dev, sg_dma_address(sg),
					   sg_dma_len(sg), direction);
	}
}
#define ib_dma_sync_sg_for_cpu	rds_ib_dma_sync_sg_for_cpu

static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
						 struct scatterlist *sglist,
						 unsigned int sg_dma_len,
						 int direction)
{
	struct scatterlist *sg;
	unsigned int i;

	for_each_sg(sglist, sg, sg_dma_len, i) {
		ib_dma_sync_single_for_device(dev, sg_dma_address(sg),
					      sg_dma_len(sg), direction);
	}
}
#define ib_dma_sync_sg_for_device	rds_ib_dma_sync_sg_for_device


/* ib.c */
extern struct rds_transport rds_ib_transport;
struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device);
void rds_ib_dev_put(struct rds_ib_device *rds_ibdev);
extern struct ib_client rds_ib_client;

extern unsigned int rds_ib_retry_count;

extern spinlock_t ib_nodev_conns_lock;
extern struct list_head ib_nodev_conns;

/* ib_cm.c */
int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp);
void rds_ib_conn_free(void *arg);
int rds_ib_conn_path_connect(struct rds_conn_path *cp);
void rds_ib_conn_path_shutdown(struct rds_conn_path *cp);
void rds_ib_state_change(struct sock *sk);
int rds_ib_listen_init(void);
void rds_ib_listen_stop(void);
__printf(2, 3)
void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...);
int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
			     struct rdma_cm_event *event, bool isv6);
int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6);
void rds_ib_cm_connect_complete(struct rds_connection *conn,
				struct rdma_cm_event *event);


#define rds_ib_conn_error(conn, fmt...) \
	__rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)

/* ib_rdma.c */
int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev,
			 struct in6_addr *ipaddr);
void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
void rds_ib_destroy_nodev_conns(void);
void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc);

/* ib_recv.c */
int rds_ib_recv_init(void);
void rds_ib_recv_exit(void);
int rds_ib_recv_path(struct rds_conn_path *conn);
int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic, gfp_t gfp);
void rds_ib_recv_free_caches(struct rds_ib_connection *ic);
void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp);
void rds_ib_inc_free(struct rds_incoming *inc);
int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc,
			     struct rds_ib_ack_state *state);
void rds_ib_recv_tasklet_fn(unsigned long data);
void rds_ib_recv_init_ring(struct rds_ib_connection *ic);
void rds_ib_recv_clear_ring(struct rds_ib_connection *ic);
void rds_ib_recv_init_ack(struct rds_ib_connection *ic);
void rds_ib_attempt_ack(struct rds_ib_connection *ic);
void rds_ib_ack_send_complete(struct rds_ib_connection *ic);
u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic);
void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required);

/* ib_ring.c */
void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr);
void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr);
u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos);
void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val);
void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val);
int rds_ib_ring_empty(struct rds_ib_work_ring *ring);
int rds_ib_ring_low(struct rds_ib_work_ring *ring);
u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring);
u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
extern wait_queue_head_t rds_ib_ring_empty_wait;

/* ib_send.c */
void rds_ib_xmit_path_complete(struct rds_conn_path *cp);
int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
		unsigned int hdr_off, unsigned int sg, unsigned int off);
void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc);
void rds_ib_send_init_ring(struct rds_ib_connection *ic);
void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
			     u32 *adv_credits, int need_posted, int max_posted);
int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op);

/* ib_stats.c */
DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_ib_statistics, rds_ib_stats);
#define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
#define rds_ib_stats_add(member, count) \
		rds_stats_add_which(rds_ib_stats, member, count)
unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
				    unsigned int avail);

/* ib_sysctl.c */
int rds_ib_sysctl_init(void);
void rds_ib_sysctl_exit(void);
extern unsigned long rds_ib_sysctl_max_send_wr;
extern unsigned long rds_ib_sysctl_max_recv_wr;
extern unsigned long rds_ib_sysctl_max_unsig_wrs;
extern unsigned long rds_ib_sysctl_max_unsig_bytes;
extern unsigned long rds_ib_sysctl_max_recv_allocation;
extern unsigned int rds_ib_sysctl_flow_control;

#endif
Commit	Line	Data
b2441318	1	/* SPDX-License-Identifier: GPL-2.0 */
ec16227e AG	2	#ifndef _RDS_IB_H
	3	#define _RDS_IB_H
	4
	5	#include <rdma/ib_verbs.h>
	6	#include <rdma/rdma_cm.h>
a6b7a407	7	#include <linux/interrupt.h>
e4c52c98 AG	8	#include <linux/pci.h>
e4c52c98 AG	9	#include <linux/slab.h>
ec16227e AG	10	#include "rds.h"
	11	#include "rdma_transport.h"
	12
ec16227e AG	13	#define RDS_IB_MAX_SGE 8
	14	#define RDS_IB_RECV_SGE 2
	15
	16	#define RDS_IB_DEFAULT_RECV_WR 1024
	17	#define RDS_IB_DEFAULT_SEND_WR 256
a5520788	18	#define RDS_IB_DEFAULT_FR_WR 512
ec16227e	19
fab8688d	20	#define RDS_IB_DEFAULT_RETRY_COUNT 1
3ba23ade	21
ec16227e AG	22	#define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */
ec16227e AG	23
33244125 CM	24	#define RDS_IB_RECYCLE_BATCH_COUNT 32
33244125 CM	25
f4f943c9 SS	26	#define RDS_IB_WC_MAX 32
f4f943c9 SS	27
ea819867	28	extern struct rw_semaphore rds_ib_devices_lock;
ec16227e AG	29	extern struct list_head rds_ib_devices;
	30
	31	/*
	32	* IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
	33	* try and minimize the amount of memory tied up both the device and
	34	* socket receive queues.
	35	*/
ec16227e AG	36	struct rds_page_frag {
ec16227e AG	37	struct list_head f_item;
33244125	38	struct list_head f_cache_entry;
0b088e00	39	struct scatterlist f_sg;
ec16227e AG	40	};
	41
	42	struct rds_ib_incoming {
	43	struct list_head ii_frags;
33244125	44	struct list_head ii_cache_entry;
ec16227e AG	45	struct rds_incoming ii_inc;
	46	};
	47
33244125 CM	48	struct rds_ib_cache_head {
	49	struct list_head *first;
	50	unsigned long count;
	51	};
	52
	53	struct rds_ib_refill_cache {
ae4b46e9	54	struct rds_ib_cache_head __percpu *percpu;
33244125 CM	55	struct list_head *xfer;
	56	struct list_head *ready;
	57	};
	58
eee2fa6a KCP	59	/* This is the common structure for the IB private data exchange in setting up
	60	* an RDS connection. The exchange is different for IPv4 and IPv6 connections.
	61	* The reason is that the address size is different and the addresses
	62	* exchanged are in the beginning of the structure. Hence it is not possible
	63	* for interoperability if same structure is used.
	64	*/
	65	struct rds_ib_conn_priv_cmn {
	66	u8 ricpc_protocol_major;
	67	u8 ricpc_protocol_minor;
	68	__be16 ricpc_protocol_minor_mask; /* bitmask */
fd261ce6 SS	69	u8 ricpc_dp_toss;
	70	u8 ripc_reserved1;
	71	__be16 ripc_reserved2;
eee2fa6a KCP	72	__be64 ricpc_ack_seq;
	73	__be32 ricpc_credit; /* non-zero enables flow ctl */
	74	};
	75
ec16227e AG	76	struct rds_ib_connect_private {
ec16227e AG	77	/* Add new fields at the end, and don't permute existing fields. */
eee2fa6a KCP	78	__be32 dp_saddr;
	79	__be32 dp_daddr;
	80	struct rds_ib_conn_priv_cmn dp_cmn;
	81	};
	82
	83	struct rds6_ib_connect_private {
	84	/* Add new fields at the end, and don't permute existing fields. */
	85	struct in6_addr dp_saddr;
	86	struct in6_addr dp_daddr;
	87	struct rds_ib_conn_priv_cmn dp_cmn;
	88	};
	89
	90	#define dp_protocol_major dp_cmn.ricpc_protocol_major
	91	#define dp_protocol_minor dp_cmn.ricpc_protocol_minor
	92	#define dp_protocol_minor_mask dp_cmn.ricpc_protocol_minor_mask
	93	#define dp_ack_seq dp_cmn.ricpc_ack_seq
	94	#define dp_credit dp_cmn.ricpc_credit
	95
	96	union rds_ib_conn_priv {
	97	struct rds_ib_connect_private ricp_v4;
	98	struct rds6_ib_connect_private ricp_v6;
ec16227e AG	99	};
	100
	101	struct rds_ib_send_work {
ff3d7d36	102	void *s_op;
e622f2f4 CH	103	union {
	104	struct ib_send_wr s_wr;
	105	struct ib_rdma_wr s_rdma_wr;
	106	struct ib_atomic_wr s_atomic_wr;
	107	};
ec16227e AG	108	struct ib_sge s_sge[RDS_IB_MAX_SGE];
	109	unsigned long s_queued;
	110	};
	111
	112	struct rds_ib_recv_work {
	113	struct rds_ib_incoming *r_ibinc;
	114	struct rds_page_frag *r_frag;
	115	struct ib_recv_wr r_wr;
	116	struct ib_sge r_sge[2];
	117	};
	118
	119	struct rds_ib_work_ring {
	120	u32 w_nr;
	121	u32 w_alloc_ptr;
	122	u32 w_alloc_ctr;
	123	u32 w_free_ptr;
	124	atomic_t w_free_ctr;
	125	};
	126
f4f943c9 SS	127	/* Rings are posted with all the allocations they'll need to queue the
	128	* incoming message to the receiving socket so this can't fail.
	129	* All fragments start with a header, so we can make sure we're not receiving
	130	* garbage, and we can tell a small 8 byte fragment from an ACK frame.
	131	*/
	132	struct rds_ib_ack_state {
	133	u64 ack_next;
	134	u64 ack_recv;
	135	unsigned int ack_required:1;
	136	unsigned int ack_next_valid:1;
	137	unsigned int ack_recv_valid:1;
	138	};
	139
	140
ec16227e AG	141	struct rds_ib_device;
	142
	143	struct rds_ib_connection {
	144
	145	struct list_head ib_node;
	146	struct rds_ib_device *rds_ibdev;
	147	struct rds_connection *conn;
	148
	149	/* alphabet soup, IBTA style */
	150	struct rdma_cm_id *i_cm_id;
	151	struct ib_pd *i_pd;
ec16227e AG	152	struct ib_cq *i_send_cq;
ec16227e AG	153	struct ib_cq *i_recv_cq;
0c28c045	154	struct ib_wc i_send_wc[RDS_IB_WC_MAX];
f4f943c9 SS	155	struct ib_wc i_recv_wc[RDS_IB_WC_MAX];
f4f943c9 SS	156
ad6832f9	157	/* To control the number of wrs from fastreg */
ad6832f9	158	atomic_t i_fastreg_wrs;
3a2886cc	159	atomic_t i_fastreg_inuse_count;
ad6832f9	160
f4f943c9	161	/* interrupt handling */
0c28c045	162	struct tasklet_struct i_send_tasklet;
f4f943c9	163	struct tasklet_struct i_recv_tasklet;
ec16227e AG	164
	165	/* tx */
	166	struct rds_ib_work_ring i_send_ring;
ff3d7d36	167	struct rm_data_op *i_data_op;
ec16227e	168	struct rds_header *i_send_hdrs;
d43dbacf	169	dma_addr_t i_send_hdrs_dma;
ec16227e	170	struct rds_ib_send_work *i_sends;
f046011c	171	atomic_t i_signaled_sends;
ec16227e AG	172
	173	/* rx */
	174	struct mutex i_recv_mutex;
	175	struct rds_ib_work_ring i_recv_ring;
	176	struct rds_ib_incoming *i_ibinc;
	177	u32 i_recv_data_rem;
	178	struct rds_header *i_recv_hdrs;
d43dbacf	179	dma_addr_t i_recv_hdrs_dma;
ec16227e	180	struct rds_ib_recv_work *i_recvs;
ec16227e	181	u64 i_ack_recv; /* last ACK received */
33244125 CM	182	struct rds_ib_refill_cache i_cache_incs;
33244125 CM	183	struct rds_ib_refill_cache i_cache_frags;
09b2b8f5	184	atomic_t i_cache_allocs;
ec16227e AG	185
	186	/* sending acks */
	187	unsigned long i_ack_flags;
8cbd9606 AG	188	#ifdef KERNEL_HAS_ATOMIC64
	189	atomic64_t i_ack_next; /* next ACK to send */
	190	#else
	191	spinlock_t i_ack_lock; /* protect i_ack_next */
ec16227e	192	u64 i_ack_next; /* next ACK to send */
8cbd9606	193	#endif
ec16227e AG	194	struct rds_header *i_ack;
	195	struct ib_send_wr i_ack_wr;
	196	struct ib_sge i_ack_sge;
d43dbacf	197	dma_addr_t i_ack_dma;
ec16227e AG	198	unsigned long i_ack_queued;
	199
	200	/* Flow control related information
	201	*
	202	* Our algorithm uses a pair variables that we need to access
	203	* atomically - one for the send credits, and one posted
	204	* recv credits we need to transfer to remote.
	205	* Rather than protect them using a slow spinlock, we put both into
	206	* a single atomic_t and update it using cmpxchg
	207	*/
	208	atomic_t i_credits;
	209
	210	/* Protocol version specific information */
	211	unsigned int i_flowctl:1; /* enable/disable flow ctl */
	212
	213	/* Batched completions */
	214	unsigned int i_unsignaled_wrs;
581d53c9 SS	215
	216	/* Endpoint role in connection */
	217	bool i_active_side;
cf657269	218	atomic_t i_cq_quiesce;
be2f76ea SS	219
	220	/* Send/Recv vectors */
	221	int i_scq_vector;
	222	int i_rcq_vector;
e0e6d062	223	u8 i_sl;
ec16227e AG	224	};
	225
	226	/* This assumes that atomic_t is at least 32 bits */
	227	#define IB_GET_SEND_CREDITS(v) ((v) & 0xffff)
	228	#define IB_GET_POST_CREDITS(v) ((v) >> 16)
	229	#define IB_SET_SEND_CREDITS(v) ((v) & 0xffff)
	230	#define IB_SET_POST_CREDITS(v) ((v) << 16)
	231
	232	struct rds_ib_ipaddr {
	233	struct list_head list;
	234	__be32 ipaddr;
59fe4606	235	struct rcu_head rcu;
ec16227e AG	236	};
ec16227e AG	237
06766513 SS	238	enum {
	239	RDS_IB_MR_8K_POOL,
	240	RDS_IB_MR_1M_POOL,
	241	};
	242
ec16227e AG	243	struct rds_ib_device {
	244	struct list_head list;
	245	struct list_head ipaddr_list;
	246	struct list_head conn_list;
	247	struct ib_device *dev;
	248	struct ib_pd *pd;
2cb2912d	249	bool use_fastreg;
2cb2912d	250
f6df683f	251	unsigned int max_mrs;
06766513 SS	252	struct rds_ib_mr_pool *mr_1m_pool;
	253	struct rds_ib_mr_pool *mr_8k_pool;
	254	unsigned int fmr_max_remaps;
f6df683f	255	unsigned int max_8k_mrs;
f6df683f	256	unsigned int max_1m_mrs;
ec16227e AG	257	int max_sge;
ec16227e AG	258	unsigned int max_wrs;
40589e74 AG	259	unsigned int max_initiator_depth;
40589e74 AG	260	unsigned int max_responder_resources;
ec16227e	261	spinlock_t spinlock; /* protect the above */
50d61ff7	262	refcount_t refcount;
3e0249f9	263	struct work_struct free_work;
be2f76ea	264	int *vector_load;
ec16227e AG	265	};
ec16227e AG	266
5f68dcaf	267	#define ibdev_to_node(ibdev) dev_to_node((ibdev)->dev.parent)
e4c52c98 AG	268	#define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev)
e4c52c98 AG	269
ec16227e AG	270	/* bits for i_ack_flags */
	271	#define IB_ACK_IN_FLIGHT 0
	272	#define IB_ACK_REQUESTED 1
	273
	274	/* Magic WR_ID for ACKs */
	275	#define RDS_IB_ACK_WR_ID (~(u64) 0)
	276
	277	struct rds_ib_statistics {
	278	uint64_t s_ib_connect_raced;
	279	uint64_t s_ib_listen_closed_stale;
f4f943c9 SS	280	uint64_t s_ib_evt_handler_call;
f4f943c9 SS	281	uint64_t s_ib_tasklet_call;
ec16227e AG	282	uint64_t s_ib_tx_cq_event;
	283	uint64_t s_ib_tx_ring_full;
	284	uint64_t s_ib_tx_throttle;
	285	uint64_t s_ib_tx_sg_mapping_failure;
	286	uint64_t s_ib_tx_stalled;
	287	uint64_t s_ib_tx_credit_updates;
ec16227e AG	288	uint64_t s_ib_rx_cq_event;
	289	uint64_t s_ib_rx_ring_empty;
	290	uint64_t s_ib_rx_refill_from_cq;
	291	uint64_t s_ib_rx_refill_from_thread;
	292	uint64_t s_ib_rx_alloc_limit;
09b2b8f5 SS	293	uint64_t s_ib_rx_total_frags;
09b2b8f5 SS	294	uint64_t s_ib_rx_total_incs;
ec16227e AG	295	uint64_t s_ib_rx_credit_updates;
	296	uint64_t s_ib_ack_sent;
	297	uint64_t s_ib_ack_send_failure;
	298	uint64_t s_ib_ack_send_delayed;
	299	uint64_t s_ib_ack_send_piggybacked;
	300	uint64_t s_ib_ack_received;
06766513 SS	301	uint64_t s_ib_rdma_mr_8k_alloc;
	302	uint64_t s_ib_rdma_mr_8k_free;
	303	uint64_t s_ib_rdma_mr_8k_used;
	304	uint64_t s_ib_rdma_mr_8k_pool_flush;
	305	uint64_t s_ib_rdma_mr_8k_pool_wait;
	306	uint64_t s_ib_rdma_mr_8k_pool_depleted;
	307	uint64_t s_ib_rdma_mr_1m_alloc;
	308	uint64_t s_ib_rdma_mr_1m_free;
	309	uint64_t s_ib_rdma_mr_1m_used;
	310	uint64_t s_ib_rdma_mr_1m_pool_flush;
	311	uint64_t s_ib_rdma_mr_1m_pool_wait;
	312	uint64_t s_ib_rdma_mr_1m_pool_depleted;
db42753a	313	uint64_t s_ib_rdma_mr_8k_reused;
db42753a	314	uint64_t s_ib_rdma_mr_1m_reused;
51e2cba8 AG	315	uint64_t s_ib_atomic_cswp;
51e2cba8 AG	316	uint64_t s_ib_atomic_fadd;
09b2b8f5 SS	317	uint64_t s_ib_recv_added_to_cache;
09b2b8f5 SS	318	uint64_t s_ib_recv_removed_from_cache;
ec16227e AG	319	};
	320
	321	extern struct workqueue_struct *rds_ib_wq;
	322
	323	/*
	324	* Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
	325	* doesn't define it.
	326	*/
	327	static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev,
d2a9ec64 FF	328	struct scatterlist *sglist,
	329	unsigned int sg_dma_len,
	330	int direction)
ec16227e	331	{
d2a9ec64	332	struct scatterlist *sg;
ec16227e AG	333	unsigned int i;
ec16227e AG	334
d2a9ec64	335	for_each_sg(sglist, sg, sg_dma_len, i) {
a163afc8 BVA	336	ib_dma_sync_single_for_cpu(dev, sg_dma_address(sg),
a163afc8 BVA	337	sg_dma_len(sg), direction);
ec16227e AG	338	}
	339	}
	340	#define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu
	341
	342	static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
d2a9ec64 FF	343	struct scatterlist *sglist,
	344	unsigned int sg_dma_len,
	345	int direction)
ec16227e	346	{
d2a9ec64	347	struct scatterlist *sg;
ec16227e AG	348	unsigned int i;
ec16227e AG	349
d2a9ec64	350	for_each_sg(sglist, sg, sg_dma_len, i) {
a163afc8 BVA	351	ib_dma_sync_single_for_device(dev, sg_dma_address(sg),
a163afc8 BVA	352	sg_dma_len(sg), direction);
ec16227e AG	353	}
	354	}
	355	#define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device
	356
	357
	358	/* ib.c */
	359	extern struct rds_transport rds_ib_transport;
3e0249f9 ZB	360	struct rds_ib_device rds_ib_get_client_data(struct ib_device device);
3e0249f9 ZB	361	void rds_ib_dev_put(struct rds_ib_device *rds_ibdev);
ec16227e AG	362	extern struct ib_client rds_ib_client;
ec16227e AG	363
3ba23ade	364	extern unsigned int rds_ib_retry_count;
ec16227e AG	365
	366	extern spinlock_t ib_nodev_conns_lock;
	367	extern struct list_head ib_nodev_conns;
	368
	369	/* ib_cm.c */
	370	int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp);
	371	void rds_ib_conn_free(void *arg);
b04e8554	372	int rds_ib_conn_path_connect(struct rds_conn_path *cp);
226f7a7d	373	void rds_ib_conn_path_shutdown(struct rds_conn_path *cp);
ec16227e	374	void rds_ib_state_change(struct sock *sk);
ef87b7ea	375	int rds_ib_listen_init(void);
ec16227e	376	void rds_ib_listen_stop(void);
6cdaf03f	377	__printf(2, 3)
ec16227e AG	378	void __rds_ib_conn_error(struct rds_connection conn, const char , ...);
ec16227e AG	379	int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
eee2fa6a KCP	380	struct rdma_cm_event *event, bool isv6);
eee2fa6a KCP	381	int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6);
ec16227e AG	382	void rds_ib_cm_connect_complete(struct rds_connection *conn,
	383	struct rdma_cm_event *event);
	384
	385
	386	#define rds_ib_conn_error(conn, fmt...) \
	387	__rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)
	388
	389	/* ib_rdma.c */
eee2fa6a KCP	390	int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev,
eee2fa6a KCP	391	struct in6_addr *ipaddr);
745cbcca AG	392	void rds_ib_add_conn(struct rds_ib_device rds_ibdev, struct rds_connection conn);
745cbcca AG	393	void rds_ib_remove_conn(struct rds_ib_device rds_ibdev, struct rds_connection conn);
8aeb1ba6	394	void rds_ib_destroy_nodev_conns(void);
1659185f	395	void rds_ib_mr_cqe_handler(struct rds_ib_connection ic, struct ib_wc wc);
ec16227e AG	396
ec16227e AG	397	/* ib_recv.c */
ef87b7ea	398	int rds_ib_recv_init(void);
ec16227e	399	void rds_ib_recv_exit(void);
2da43c4a	400	int rds_ib_recv_path(struct rds_conn_path *conn);
f394ad28	401	int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic, gfp_t gfp);
33244125	402	void rds_ib_recv_free_caches(struct rds_ib_connection *ic);
73ce4317	403	void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp);
ec16227e	404	void rds_ib_inc_free(struct rds_incoming *inc);
c310e72c	405	int rds_ib_inc_copy_to_user(struct rds_incoming inc, struct iov_iter to);
f4f943c9 SS	406	void rds_ib_recv_cqe_handler(struct rds_ib_connection ic, struct ib_wc wc,
f4f943c9 SS	407	struct rds_ib_ack_state *state);
d521b63b	408	void rds_ib_recv_tasklet_fn(unsigned long data);
ec16227e AG	409	void rds_ib_recv_init_ring(struct rds_ib_connection *ic);
	410	void rds_ib_recv_clear_ring(struct rds_ib_connection *ic);
	411	void rds_ib_recv_init_ack(struct rds_ib_connection *ic);
	412	void rds_ib_attempt_ack(struct rds_ib_connection *ic);
	413	void rds_ib_ack_send_complete(struct rds_ib_connection *ic);
	414	u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic);
f4f943c9	415	void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required);
ec16227e AG	416
	417	/* ib_ring.c */
	418	void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr);
	419	void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr);
	420	u32 rds_ib_ring_alloc(struct rds_ib_work_ring ring, u32 val, u32 pos);
	421	void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val);
	422	void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val);
	423	int rds_ib_ring_empty(struct rds_ib_work_ring *ring);
	424	int rds_ib_ring_low(struct rds_ib_work_ring *ring);
	425	u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring);
	426	u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
	427	extern wait_queue_head_t rds_ib_ring_empty_wait;
	428
	429	/* ib_send.c */
226f7a7d	430	void rds_ib_xmit_path_complete(struct rds_conn_path *cp);
ec16227e AG	431	int rds_ib_xmit(struct rds_connection conn, struct rds_message rm,
ec16227e AG	432	unsigned int hdr_off, unsigned int sg, unsigned int off);
0c28c045	433	void rds_ib_send_cqe_handler(struct rds_ib_connection ic, struct ib_wc wc);
ec16227e AG	434	void rds_ib_send_init_ring(struct rds_ib_connection *ic);
ec16227e AG	435	void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
f8b3aaf2	436	int rds_ib_xmit_rdma(struct rds_connection conn, struct rm_rdma_op op);
ec16227e AG	437	void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
	438	void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
	439	int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
7b70d033	440	u32 *adv_credits, int need_posted, int max_posted);
ff3d7d36	441	int rds_ib_xmit_atomic(struct rds_connection conn, struct rm_atomic_op op);
ec16227e AG	442
ec16227e AG	443	/* ib_stats.c */
16fdf8ba	444	DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_ib_statistics, rds_ib_stats);
ec16227e	445	#define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
09b2b8f5 SS	446	#define rds_ib_stats_add(member, count) \
09b2b8f5 SS	447	rds_stats_add_which(rds_ib_stats, member, count)
ec16227e AG	448	unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
	449	unsigned int avail);
	450
	451	/* ib_sysctl.c */
ef87b7ea	452	int rds_ib_sysctl_init(void);
ec16227e AG	453	void rds_ib_sysctl_exit(void);
	454	extern unsigned long rds_ib_sysctl_max_send_wr;
	455	extern unsigned long rds_ib_sysctl_max_recv_wr;
	456	extern unsigned long rds_ib_sysctl_max_unsig_wrs;
	457	extern unsigned long rds_ib_sysctl_max_unsig_bytes;
	458	extern unsigned long rds_ib_sysctl_max_recv_allocation;
	459	extern unsigned int rds_ib_sysctl_flow_control;
ec16227e	460
ec16227e	461	#endif