xref: /linux/net/rds/rds.h (revision e58e871becec2d3b04ed91c0c16fe8deac9c9dfa)
1 #ifndef _RDS_RDS_H
2 #define _RDS_RDS_H
3 
4 #include <net/sock.h>
5 #include <linux/scatterlist.h>
6 #include <linux/highmem.h>
7 #include <rdma/rdma_cm.h>
8 #include <linux/mutex.h>
9 #include <linux/rds.h>
10 #include <linux/rhashtable.h>
11 
12 #include "info.h"
13 
14 /*
15  * RDS Network protocol version
16  */
17 #define RDS_PROTOCOL_3_0	0x0300
18 #define RDS_PROTOCOL_3_1	0x0301
19 #define RDS_PROTOCOL_VERSION	RDS_PROTOCOL_3_1
20 #define RDS_PROTOCOL_MAJOR(v)	((v) >> 8)
21 #define RDS_PROTOCOL_MINOR(v)	((v) & 255)
22 #define RDS_PROTOCOL(maj, min)	(((maj) << 8) | min)
23 
24 /*
25  * XXX randomly chosen, but at least seems to be unused:
26  * #               18464-18768 Unassigned
27  * We should do better.  We want a reserved port to discourage unpriv'ed
28  * userspace from listening.
29  */
30 #define RDS_PORT	18634
31 
32 #ifdef ATOMIC64_INIT
33 #define KERNEL_HAS_ATOMIC64
34 #endif
35 
36 #ifdef RDS_DEBUG
37 #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
38 #else
39 /* sigh, pr_debug() causes unused variable warnings */
40 static inline __printf(1, 2)
41 void rdsdebug(char *fmt, ...)
42 {
43 }
44 #endif
45 
46 /* XXX is there one of these somewhere? */
47 #define ceil(x, y) \
48 	({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; })
49 
50 #define RDS_FRAG_SHIFT	12
51 #define RDS_FRAG_SIZE	((unsigned int)(1 << RDS_FRAG_SHIFT))
52 
53 /* Used to limit both RDMA and non-RDMA RDS message to 1MB */
54 #define RDS_MAX_MSG_SIZE	((unsigned int)(1 << 20))
55 
56 #define RDS_CONG_MAP_BYTES	(65536 / 8)
57 #define RDS_CONG_MAP_PAGES	(PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
58 #define RDS_CONG_MAP_PAGE_BITS	(PAGE_SIZE * 8)
59 
60 struct rds_cong_map {
61 	struct rb_node		m_rb_node;
62 	__be32			m_addr;
63 	wait_queue_head_t	m_waitq;
64 	struct list_head	m_conn_list;
65 	unsigned long		m_page_addrs[RDS_CONG_MAP_PAGES];
66 };
67 
68 
69 /*
70  * This is how we will track the connection state:
71  * A connection is always in one of the following
72  * states. Updates to the state are atomic and imply
73  * a memory barrier.
74  */
75 enum {
76 	RDS_CONN_DOWN = 0,
77 	RDS_CONN_CONNECTING,
78 	RDS_CONN_DISCONNECTING,
79 	RDS_CONN_UP,
80 	RDS_CONN_RESETTING,
81 	RDS_CONN_ERROR,
82 };
83 
84 /* Bits for c_flags */
85 #define RDS_LL_SEND_FULL	0
86 #define RDS_RECONNECT_PENDING	1
87 #define RDS_IN_XMIT		2
88 #define RDS_RECV_REFILL		3
89 
90 /* Max number of multipaths per RDS connection. Must be a power of 2 */
91 #define	RDS_MPATH_WORKERS	8
92 #define	RDS_MPATH_HASH(rs, n) (jhash_1word((rs)->rs_bound_port, \
93 			       (rs)->rs_hash_initval) & ((n) - 1))
94 
95 /* Per mpath connection state */
96 struct rds_conn_path {
97 	struct rds_connection	*cp_conn;
98 	struct rds_message	*cp_xmit_rm;
99 	unsigned long		cp_xmit_sg;
100 	unsigned int		cp_xmit_hdr_off;
101 	unsigned int		cp_xmit_data_off;
102 	unsigned int		cp_xmit_atomic_sent;
103 	unsigned int		cp_xmit_rdma_sent;
104 	unsigned int		cp_xmit_data_sent;
105 
106 	spinlock_t		cp_lock;		/* protect msg queues */
107 	u64			cp_next_tx_seq;
108 	struct list_head	cp_send_queue;
109 	struct list_head	cp_retrans;
110 
111 	u64			cp_next_rx_seq;
112 
113 	void			*cp_transport_data;
114 
115 	atomic_t		cp_state;
116 	unsigned long		cp_send_gen;
117 	unsigned long		cp_flags;
118 	unsigned long		cp_reconnect_jiffies;
119 	struct delayed_work	cp_send_w;
120 	struct delayed_work	cp_recv_w;
121 	struct delayed_work	cp_conn_w;
122 	struct work_struct	cp_down_w;
123 	struct mutex		cp_cm_lock;	/* protect cp_state & cm */
124 	wait_queue_head_t	cp_waitq;
125 
126 	unsigned int		cp_unacked_packets;
127 	unsigned int		cp_unacked_bytes;
128 	unsigned int		cp_outgoing:1,
129 				cp_pad_to_32:31;
130 	unsigned int		cp_index;
131 };
132 
133 /* One rds_connection per RDS address pair */
134 struct rds_connection {
135 	struct hlist_node	c_hash_node;
136 	__be32			c_laddr;
137 	__be32			c_faddr;
138 	unsigned int		c_loopback:1,
139 				c_ping_triggered:1,
140 				c_pad_to_32:30;
141 	int			c_npaths;
142 	struct rds_connection	*c_passive;
143 	struct rds_transport	*c_trans;
144 
145 	struct rds_cong_map	*c_lcong;
146 	struct rds_cong_map	*c_fcong;
147 
148 	/* Protocol version */
149 	unsigned int		c_version;
150 	struct net		*c_net;
151 
152 	struct list_head	c_map_item;
153 	unsigned long		c_map_queued;
154 
155 	struct rds_conn_path	c_path[RDS_MPATH_WORKERS];
156 	wait_queue_head_t	c_hs_waitq; /* handshake waitq */
157 
158 	u32			c_my_gen_num;
159 	u32			c_peer_gen_num;
160 };
161 
162 static inline
163 struct net *rds_conn_net(struct rds_connection *conn)
164 {
165 	return conn->c_net;
166 }
167 
168 static inline
169 void rds_conn_net_set(struct rds_connection *conn, struct net *net)
170 {
171 	conn->c_net = get_net(net);
172 }
173 
174 #define RDS_FLAG_CONG_BITMAP	0x01
175 #define RDS_FLAG_ACK_REQUIRED	0x02
176 #define RDS_FLAG_RETRANSMITTED	0x04
177 #define RDS_MAX_ADV_CREDIT	255
178 
179 /* RDS_FLAG_PROBE_PORT is the reserved sport used for sending a ping
180  * probe to exchange control information before establishing a connection.
181  * Currently the control information that is exchanged is the number of
182  * supported paths. If the peer is a legacy (older kernel revision) peer,
183  * it would return a pong message without additional control information
184  * that would then alert the sender that the peer was an older rev.
185  */
186 #define RDS_FLAG_PROBE_PORT	1
187 #define	RDS_HS_PROBE(sport, dport) \
188 		((sport == RDS_FLAG_PROBE_PORT && dport == 0) || \
189 		 (sport == 0 && dport == RDS_FLAG_PROBE_PORT))
190 /*
191  * Maximum space available for extension headers.
192  */
193 #define RDS_HEADER_EXT_SPACE	16
194 
195 struct rds_header {
196 	__be64	h_sequence;
197 	__be64	h_ack;
198 	__be32	h_len;
199 	__be16	h_sport;
200 	__be16	h_dport;
201 	u8	h_flags;
202 	u8	h_credit;
203 	u8	h_padding[4];
204 	__sum16	h_csum;
205 
206 	u8	h_exthdr[RDS_HEADER_EXT_SPACE];
207 };
208 
209 /*
210  * Reserved - indicates end of extensions
211  */
212 #define RDS_EXTHDR_NONE		0
213 
214 /*
215  * This extension header is included in the very
216  * first message that is sent on a new connection,
217  * and identifies the protocol level. This will help
218  * rolling updates if a future change requires breaking
219  * the protocol.
220  * NB: This is no longer true for IB, where we do a version
221  * negotiation during the connection setup phase (protocol
222  * version information is included in the RDMA CM private data).
223  */
224 #define RDS_EXTHDR_VERSION	1
225 struct rds_ext_header_version {
226 	__be32			h_version;
227 };
228 
229 /*
230  * This extension header is included in the RDS message
231  * chasing an RDMA operation.
232  */
233 #define RDS_EXTHDR_RDMA		2
234 struct rds_ext_header_rdma {
235 	__be32			h_rdma_rkey;
236 };
237 
238 /*
239  * This extension header tells the peer about the
240  * destination <R_Key,offset> of the requested RDMA
241  * operation.
242  */
243 #define RDS_EXTHDR_RDMA_DEST	3
244 struct rds_ext_header_rdma_dest {
245 	__be32			h_rdma_rkey;
246 	__be32			h_rdma_offset;
247 };
248 
249 /* Extension header announcing number of paths.
250  * Implicit length = 2 bytes.
251  */
252 #define RDS_EXTHDR_NPATHS	5
253 #define RDS_EXTHDR_GEN_NUM	6
254 
255 #define __RDS_EXTHDR_MAX	16 /* for now */
256 #define RDS_RX_MAX_TRACES	(RDS_MSG_RX_DGRAM_TRACE_MAX + 1)
257 #define	RDS_MSG_RX_HDR		0
258 #define	RDS_MSG_RX_START	1
259 #define	RDS_MSG_RX_END		2
260 #define	RDS_MSG_RX_CMSG		3
261 
262 struct rds_incoming {
263 	atomic_t		i_refcount;
264 	struct list_head	i_item;
265 	struct rds_connection	*i_conn;
266 	struct rds_conn_path	*i_conn_path;
267 	struct rds_header	i_hdr;
268 	unsigned long		i_rx_jiffies;
269 	__be32			i_saddr;
270 
271 	rds_rdma_cookie_t	i_rdma_cookie;
272 	struct timeval		i_rx_tstamp;
273 	u64			i_rx_lat_trace[RDS_RX_MAX_TRACES];
274 };
275 
276 struct rds_mr {
277 	struct rb_node		r_rb_node;
278 	atomic_t		r_refcount;
279 	u32			r_key;
280 
281 	/* A copy of the creation flags */
282 	unsigned int		r_use_once:1;
283 	unsigned int		r_invalidate:1;
284 	unsigned int		r_write:1;
285 
286 	/* This is for RDS_MR_DEAD.
287 	 * It would be nice & consistent to make this part of the above
288 	 * bit field here, but we need to use test_and_set_bit.
289 	 */
290 	unsigned long		r_state;
291 	struct rds_sock		*r_sock; /* back pointer to the socket that owns us */
292 	struct rds_transport	*r_trans;
293 	void			*r_trans_private;
294 };
295 
296 /* Flags for mr->r_state */
297 #define RDS_MR_DEAD		0
298 
299 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
300 {
301 	return r_key | (((u64) offset) << 32);
302 }
303 
304 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
305 {
306 	return cookie;
307 }
308 
309 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
310 {
311 	return cookie >> 32;
312 }
313 
314 /* atomic operation types */
315 #define RDS_ATOMIC_TYPE_CSWP		0
316 #define RDS_ATOMIC_TYPE_FADD		1
317 
318 /*
319  * m_sock_item and m_conn_item are on lists that are serialized under
320  * conn->c_lock.  m_sock_item has additional meaning in that once it is empty
321  * the message will not be put back on the retransmit list after being sent.
322  * messages that are canceled while being sent rely on this.
323  *
324  * m_inc is used by loopback so that it can pass an incoming message straight
325  * back up into the rx path.  It embeds a wire header which is also used by
326  * the send path, which is kind of awkward.
327  *
328  * m_sock_item indicates the message's presence on a socket's send or receive
329  * queue.  m_rs will point to that socket.
330  *
331  * m_daddr is used by cancellation to prune messages to a given destination.
332  *
333  * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
334  * nesting.  As paths iterate over messages on a sock, or conn, they must
335  * also lock the conn, or sock, to remove the message from those lists too.
336  * Testing the flag to determine if the message is still on the lists lets
337  * us avoid testing the list_head directly.  That means each path can use
338  * the message's list_head to keep it on a local list while juggling locks
339  * without confusing the other path.
340  *
341  * m_ack_seq is an optional field set by transports who need a different
342  * sequence number range to invalidate.  They can use this in a callback
343  * that they pass to rds_send_drop_acked() to see if each message has been
344  * acked.  The HAS_ACK_SEQ flag can be used to detect messages which haven't
345  * had ack_seq set yet.
346  */
347 #define RDS_MSG_ON_SOCK		1
348 #define RDS_MSG_ON_CONN		2
349 #define RDS_MSG_HAS_ACK_SEQ	3
350 #define RDS_MSG_ACK_REQUIRED	4
351 #define RDS_MSG_RETRANSMITTED	5
352 #define RDS_MSG_MAPPED		6
353 #define RDS_MSG_PAGEVEC		7
354 #define RDS_MSG_FLUSH		8
355 
356 struct rds_message {
357 	atomic_t		m_refcount;
358 	struct list_head	m_sock_item;
359 	struct list_head	m_conn_item;
360 	struct rds_incoming	m_inc;
361 	u64			m_ack_seq;
362 	__be32			m_daddr;
363 	unsigned long		m_flags;
364 
365 	/* Never access m_rs without holding m_rs_lock.
366 	 * Lock nesting is
367 	 *  rm->m_rs_lock
368 	 *   -> rs->rs_lock
369 	 */
370 	spinlock_t		m_rs_lock;
371 	wait_queue_head_t	m_flush_wait;
372 
373 	struct rds_sock		*m_rs;
374 
375 	/* cookie to send to remote, in rds header */
376 	rds_rdma_cookie_t	m_rdma_cookie;
377 
378 	unsigned int		m_used_sgs;
379 	unsigned int		m_total_sgs;
380 
381 	void			*m_final_op;
382 
383 	struct {
384 		struct rm_atomic_op {
385 			int			op_type;
386 			union {
387 				struct {
388 					uint64_t	compare;
389 					uint64_t	swap;
390 					uint64_t	compare_mask;
391 					uint64_t	swap_mask;
392 				} op_m_cswp;
393 				struct {
394 					uint64_t	add;
395 					uint64_t	nocarry_mask;
396 				} op_m_fadd;
397 			};
398 
399 			u32			op_rkey;
400 			u64			op_remote_addr;
401 			unsigned int		op_notify:1;
402 			unsigned int		op_recverr:1;
403 			unsigned int		op_mapped:1;
404 			unsigned int		op_silent:1;
405 			unsigned int		op_active:1;
406 			struct scatterlist	*op_sg;
407 			struct rds_notifier	*op_notifier;
408 
409 			struct rds_mr		*op_rdma_mr;
410 		} atomic;
411 		struct rm_rdma_op {
412 			u32			op_rkey;
413 			u64			op_remote_addr;
414 			unsigned int		op_write:1;
415 			unsigned int		op_fence:1;
416 			unsigned int		op_notify:1;
417 			unsigned int		op_recverr:1;
418 			unsigned int		op_mapped:1;
419 			unsigned int		op_silent:1;
420 			unsigned int		op_active:1;
421 			unsigned int		op_bytes;
422 			unsigned int		op_nents;
423 			unsigned int		op_count;
424 			struct scatterlist	*op_sg;
425 			struct rds_notifier	*op_notifier;
426 
427 			struct rds_mr		*op_rdma_mr;
428 		} rdma;
429 		struct rm_data_op {
430 			unsigned int		op_active:1;
431 			unsigned int		op_notify:1;
432 			unsigned int		op_nents;
433 			unsigned int		op_count;
434 			unsigned int		op_dmasg;
435 			unsigned int		op_dmaoff;
436 			struct scatterlist	*op_sg;
437 		} data;
438 	};
439 };
440 
441 /*
442  * The RDS notifier is used (optionally) to tell the application about
443  * completed RDMA operations. Rather than keeping the whole rds message
444  * around on the queue, we allocate a small notifier that is put on the
445  * socket's notifier_list. Notifications are delivered to the application
446  * through control messages.
447  */
448 struct rds_notifier {
449 	struct list_head	n_list;
450 	uint64_t		n_user_token;
451 	int			n_status;
452 };
453 
454 /**
455  * struct rds_transport -  transport specific behavioural hooks
456  *
457  * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
458  *        part of a message.  The caller serializes on the send_sem so this
459  *        doesn't need to be reentrant for a given conn.  The header must be
460  *        sent before the data payload.  .xmit must be prepared to send a
461  *        message with no data payload.  .xmit should return the number of
462  *        bytes that were sent down the connection, including header bytes.
463  *        Returning 0 tells the caller that it doesn't need to perform any
464  *        additional work now.  This is usually the case when the transport has
465  *        filled the sending queue for its connection and will handle
466  *        triggering the rds thread to continue the send when space becomes
467  *        available.  Returning -EAGAIN tells the caller to retry the send
468  *        immediately.  Returning -ENOMEM tells the caller to retry the send at
469  *        some point in the future.
470  *
471  * @conn_shutdown: conn_shutdown stops traffic on the given connection.  Once
472  *                 it returns the connection can not call rds_recv_incoming().
473  *                 This will only be called once after conn_connect returns
474  *                 non-zero success and will The caller serializes this with
475  *                 the send and connecting paths (xmit_* and conn_*).  The
476  *                 transport is responsible for other serialization, including
477  *                 rds_recv_incoming().  This is called in process context but
478  *                 should try hard not to block.
479  */
480 
481 struct rds_transport {
482 	char			t_name[TRANSNAMSIZ];
483 	struct list_head	t_item;
484 	struct module		*t_owner;
485 	unsigned int		t_prefer_loopback:1,
486 				t_mp_capable:1;
487 	unsigned int		t_type;
488 
489 	int (*laddr_check)(struct net *net, __be32 addr);
490 	int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
491 	void (*conn_free)(void *data);
492 	int (*conn_path_connect)(struct rds_conn_path *cp);
493 	void (*conn_path_shutdown)(struct rds_conn_path *conn);
494 	void (*xmit_path_prepare)(struct rds_conn_path *cp);
495 	void (*xmit_path_complete)(struct rds_conn_path *cp);
496 	int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
497 		    unsigned int hdr_off, unsigned int sg, unsigned int off);
498 	int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
499 	int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
500 	int (*recv_path)(struct rds_conn_path *cp);
501 	int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
502 	void (*inc_free)(struct rds_incoming *inc);
503 
504 	int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
505 				 struct rdma_cm_event *event);
506 	int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
507 	void (*cm_connect_complete)(struct rds_connection *conn,
508 				    struct rdma_cm_event *event);
509 
510 	unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
511 					unsigned int avail);
512 	void (*exit)(void);
513 	void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
514 			struct rds_sock *rs, u32 *key_ret);
515 	void (*sync_mr)(void *trans_private, int direction);
516 	void (*free_mr)(void *trans_private, int invalidate);
517 	void (*flush_mrs)(void);
518 };
519 
520 struct rds_sock {
521 	struct sock		rs_sk;
522 
523 	u64			rs_user_addr;
524 	u64			rs_user_bytes;
525 
526 	/*
527 	 * bound_addr used for both incoming and outgoing, no INADDR_ANY
528 	 * support.
529 	 */
530 	struct rhash_head	rs_bound_node;
531 	u64			rs_bound_key;
532 	__be32			rs_bound_addr;
533 	__be32			rs_conn_addr;
534 	__be16			rs_bound_port;
535 	__be16			rs_conn_port;
536 	struct rds_transport    *rs_transport;
537 
538 	/*
539 	 * rds_sendmsg caches the conn it used the last time around.
540 	 * This helps avoid costly lookups.
541 	 */
542 	struct rds_connection	*rs_conn;
543 
544 	/* flag indicating we were congested or not */
545 	int			rs_congested;
546 	/* seen congestion (ENOBUFS) when sending? */
547 	int			rs_seen_congestion;
548 
549 	/* rs_lock protects all these adjacent members before the newline */
550 	spinlock_t		rs_lock;
551 	struct list_head	rs_send_queue;
552 	u32			rs_snd_bytes;
553 	int			rs_rcv_bytes;
554 	struct list_head	rs_notify_queue;	/* currently used for failed RDMAs */
555 
556 	/* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
557 	 * to decide whether the application should be woken up.
558 	 * If not set, we use rs_cong_track to find out whether a cong map
559 	 * update arrived.
560 	 */
561 	uint64_t		rs_cong_mask;
562 	uint64_t		rs_cong_notify;
563 	struct list_head	rs_cong_list;
564 	unsigned long		rs_cong_track;
565 
566 	/*
567 	 * rs_recv_lock protects the receive queue, and is
568 	 * used to serialize with rds_release.
569 	 */
570 	rwlock_t		rs_recv_lock;
571 	struct list_head	rs_recv_queue;
572 
573 	/* just for stats reporting */
574 	struct list_head	rs_item;
575 
576 	/* these have their own lock */
577 	spinlock_t		rs_rdma_lock;
578 	struct rb_root		rs_rdma_keys;
579 
580 	/* Socket options - in case there will be more */
581 	unsigned char		rs_recverr,
582 				rs_cong_monitor;
583 	u32			rs_hash_initval;
584 
585 	/* Socket receive path trace points*/
586 	u8			rs_rx_traces;
587 	u8			rs_rx_trace[RDS_MSG_RX_DGRAM_TRACE_MAX];
588 };
589 
590 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
591 {
592 	return container_of(sk, struct rds_sock, rs_sk);
593 }
594 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
595 {
596 	return &rs->rs_sk;
597 }
598 
599 /*
600  * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
601  * to account for overhead.  We don't account for overhead, we just apply
602  * the number of payload bytes to the specified value.
603  */
604 static inline int rds_sk_sndbuf(struct rds_sock *rs)
605 {
606 	return rds_rs_to_sk(rs)->sk_sndbuf / 2;
607 }
608 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
609 {
610 	return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
611 }
612 
613 struct rds_statistics {
614 	uint64_t	s_conn_reset;
615 	uint64_t	s_recv_drop_bad_checksum;
616 	uint64_t	s_recv_drop_old_seq;
617 	uint64_t	s_recv_drop_no_sock;
618 	uint64_t	s_recv_drop_dead_sock;
619 	uint64_t	s_recv_deliver_raced;
620 	uint64_t	s_recv_delivered;
621 	uint64_t	s_recv_queued;
622 	uint64_t	s_recv_immediate_retry;
623 	uint64_t	s_recv_delayed_retry;
624 	uint64_t	s_recv_ack_required;
625 	uint64_t	s_recv_rdma_bytes;
626 	uint64_t	s_recv_ping;
627 	uint64_t	s_send_queue_empty;
628 	uint64_t	s_send_queue_full;
629 	uint64_t	s_send_lock_contention;
630 	uint64_t	s_send_lock_queue_raced;
631 	uint64_t	s_send_immediate_retry;
632 	uint64_t	s_send_delayed_retry;
633 	uint64_t	s_send_drop_acked;
634 	uint64_t	s_send_ack_required;
635 	uint64_t	s_send_queued;
636 	uint64_t	s_send_rdma;
637 	uint64_t	s_send_rdma_bytes;
638 	uint64_t	s_send_pong;
639 	uint64_t	s_page_remainder_hit;
640 	uint64_t	s_page_remainder_miss;
641 	uint64_t	s_copy_to_user;
642 	uint64_t	s_copy_from_user;
643 	uint64_t	s_cong_update_queued;
644 	uint64_t	s_cong_update_received;
645 	uint64_t	s_cong_send_error;
646 	uint64_t	s_cong_send_blocked;
647 	uint64_t	s_recv_bytes_added_to_socket;
648 	uint64_t	s_recv_bytes_removed_from_socket;
649 
650 };
651 
652 /* af_rds.c */
653 void rds_sock_addref(struct rds_sock *rs);
654 void rds_sock_put(struct rds_sock *rs);
655 void rds_wake_sk_sleep(struct rds_sock *rs);
656 static inline void __rds_wake_sk_sleep(struct sock *sk)
657 {
658 	wait_queue_head_t *waitq = sk_sleep(sk);
659 
660 	if (!sock_flag(sk, SOCK_DEAD) && waitq)
661 		wake_up(waitq);
662 }
663 extern wait_queue_head_t rds_poll_waitq;
664 
665 
666 /* bind.c */
667 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
668 void rds_remove_bound(struct rds_sock *rs);
669 struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
670 int rds_bind_lock_init(void);
671 void rds_bind_lock_destroy(void);
672 
673 /* cong.c */
674 int rds_cong_get_maps(struct rds_connection *conn);
675 void rds_cong_add_conn(struct rds_connection *conn);
676 void rds_cong_remove_conn(struct rds_connection *conn);
677 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
678 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
679 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
680 void rds_cong_queue_updates(struct rds_cong_map *map);
681 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
682 int rds_cong_updated_since(unsigned long *recent);
683 void rds_cong_add_socket(struct rds_sock *);
684 void rds_cong_remove_socket(struct rds_sock *);
685 void rds_cong_exit(void);
686 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
687 
688 /* conn.c */
689 extern u32 rds_gen_num;
690 int rds_conn_init(void);
691 void rds_conn_exit(void);
692 struct rds_connection *rds_conn_create(struct net *net,
693 				       __be32 laddr, __be32 faddr,
694 				       struct rds_transport *trans, gfp_t gfp);
695 struct rds_connection *rds_conn_create_outgoing(struct net *net,
696 						__be32 laddr, __be32 faddr,
697 			       struct rds_transport *trans, gfp_t gfp);
698 void rds_conn_shutdown(struct rds_conn_path *cpath);
699 void rds_conn_destroy(struct rds_connection *conn);
700 void rds_conn_drop(struct rds_connection *conn);
701 void rds_conn_path_drop(struct rds_conn_path *cpath);
702 void rds_conn_connect_if_down(struct rds_connection *conn);
703 void rds_conn_path_connect_if_down(struct rds_conn_path *cp);
704 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
705 			  struct rds_info_iterator *iter,
706 			  struct rds_info_lengths *lens,
707 			  int (*visitor)(struct rds_connection *, void *),
708 			  size_t item_len);
709 
710 __printf(2, 3)
711 void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...);
712 #define rds_conn_path_error(cp, fmt...) \
713 	__rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt)
714 
715 static inline int
716 rds_conn_path_transition(struct rds_conn_path *cp, int old, int new)
717 {
718 	return atomic_cmpxchg(&cp->cp_state, old, new) == old;
719 }
720 
721 static inline int
722 rds_conn_transition(struct rds_connection *conn, int old, int new)
723 {
724 	WARN_ON(conn->c_trans->t_mp_capable);
725 	return rds_conn_path_transition(&conn->c_path[0], old, new);
726 }
727 
728 static inline int
729 rds_conn_path_state(struct rds_conn_path *cp)
730 {
731 	return atomic_read(&cp->cp_state);
732 }
733 
734 static inline int
735 rds_conn_state(struct rds_connection *conn)
736 {
737 	WARN_ON(conn->c_trans->t_mp_capable);
738 	return rds_conn_path_state(&conn->c_path[0]);
739 }
740 
741 static inline int
742 rds_conn_path_up(struct rds_conn_path *cp)
743 {
744 	return atomic_read(&cp->cp_state) == RDS_CONN_UP;
745 }
746 
747 static inline int
748 rds_conn_up(struct rds_connection *conn)
749 {
750 	WARN_ON(conn->c_trans->t_mp_capable);
751 	return rds_conn_path_up(&conn->c_path[0]);
752 }
753 
754 static inline int
755 rds_conn_path_connecting(struct rds_conn_path *cp)
756 {
757 	return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING;
758 }
759 
760 static inline int
761 rds_conn_connecting(struct rds_connection *conn)
762 {
763 	WARN_ON(conn->c_trans->t_mp_capable);
764 	return rds_conn_path_connecting(&conn->c_path[0]);
765 }
766 
767 /* message.c */
768 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
769 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
770 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from);
771 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
772 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
773 				 __be16 dport, u64 seq);
774 int rds_message_add_extension(struct rds_header *hdr,
775 			      unsigned int type, const void *data, unsigned int len);
776 int rds_message_next_extension(struct rds_header *hdr,
777 			       unsigned int *pos, void *buf, unsigned int *buflen);
778 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
779 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
780 void rds_message_inc_free(struct rds_incoming *inc);
781 void rds_message_addref(struct rds_message *rm);
782 void rds_message_put(struct rds_message *rm);
783 void rds_message_wait(struct rds_message *rm);
784 void rds_message_unmapped(struct rds_message *rm);
785 
786 static inline void rds_message_make_checksum(struct rds_header *hdr)
787 {
788 	hdr->h_csum = 0;
789 	hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
790 }
791 
792 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
793 {
794 	return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
795 }
796 
797 
798 /* page.c */
799 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
800 			     gfp_t gfp);
801 void rds_page_exit(void);
802 
803 /* recv.c */
804 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
805 		  __be32 saddr);
806 void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn,
807 		       __be32 saddr);
808 void rds_inc_put(struct rds_incoming *inc);
809 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
810 		       struct rds_incoming *inc, gfp_t gfp);
811 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
812 		int msg_flags);
813 void rds_clear_recv_queue(struct rds_sock *rs);
814 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
815 void rds_inc_info_copy(struct rds_incoming *inc,
816 		       struct rds_info_iterator *iter,
817 		       __be32 saddr, __be32 daddr, int flip);
818 
819 /* send.c */
820 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
821 void rds_send_path_reset(struct rds_conn_path *conn);
822 int rds_send_xmit(struct rds_conn_path *cp);
823 struct sockaddr_in;
824 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
825 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
826 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
827 			 is_acked_func is_acked);
828 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
829 			      is_acked_func is_acked);
830 int rds_send_pong(struct rds_conn_path *cp, __be16 dport);
831 
832 /* rdma.c */
833 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
834 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
835 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
836 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
837 void rds_rdma_drop_keys(struct rds_sock *rs);
838 int rds_rdma_extra_size(struct rds_rdma_args *args);
839 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
840 			  struct cmsghdr *cmsg);
841 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
842 			  struct cmsghdr *cmsg);
843 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
844 			  struct cmsghdr *cmsg);
845 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
846 			  struct cmsghdr *cmsg);
847 void rds_rdma_free_op(struct rm_rdma_op *ro);
848 void rds_atomic_free_op(struct rm_atomic_op *ao);
849 void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
850 void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
851 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
852 		    struct cmsghdr *cmsg);
853 
854 void __rds_put_mr_final(struct rds_mr *mr);
855 static inline void rds_mr_put(struct rds_mr *mr)
856 {
857 	if (atomic_dec_and_test(&mr->r_refcount))
858 		__rds_put_mr_final(mr);
859 }
860 
861 /* stats.c */
862 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
863 #define rds_stats_inc_which(which, member) do {		\
864 	per_cpu(which, get_cpu()).member++;		\
865 	put_cpu();					\
866 } while (0)
867 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
868 #define rds_stats_add_which(which, member, count) do {		\
869 	per_cpu(which, get_cpu()).member += count;	\
870 	put_cpu();					\
871 } while (0)
872 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
873 int rds_stats_init(void);
874 void rds_stats_exit(void);
875 void rds_stats_info_copy(struct rds_info_iterator *iter,
876 			 uint64_t *values, const char *const *names,
877 			 size_t nr);
878 
879 /* sysctl.c */
880 int rds_sysctl_init(void);
881 void rds_sysctl_exit(void);
882 extern unsigned long rds_sysctl_sndbuf_min;
883 extern unsigned long rds_sysctl_sndbuf_default;
884 extern unsigned long rds_sysctl_sndbuf_max;
885 extern unsigned long rds_sysctl_reconnect_min_jiffies;
886 extern unsigned long rds_sysctl_reconnect_max_jiffies;
887 extern unsigned int  rds_sysctl_max_unacked_packets;
888 extern unsigned int  rds_sysctl_max_unacked_bytes;
889 extern unsigned int  rds_sysctl_ping_enable;
890 extern unsigned long rds_sysctl_trace_flags;
891 extern unsigned int  rds_sysctl_trace_level;
892 
893 /* threads.c */
894 int rds_threads_init(void);
895 void rds_threads_exit(void);
896 extern struct workqueue_struct *rds_wq;
897 void rds_queue_reconnect(struct rds_conn_path *cp);
898 void rds_connect_worker(struct work_struct *);
899 void rds_shutdown_worker(struct work_struct *);
900 void rds_send_worker(struct work_struct *);
901 void rds_recv_worker(struct work_struct *);
902 void rds_connect_path_complete(struct rds_conn_path *conn, int curr);
903 void rds_connect_complete(struct rds_connection *conn);
904 
905 /* transport.c */
906 void rds_trans_register(struct rds_transport *trans);
907 void rds_trans_unregister(struct rds_transport *trans);
908 struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr);
909 void rds_trans_put(struct rds_transport *trans);
910 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
911 				       unsigned int avail);
912 struct rds_transport *rds_trans_get(int t_type);
913 int rds_trans_init(void);
914 void rds_trans_exit(void);
915 
916 #endif
917