xref: /openbmc/linux/net/rds/rds.h (revision b8bb76713ec50df2f11efee386e16f93d51e1076)
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 
11 #include "info.h"
12 
13 /*
14  * RDS Network protocol version
15  */
16 #define RDS_PROTOCOL_3_0	0x0300
17 #define RDS_PROTOCOL_3_1	0x0301
18 #define RDS_PROTOCOL_VERSION	RDS_PROTOCOL_3_1
19 #define RDS_PROTOCOL_MAJOR(v)	((v) >> 8)
20 #define RDS_PROTOCOL_MINOR(v)	((v) & 255)
21 #define RDS_PROTOCOL(maj, min)	(((maj) << 8) | min)
22 
23 /*
24  * XXX randomly chosen, but at least seems to be unused:
25  * #               18464-18768 Unassigned
26  * We should do better.  We want a reserved port to discourage unpriv'ed
27  * userspace from listening.
28  */
29 #define RDS_PORT	18634
30 
31 #ifdef DEBUG
32 #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
33 #else
34 /* sigh, pr_debug() causes unused variable warnings */
35 static inline void __attribute__ ((format (printf, 1, 2)))
36 rdsdebug(char *fmt, ...)
37 {
38 }
39 #endif
40 
41 /* XXX is there one of these somewhere? */
42 #define ceil(x, y) \
43 	({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; })
44 
45 #define RDS_FRAG_SHIFT	12
46 #define RDS_FRAG_SIZE	((unsigned int)(1 << RDS_FRAG_SHIFT))
47 
48 #define RDS_CONG_MAP_BYTES	(65536 / 8)
49 #define RDS_CONG_MAP_LONGS	(RDS_CONG_MAP_BYTES / sizeof(unsigned long))
50 #define RDS_CONG_MAP_PAGES	(PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
51 #define RDS_CONG_MAP_PAGE_BITS	(PAGE_SIZE * 8)
52 
53 struct rds_cong_map {
54 	struct rb_node		m_rb_node;
55 	__be32			m_addr;
56 	wait_queue_head_t	m_waitq;
57 	struct list_head	m_conn_list;
58 	unsigned long		m_page_addrs[RDS_CONG_MAP_PAGES];
59 };
60 
61 
62 /*
63  * This is how we will track the connection state:
64  * A connection is always in one of the following
65  * states. Updates to the state are atomic and imply
66  * a memory barrier.
67  */
68 enum {
69 	RDS_CONN_DOWN = 0,
70 	RDS_CONN_CONNECTING,
71 	RDS_CONN_DISCONNECTING,
72 	RDS_CONN_UP,
73 	RDS_CONN_ERROR,
74 };
75 
76 /* Bits for c_flags */
77 #define RDS_LL_SEND_FULL	0
78 #define RDS_RECONNECT_PENDING	1
79 
80 struct rds_connection {
81 	struct hlist_node	c_hash_node;
82 	__be32			c_laddr;
83 	__be32			c_faddr;
84 	unsigned int		c_loopback:1;
85 	struct rds_connection	*c_passive;
86 
87 	struct rds_cong_map	*c_lcong;
88 	struct rds_cong_map	*c_fcong;
89 
90 	struct mutex		c_send_lock;	/* protect send ring */
91 	struct rds_message	*c_xmit_rm;
92 	unsigned long		c_xmit_sg;
93 	unsigned int		c_xmit_hdr_off;
94 	unsigned int		c_xmit_data_off;
95 	unsigned int		c_xmit_rdma_sent;
96 
97 	spinlock_t		c_lock;		/* protect msg queues */
98 	u64			c_next_tx_seq;
99 	struct list_head	c_send_queue;
100 	struct list_head	c_retrans;
101 
102 	u64			c_next_rx_seq;
103 
104 	struct rds_transport	*c_trans;
105 	void			*c_transport_data;
106 
107 	atomic_t		c_state;
108 	unsigned long		c_flags;
109 	unsigned long		c_reconnect_jiffies;
110 	struct delayed_work	c_send_w;
111 	struct delayed_work	c_recv_w;
112 	struct delayed_work	c_conn_w;
113 	struct work_struct	c_down_w;
114 	struct mutex		c_cm_lock;	/* protect conn state & cm */
115 
116 	struct list_head	c_map_item;
117 	unsigned long		c_map_queued;
118 	unsigned long		c_map_offset;
119 	unsigned long		c_map_bytes;
120 
121 	unsigned int		c_unacked_packets;
122 	unsigned int		c_unacked_bytes;
123 
124 	/* Protocol version */
125 	unsigned int		c_version;
126 };
127 
128 #define RDS_FLAG_CONG_BITMAP	0x01
129 #define RDS_FLAG_ACK_REQUIRED	0x02
130 #define RDS_FLAG_RETRANSMITTED	0x04
131 #define RDS_MAX_ADV_CREDIT	127
132 
133 /*
134  * Maximum space available for extension headers.
135  */
136 #define RDS_HEADER_EXT_SPACE	16
137 
138 struct rds_header {
139 	__be64	h_sequence;
140 	__be64	h_ack;
141 	__be32	h_len;
142 	__be16	h_sport;
143 	__be16	h_dport;
144 	u8	h_flags;
145 	u8	h_credit;
146 	u8	h_padding[4];
147 	__sum16	h_csum;
148 
149 	u8	h_exthdr[RDS_HEADER_EXT_SPACE];
150 };
151 
152 /*
153  * Reserved - indicates end of extensions
154  */
155 #define RDS_EXTHDR_NONE		0
156 
157 /*
158  * This extension header is included in the very
159  * first message that is sent on a new connection,
160  * and identifies the protocol level. This will help
161  * rolling updates if a future change requires breaking
162  * the protocol.
163  * NB: This is no longer true for IB, where we do a version
164  * negotiation during the connection setup phase (protocol
165  * version information is included in the RDMA CM private data).
166  */
167 #define RDS_EXTHDR_VERSION	1
168 struct rds_ext_header_version {
169 	__be32			h_version;
170 };
171 
172 /*
173  * This extension header is included in the RDS message
174  * chasing an RDMA operation.
175  */
176 #define RDS_EXTHDR_RDMA		2
177 struct rds_ext_header_rdma {
178 	__be32			h_rdma_rkey;
179 };
180 
181 /*
182  * This extension header tells the peer about the
183  * destination <R_Key,offset> of the requested RDMA
184  * operation.
185  */
186 #define RDS_EXTHDR_RDMA_DEST	3
187 struct rds_ext_header_rdma_dest {
188 	__be32			h_rdma_rkey;
189 	__be32			h_rdma_offset;
190 };
191 
192 #define __RDS_EXTHDR_MAX	16 /* for now */
193 
194 struct rds_incoming {
195 	atomic_t		i_refcount;
196 	struct list_head	i_item;
197 	struct rds_connection	*i_conn;
198 	struct rds_header	i_hdr;
199 	unsigned long		i_rx_jiffies;
200 	__be32			i_saddr;
201 
202 	rds_rdma_cookie_t	i_rdma_cookie;
203 };
204 
205 /*
206  * m_sock_item and m_conn_item are on lists that are serialized under
207  * conn->c_lock.  m_sock_item has additional meaning in that once it is empty
208  * the message will not be put back on the retransmit list after being sent.
209  * messages that are canceled while being sent rely on this.
210  *
211  * m_inc is used by loopback so that it can pass an incoming message straight
212  * back up into the rx path.  It embeds a wire header which is also used by
213  * the send path, which is kind of awkward.
214  *
215  * m_sock_item indicates the message's presence on a socket's send or receive
216  * queue.  m_rs will point to that socket.
217  *
218  * m_daddr is used by cancellation to prune messages to a given destination.
219  *
220  * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
221  * nesting.  As paths iterate over messages on a sock, or conn, they must
222  * also lock the conn, or sock, to remove the message from those lists too.
223  * Testing the flag to determine if the message is still on the lists lets
224  * us avoid testing the list_head directly.  That means each path can use
225  * the message's list_head to keep it on a local list while juggling locks
226  * without confusing the other path.
227  *
228  * m_ack_seq is an optional field set by transports who need a different
229  * sequence number range to invalidate.  They can use this in a callback
230  * that they pass to rds_send_drop_acked() to see if each message has been
231  * acked.  The HAS_ACK_SEQ flag can be used to detect messages which haven't
232  * had ack_seq set yet.
233  */
234 #define RDS_MSG_ON_SOCK		1
235 #define RDS_MSG_ON_CONN		2
236 #define RDS_MSG_HAS_ACK_SEQ	3
237 #define RDS_MSG_ACK_REQUIRED	4
238 #define RDS_MSG_RETRANSMITTED	5
239 #define RDS_MSG_MAPPED		6
240 #define RDS_MSG_PAGEVEC		7
241 
242 struct rds_message {
243 	atomic_t		m_refcount;
244 	struct list_head	m_sock_item;
245 	struct list_head	m_conn_item;
246 	struct rds_incoming	m_inc;
247 	u64			m_ack_seq;
248 	__be32			m_daddr;
249 	unsigned long		m_flags;
250 
251 	/* Never access m_rs without holding m_rs_lock.
252 	 * Lock nesting is
253 	 *  rm->m_rs_lock
254 	 *   -> rs->rs_lock
255 	 */
256 	spinlock_t		m_rs_lock;
257 	struct rds_sock		*m_rs;
258 	struct rds_rdma_op	*m_rdma_op;
259 	rds_rdma_cookie_t	m_rdma_cookie;
260 	struct rds_mr		*m_rdma_mr;
261 	unsigned int		m_nents;
262 	unsigned int		m_count;
263 	struct scatterlist	m_sg[0];
264 };
265 
266 /*
267  * The RDS notifier is used (optionally) to tell the application about
268  * completed RDMA operations. Rather than keeping the whole rds message
269  * around on the queue, we allocate a small notifier that is put on the
270  * socket's notifier_list. Notifications are delivered to the application
271  * through control messages.
272  */
273 struct rds_notifier {
274 	struct list_head	n_list;
275 	uint64_t		n_user_token;
276 	int			n_status;
277 };
278 
279 /**
280  * struct rds_transport -  transport specific behavioural hooks
281  *
282  * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
283  *        part of a message.  The caller serializes on the send_sem so this
284  *        doesn't need to be reentrant for a given conn.  The header must be
285  *        sent before the data payload.  .xmit must be prepared to send a
286  *        message with no data payload.  .xmit should return the number of
287  *        bytes that were sent down the connection, including header bytes.
288  *        Returning 0 tells the caller that it doesn't need to perform any
289  *        additional work now.  This is usually the case when the transport has
290  *        filled the sending queue for its connection and will handle
291  *        triggering the rds thread to continue the send when space becomes
292  *        available.  Returning -EAGAIN tells the caller to retry the send
293  *        immediately.  Returning -ENOMEM tells the caller to retry the send at
294  *        some point in the future.
295  *
296  * @conn_shutdown: conn_shutdown stops traffic on the given connection.  Once
297  *                 it returns the connection can not call rds_recv_incoming().
298  *                 This will only be called once after conn_connect returns
299  *                 non-zero success and will The caller serializes this with
300  *                 the send and connecting paths (xmit_* and conn_*).  The
301  *                 transport is responsible for other serialization, including
302  *                 rds_recv_incoming().  This is called in process context but
303  *                 should try hard not to block.
304  *
305  * @xmit_cong_map: This asks the transport to send the local bitmap down the
306  * 		   given connection.  XXX get a better story about the bitmap
307  * 		   flag and header.
308  */
309 
310 struct rds_transport {
311 	char			t_name[TRANSNAMSIZ];
312 	struct list_head	t_item;
313 	struct module		*t_owner;
314 	unsigned int		t_prefer_loopback:1;
315 
316 	int (*laddr_check)(__be32 addr);
317 	int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
318 	void (*conn_free)(void *data);
319 	int (*conn_connect)(struct rds_connection *conn);
320 	void (*conn_shutdown)(struct rds_connection *conn);
321 	void (*xmit_prepare)(struct rds_connection *conn);
322 	void (*xmit_complete)(struct rds_connection *conn);
323 	int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
324 		    unsigned int hdr_off, unsigned int sg, unsigned int off);
325 	int (*xmit_cong_map)(struct rds_connection *conn,
326 			     struct rds_cong_map *map, unsigned long offset);
327 	int (*xmit_rdma)(struct rds_connection *conn, struct rds_rdma_op *op);
328 	int (*recv)(struct rds_connection *conn);
329 	int (*inc_copy_to_user)(struct rds_incoming *inc, struct iovec *iov,
330 				size_t size);
331 	void (*inc_purge)(struct rds_incoming *inc);
332 	void (*inc_free)(struct rds_incoming *inc);
333 
334 	int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
335 				 struct rdma_cm_event *event);
336 	int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
337 	void (*cm_connect_complete)(struct rds_connection *conn,
338 				    struct rdma_cm_event *event);
339 
340 	unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
341 					unsigned int avail);
342 	void (*exit)(void);
343 	void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
344 			struct rds_sock *rs, u32 *key_ret);
345 	void (*sync_mr)(void *trans_private, int direction);
346 	void (*free_mr)(void *trans_private, int invalidate);
347 	void (*flush_mrs)(void);
348 };
349 
350 struct rds_sock {
351 	struct sock		rs_sk;
352 
353 	u64			rs_user_addr;
354 	u64			rs_user_bytes;
355 
356 	/*
357 	 * bound_addr used for both incoming and outgoing, no INADDR_ANY
358 	 * support.
359 	 */
360 	struct rb_node		rs_bound_node;
361 	__be32			rs_bound_addr;
362 	__be32			rs_conn_addr;
363 	__be16			rs_bound_port;
364 	__be16			rs_conn_port;
365 
366 	/*
367 	 * This is only used to communicate the transport between bind and
368 	 * initiating connections.  All other trans use is referenced through
369 	 * the connection.
370 	 */
371 	struct rds_transport    *rs_transport;
372 
373 	/*
374 	 * rds_sendmsg caches the conn it used the last time around.
375 	 * This helps avoid costly lookups.
376 	 */
377 	struct rds_connection	*rs_conn;
378 
379 	/* flag indicating we were congested or not */
380 	int			rs_congested;
381 
382 	/* rs_lock protects all these adjacent members before the newline */
383 	spinlock_t		rs_lock;
384 	struct list_head	rs_send_queue;
385 	u32			rs_snd_bytes;
386 	int			rs_rcv_bytes;
387 	struct list_head	rs_notify_queue;	/* currently used for failed RDMAs */
388 
389 	/* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
390 	 * to decide whether the application should be woken up.
391 	 * If not set, we use rs_cong_track to find out whether a cong map
392 	 * update arrived.
393 	 */
394 	uint64_t		rs_cong_mask;
395 	uint64_t		rs_cong_notify;
396 	struct list_head	rs_cong_list;
397 	unsigned long		rs_cong_track;
398 
399 	/*
400 	 * rs_recv_lock protects the receive queue, and is
401 	 * used to serialize with rds_release.
402 	 */
403 	rwlock_t		rs_recv_lock;
404 	struct list_head	rs_recv_queue;
405 
406 	/* just for stats reporting */
407 	struct list_head	rs_item;
408 
409 	/* these have their own lock */
410 	spinlock_t		rs_rdma_lock;
411 	struct rb_root		rs_rdma_keys;
412 
413 	/* Socket options - in case there will be more */
414 	unsigned char		rs_recverr,
415 				rs_cong_monitor;
416 };
417 
418 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
419 {
420 	return container_of(sk, struct rds_sock, rs_sk);
421 }
422 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
423 {
424 	return &rs->rs_sk;
425 }
426 
427 /*
428  * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
429  * to account for overhead.  We don't account for overhead, we just apply
430  * the number of payload bytes to the specified value.
431  */
432 static inline int rds_sk_sndbuf(struct rds_sock *rs)
433 {
434 	return rds_rs_to_sk(rs)->sk_sndbuf / 2;
435 }
436 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
437 {
438 	return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
439 }
440 
441 struct rds_statistics {
442 	uint64_t	s_conn_reset;
443 	uint64_t	s_recv_drop_bad_checksum;
444 	uint64_t	s_recv_drop_old_seq;
445 	uint64_t	s_recv_drop_no_sock;
446 	uint64_t	s_recv_drop_dead_sock;
447 	uint64_t	s_recv_deliver_raced;
448 	uint64_t	s_recv_delivered;
449 	uint64_t	s_recv_queued;
450 	uint64_t	s_recv_immediate_retry;
451 	uint64_t	s_recv_delayed_retry;
452 	uint64_t	s_recv_ack_required;
453 	uint64_t	s_recv_rdma_bytes;
454 	uint64_t	s_recv_ping;
455 	uint64_t	s_send_queue_empty;
456 	uint64_t	s_send_queue_full;
457 	uint64_t	s_send_sem_contention;
458 	uint64_t	s_send_sem_queue_raced;
459 	uint64_t	s_send_immediate_retry;
460 	uint64_t	s_send_delayed_retry;
461 	uint64_t	s_send_drop_acked;
462 	uint64_t	s_send_ack_required;
463 	uint64_t	s_send_queued;
464 	uint64_t	s_send_rdma;
465 	uint64_t	s_send_rdma_bytes;
466 	uint64_t	s_send_pong;
467 	uint64_t	s_page_remainder_hit;
468 	uint64_t	s_page_remainder_miss;
469 	uint64_t	s_copy_to_user;
470 	uint64_t	s_copy_from_user;
471 	uint64_t	s_cong_update_queued;
472 	uint64_t	s_cong_update_received;
473 	uint64_t	s_cong_send_error;
474 	uint64_t	s_cong_send_blocked;
475 };
476 
477 /* af_rds.c */
478 void rds_sock_addref(struct rds_sock *rs);
479 void rds_sock_put(struct rds_sock *rs);
480 void rds_wake_sk_sleep(struct rds_sock *rs);
481 static inline void __rds_wake_sk_sleep(struct sock *sk)
482 {
483 	wait_queue_head_t *waitq = sk->sk_sleep;
484 
485 	if (!sock_flag(sk, SOCK_DEAD) && waitq)
486 		wake_up(waitq);
487 }
488 extern wait_queue_head_t rds_poll_waitq;
489 
490 
491 /* bind.c */
492 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
493 void rds_remove_bound(struct rds_sock *rs);
494 struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
495 
496 /* cong.c */
497 int rds_cong_get_maps(struct rds_connection *conn);
498 void rds_cong_add_conn(struct rds_connection *conn);
499 void rds_cong_remove_conn(struct rds_connection *conn);
500 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
501 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
502 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
503 void rds_cong_queue_updates(struct rds_cong_map *map);
504 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
505 int rds_cong_updated_since(unsigned long *recent);
506 void rds_cong_add_socket(struct rds_sock *);
507 void rds_cong_remove_socket(struct rds_sock *);
508 void rds_cong_exit(void);
509 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
510 
511 /* conn.c */
512 int __init rds_conn_init(void);
513 void rds_conn_exit(void);
514 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
515 				       struct rds_transport *trans, gfp_t gfp);
516 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
517 			       struct rds_transport *trans, gfp_t gfp);
518 void rds_conn_destroy(struct rds_connection *conn);
519 void rds_conn_reset(struct rds_connection *conn);
520 void rds_conn_drop(struct rds_connection *conn);
521 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
522 			  struct rds_info_iterator *iter,
523 			  struct rds_info_lengths *lens,
524 			  int (*visitor)(struct rds_connection *, void *),
525 			  size_t item_len);
526 void __rds_conn_error(struct rds_connection *conn, const char *, ...)
527 				__attribute__ ((format (printf, 2, 3)));
528 #define rds_conn_error(conn, fmt...) \
529 	__rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
530 
531 static inline int
532 rds_conn_transition(struct rds_connection *conn, int old, int new)
533 {
534 	return atomic_cmpxchg(&conn->c_state, old, new) == old;
535 }
536 
537 static inline int
538 rds_conn_state(struct rds_connection *conn)
539 {
540 	return atomic_read(&conn->c_state);
541 }
542 
543 static inline int
544 rds_conn_up(struct rds_connection *conn)
545 {
546 	return atomic_read(&conn->c_state) == RDS_CONN_UP;
547 }
548 
549 static inline int
550 rds_conn_connecting(struct rds_connection *conn)
551 {
552 	return atomic_read(&conn->c_state) == RDS_CONN_CONNECTING;
553 }
554 
555 /* message.c */
556 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
557 struct rds_message *rds_message_copy_from_user(struct iovec *first_iov,
558 					       size_t total_len);
559 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
560 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
561 				 __be16 dport, u64 seq);
562 int rds_message_add_extension(struct rds_header *hdr,
563 			      unsigned int type, const void *data, unsigned int len);
564 int rds_message_next_extension(struct rds_header *hdr,
565 			       unsigned int *pos, void *buf, unsigned int *buflen);
566 int rds_message_add_version_extension(struct rds_header *hdr, unsigned int version);
567 int rds_message_get_version_extension(struct rds_header *hdr, unsigned int *version);
568 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
569 int rds_message_inc_copy_to_user(struct rds_incoming *inc,
570 				 struct iovec *first_iov, size_t size);
571 void rds_message_inc_purge(struct rds_incoming *inc);
572 void rds_message_inc_free(struct rds_incoming *inc);
573 void rds_message_addref(struct rds_message *rm);
574 void rds_message_put(struct rds_message *rm);
575 void rds_message_wait(struct rds_message *rm);
576 void rds_message_unmapped(struct rds_message *rm);
577 
578 static inline void rds_message_make_checksum(struct rds_header *hdr)
579 {
580 	hdr->h_csum = 0;
581 	hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
582 }
583 
584 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
585 {
586 	return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
587 }
588 
589 
590 /* page.c */
591 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
592 			     gfp_t gfp);
593 int rds_page_copy_user(struct page *page, unsigned long offset,
594 		       void __user *ptr, unsigned long bytes,
595 		       int to_user);
596 #define rds_page_copy_to_user(page, offset, ptr, bytes) \
597 	rds_page_copy_user(page, offset, ptr, bytes, 1)
598 #define rds_page_copy_from_user(page, offset, ptr, bytes) \
599 	rds_page_copy_user(page, offset, ptr, bytes, 0)
600 void rds_page_exit(void);
601 
602 /* recv.c */
603 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
604 		  __be32 saddr);
605 void rds_inc_addref(struct rds_incoming *inc);
606 void rds_inc_put(struct rds_incoming *inc);
607 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
608 		       struct rds_incoming *inc, gfp_t gfp, enum km_type km);
609 int rds_recvmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
610 		size_t size, int msg_flags);
611 void rds_clear_recv_queue(struct rds_sock *rs);
612 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
613 void rds_inc_info_copy(struct rds_incoming *inc,
614 		       struct rds_info_iterator *iter,
615 		       __be32 saddr, __be32 daddr, int flip);
616 
617 /* send.c */
618 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
619 		size_t payload_len);
620 void rds_send_reset(struct rds_connection *conn);
621 int rds_send_xmit(struct rds_connection *conn);
622 struct sockaddr_in;
623 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
624 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
625 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
626 			 is_acked_func is_acked);
627 int rds_send_acked_before(struct rds_connection *conn, u64 seq);
628 void rds_send_remove_from_sock(struct list_head *messages, int status);
629 int rds_send_pong(struct rds_connection *conn, __be16 dport);
630 struct rds_message *rds_send_get_message(struct rds_connection *,
631 					 struct rds_rdma_op *);
632 
633 /* rdma.c */
634 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
635 
636 /* stats.c */
637 DECLARE_PER_CPU(struct rds_statistics, rds_stats);
638 #define rds_stats_inc_which(which, member) do {		\
639 	per_cpu(which, get_cpu()).member++;		\
640 	put_cpu();					\
641 } while (0)
642 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
643 #define rds_stats_add_which(which, member, count) do {		\
644 	per_cpu(which, get_cpu()).member += count;	\
645 	put_cpu();					\
646 } while (0)
647 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
648 int __init rds_stats_init(void);
649 void rds_stats_exit(void);
650 void rds_stats_info_copy(struct rds_info_iterator *iter,
651 			 uint64_t *values, char **names, size_t nr);
652 
653 /* sysctl.c */
654 int __init rds_sysctl_init(void);
655 void rds_sysctl_exit(void);
656 extern unsigned long rds_sysctl_sndbuf_min;
657 extern unsigned long rds_sysctl_sndbuf_default;
658 extern unsigned long rds_sysctl_sndbuf_max;
659 extern unsigned long rds_sysctl_reconnect_min_jiffies;
660 extern unsigned long rds_sysctl_reconnect_max_jiffies;
661 extern unsigned int  rds_sysctl_max_unacked_packets;
662 extern unsigned int  rds_sysctl_max_unacked_bytes;
663 extern unsigned int  rds_sysctl_ping_enable;
664 extern unsigned long rds_sysctl_trace_flags;
665 extern unsigned int  rds_sysctl_trace_level;
666 
667 /* threads.c */
668 int __init rds_threads_init(void);
669 void rds_threads_exit(void);
670 extern struct workqueue_struct *rds_wq;
671 void rds_connect_worker(struct work_struct *);
672 void rds_shutdown_worker(struct work_struct *);
673 void rds_send_worker(struct work_struct *);
674 void rds_recv_worker(struct work_struct *);
675 void rds_connect_complete(struct rds_connection *conn);
676 
677 /* transport.c */
678 int rds_trans_register(struct rds_transport *trans);
679 void rds_trans_unregister(struct rds_transport *trans);
680 struct rds_transport *rds_trans_get_preferred(__be32 addr);
681 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
682 				       unsigned int avail);
683 int __init rds_trans_init(void);
684 void rds_trans_exit(void);
685 
686 #endif
687