xref: /linux/net/rds/rds.h (revision 8c749ce93ee69e789e46b3be98de9e0cbfcf8ed8)
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 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 #define RDS_CONG_MAP_BYTES	(65536 / 8)
54 #define RDS_CONG_MAP_PAGES	(PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
55 #define RDS_CONG_MAP_PAGE_BITS	(PAGE_SIZE * 8)
56 
57 struct rds_cong_map {
58 	struct rb_node		m_rb_node;
59 	__be32			m_addr;
60 	wait_queue_head_t	m_waitq;
61 	struct list_head	m_conn_list;
62 	unsigned long		m_page_addrs[RDS_CONG_MAP_PAGES];
63 };
64 
65 
66 /*
67  * This is how we will track the connection state:
68  * A connection is always in one of the following
69  * states. Updates to the state are atomic and imply
70  * a memory barrier.
71  */
72 enum {
73 	RDS_CONN_DOWN = 0,
74 	RDS_CONN_CONNECTING,
75 	RDS_CONN_DISCONNECTING,
76 	RDS_CONN_UP,
77 	RDS_CONN_ERROR,
78 };
79 
80 /* Bits for c_flags */
81 #define RDS_LL_SEND_FULL	0
82 #define RDS_RECONNECT_PENDING	1
83 #define RDS_IN_XMIT		2
84 #define RDS_RECV_REFILL		3
85 
86 struct rds_connection {
87 	struct hlist_node	c_hash_node;
88 	__be32			c_laddr;
89 	__be32			c_faddr;
90 	unsigned int		c_loopback:1,
91 				c_outgoing:1,
92 				c_pad_to_32:30;
93 	struct rds_connection	*c_passive;
94 
95 	struct rds_cong_map	*c_lcong;
96 	struct rds_cong_map	*c_fcong;
97 
98 	struct rds_message	*c_xmit_rm;
99 	unsigned long		c_xmit_sg;
100 	unsigned int		c_xmit_hdr_off;
101 	unsigned int		c_xmit_data_off;
102 	unsigned int		c_xmit_atomic_sent;
103 	unsigned int		c_xmit_rdma_sent;
104 	unsigned int		c_xmit_data_sent;
105 
106 	spinlock_t		c_lock;		/* protect msg queues */
107 	u64			c_next_tx_seq;
108 	struct list_head	c_send_queue;
109 	struct list_head	c_retrans;
110 
111 	u64			c_next_rx_seq;
112 
113 	struct rds_transport	*c_trans;
114 	void			*c_transport_data;
115 
116 	atomic_t		c_state;
117 	unsigned long		c_send_gen;
118 	unsigned long		c_flags;
119 	unsigned long		c_reconnect_jiffies;
120 	struct delayed_work	c_send_w;
121 	struct delayed_work	c_recv_w;
122 	struct delayed_work	c_conn_w;
123 	struct work_struct	c_down_w;
124 	struct mutex		c_cm_lock;	/* protect conn state & cm */
125 	wait_queue_head_t	c_waitq;
126 
127 	struct list_head	c_map_item;
128 	unsigned long		c_map_queued;
129 
130 	unsigned int		c_unacked_packets;
131 	unsigned int		c_unacked_bytes;
132 
133 	/* Protocol version */
134 	unsigned int		c_version;
135 	possible_net_t		c_net;
136 };
137 
138 static inline
139 struct net *rds_conn_net(struct rds_connection *conn)
140 {
141 	return read_pnet(&conn->c_net);
142 }
143 
144 static inline
145 void rds_conn_net_set(struct rds_connection *conn, struct net *net)
146 {
147 	write_pnet(&conn->c_net, net);
148 }
149 
150 #define RDS_FLAG_CONG_BITMAP	0x01
151 #define RDS_FLAG_ACK_REQUIRED	0x02
152 #define RDS_FLAG_RETRANSMITTED	0x04
153 #define RDS_MAX_ADV_CREDIT	255
154 
155 /*
156  * Maximum space available for extension headers.
157  */
158 #define RDS_HEADER_EXT_SPACE	16
159 
160 struct rds_header {
161 	__be64	h_sequence;
162 	__be64	h_ack;
163 	__be32	h_len;
164 	__be16	h_sport;
165 	__be16	h_dport;
166 	u8	h_flags;
167 	u8	h_credit;
168 	u8	h_padding[4];
169 	__sum16	h_csum;
170 
171 	u8	h_exthdr[RDS_HEADER_EXT_SPACE];
172 };
173 
174 /*
175  * Reserved - indicates end of extensions
176  */
177 #define RDS_EXTHDR_NONE		0
178 
179 /*
180  * This extension header is included in the very
181  * first message that is sent on a new connection,
182  * and identifies the protocol level. This will help
183  * rolling updates if a future change requires breaking
184  * the protocol.
185  * NB: This is no longer true for IB, where we do a version
186  * negotiation during the connection setup phase (protocol
187  * version information is included in the RDMA CM private data).
188  */
189 #define RDS_EXTHDR_VERSION	1
190 struct rds_ext_header_version {
191 	__be32			h_version;
192 };
193 
194 /*
195  * This extension header is included in the RDS message
196  * chasing an RDMA operation.
197  */
198 #define RDS_EXTHDR_RDMA		2
199 struct rds_ext_header_rdma {
200 	__be32			h_rdma_rkey;
201 };
202 
203 /*
204  * This extension header tells the peer about the
205  * destination <R_Key,offset> of the requested RDMA
206  * operation.
207  */
208 #define RDS_EXTHDR_RDMA_DEST	3
209 struct rds_ext_header_rdma_dest {
210 	__be32			h_rdma_rkey;
211 	__be32			h_rdma_offset;
212 };
213 
214 #define __RDS_EXTHDR_MAX	16 /* for now */
215 
216 struct rds_incoming {
217 	atomic_t		i_refcount;
218 	struct list_head	i_item;
219 	struct rds_connection	*i_conn;
220 	struct rds_header	i_hdr;
221 	unsigned long		i_rx_jiffies;
222 	__be32			i_saddr;
223 
224 	rds_rdma_cookie_t	i_rdma_cookie;
225 };
226 
227 struct rds_mr {
228 	struct rb_node		r_rb_node;
229 	atomic_t		r_refcount;
230 	u32			r_key;
231 
232 	/* A copy of the creation flags */
233 	unsigned int		r_use_once:1;
234 	unsigned int		r_invalidate:1;
235 	unsigned int		r_write:1;
236 
237 	/* This is for RDS_MR_DEAD.
238 	 * It would be nice & consistent to make this part of the above
239 	 * bit field here, but we need to use test_and_set_bit.
240 	 */
241 	unsigned long		r_state;
242 	struct rds_sock		*r_sock; /* back pointer to the socket that owns us */
243 	struct rds_transport	*r_trans;
244 	void			*r_trans_private;
245 };
246 
247 /* Flags for mr->r_state */
248 #define RDS_MR_DEAD		0
249 
250 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
251 {
252 	return r_key | (((u64) offset) << 32);
253 }
254 
255 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
256 {
257 	return cookie;
258 }
259 
260 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
261 {
262 	return cookie >> 32;
263 }
264 
265 /* atomic operation types */
266 #define RDS_ATOMIC_TYPE_CSWP		0
267 #define RDS_ATOMIC_TYPE_FADD		1
268 
269 /*
270  * m_sock_item and m_conn_item are on lists that are serialized under
271  * conn->c_lock.  m_sock_item has additional meaning in that once it is empty
272  * the message will not be put back on the retransmit list after being sent.
273  * messages that are canceled while being sent rely on this.
274  *
275  * m_inc is used by loopback so that it can pass an incoming message straight
276  * back up into the rx path.  It embeds a wire header which is also used by
277  * the send path, which is kind of awkward.
278  *
279  * m_sock_item indicates the message's presence on a socket's send or receive
280  * queue.  m_rs will point to that socket.
281  *
282  * m_daddr is used by cancellation to prune messages to a given destination.
283  *
284  * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
285  * nesting.  As paths iterate over messages on a sock, or conn, they must
286  * also lock the conn, or sock, to remove the message from those lists too.
287  * Testing the flag to determine if the message is still on the lists lets
288  * us avoid testing the list_head directly.  That means each path can use
289  * the message's list_head to keep it on a local list while juggling locks
290  * without confusing the other path.
291  *
292  * m_ack_seq is an optional field set by transports who need a different
293  * sequence number range to invalidate.  They can use this in a callback
294  * that they pass to rds_send_drop_acked() to see if each message has been
295  * acked.  The HAS_ACK_SEQ flag can be used to detect messages which haven't
296  * had ack_seq set yet.
297  */
298 #define RDS_MSG_ON_SOCK		1
299 #define RDS_MSG_ON_CONN		2
300 #define RDS_MSG_HAS_ACK_SEQ	3
301 #define RDS_MSG_ACK_REQUIRED	4
302 #define RDS_MSG_RETRANSMITTED	5
303 #define RDS_MSG_MAPPED		6
304 #define RDS_MSG_PAGEVEC		7
305 
306 struct rds_message {
307 	atomic_t		m_refcount;
308 	struct list_head	m_sock_item;
309 	struct list_head	m_conn_item;
310 	struct rds_incoming	m_inc;
311 	u64			m_ack_seq;
312 	__be32			m_daddr;
313 	unsigned long		m_flags;
314 
315 	/* Never access m_rs without holding m_rs_lock.
316 	 * Lock nesting is
317 	 *  rm->m_rs_lock
318 	 *   -> rs->rs_lock
319 	 */
320 	spinlock_t		m_rs_lock;
321 	wait_queue_head_t	m_flush_wait;
322 
323 	struct rds_sock		*m_rs;
324 
325 	/* cookie to send to remote, in rds header */
326 	rds_rdma_cookie_t	m_rdma_cookie;
327 
328 	unsigned int		m_used_sgs;
329 	unsigned int		m_total_sgs;
330 
331 	void			*m_final_op;
332 
333 	struct {
334 		struct rm_atomic_op {
335 			int			op_type;
336 			union {
337 				struct {
338 					uint64_t	compare;
339 					uint64_t	swap;
340 					uint64_t	compare_mask;
341 					uint64_t	swap_mask;
342 				} op_m_cswp;
343 				struct {
344 					uint64_t	add;
345 					uint64_t	nocarry_mask;
346 				} op_m_fadd;
347 			};
348 
349 			u32			op_rkey;
350 			u64			op_remote_addr;
351 			unsigned int		op_notify:1;
352 			unsigned int		op_recverr:1;
353 			unsigned int		op_mapped:1;
354 			unsigned int		op_silent:1;
355 			unsigned int		op_active:1;
356 			struct scatterlist	*op_sg;
357 			struct rds_notifier	*op_notifier;
358 
359 			struct rds_mr		*op_rdma_mr;
360 		} atomic;
361 		struct rm_rdma_op {
362 			u32			op_rkey;
363 			u64			op_remote_addr;
364 			unsigned int		op_write:1;
365 			unsigned int		op_fence:1;
366 			unsigned int		op_notify:1;
367 			unsigned int		op_recverr:1;
368 			unsigned int		op_mapped:1;
369 			unsigned int		op_silent:1;
370 			unsigned int		op_active:1;
371 			unsigned int		op_bytes;
372 			unsigned int		op_nents;
373 			unsigned int		op_count;
374 			struct scatterlist	*op_sg;
375 			struct rds_notifier	*op_notifier;
376 
377 			struct rds_mr		*op_rdma_mr;
378 		} rdma;
379 		struct rm_data_op {
380 			unsigned int		op_active:1;
381 			unsigned int		op_nents;
382 			unsigned int		op_count;
383 			unsigned int		op_dmasg;
384 			unsigned int		op_dmaoff;
385 			struct scatterlist	*op_sg;
386 		} data;
387 	};
388 };
389 
390 /*
391  * The RDS notifier is used (optionally) to tell the application about
392  * completed RDMA operations. Rather than keeping the whole rds message
393  * around on the queue, we allocate a small notifier that is put on the
394  * socket's notifier_list. Notifications are delivered to the application
395  * through control messages.
396  */
397 struct rds_notifier {
398 	struct list_head	n_list;
399 	uint64_t		n_user_token;
400 	int			n_status;
401 };
402 
403 /**
404  * struct rds_transport -  transport specific behavioural hooks
405  *
406  * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
407  *        part of a message.  The caller serializes on the send_sem so this
408  *        doesn't need to be reentrant for a given conn.  The header must be
409  *        sent before the data payload.  .xmit must be prepared to send a
410  *        message with no data payload.  .xmit should return the number of
411  *        bytes that were sent down the connection, including header bytes.
412  *        Returning 0 tells the caller that it doesn't need to perform any
413  *        additional work now.  This is usually the case when the transport has
414  *        filled the sending queue for its connection and will handle
415  *        triggering the rds thread to continue the send when space becomes
416  *        available.  Returning -EAGAIN tells the caller to retry the send
417  *        immediately.  Returning -ENOMEM tells the caller to retry the send at
418  *        some point in the future.
419  *
420  * @conn_shutdown: conn_shutdown stops traffic on the given connection.  Once
421  *                 it returns the connection can not call rds_recv_incoming().
422  *                 This will only be called once after conn_connect returns
423  *                 non-zero success and will The caller serializes this with
424  *                 the send and connecting paths (xmit_* and conn_*).  The
425  *                 transport is responsible for other serialization, including
426  *                 rds_recv_incoming().  This is called in process context but
427  *                 should try hard not to block.
428  */
429 
430 struct rds_transport {
431 	char			t_name[TRANSNAMSIZ];
432 	struct list_head	t_item;
433 	struct module		*t_owner;
434 	unsigned int		t_prefer_loopback:1;
435 	unsigned int		t_type;
436 
437 	int (*laddr_check)(struct net *net, __be32 addr);
438 	int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
439 	void (*conn_free)(void *data);
440 	int (*conn_connect)(struct rds_connection *conn);
441 	void (*conn_shutdown)(struct rds_connection *conn);
442 	void (*xmit_prepare)(struct rds_connection *conn);
443 	void (*xmit_complete)(struct rds_connection *conn);
444 	int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
445 		    unsigned int hdr_off, unsigned int sg, unsigned int off);
446 	int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
447 	int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
448 	int (*recv)(struct rds_connection *conn);
449 	int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
450 	void (*inc_free)(struct rds_incoming *inc);
451 
452 	int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
453 				 struct rdma_cm_event *event);
454 	int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
455 	void (*cm_connect_complete)(struct rds_connection *conn,
456 				    struct rdma_cm_event *event);
457 
458 	unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
459 					unsigned int avail);
460 	void (*exit)(void);
461 	void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
462 			struct rds_sock *rs, u32 *key_ret);
463 	void (*sync_mr)(void *trans_private, int direction);
464 	void (*free_mr)(void *trans_private, int invalidate);
465 	void (*flush_mrs)(void);
466 };
467 
468 struct rds_sock {
469 	struct sock		rs_sk;
470 
471 	u64			rs_user_addr;
472 	u64			rs_user_bytes;
473 
474 	/*
475 	 * bound_addr used for both incoming and outgoing, no INADDR_ANY
476 	 * support.
477 	 */
478 	struct rhash_head	rs_bound_node;
479 	u64			rs_bound_key;
480 	__be32			rs_bound_addr;
481 	__be32			rs_conn_addr;
482 	__be16			rs_bound_port;
483 	__be16			rs_conn_port;
484 	struct rds_transport    *rs_transport;
485 
486 	/*
487 	 * rds_sendmsg caches the conn it used the last time around.
488 	 * This helps avoid costly lookups.
489 	 */
490 	struct rds_connection	*rs_conn;
491 
492 	/* flag indicating we were congested or not */
493 	int			rs_congested;
494 	/* seen congestion (ENOBUFS) when sending? */
495 	int			rs_seen_congestion;
496 
497 	/* rs_lock protects all these adjacent members before the newline */
498 	spinlock_t		rs_lock;
499 	struct list_head	rs_send_queue;
500 	u32			rs_snd_bytes;
501 	int			rs_rcv_bytes;
502 	struct list_head	rs_notify_queue;	/* currently used for failed RDMAs */
503 
504 	/* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
505 	 * to decide whether the application should be woken up.
506 	 * If not set, we use rs_cong_track to find out whether a cong map
507 	 * update arrived.
508 	 */
509 	uint64_t		rs_cong_mask;
510 	uint64_t		rs_cong_notify;
511 	struct list_head	rs_cong_list;
512 	unsigned long		rs_cong_track;
513 
514 	/*
515 	 * rs_recv_lock protects the receive queue, and is
516 	 * used to serialize with rds_release.
517 	 */
518 	rwlock_t		rs_recv_lock;
519 	struct list_head	rs_recv_queue;
520 
521 	/* just for stats reporting */
522 	struct list_head	rs_item;
523 
524 	/* these have their own lock */
525 	spinlock_t		rs_rdma_lock;
526 	struct rb_root		rs_rdma_keys;
527 
528 	/* Socket options - in case there will be more */
529 	unsigned char		rs_recverr,
530 				rs_cong_monitor;
531 };
532 
533 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
534 {
535 	return container_of(sk, struct rds_sock, rs_sk);
536 }
537 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
538 {
539 	return &rs->rs_sk;
540 }
541 
542 /*
543  * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
544  * to account for overhead.  We don't account for overhead, we just apply
545  * the number of payload bytes to the specified value.
546  */
547 static inline int rds_sk_sndbuf(struct rds_sock *rs)
548 {
549 	return rds_rs_to_sk(rs)->sk_sndbuf / 2;
550 }
551 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
552 {
553 	return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
554 }
555 
556 struct rds_statistics {
557 	uint64_t	s_conn_reset;
558 	uint64_t	s_recv_drop_bad_checksum;
559 	uint64_t	s_recv_drop_old_seq;
560 	uint64_t	s_recv_drop_no_sock;
561 	uint64_t	s_recv_drop_dead_sock;
562 	uint64_t	s_recv_deliver_raced;
563 	uint64_t	s_recv_delivered;
564 	uint64_t	s_recv_queued;
565 	uint64_t	s_recv_immediate_retry;
566 	uint64_t	s_recv_delayed_retry;
567 	uint64_t	s_recv_ack_required;
568 	uint64_t	s_recv_rdma_bytes;
569 	uint64_t	s_recv_ping;
570 	uint64_t	s_send_queue_empty;
571 	uint64_t	s_send_queue_full;
572 	uint64_t	s_send_lock_contention;
573 	uint64_t	s_send_lock_queue_raced;
574 	uint64_t	s_send_immediate_retry;
575 	uint64_t	s_send_delayed_retry;
576 	uint64_t	s_send_drop_acked;
577 	uint64_t	s_send_ack_required;
578 	uint64_t	s_send_queued;
579 	uint64_t	s_send_rdma;
580 	uint64_t	s_send_rdma_bytes;
581 	uint64_t	s_send_pong;
582 	uint64_t	s_page_remainder_hit;
583 	uint64_t	s_page_remainder_miss;
584 	uint64_t	s_copy_to_user;
585 	uint64_t	s_copy_from_user;
586 	uint64_t	s_cong_update_queued;
587 	uint64_t	s_cong_update_received;
588 	uint64_t	s_cong_send_error;
589 	uint64_t	s_cong_send_blocked;
590 };
591 
592 /* af_rds.c */
593 void rds_sock_addref(struct rds_sock *rs);
594 void rds_sock_put(struct rds_sock *rs);
595 void rds_wake_sk_sleep(struct rds_sock *rs);
596 static inline void __rds_wake_sk_sleep(struct sock *sk)
597 {
598 	wait_queue_head_t *waitq = sk_sleep(sk);
599 
600 	if (!sock_flag(sk, SOCK_DEAD) && waitq)
601 		wake_up(waitq);
602 }
603 extern wait_queue_head_t rds_poll_waitq;
604 
605 
606 /* bind.c */
607 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
608 void rds_remove_bound(struct rds_sock *rs);
609 struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
610 int rds_bind_lock_init(void);
611 void rds_bind_lock_destroy(void);
612 
613 /* cong.c */
614 int rds_cong_get_maps(struct rds_connection *conn);
615 void rds_cong_add_conn(struct rds_connection *conn);
616 void rds_cong_remove_conn(struct rds_connection *conn);
617 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
618 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
619 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
620 void rds_cong_queue_updates(struct rds_cong_map *map);
621 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
622 int rds_cong_updated_since(unsigned long *recent);
623 void rds_cong_add_socket(struct rds_sock *);
624 void rds_cong_remove_socket(struct rds_sock *);
625 void rds_cong_exit(void);
626 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
627 
628 /* conn.c */
629 int rds_conn_init(void);
630 void rds_conn_exit(void);
631 struct rds_connection *rds_conn_create(struct net *net,
632 				       __be32 laddr, __be32 faddr,
633 				       struct rds_transport *trans, gfp_t gfp);
634 struct rds_connection *rds_conn_create_outgoing(struct net *net,
635 						__be32 laddr, __be32 faddr,
636 			       struct rds_transport *trans, gfp_t gfp);
637 void rds_conn_shutdown(struct rds_connection *conn);
638 void rds_conn_destroy(struct rds_connection *conn);
639 void rds_conn_drop(struct rds_connection *conn);
640 void rds_conn_connect_if_down(struct rds_connection *conn);
641 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
642 			  struct rds_info_iterator *iter,
643 			  struct rds_info_lengths *lens,
644 			  int (*visitor)(struct rds_connection *, void *),
645 			  size_t item_len);
646 __printf(2, 3)
647 void __rds_conn_error(struct rds_connection *conn, const char *, ...);
648 #define rds_conn_error(conn, fmt...) \
649 	__rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
650 
651 static inline int
652 rds_conn_transition(struct rds_connection *conn, int old, int new)
653 {
654 	return atomic_cmpxchg(&conn->c_state, old, new) == old;
655 }
656 
657 static inline int
658 rds_conn_state(struct rds_connection *conn)
659 {
660 	return atomic_read(&conn->c_state);
661 }
662 
663 static inline int
664 rds_conn_up(struct rds_connection *conn)
665 {
666 	return atomic_read(&conn->c_state) == RDS_CONN_UP;
667 }
668 
669 static inline int
670 rds_conn_connecting(struct rds_connection *conn)
671 {
672 	return atomic_read(&conn->c_state) == RDS_CONN_CONNECTING;
673 }
674 
675 /* message.c */
676 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
677 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
678 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from);
679 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
680 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
681 				 __be16 dport, u64 seq);
682 int rds_message_add_extension(struct rds_header *hdr,
683 			      unsigned int type, const void *data, unsigned int len);
684 int rds_message_next_extension(struct rds_header *hdr,
685 			       unsigned int *pos, void *buf, unsigned int *buflen);
686 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
687 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
688 void rds_message_inc_free(struct rds_incoming *inc);
689 void rds_message_addref(struct rds_message *rm);
690 void rds_message_put(struct rds_message *rm);
691 void rds_message_wait(struct rds_message *rm);
692 void rds_message_unmapped(struct rds_message *rm);
693 
694 static inline void rds_message_make_checksum(struct rds_header *hdr)
695 {
696 	hdr->h_csum = 0;
697 	hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
698 }
699 
700 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
701 {
702 	return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
703 }
704 
705 
706 /* page.c */
707 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
708 			     gfp_t gfp);
709 int rds_page_copy_user(struct page *page, unsigned long offset,
710 		       void __user *ptr, unsigned long bytes,
711 		       int to_user);
712 #define rds_page_copy_to_user(page, offset, ptr, bytes) \
713 	rds_page_copy_user(page, offset, ptr, bytes, 1)
714 #define rds_page_copy_from_user(page, offset, ptr, bytes) \
715 	rds_page_copy_user(page, offset, ptr, bytes, 0)
716 void rds_page_exit(void);
717 
718 /* recv.c */
719 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
720 		  __be32 saddr);
721 void rds_inc_put(struct rds_incoming *inc);
722 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
723 		       struct rds_incoming *inc, gfp_t gfp);
724 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
725 		int msg_flags);
726 void rds_clear_recv_queue(struct rds_sock *rs);
727 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
728 void rds_inc_info_copy(struct rds_incoming *inc,
729 		       struct rds_info_iterator *iter,
730 		       __be32 saddr, __be32 daddr, int flip);
731 
732 /* send.c */
733 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
734 void rds_send_reset(struct rds_connection *conn);
735 int rds_send_xmit(struct rds_connection *conn);
736 struct sockaddr_in;
737 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
738 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
739 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
740 			 is_acked_func is_acked);
741 int rds_send_pong(struct rds_connection *conn, __be16 dport);
742 struct rds_message *rds_send_get_message(struct rds_connection *,
743 					 struct rm_rdma_op *);
744 
745 /* rdma.c */
746 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
747 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
748 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
749 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
750 void rds_rdma_drop_keys(struct rds_sock *rs);
751 int rds_rdma_extra_size(struct rds_rdma_args *args);
752 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
753 			  struct cmsghdr *cmsg);
754 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
755 			  struct cmsghdr *cmsg);
756 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
757 			  struct cmsghdr *cmsg);
758 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
759 			  struct cmsghdr *cmsg);
760 void rds_rdma_free_op(struct rm_rdma_op *ro);
761 void rds_atomic_free_op(struct rm_atomic_op *ao);
762 void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
763 void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
764 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
765 		    struct cmsghdr *cmsg);
766 
767 void __rds_put_mr_final(struct rds_mr *mr);
768 static inline void rds_mr_put(struct rds_mr *mr)
769 {
770 	if (atomic_dec_and_test(&mr->r_refcount))
771 		__rds_put_mr_final(mr);
772 }
773 
774 /* stats.c */
775 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
776 #define rds_stats_inc_which(which, member) do {		\
777 	per_cpu(which, get_cpu()).member++;		\
778 	put_cpu();					\
779 } while (0)
780 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
781 #define rds_stats_add_which(which, member, count) do {		\
782 	per_cpu(which, get_cpu()).member += count;	\
783 	put_cpu();					\
784 } while (0)
785 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
786 int rds_stats_init(void);
787 void rds_stats_exit(void);
788 void rds_stats_info_copy(struct rds_info_iterator *iter,
789 			 uint64_t *values, const char *const *names,
790 			 size_t nr);
791 
792 /* sysctl.c */
793 int rds_sysctl_init(void);
794 void rds_sysctl_exit(void);
795 extern unsigned long rds_sysctl_sndbuf_min;
796 extern unsigned long rds_sysctl_sndbuf_default;
797 extern unsigned long rds_sysctl_sndbuf_max;
798 extern unsigned long rds_sysctl_reconnect_min_jiffies;
799 extern unsigned long rds_sysctl_reconnect_max_jiffies;
800 extern unsigned int  rds_sysctl_max_unacked_packets;
801 extern unsigned int  rds_sysctl_max_unacked_bytes;
802 extern unsigned int  rds_sysctl_ping_enable;
803 extern unsigned long rds_sysctl_trace_flags;
804 extern unsigned int  rds_sysctl_trace_level;
805 
806 /* threads.c */
807 int rds_threads_init(void);
808 void rds_threads_exit(void);
809 extern struct workqueue_struct *rds_wq;
810 void rds_queue_reconnect(struct rds_connection *conn);
811 void rds_connect_worker(struct work_struct *);
812 void rds_shutdown_worker(struct work_struct *);
813 void rds_send_worker(struct work_struct *);
814 void rds_recv_worker(struct work_struct *);
815 void rds_connect_complete(struct rds_connection *conn);
816 
817 /* transport.c */
818 int rds_trans_register(struct rds_transport *trans);
819 void rds_trans_unregister(struct rds_transport *trans);
820 struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr);
821 void rds_trans_put(struct rds_transport *trans);
822 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
823 				       unsigned int avail);
824 struct rds_transport *rds_trans_get(int t_type);
825 int rds_trans_init(void);
826 void rds_trans_exit(void);
827 
828 #endif
829