xref: /openbmc/linux/net/rds/rds.h (revision fb1b3dc43dabd4bf7b57b3d63fd2875a499c81f0)
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_RESETTING,
78 	RDS_CONN_ERROR,
79 };
80 
81 /* Bits for c_flags */
82 #define RDS_LL_SEND_FULL	0
83 #define RDS_RECONNECT_PENDING	1
84 #define RDS_IN_XMIT		2
85 #define RDS_RECV_REFILL		3
86 
87 /* Max number of multipaths per RDS connection. Must be a power of 2 */
88 #define	RDS_MPATH_WORKERS	1
89 
90 /* Per mpath connection state */
91 struct rds_conn_path {
92 	struct rds_connection	*cp_conn;
93 	struct rds_message	*cp_xmit_rm;
94 	unsigned long		cp_xmit_sg;
95 	unsigned int		cp_xmit_hdr_off;
96 	unsigned int		cp_xmit_data_off;
97 	unsigned int		cp_xmit_atomic_sent;
98 	unsigned int		cp_xmit_rdma_sent;
99 	unsigned int		cp_xmit_data_sent;
100 
101 	spinlock_t		cp_lock;		/* protect msg queues */
102 	u64			cp_next_tx_seq;
103 	struct list_head	cp_send_queue;
104 	struct list_head	cp_retrans;
105 
106 	u64			cp_next_rx_seq;
107 
108 	void			*cp_transport_data;
109 
110 	atomic_t		cp_state;
111 	unsigned long		cp_send_gen;
112 	unsigned long		cp_flags;
113 	unsigned long		cp_reconnect_jiffies;
114 	struct delayed_work	cp_send_w;
115 	struct delayed_work	cp_recv_w;
116 	struct delayed_work	cp_conn_w;
117 	struct work_struct	cp_down_w;
118 	struct mutex		cp_cm_lock;	/* protect cp_state & cm */
119 	wait_queue_head_t	cp_waitq;
120 
121 	unsigned int		cp_unacked_packets;
122 	unsigned int		cp_unacked_bytes;
123 	unsigned int		cp_outgoing:1,
124 				cp_pad_to_32:31;
125 	unsigned int		cp_index;
126 };
127 
128 /* One rds_connection per RDS address pair */
129 struct rds_connection {
130 	struct hlist_node	c_hash_node;
131 	__be32			c_laddr;
132 	__be32			c_faddr;
133 	unsigned int		c_loopback:1,
134 				c_pad_to_32:31;
135 	int			c_npaths;
136 	struct rds_connection	*c_passive;
137 	struct rds_transport	*c_trans;
138 
139 	struct rds_cong_map	*c_lcong;
140 	struct rds_cong_map	*c_fcong;
141 
142 	/* Protocol version */
143 	unsigned int		c_version;
144 	possible_net_t		c_net;
145 
146 	struct list_head	c_map_item;
147 	unsigned long		c_map_queued;
148 
149 	struct rds_conn_path	c_path[RDS_MPATH_WORKERS];
150 };
151 
152 static inline
153 struct net *rds_conn_net(struct rds_connection *conn)
154 {
155 	return read_pnet(&conn->c_net);
156 }
157 
158 static inline
159 void rds_conn_net_set(struct rds_connection *conn, struct net *net)
160 {
161 	write_pnet(&conn->c_net, net);
162 }
163 
164 #define RDS_FLAG_CONG_BITMAP	0x01
165 #define RDS_FLAG_ACK_REQUIRED	0x02
166 #define RDS_FLAG_RETRANSMITTED	0x04
167 #define RDS_MAX_ADV_CREDIT	255
168 
169 /*
170  * Maximum space available for extension headers.
171  */
172 #define RDS_HEADER_EXT_SPACE	16
173 
174 struct rds_header {
175 	__be64	h_sequence;
176 	__be64	h_ack;
177 	__be32	h_len;
178 	__be16	h_sport;
179 	__be16	h_dport;
180 	u8	h_flags;
181 	u8	h_credit;
182 	u8	h_padding[4];
183 	__sum16	h_csum;
184 
185 	u8	h_exthdr[RDS_HEADER_EXT_SPACE];
186 };
187 
188 /*
189  * Reserved - indicates end of extensions
190  */
191 #define RDS_EXTHDR_NONE		0
192 
193 /*
194  * This extension header is included in the very
195  * first message that is sent on a new connection,
196  * and identifies the protocol level. This will help
197  * rolling updates if a future change requires breaking
198  * the protocol.
199  * NB: This is no longer true for IB, where we do a version
200  * negotiation during the connection setup phase (protocol
201  * version information is included in the RDMA CM private data).
202  */
203 #define RDS_EXTHDR_VERSION	1
204 struct rds_ext_header_version {
205 	__be32			h_version;
206 };
207 
208 /*
209  * This extension header is included in the RDS message
210  * chasing an RDMA operation.
211  */
212 #define RDS_EXTHDR_RDMA		2
213 struct rds_ext_header_rdma {
214 	__be32			h_rdma_rkey;
215 };
216 
217 /*
218  * This extension header tells the peer about the
219  * destination <R_Key,offset> of the requested RDMA
220  * operation.
221  */
222 #define RDS_EXTHDR_RDMA_DEST	3
223 struct rds_ext_header_rdma_dest {
224 	__be32			h_rdma_rkey;
225 	__be32			h_rdma_offset;
226 };
227 
228 #define __RDS_EXTHDR_MAX	16 /* for now */
229 
230 struct rds_incoming {
231 	atomic_t		i_refcount;
232 	struct list_head	i_item;
233 	struct rds_connection	*i_conn;
234 	struct rds_conn_path	*i_conn_path;
235 	struct rds_header	i_hdr;
236 	unsigned long		i_rx_jiffies;
237 	__be32			i_saddr;
238 
239 	rds_rdma_cookie_t	i_rdma_cookie;
240 	struct timeval		i_rx_tstamp;
241 };
242 
243 struct rds_mr {
244 	struct rb_node		r_rb_node;
245 	atomic_t		r_refcount;
246 	u32			r_key;
247 
248 	/* A copy of the creation flags */
249 	unsigned int		r_use_once:1;
250 	unsigned int		r_invalidate:1;
251 	unsigned int		r_write:1;
252 
253 	/* This is for RDS_MR_DEAD.
254 	 * It would be nice & consistent to make this part of the above
255 	 * bit field here, but we need to use test_and_set_bit.
256 	 */
257 	unsigned long		r_state;
258 	struct rds_sock		*r_sock; /* back pointer to the socket that owns us */
259 	struct rds_transport	*r_trans;
260 	void			*r_trans_private;
261 };
262 
263 /* Flags for mr->r_state */
264 #define RDS_MR_DEAD		0
265 
266 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
267 {
268 	return r_key | (((u64) offset) << 32);
269 }
270 
271 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
272 {
273 	return cookie;
274 }
275 
276 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
277 {
278 	return cookie >> 32;
279 }
280 
281 /* atomic operation types */
282 #define RDS_ATOMIC_TYPE_CSWP		0
283 #define RDS_ATOMIC_TYPE_FADD		1
284 
285 /*
286  * m_sock_item and m_conn_item are on lists that are serialized under
287  * conn->c_lock.  m_sock_item has additional meaning in that once it is empty
288  * the message will not be put back on the retransmit list after being sent.
289  * messages that are canceled while being sent rely on this.
290  *
291  * m_inc is used by loopback so that it can pass an incoming message straight
292  * back up into the rx path.  It embeds a wire header which is also used by
293  * the send path, which is kind of awkward.
294  *
295  * m_sock_item indicates the message's presence on a socket's send or receive
296  * queue.  m_rs will point to that socket.
297  *
298  * m_daddr is used by cancellation to prune messages to a given destination.
299  *
300  * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
301  * nesting.  As paths iterate over messages on a sock, or conn, they must
302  * also lock the conn, or sock, to remove the message from those lists too.
303  * Testing the flag to determine if the message is still on the lists lets
304  * us avoid testing the list_head directly.  That means each path can use
305  * the message's list_head to keep it on a local list while juggling locks
306  * without confusing the other path.
307  *
308  * m_ack_seq is an optional field set by transports who need a different
309  * sequence number range to invalidate.  They can use this in a callback
310  * that they pass to rds_send_drop_acked() to see if each message has been
311  * acked.  The HAS_ACK_SEQ flag can be used to detect messages which haven't
312  * had ack_seq set yet.
313  */
314 #define RDS_MSG_ON_SOCK		1
315 #define RDS_MSG_ON_CONN		2
316 #define RDS_MSG_HAS_ACK_SEQ	3
317 #define RDS_MSG_ACK_REQUIRED	4
318 #define RDS_MSG_RETRANSMITTED	5
319 #define RDS_MSG_MAPPED		6
320 #define RDS_MSG_PAGEVEC		7
321 
322 struct rds_message {
323 	atomic_t		m_refcount;
324 	struct list_head	m_sock_item;
325 	struct list_head	m_conn_item;
326 	struct rds_incoming	m_inc;
327 	u64			m_ack_seq;
328 	__be32			m_daddr;
329 	unsigned long		m_flags;
330 
331 	/* Never access m_rs without holding m_rs_lock.
332 	 * Lock nesting is
333 	 *  rm->m_rs_lock
334 	 *   -> rs->rs_lock
335 	 */
336 	spinlock_t		m_rs_lock;
337 	wait_queue_head_t	m_flush_wait;
338 
339 	struct rds_sock		*m_rs;
340 
341 	/* cookie to send to remote, in rds header */
342 	rds_rdma_cookie_t	m_rdma_cookie;
343 
344 	unsigned int		m_used_sgs;
345 	unsigned int		m_total_sgs;
346 
347 	void			*m_final_op;
348 
349 	struct {
350 		struct rm_atomic_op {
351 			int			op_type;
352 			union {
353 				struct {
354 					uint64_t	compare;
355 					uint64_t	swap;
356 					uint64_t	compare_mask;
357 					uint64_t	swap_mask;
358 				} op_m_cswp;
359 				struct {
360 					uint64_t	add;
361 					uint64_t	nocarry_mask;
362 				} op_m_fadd;
363 			};
364 
365 			u32			op_rkey;
366 			u64			op_remote_addr;
367 			unsigned int		op_notify:1;
368 			unsigned int		op_recverr:1;
369 			unsigned int		op_mapped:1;
370 			unsigned int		op_silent:1;
371 			unsigned int		op_active:1;
372 			struct scatterlist	*op_sg;
373 			struct rds_notifier	*op_notifier;
374 
375 			struct rds_mr		*op_rdma_mr;
376 		} atomic;
377 		struct rm_rdma_op {
378 			u32			op_rkey;
379 			u64			op_remote_addr;
380 			unsigned int		op_write:1;
381 			unsigned int		op_fence:1;
382 			unsigned int		op_notify:1;
383 			unsigned int		op_recverr:1;
384 			unsigned int		op_mapped:1;
385 			unsigned int		op_silent:1;
386 			unsigned int		op_active:1;
387 			unsigned int		op_bytes;
388 			unsigned int		op_nents;
389 			unsigned int		op_count;
390 			struct scatterlist	*op_sg;
391 			struct rds_notifier	*op_notifier;
392 
393 			struct rds_mr		*op_rdma_mr;
394 		} rdma;
395 		struct rm_data_op {
396 			unsigned int		op_active:1;
397 			unsigned int		op_nents;
398 			unsigned int		op_count;
399 			unsigned int		op_dmasg;
400 			unsigned int		op_dmaoff;
401 			struct scatterlist	*op_sg;
402 		} data;
403 	};
404 };
405 
406 /*
407  * The RDS notifier is used (optionally) to tell the application about
408  * completed RDMA operations. Rather than keeping the whole rds message
409  * around on the queue, we allocate a small notifier that is put on the
410  * socket's notifier_list. Notifications are delivered to the application
411  * through control messages.
412  */
413 struct rds_notifier {
414 	struct list_head	n_list;
415 	uint64_t		n_user_token;
416 	int			n_status;
417 };
418 
419 /**
420  * struct rds_transport -  transport specific behavioural hooks
421  *
422  * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
423  *        part of a message.  The caller serializes on the send_sem so this
424  *        doesn't need to be reentrant for a given conn.  The header must be
425  *        sent before the data payload.  .xmit must be prepared to send a
426  *        message with no data payload.  .xmit should return the number of
427  *        bytes that were sent down the connection, including header bytes.
428  *        Returning 0 tells the caller that it doesn't need to perform any
429  *        additional work now.  This is usually the case when the transport has
430  *        filled the sending queue for its connection and will handle
431  *        triggering the rds thread to continue the send when space becomes
432  *        available.  Returning -EAGAIN tells the caller to retry the send
433  *        immediately.  Returning -ENOMEM tells the caller to retry the send at
434  *        some point in the future.
435  *
436  * @conn_shutdown: conn_shutdown stops traffic on the given connection.  Once
437  *                 it returns the connection can not call rds_recv_incoming().
438  *                 This will only be called once after conn_connect returns
439  *                 non-zero success and will The caller serializes this with
440  *                 the send and connecting paths (xmit_* and conn_*).  The
441  *                 transport is responsible for other serialization, including
442  *                 rds_recv_incoming().  This is called in process context but
443  *                 should try hard not to block.
444  */
445 
446 struct rds_transport {
447 	char			t_name[TRANSNAMSIZ];
448 	struct list_head	t_item;
449 	struct module		*t_owner;
450 	unsigned int		t_prefer_loopback:1,
451 				t_mp_capable:1;
452 	unsigned int		t_type;
453 
454 	int (*laddr_check)(struct net *net, __be32 addr);
455 	int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
456 	void (*conn_free)(void *data);
457 	int (*conn_connect)(struct rds_connection *conn);
458 	void (*conn_shutdown)(struct rds_connection *conn);
459 	void (*xmit_prepare)(struct rds_connection *conn);
460 	void (*xmit_path_prepare)(struct rds_conn_path *cp);
461 	void (*xmit_complete)(struct rds_connection *conn);
462 	void (*xmit_path_complete)(struct rds_conn_path *cp);
463 	int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
464 		    unsigned int hdr_off, unsigned int sg, unsigned int off);
465 	int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
466 	int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
467 	int (*recv)(struct rds_connection *conn);
468 	int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
469 	void (*inc_free)(struct rds_incoming *inc);
470 
471 	int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
472 				 struct rdma_cm_event *event);
473 	int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
474 	void (*cm_connect_complete)(struct rds_connection *conn,
475 				    struct rdma_cm_event *event);
476 
477 	unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
478 					unsigned int avail);
479 	void (*exit)(void);
480 	void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
481 			struct rds_sock *rs, u32 *key_ret);
482 	void (*sync_mr)(void *trans_private, int direction);
483 	void (*free_mr)(void *trans_private, int invalidate);
484 	void (*flush_mrs)(void);
485 };
486 
487 struct rds_sock {
488 	struct sock		rs_sk;
489 
490 	u64			rs_user_addr;
491 	u64			rs_user_bytes;
492 
493 	/*
494 	 * bound_addr used for both incoming and outgoing, no INADDR_ANY
495 	 * support.
496 	 */
497 	struct rhash_head	rs_bound_node;
498 	u64			rs_bound_key;
499 	__be32			rs_bound_addr;
500 	__be32			rs_conn_addr;
501 	__be16			rs_bound_port;
502 	__be16			rs_conn_port;
503 	struct rds_transport    *rs_transport;
504 
505 	/*
506 	 * rds_sendmsg caches the conn it used the last time around.
507 	 * This helps avoid costly lookups.
508 	 */
509 	struct rds_connection	*rs_conn;
510 
511 	/* flag indicating we were congested or not */
512 	int			rs_congested;
513 	/* seen congestion (ENOBUFS) when sending? */
514 	int			rs_seen_congestion;
515 
516 	/* rs_lock protects all these adjacent members before the newline */
517 	spinlock_t		rs_lock;
518 	struct list_head	rs_send_queue;
519 	u32			rs_snd_bytes;
520 	int			rs_rcv_bytes;
521 	struct list_head	rs_notify_queue;	/* currently used for failed RDMAs */
522 
523 	/* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
524 	 * to decide whether the application should be woken up.
525 	 * If not set, we use rs_cong_track to find out whether a cong map
526 	 * update arrived.
527 	 */
528 	uint64_t		rs_cong_mask;
529 	uint64_t		rs_cong_notify;
530 	struct list_head	rs_cong_list;
531 	unsigned long		rs_cong_track;
532 
533 	/*
534 	 * rs_recv_lock protects the receive queue, and is
535 	 * used to serialize with rds_release.
536 	 */
537 	rwlock_t		rs_recv_lock;
538 	struct list_head	rs_recv_queue;
539 
540 	/* just for stats reporting */
541 	struct list_head	rs_item;
542 
543 	/* these have their own lock */
544 	spinlock_t		rs_rdma_lock;
545 	struct rb_root		rs_rdma_keys;
546 
547 	/* Socket options - in case there will be more */
548 	unsigned char		rs_recverr,
549 				rs_cong_monitor;
550 };
551 
552 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
553 {
554 	return container_of(sk, struct rds_sock, rs_sk);
555 }
556 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
557 {
558 	return &rs->rs_sk;
559 }
560 
561 /*
562  * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
563  * to account for overhead.  We don't account for overhead, we just apply
564  * the number of payload bytes to the specified value.
565  */
566 static inline int rds_sk_sndbuf(struct rds_sock *rs)
567 {
568 	return rds_rs_to_sk(rs)->sk_sndbuf / 2;
569 }
570 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
571 {
572 	return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
573 }
574 
575 struct rds_statistics {
576 	uint64_t	s_conn_reset;
577 	uint64_t	s_recv_drop_bad_checksum;
578 	uint64_t	s_recv_drop_old_seq;
579 	uint64_t	s_recv_drop_no_sock;
580 	uint64_t	s_recv_drop_dead_sock;
581 	uint64_t	s_recv_deliver_raced;
582 	uint64_t	s_recv_delivered;
583 	uint64_t	s_recv_queued;
584 	uint64_t	s_recv_immediate_retry;
585 	uint64_t	s_recv_delayed_retry;
586 	uint64_t	s_recv_ack_required;
587 	uint64_t	s_recv_rdma_bytes;
588 	uint64_t	s_recv_ping;
589 	uint64_t	s_send_queue_empty;
590 	uint64_t	s_send_queue_full;
591 	uint64_t	s_send_lock_contention;
592 	uint64_t	s_send_lock_queue_raced;
593 	uint64_t	s_send_immediate_retry;
594 	uint64_t	s_send_delayed_retry;
595 	uint64_t	s_send_drop_acked;
596 	uint64_t	s_send_ack_required;
597 	uint64_t	s_send_queued;
598 	uint64_t	s_send_rdma;
599 	uint64_t	s_send_rdma_bytes;
600 	uint64_t	s_send_pong;
601 	uint64_t	s_page_remainder_hit;
602 	uint64_t	s_page_remainder_miss;
603 	uint64_t	s_copy_to_user;
604 	uint64_t	s_copy_from_user;
605 	uint64_t	s_cong_update_queued;
606 	uint64_t	s_cong_update_received;
607 	uint64_t	s_cong_send_error;
608 	uint64_t	s_cong_send_blocked;
609 };
610 
611 /* af_rds.c */
612 void rds_sock_addref(struct rds_sock *rs);
613 void rds_sock_put(struct rds_sock *rs);
614 void rds_wake_sk_sleep(struct rds_sock *rs);
615 static inline void __rds_wake_sk_sleep(struct sock *sk)
616 {
617 	wait_queue_head_t *waitq = sk_sleep(sk);
618 
619 	if (!sock_flag(sk, SOCK_DEAD) && waitq)
620 		wake_up(waitq);
621 }
622 extern wait_queue_head_t rds_poll_waitq;
623 
624 
625 /* bind.c */
626 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
627 void rds_remove_bound(struct rds_sock *rs);
628 struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
629 int rds_bind_lock_init(void);
630 void rds_bind_lock_destroy(void);
631 
632 /* cong.c */
633 int rds_cong_get_maps(struct rds_connection *conn);
634 void rds_cong_add_conn(struct rds_connection *conn);
635 void rds_cong_remove_conn(struct rds_connection *conn);
636 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
637 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
638 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
639 void rds_cong_queue_updates(struct rds_cong_map *map);
640 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
641 int rds_cong_updated_since(unsigned long *recent);
642 void rds_cong_add_socket(struct rds_sock *);
643 void rds_cong_remove_socket(struct rds_sock *);
644 void rds_cong_exit(void);
645 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
646 
647 /* conn.c */
648 int rds_conn_init(void);
649 void rds_conn_exit(void);
650 struct rds_connection *rds_conn_create(struct net *net,
651 				       __be32 laddr, __be32 faddr,
652 				       struct rds_transport *trans, gfp_t gfp);
653 struct rds_connection *rds_conn_create_outgoing(struct net *net,
654 						__be32 laddr, __be32 faddr,
655 			       struct rds_transport *trans, gfp_t gfp);
656 void rds_conn_shutdown(struct rds_connection *conn);
657 void rds_conn_destroy(struct rds_connection *conn);
658 void rds_conn_drop(struct rds_connection *conn);
659 void rds_conn_path_drop(struct rds_conn_path *cpath);
660 void rds_conn_connect_if_down(struct rds_connection *conn);
661 void rds_conn_path_connect_if_down(struct rds_conn_path *cp);
662 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
663 			  struct rds_info_iterator *iter,
664 			  struct rds_info_lengths *lens,
665 			  int (*visitor)(struct rds_connection *, void *),
666 			  size_t item_len);
667 __printf(2, 3)
668 void __rds_conn_error(struct rds_connection *conn, const char *, ...);
669 #define rds_conn_error(conn, fmt...) \
670 	__rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
671 
672 void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...);
673 #define rds_conn_path_error(cp, fmt...) \
674 	__rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt)
675 
676 static inline int
677 rds_conn_path_transition(struct rds_conn_path *cp, int old, int new)
678 {
679 	return atomic_cmpxchg(&cp->cp_state, old, new) == old;
680 }
681 
682 static inline int
683 rds_conn_transition(struct rds_connection *conn, int old, int new)
684 {
685 	WARN_ON(conn->c_trans->t_mp_capable);
686 	return rds_conn_path_transition(&conn->c_path[0], old, new);
687 }
688 
689 static inline int
690 rds_conn_path_state(struct rds_conn_path *cp)
691 {
692 	return atomic_read(&cp->cp_state);
693 }
694 
695 static inline int
696 rds_conn_state(struct rds_connection *conn)
697 {
698 	WARN_ON(conn->c_trans->t_mp_capable);
699 	return rds_conn_path_state(&conn->c_path[0]);
700 }
701 
702 static inline int
703 rds_conn_path_up(struct rds_conn_path *cp)
704 {
705 	return atomic_read(&cp->cp_state) == RDS_CONN_UP;
706 }
707 
708 static inline int
709 rds_conn_up(struct rds_connection *conn)
710 {
711 	WARN_ON(conn->c_trans->t_mp_capable);
712 	return rds_conn_path_up(&conn->c_path[0]);
713 }
714 
715 static inline int
716 rds_conn_path_connecting(struct rds_conn_path *cp)
717 {
718 	return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING;
719 }
720 
721 static inline int
722 rds_conn_connecting(struct rds_connection *conn)
723 {
724 	WARN_ON(conn->c_trans->t_mp_capable);
725 	return rds_conn_path_connecting(&conn->c_path[0]);
726 }
727 
728 /* message.c */
729 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
730 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
731 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from);
732 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
733 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
734 				 __be16 dport, u64 seq);
735 int rds_message_add_extension(struct rds_header *hdr,
736 			      unsigned int type, const void *data, unsigned int len);
737 int rds_message_next_extension(struct rds_header *hdr,
738 			       unsigned int *pos, void *buf, unsigned int *buflen);
739 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
740 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
741 void rds_message_inc_free(struct rds_incoming *inc);
742 void rds_message_addref(struct rds_message *rm);
743 void rds_message_put(struct rds_message *rm);
744 void rds_message_wait(struct rds_message *rm);
745 void rds_message_unmapped(struct rds_message *rm);
746 
747 static inline void rds_message_make_checksum(struct rds_header *hdr)
748 {
749 	hdr->h_csum = 0;
750 	hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
751 }
752 
753 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
754 {
755 	return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
756 }
757 
758 
759 /* page.c */
760 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
761 			     gfp_t gfp);
762 int rds_page_copy_user(struct page *page, unsigned long offset,
763 		       void __user *ptr, unsigned long bytes,
764 		       int to_user);
765 #define rds_page_copy_to_user(page, offset, ptr, bytes) \
766 	rds_page_copy_user(page, offset, ptr, bytes, 1)
767 #define rds_page_copy_from_user(page, offset, ptr, bytes) \
768 	rds_page_copy_user(page, offset, ptr, bytes, 0)
769 void rds_page_exit(void);
770 
771 /* recv.c */
772 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
773 		  __be32 saddr);
774 void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn,
775 		       __be32 saddr);
776 void rds_inc_put(struct rds_incoming *inc);
777 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
778 		       struct rds_incoming *inc, gfp_t gfp);
779 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
780 		int msg_flags);
781 void rds_clear_recv_queue(struct rds_sock *rs);
782 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
783 void rds_inc_info_copy(struct rds_incoming *inc,
784 		       struct rds_info_iterator *iter,
785 		       __be32 saddr, __be32 daddr, int flip);
786 
787 /* send.c */
788 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
789 void rds_send_reset(struct rds_connection *conn);
790 int rds_send_xmit(struct rds_conn_path *cp);
791 struct sockaddr_in;
792 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
793 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
794 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
795 			 is_acked_func is_acked);
796 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
797 			      is_acked_func is_acked);
798 int rds_send_pong(struct rds_conn_path *cp, __be16 dport);
799 
800 /* rdma.c */
801 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
802 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
803 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
804 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
805 void rds_rdma_drop_keys(struct rds_sock *rs);
806 int rds_rdma_extra_size(struct rds_rdma_args *args);
807 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
808 			  struct cmsghdr *cmsg);
809 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
810 			  struct cmsghdr *cmsg);
811 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
812 			  struct cmsghdr *cmsg);
813 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
814 			  struct cmsghdr *cmsg);
815 void rds_rdma_free_op(struct rm_rdma_op *ro);
816 void rds_atomic_free_op(struct rm_atomic_op *ao);
817 void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
818 void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
819 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
820 		    struct cmsghdr *cmsg);
821 
822 void __rds_put_mr_final(struct rds_mr *mr);
823 static inline void rds_mr_put(struct rds_mr *mr)
824 {
825 	if (atomic_dec_and_test(&mr->r_refcount))
826 		__rds_put_mr_final(mr);
827 }
828 
829 /* stats.c */
830 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
831 #define rds_stats_inc_which(which, member) do {		\
832 	per_cpu(which, get_cpu()).member++;		\
833 	put_cpu();					\
834 } while (0)
835 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
836 #define rds_stats_add_which(which, member, count) do {		\
837 	per_cpu(which, get_cpu()).member += count;	\
838 	put_cpu();					\
839 } while (0)
840 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
841 int rds_stats_init(void);
842 void rds_stats_exit(void);
843 void rds_stats_info_copy(struct rds_info_iterator *iter,
844 			 uint64_t *values, const char *const *names,
845 			 size_t nr);
846 
847 /* sysctl.c */
848 int rds_sysctl_init(void);
849 void rds_sysctl_exit(void);
850 extern unsigned long rds_sysctl_sndbuf_min;
851 extern unsigned long rds_sysctl_sndbuf_default;
852 extern unsigned long rds_sysctl_sndbuf_max;
853 extern unsigned long rds_sysctl_reconnect_min_jiffies;
854 extern unsigned long rds_sysctl_reconnect_max_jiffies;
855 extern unsigned int  rds_sysctl_max_unacked_packets;
856 extern unsigned int  rds_sysctl_max_unacked_bytes;
857 extern unsigned int  rds_sysctl_ping_enable;
858 extern unsigned long rds_sysctl_trace_flags;
859 extern unsigned int  rds_sysctl_trace_level;
860 
861 /* threads.c */
862 int rds_threads_init(void);
863 void rds_threads_exit(void);
864 extern struct workqueue_struct *rds_wq;
865 void rds_queue_reconnect(struct rds_conn_path *cp);
866 void rds_connect_worker(struct work_struct *);
867 void rds_shutdown_worker(struct work_struct *);
868 void rds_send_worker(struct work_struct *);
869 void rds_recv_worker(struct work_struct *);
870 void rds_connect_path_complete(struct rds_conn_path *conn, int curr);
871 void rds_connect_complete(struct rds_connection *conn);
872 
873 /* transport.c */
874 int rds_trans_register(struct rds_transport *trans);
875 void rds_trans_unregister(struct rds_transport *trans);
876 struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr);
877 void rds_trans_put(struct rds_transport *trans);
878 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
879 				       unsigned int avail);
880 struct rds_transport *rds_trans_get(int t_type);
881 int rds_trans_init(void);
882 void rds_trans_exit(void);
883 
884 #endif
885