xref: /openbmc/linux/net/rds/rds.h (revision 77a87824)
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	8
89 #define	RDS_MPATH_HASH(rs, n) (jhash_1word((rs)->rs_bound_port, \
90 			       (rs)->rs_hash_initval) & ((n) - 1))
91 
92 /* Per mpath connection state */
93 struct rds_conn_path {
94 	struct rds_connection	*cp_conn;
95 	struct rds_message	*cp_xmit_rm;
96 	unsigned long		cp_xmit_sg;
97 	unsigned int		cp_xmit_hdr_off;
98 	unsigned int		cp_xmit_data_off;
99 	unsigned int		cp_xmit_atomic_sent;
100 	unsigned int		cp_xmit_rdma_sent;
101 	unsigned int		cp_xmit_data_sent;
102 
103 	spinlock_t		cp_lock;		/* protect msg queues */
104 	u64			cp_next_tx_seq;
105 	struct list_head	cp_send_queue;
106 	struct list_head	cp_retrans;
107 
108 	u64			cp_next_rx_seq;
109 
110 	void			*cp_transport_data;
111 
112 	atomic_t		cp_state;
113 	unsigned long		cp_send_gen;
114 	unsigned long		cp_flags;
115 	unsigned long		cp_reconnect_jiffies;
116 	struct delayed_work	cp_send_w;
117 	struct delayed_work	cp_recv_w;
118 	struct delayed_work	cp_conn_w;
119 	struct work_struct	cp_down_w;
120 	struct mutex		cp_cm_lock;	/* protect cp_state & cm */
121 	wait_queue_head_t	cp_waitq;
122 
123 	unsigned int		cp_unacked_packets;
124 	unsigned int		cp_unacked_bytes;
125 	unsigned int		cp_outgoing:1,
126 				cp_pad_to_32:31;
127 	unsigned int		cp_index;
128 };
129 
130 /* One rds_connection per RDS address pair */
131 struct rds_connection {
132 	struct hlist_node	c_hash_node;
133 	__be32			c_laddr;
134 	__be32			c_faddr;
135 	unsigned int		c_loopback:1,
136 				c_ping_triggered:1,
137 				c_pad_to_32:30;
138 	int			c_npaths;
139 	struct rds_connection	*c_passive;
140 	struct rds_transport	*c_trans;
141 
142 	struct rds_cong_map	*c_lcong;
143 	struct rds_cong_map	*c_fcong;
144 
145 	/* Protocol version */
146 	unsigned int		c_version;
147 	possible_net_t		c_net;
148 
149 	struct list_head	c_map_item;
150 	unsigned long		c_map_queued;
151 
152 	struct rds_conn_path	c_path[RDS_MPATH_WORKERS];
153 	wait_queue_head_t	c_hs_waitq; /* handshake waitq */
154 };
155 
156 static inline
157 struct net *rds_conn_net(struct rds_connection *conn)
158 {
159 	return read_pnet(&conn->c_net);
160 }
161 
162 static inline
163 void rds_conn_net_set(struct rds_connection *conn, struct net *net)
164 {
165 	write_pnet(&conn->c_net, net);
166 }
167 
168 #define RDS_FLAG_CONG_BITMAP	0x01
169 #define RDS_FLAG_ACK_REQUIRED	0x02
170 #define RDS_FLAG_RETRANSMITTED	0x04
171 #define RDS_MAX_ADV_CREDIT	255
172 
173 /* RDS_FLAG_PROBE_PORT is the reserved sport used for sending a ping
174  * probe to exchange control information before establishing a connection.
175  * Currently the control information that is exchanged is the number of
176  * supported paths. If the peer is a legacy (older kernel revision) peer,
177  * it would return a pong message without additional control information
178  * that would then alert the sender that the peer was an older rev.
179  */
180 #define RDS_FLAG_PROBE_PORT	1
181 #define	RDS_HS_PROBE(sport, dport) \
182 		((sport == RDS_FLAG_PROBE_PORT && dport == 0) || \
183 		 (sport == 0 && dport == RDS_FLAG_PROBE_PORT))
184 /*
185  * Maximum space available for extension headers.
186  */
187 #define RDS_HEADER_EXT_SPACE	16
188 
189 struct rds_header {
190 	__be64	h_sequence;
191 	__be64	h_ack;
192 	__be32	h_len;
193 	__be16	h_sport;
194 	__be16	h_dport;
195 	u8	h_flags;
196 	u8	h_credit;
197 	u8	h_padding[4];
198 	__sum16	h_csum;
199 
200 	u8	h_exthdr[RDS_HEADER_EXT_SPACE];
201 };
202 
203 /*
204  * Reserved - indicates end of extensions
205  */
206 #define RDS_EXTHDR_NONE		0
207 
208 /*
209  * This extension header is included in the very
210  * first message that is sent on a new connection,
211  * and identifies the protocol level. This will help
212  * rolling updates if a future change requires breaking
213  * the protocol.
214  * NB: This is no longer true for IB, where we do a version
215  * negotiation during the connection setup phase (protocol
216  * version information is included in the RDMA CM private data).
217  */
218 #define RDS_EXTHDR_VERSION	1
219 struct rds_ext_header_version {
220 	__be32			h_version;
221 };
222 
223 /*
224  * This extension header is included in the RDS message
225  * chasing an RDMA operation.
226  */
227 #define RDS_EXTHDR_RDMA		2
228 struct rds_ext_header_rdma {
229 	__be32			h_rdma_rkey;
230 };
231 
232 /*
233  * This extension header tells the peer about the
234  * destination <R_Key,offset> of the requested RDMA
235  * operation.
236  */
237 #define RDS_EXTHDR_RDMA_DEST	3
238 struct rds_ext_header_rdma_dest {
239 	__be32			h_rdma_rkey;
240 	__be32			h_rdma_offset;
241 };
242 
243 /* Extension header announcing number of paths.
244  * Implicit length = 2 bytes.
245  */
246 #define RDS_EXTHDR_NPATHS	4
247 
248 #define __RDS_EXTHDR_MAX	16 /* for now */
249 
250 struct rds_incoming {
251 	atomic_t		i_refcount;
252 	struct list_head	i_item;
253 	struct rds_connection	*i_conn;
254 	struct rds_conn_path	*i_conn_path;
255 	struct rds_header	i_hdr;
256 	unsigned long		i_rx_jiffies;
257 	__be32			i_saddr;
258 
259 	rds_rdma_cookie_t	i_rdma_cookie;
260 	struct timeval		i_rx_tstamp;
261 };
262 
263 struct rds_mr {
264 	struct rb_node		r_rb_node;
265 	atomic_t		r_refcount;
266 	u32			r_key;
267 
268 	/* A copy of the creation flags */
269 	unsigned int		r_use_once:1;
270 	unsigned int		r_invalidate:1;
271 	unsigned int		r_write:1;
272 
273 	/* This is for RDS_MR_DEAD.
274 	 * It would be nice & consistent to make this part of the above
275 	 * bit field here, but we need to use test_and_set_bit.
276 	 */
277 	unsigned long		r_state;
278 	struct rds_sock		*r_sock; /* back pointer to the socket that owns us */
279 	struct rds_transport	*r_trans;
280 	void			*r_trans_private;
281 };
282 
283 /* Flags for mr->r_state */
284 #define RDS_MR_DEAD		0
285 
286 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
287 {
288 	return r_key | (((u64) offset) << 32);
289 }
290 
291 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
292 {
293 	return cookie;
294 }
295 
296 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
297 {
298 	return cookie >> 32;
299 }
300 
301 /* atomic operation types */
302 #define RDS_ATOMIC_TYPE_CSWP		0
303 #define RDS_ATOMIC_TYPE_FADD		1
304 
305 /*
306  * m_sock_item and m_conn_item are on lists that are serialized under
307  * conn->c_lock.  m_sock_item has additional meaning in that once it is empty
308  * the message will not be put back on the retransmit list after being sent.
309  * messages that are canceled while being sent rely on this.
310  *
311  * m_inc is used by loopback so that it can pass an incoming message straight
312  * back up into the rx path.  It embeds a wire header which is also used by
313  * the send path, which is kind of awkward.
314  *
315  * m_sock_item indicates the message's presence on a socket's send or receive
316  * queue.  m_rs will point to that socket.
317  *
318  * m_daddr is used by cancellation to prune messages to a given destination.
319  *
320  * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
321  * nesting.  As paths iterate over messages on a sock, or conn, they must
322  * also lock the conn, or sock, to remove the message from those lists too.
323  * Testing the flag to determine if the message is still on the lists lets
324  * us avoid testing the list_head directly.  That means each path can use
325  * the message's list_head to keep it on a local list while juggling locks
326  * without confusing the other path.
327  *
328  * m_ack_seq is an optional field set by transports who need a different
329  * sequence number range to invalidate.  They can use this in a callback
330  * that they pass to rds_send_drop_acked() to see if each message has been
331  * acked.  The HAS_ACK_SEQ flag can be used to detect messages which haven't
332  * had ack_seq set yet.
333  */
334 #define RDS_MSG_ON_SOCK		1
335 #define RDS_MSG_ON_CONN		2
336 #define RDS_MSG_HAS_ACK_SEQ	3
337 #define RDS_MSG_ACK_REQUIRED	4
338 #define RDS_MSG_RETRANSMITTED	5
339 #define RDS_MSG_MAPPED		6
340 #define RDS_MSG_PAGEVEC		7
341 
342 struct rds_message {
343 	atomic_t		m_refcount;
344 	struct list_head	m_sock_item;
345 	struct list_head	m_conn_item;
346 	struct rds_incoming	m_inc;
347 	u64			m_ack_seq;
348 	__be32			m_daddr;
349 	unsigned long		m_flags;
350 
351 	/* Never access m_rs without holding m_rs_lock.
352 	 * Lock nesting is
353 	 *  rm->m_rs_lock
354 	 *   -> rs->rs_lock
355 	 */
356 	spinlock_t		m_rs_lock;
357 	wait_queue_head_t	m_flush_wait;
358 
359 	struct rds_sock		*m_rs;
360 
361 	/* cookie to send to remote, in rds header */
362 	rds_rdma_cookie_t	m_rdma_cookie;
363 
364 	unsigned int		m_used_sgs;
365 	unsigned int		m_total_sgs;
366 
367 	void			*m_final_op;
368 
369 	struct {
370 		struct rm_atomic_op {
371 			int			op_type;
372 			union {
373 				struct {
374 					uint64_t	compare;
375 					uint64_t	swap;
376 					uint64_t	compare_mask;
377 					uint64_t	swap_mask;
378 				} op_m_cswp;
379 				struct {
380 					uint64_t	add;
381 					uint64_t	nocarry_mask;
382 				} op_m_fadd;
383 			};
384 
385 			u32			op_rkey;
386 			u64			op_remote_addr;
387 			unsigned int		op_notify:1;
388 			unsigned int		op_recverr:1;
389 			unsigned int		op_mapped:1;
390 			unsigned int		op_silent:1;
391 			unsigned int		op_active:1;
392 			struct scatterlist	*op_sg;
393 			struct rds_notifier	*op_notifier;
394 
395 			struct rds_mr		*op_rdma_mr;
396 		} atomic;
397 		struct rm_rdma_op {
398 			u32			op_rkey;
399 			u64			op_remote_addr;
400 			unsigned int		op_write:1;
401 			unsigned int		op_fence:1;
402 			unsigned int		op_notify:1;
403 			unsigned int		op_recverr:1;
404 			unsigned int		op_mapped:1;
405 			unsigned int		op_silent:1;
406 			unsigned int		op_active:1;
407 			unsigned int		op_bytes;
408 			unsigned int		op_nents;
409 			unsigned int		op_count;
410 			struct scatterlist	*op_sg;
411 			struct rds_notifier	*op_notifier;
412 
413 			struct rds_mr		*op_rdma_mr;
414 		} rdma;
415 		struct rm_data_op {
416 			unsigned int		op_active:1;
417 			unsigned int		op_nents;
418 			unsigned int		op_count;
419 			unsigned int		op_dmasg;
420 			unsigned int		op_dmaoff;
421 			struct scatterlist	*op_sg;
422 		} data;
423 	};
424 };
425 
426 /*
427  * The RDS notifier is used (optionally) to tell the application about
428  * completed RDMA operations. Rather than keeping the whole rds message
429  * around on the queue, we allocate a small notifier that is put on the
430  * socket's notifier_list. Notifications are delivered to the application
431  * through control messages.
432  */
433 struct rds_notifier {
434 	struct list_head	n_list;
435 	uint64_t		n_user_token;
436 	int			n_status;
437 };
438 
439 /**
440  * struct rds_transport -  transport specific behavioural hooks
441  *
442  * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
443  *        part of a message.  The caller serializes on the send_sem so this
444  *        doesn't need to be reentrant for a given conn.  The header must be
445  *        sent before the data payload.  .xmit must be prepared to send a
446  *        message with no data payload.  .xmit should return the number of
447  *        bytes that were sent down the connection, including header bytes.
448  *        Returning 0 tells the caller that it doesn't need to perform any
449  *        additional work now.  This is usually the case when the transport has
450  *        filled the sending queue for its connection and will handle
451  *        triggering the rds thread to continue the send when space becomes
452  *        available.  Returning -EAGAIN tells the caller to retry the send
453  *        immediately.  Returning -ENOMEM tells the caller to retry the send at
454  *        some point in the future.
455  *
456  * @conn_shutdown: conn_shutdown stops traffic on the given connection.  Once
457  *                 it returns the connection can not call rds_recv_incoming().
458  *                 This will only be called once after conn_connect returns
459  *                 non-zero success and will The caller serializes this with
460  *                 the send and connecting paths (xmit_* and conn_*).  The
461  *                 transport is responsible for other serialization, including
462  *                 rds_recv_incoming().  This is called in process context but
463  *                 should try hard not to block.
464  */
465 
466 struct rds_transport {
467 	char			t_name[TRANSNAMSIZ];
468 	struct list_head	t_item;
469 	struct module		*t_owner;
470 	unsigned int		t_prefer_loopback:1,
471 				t_mp_capable:1;
472 	unsigned int		t_type;
473 
474 	int (*laddr_check)(struct net *net, __be32 addr);
475 	int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
476 	void (*conn_free)(void *data);
477 	int (*conn_path_connect)(struct rds_conn_path *cp);
478 	void (*conn_path_shutdown)(struct rds_conn_path *conn);
479 	void (*xmit_path_prepare)(struct rds_conn_path *cp);
480 	void (*xmit_path_complete)(struct rds_conn_path *cp);
481 	int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
482 		    unsigned int hdr_off, unsigned int sg, unsigned int off);
483 	int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
484 	int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
485 	int (*recv_path)(struct rds_conn_path *cp);
486 	int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
487 	void (*inc_free)(struct rds_incoming *inc);
488 
489 	int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
490 				 struct rdma_cm_event *event);
491 	int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
492 	void (*cm_connect_complete)(struct rds_connection *conn,
493 				    struct rdma_cm_event *event);
494 
495 	unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
496 					unsigned int avail);
497 	void (*exit)(void);
498 	void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
499 			struct rds_sock *rs, u32 *key_ret);
500 	void (*sync_mr)(void *trans_private, int direction);
501 	void (*free_mr)(void *trans_private, int invalidate);
502 	void (*flush_mrs)(void);
503 };
504 
505 struct rds_sock {
506 	struct sock		rs_sk;
507 
508 	u64			rs_user_addr;
509 	u64			rs_user_bytes;
510 
511 	/*
512 	 * bound_addr used for both incoming and outgoing, no INADDR_ANY
513 	 * support.
514 	 */
515 	struct rhash_head	rs_bound_node;
516 	u64			rs_bound_key;
517 	__be32			rs_bound_addr;
518 	__be32			rs_conn_addr;
519 	__be16			rs_bound_port;
520 	__be16			rs_conn_port;
521 	struct rds_transport    *rs_transport;
522 
523 	/*
524 	 * rds_sendmsg caches the conn it used the last time around.
525 	 * This helps avoid costly lookups.
526 	 */
527 	struct rds_connection	*rs_conn;
528 
529 	/* flag indicating we were congested or not */
530 	int			rs_congested;
531 	/* seen congestion (ENOBUFS) when sending? */
532 	int			rs_seen_congestion;
533 
534 	/* rs_lock protects all these adjacent members before the newline */
535 	spinlock_t		rs_lock;
536 	struct list_head	rs_send_queue;
537 	u32			rs_snd_bytes;
538 	int			rs_rcv_bytes;
539 	struct list_head	rs_notify_queue;	/* currently used for failed RDMAs */
540 
541 	/* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
542 	 * to decide whether the application should be woken up.
543 	 * If not set, we use rs_cong_track to find out whether a cong map
544 	 * update arrived.
545 	 */
546 	uint64_t		rs_cong_mask;
547 	uint64_t		rs_cong_notify;
548 	struct list_head	rs_cong_list;
549 	unsigned long		rs_cong_track;
550 
551 	/*
552 	 * rs_recv_lock protects the receive queue, and is
553 	 * used to serialize with rds_release.
554 	 */
555 	rwlock_t		rs_recv_lock;
556 	struct list_head	rs_recv_queue;
557 
558 	/* just for stats reporting */
559 	struct list_head	rs_item;
560 
561 	/* these have their own lock */
562 	spinlock_t		rs_rdma_lock;
563 	struct rb_root		rs_rdma_keys;
564 
565 	/* Socket options - in case there will be more */
566 	unsigned char		rs_recverr,
567 				rs_cong_monitor;
568 	u32			rs_hash_initval;
569 };
570 
571 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
572 {
573 	return container_of(sk, struct rds_sock, rs_sk);
574 }
575 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
576 {
577 	return &rs->rs_sk;
578 }
579 
580 /*
581  * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
582  * to account for overhead.  We don't account for overhead, we just apply
583  * the number of payload bytes to the specified value.
584  */
585 static inline int rds_sk_sndbuf(struct rds_sock *rs)
586 {
587 	return rds_rs_to_sk(rs)->sk_sndbuf / 2;
588 }
589 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
590 {
591 	return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
592 }
593 
594 struct rds_statistics {
595 	uint64_t	s_conn_reset;
596 	uint64_t	s_recv_drop_bad_checksum;
597 	uint64_t	s_recv_drop_old_seq;
598 	uint64_t	s_recv_drop_no_sock;
599 	uint64_t	s_recv_drop_dead_sock;
600 	uint64_t	s_recv_deliver_raced;
601 	uint64_t	s_recv_delivered;
602 	uint64_t	s_recv_queued;
603 	uint64_t	s_recv_immediate_retry;
604 	uint64_t	s_recv_delayed_retry;
605 	uint64_t	s_recv_ack_required;
606 	uint64_t	s_recv_rdma_bytes;
607 	uint64_t	s_recv_ping;
608 	uint64_t	s_send_queue_empty;
609 	uint64_t	s_send_queue_full;
610 	uint64_t	s_send_lock_contention;
611 	uint64_t	s_send_lock_queue_raced;
612 	uint64_t	s_send_immediate_retry;
613 	uint64_t	s_send_delayed_retry;
614 	uint64_t	s_send_drop_acked;
615 	uint64_t	s_send_ack_required;
616 	uint64_t	s_send_queued;
617 	uint64_t	s_send_rdma;
618 	uint64_t	s_send_rdma_bytes;
619 	uint64_t	s_send_pong;
620 	uint64_t	s_page_remainder_hit;
621 	uint64_t	s_page_remainder_miss;
622 	uint64_t	s_copy_to_user;
623 	uint64_t	s_copy_from_user;
624 	uint64_t	s_cong_update_queued;
625 	uint64_t	s_cong_update_received;
626 	uint64_t	s_cong_send_error;
627 	uint64_t	s_cong_send_blocked;
628 };
629 
630 /* af_rds.c */
631 void rds_sock_addref(struct rds_sock *rs);
632 void rds_sock_put(struct rds_sock *rs);
633 void rds_wake_sk_sleep(struct rds_sock *rs);
634 static inline void __rds_wake_sk_sleep(struct sock *sk)
635 {
636 	wait_queue_head_t *waitq = sk_sleep(sk);
637 
638 	if (!sock_flag(sk, SOCK_DEAD) && waitq)
639 		wake_up(waitq);
640 }
641 extern wait_queue_head_t rds_poll_waitq;
642 
643 
644 /* bind.c */
645 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
646 void rds_remove_bound(struct rds_sock *rs);
647 struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
648 int rds_bind_lock_init(void);
649 void rds_bind_lock_destroy(void);
650 
651 /* cong.c */
652 int rds_cong_get_maps(struct rds_connection *conn);
653 void rds_cong_add_conn(struct rds_connection *conn);
654 void rds_cong_remove_conn(struct rds_connection *conn);
655 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
656 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
657 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
658 void rds_cong_queue_updates(struct rds_cong_map *map);
659 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
660 int rds_cong_updated_since(unsigned long *recent);
661 void rds_cong_add_socket(struct rds_sock *);
662 void rds_cong_remove_socket(struct rds_sock *);
663 void rds_cong_exit(void);
664 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
665 
666 /* conn.c */
667 int rds_conn_init(void);
668 void rds_conn_exit(void);
669 struct rds_connection *rds_conn_create(struct net *net,
670 				       __be32 laddr, __be32 faddr,
671 				       struct rds_transport *trans, gfp_t gfp);
672 struct rds_connection *rds_conn_create_outgoing(struct net *net,
673 						__be32 laddr, __be32 faddr,
674 			       struct rds_transport *trans, gfp_t gfp);
675 void rds_conn_shutdown(struct rds_conn_path *cpath);
676 void rds_conn_destroy(struct rds_connection *conn);
677 void rds_conn_drop(struct rds_connection *conn);
678 void rds_conn_path_drop(struct rds_conn_path *cpath);
679 void rds_conn_connect_if_down(struct rds_connection *conn);
680 void rds_conn_path_connect_if_down(struct rds_conn_path *cp);
681 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
682 			  struct rds_info_iterator *iter,
683 			  struct rds_info_lengths *lens,
684 			  int (*visitor)(struct rds_connection *, void *),
685 			  size_t item_len);
686 __printf(2, 3)
687 void __rds_conn_error(struct rds_connection *conn, const char *, ...);
688 #define rds_conn_error(conn, fmt...) \
689 	__rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
690 
691 void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...);
692 #define rds_conn_path_error(cp, fmt...) \
693 	__rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt)
694 
695 static inline int
696 rds_conn_path_transition(struct rds_conn_path *cp, int old, int new)
697 {
698 	return atomic_cmpxchg(&cp->cp_state, old, new) == old;
699 }
700 
701 static inline int
702 rds_conn_transition(struct rds_connection *conn, int old, int new)
703 {
704 	WARN_ON(conn->c_trans->t_mp_capable);
705 	return rds_conn_path_transition(&conn->c_path[0], old, new);
706 }
707 
708 static inline int
709 rds_conn_path_state(struct rds_conn_path *cp)
710 {
711 	return atomic_read(&cp->cp_state);
712 }
713 
714 static inline int
715 rds_conn_state(struct rds_connection *conn)
716 {
717 	WARN_ON(conn->c_trans->t_mp_capable);
718 	return rds_conn_path_state(&conn->c_path[0]);
719 }
720 
721 static inline int
722 rds_conn_path_up(struct rds_conn_path *cp)
723 {
724 	return atomic_read(&cp->cp_state) == RDS_CONN_UP;
725 }
726 
727 static inline int
728 rds_conn_up(struct rds_connection *conn)
729 {
730 	WARN_ON(conn->c_trans->t_mp_capable);
731 	return rds_conn_path_up(&conn->c_path[0]);
732 }
733 
734 static inline int
735 rds_conn_path_connecting(struct rds_conn_path *cp)
736 {
737 	return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING;
738 }
739 
740 static inline int
741 rds_conn_connecting(struct rds_connection *conn)
742 {
743 	WARN_ON(conn->c_trans->t_mp_capable);
744 	return rds_conn_path_connecting(&conn->c_path[0]);
745 }
746 
747 /* message.c */
748 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
749 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
750 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from);
751 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
752 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
753 				 __be16 dport, u64 seq);
754 int rds_message_add_extension(struct rds_header *hdr,
755 			      unsigned int type, const void *data, unsigned int len);
756 int rds_message_next_extension(struct rds_header *hdr,
757 			       unsigned int *pos, void *buf, unsigned int *buflen);
758 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
759 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
760 void rds_message_inc_free(struct rds_incoming *inc);
761 void rds_message_addref(struct rds_message *rm);
762 void rds_message_put(struct rds_message *rm);
763 void rds_message_wait(struct rds_message *rm);
764 void rds_message_unmapped(struct rds_message *rm);
765 
766 static inline void rds_message_make_checksum(struct rds_header *hdr)
767 {
768 	hdr->h_csum = 0;
769 	hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
770 }
771 
772 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
773 {
774 	return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
775 }
776 
777 
778 /* page.c */
779 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
780 			     gfp_t gfp);
781 int rds_page_copy_user(struct page *page, unsigned long offset,
782 		       void __user *ptr, unsigned long bytes,
783 		       int to_user);
784 #define rds_page_copy_to_user(page, offset, ptr, bytes) \
785 	rds_page_copy_user(page, offset, ptr, bytes, 1)
786 #define rds_page_copy_from_user(page, offset, ptr, bytes) \
787 	rds_page_copy_user(page, offset, ptr, bytes, 0)
788 void rds_page_exit(void);
789 
790 /* recv.c */
791 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
792 		  __be32 saddr);
793 void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn,
794 		       __be32 saddr);
795 void rds_inc_put(struct rds_incoming *inc);
796 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
797 		       struct rds_incoming *inc, gfp_t gfp);
798 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
799 		int msg_flags);
800 void rds_clear_recv_queue(struct rds_sock *rs);
801 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
802 void rds_inc_info_copy(struct rds_incoming *inc,
803 		       struct rds_info_iterator *iter,
804 		       __be32 saddr, __be32 daddr, int flip);
805 
806 /* send.c */
807 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
808 void rds_send_path_reset(struct rds_conn_path *conn);
809 int rds_send_xmit(struct rds_conn_path *cp);
810 struct sockaddr_in;
811 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
812 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
813 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
814 			 is_acked_func is_acked);
815 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
816 			      is_acked_func is_acked);
817 int rds_send_pong(struct rds_conn_path *cp, __be16 dport);
818 
819 /* rdma.c */
820 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
821 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
822 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
823 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
824 void rds_rdma_drop_keys(struct rds_sock *rs);
825 int rds_rdma_extra_size(struct rds_rdma_args *args);
826 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
827 			  struct cmsghdr *cmsg);
828 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
829 			  struct cmsghdr *cmsg);
830 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
831 			  struct cmsghdr *cmsg);
832 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
833 			  struct cmsghdr *cmsg);
834 void rds_rdma_free_op(struct rm_rdma_op *ro);
835 void rds_atomic_free_op(struct rm_atomic_op *ao);
836 void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
837 void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
838 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
839 		    struct cmsghdr *cmsg);
840 
841 void __rds_put_mr_final(struct rds_mr *mr);
842 static inline void rds_mr_put(struct rds_mr *mr)
843 {
844 	if (atomic_dec_and_test(&mr->r_refcount))
845 		__rds_put_mr_final(mr);
846 }
847 
848 /* stats.c */
849 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
850 #define rds_stats_inc_which(which, member) do {		\
851 	per_cpu(which, get_cpu()).member++;		\
852 	put_cpu();					\
853 } while (0)
854 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
855 #define rds_stats_add_which(which, member, count) do {		\
856 	per_cpu(which, get_cpu()).member += count;	\
857 	put_cpu();					\
858 } while (0)
859 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
860 int rds_stats_init(void);
861 void rds_stats_exit(void);
862 void rds_stats_info_copy(struct rds_info_iterator *iter,
863 			 uint64_t *values, const char *const *names,
864 			 size_t nr);
865 
866 /* sysctl.c */
867 int rds_sysctl_init(void);
868 void rds_sysctl_exit(void);
869 extern unsigned long rds_sysctl_sndbuf_min;
870 extern unsigned long rds_sysctl_sndbuf_default;
871 extern unsigned long rds_sysctl_sndbuf_max;
872 extern unsigned long rds_sysctl_reconnect_min_jiffies;
873 extern unsigned long rds_sysctl_reconnect_max_jiffies;
874 extern unsigned int  rds_sysctl_max_unacked_packets;
875 extern unsigned int  rds_sysctl_max_unacked_bytes;
876 extern unsigned int  rds_sysctl_ping_enable;
877 extern unsigned long rds_sysctl_trace_flags;
878 extern unsigned int  rds_sysctl_trace_level;
879 
880 /* threads.c */
881 int rds_threads_init(void);
882 void rds_threads_exit(void);
883 extern struct workqueue_struct *rds_wq;
884 void rds_queue_reconnect(struct rds_conn_path *cp);
885 void rds_connect_worker(struct work_struct *);
886 void rds_shutdown_worker(struct work_struct *);
887 void rds_send_worker(struct work_struct *);
888 void rds_recv_worker(struct work_struct *);
889 void rds_connect_path_complete(struct rds_conn_path *conn, int curr);
890 void rds_connect_complete(struct rds_connection *conn);
891 
892 /* transport.c */
893 int rds_trans_register(struct rds_transport *trans);
894 void rds_trans_unregister(struct rds_transport *trans);
895 struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr);
896 void rds_trans_put(struct rds_transport *trans);
897 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
898 				       unsigned int avail);
899 struct rds_transport *rds_trans_get(int t_type);
900 int rds_trans_init(void);
901 void rds_trans_exit(void);
902 
903 #endif
904