xref: /openbmc/linux/net/rds/rds.h (revision bc05aa6e)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _RDS_RDS_H
3 #define _RDS_RDS_H
4 
5 #include <net/sock.h>
6 #include <linux/scatterlist.h>
7 #include <linux/highmem.h>
8 #include <rdma/rdma_cm.h>
9 #include <linux/mutex.h>
10 #include <linux/rds.h>
11 #include <linux/rhashtable.h>
12 #include <linux/refcount.h>
13 
14 #include "info.h"
15 
16 /*
17  * RDS Network protocol version
18  */
19 #define RDS_PROTOCOL_3_0	0x0300
20 #define RDS_PROTOCOL_3_1	0x0301
21 #define RDS_PROTOCOL_VERSION	RDS_PROTOCOL_3_1
22 #define RDS_PROTOCOL_MAJOR(v)	((v) >> 8)
23 #define RDS_PROTOCOL_MINOR(v)	((v) & 255)
24 #define RDS_PROTOCOL(maj, min)	(((maj) << 8) | min)
25 
26 /*
27  * XXX randomly chosen, but at least seems to be unused:
28  * #               18464-18768 Unassigned
29  * We should do better.  We want a reserved port to discourage unpriv'ed
30  * userspace from listening.
31  */
32 #define RDS_PORT	18634
33 
34 #ifdef ATOMIC64_INIT
35 #define KERNEL_HAS_ATOMIC64
36 #endif
37 
38 #ifdef RDS_DEBUG
39 #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
40 #else
41 /* sigh, pr_debug() causes unused variable warnings */
42 static inline __printf(1, 2)
43 void rdsdebug(char *fmt, ...)
44 {
45 }
46 #endif
47 
48 /* XXX is there one of these somewhere? */
49 #define ceil(x, y) \
50 	({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; })
51 
52 #define RDS_FRAG_SHIFT	12
53 #define RDS_FRAG_SIZE	((unsigned int)(1 << RDS_FRAG_SHIFT))
54 
55 /* Used to limit both RDMA and non-RDMA RDS message to 1MB */
56 #define RDS_MAX_MSG_SIZE	((unsigned int)(1 << 20))
57 
58 #define RDS_CONG_MAP_BYTES	(65536 / 8)
59 #define RDS_CONG_MAP_PAGES	(PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
60 #define RDS_CONG_MAP_PAGE_BITS	(PAGE_SIZE * 8)
61 
62 struct rds_cong_map {
63 	struct rb_node		m_rb_node;
64 	__be32			m_addr;
65 	wait_queue_head_t	m_waitq;
66 	struct list_head	m_conn_list;
67 	unsigned long		m_page_addrs[RDS_CONG_MAP_PAGES];
68 };
69 
70 
71 /*
72  * This is how we will track the connection state:
73  * A connection is always in one of the following
74  * states. Updates to the state are atomic and imply
75  * a memory barrier.
76  */
77 enum {
78 	RDS_CONN_DOWN = 0,
79 	RDS_CONN_CONNECTING,
80 	RDS_CONN_DISCONNECTING,
81 	RDS_CONN_UP,
82 	RDS_CONN_RESETTING,
83 	RDS_CONN_ERROR,
84 };
85 
86 /* Bits for c_flags */
87 #define RDS_LL_SEND_FULL	0
88 #define RDS_RECONNECT_PENDING	1
89 #define RDS_IN_XMIT		2
90 #define RDS_RECV_REFILL		3
91 #define	RDS_DESTROY_PENDING	4
92 
93 /* Max number of multipaths per RDS connection. Must be a power of 2 */
94 #define	RDS_MPATH_WORKERS	8
95 #define	RDS_MPATH_HASH(rs, n) (jhash_1word((rs)->rs_bound_port, \
96 			       (rs)->rs_hash_initval) & ((n) - 1))
97 
98 #define IS_CANONICAL(laddr, faddr) (htonl(laddr) < htonl(faddr))
99 
100 /* Per mpath connection state */
101 struct rds_conn_path {
102 	struct rds_connection	*cp_conn;
103 	struct rds_message	*cp_xmit_rm;
104 	unsigned long		cp_xmit_sg;
105 	unsigned int		cp_xmit_hdr_off;
106 	unsigned int		cp_xmit_data_off;
107 	unsigned int		cp_xmit_atomic_sent;
108 	unsigned int		cp_xmit_rdma_sent;
109 	unsigned int		cp_xmit_data_sent;
110 
111 	spinlock_t		cp_lock;		/* protect msg queues */
112 	u64			cp_next_tx_seq;
113 	struct list_head	cp_send_queue;
114 	struct list_head	cp_retrans;
115 
116 	u64			cp_next_rx_seq;
117 
118 	void			*cp_transport_data;
119 
120 	atomic_t		cp_state;
121 	unsigned long		cp_send_gen;
122 	unsigned long		cp_flags;
123 	unsigned long		cp_reconnect_jiffies;
124 	struct delayed_work	cp_send_w;
125 	struct delayed_work	cp_recv_w;
126 	struct delayed_work	cp_conn_w;
127 	struct work_struct	cp_down_w;
128 	struct mutex		cp_cm_lock;	/* protect cp_state & cm */
129 	wait_queue_head_t	cp_waitq;
130 
131 	unsigned int		cp_unacked_packets;
132 	unsigned int		cp_unacked_bytes;
133 	unsigned int		cp_index;
134 };
135 
136 /* One rds_connection per RDS address pair */
137 struct rds_connection {
138 	struct hlist_node	c_hash_node;
139 	__be32			c_laddr;
140 	__be32			c_faddr;
141 	unsigned int		c_loopback:1,
142 				c_ping_triggered:1,
143 				c_pad_to_32:30;
144 	int			c_npaths;
145 	struct rds_connection	*c_passive;
146 	struct rds_transport	*c_trans;
147 
148 	struct rds_cong_map	*c_lcong;
149 	struct rds_cong_map	*c_fcong;
150 
151 	/* Protocol version */
152 	unsigned int		c_version;
153 	possible_net_t		c_net;
154 
155 	struct list_head	c_map_item;
156 	unsigned long		c_map_queued;
157 
158 	struct rds_conn_path	*c_path;
159 	wait_queue_head_t	c_hs_waitq; /* handshake waitq */
160 
161 	u32			c_my_gen_num;
162 	u32			c_peer_gen_num;
163 };
164 
165 static inline
166 struct net *rds_conn_net(struct rds_connection *conn)
167 {
168 	return read_pnet(&conn->c_net);
169 }
170 
171 static inline
172 void rds_conn_net_set(struct rds_connection *conn, struct net *net)
173 {
174 	write_pnet(&conn->c_net, net);
175 }
176 
177 #define RDS_FLAG_CONG_BITMAP	0x01
178 #define RDS_FLAG_ACK_REQUIRED	0x02
179 #define RDS_FLAG_RETRANSMITTED	0x04
180 #define RDS_MAX_ADV_CREDIT	255
181 
182 /* RDS_FLAG_PROBE_PORT is the reserved sport used for sending a ping
183  * probe to exchange control information before establishing a connection.
184  * Currently the control information that is exchanged is the number of
185  * supported paths. If the peer is a legacy (older kernel revision) peer,
186  * it would return a pong message without additional control information
187  * that would then alert the sender that the peer was an older rev.
188  */
189 #define RDS_FLAG_PROBE_PORT	1
190 #define	RDS_HS_PROBE(sport, dport) \
191 		((sport == RDS_FLAG_PROBE_PORT && dport == 0) || \
192 		 (sport == 0 && dport == RDS_FLAG_PROBE_PORT))
193 /*
194  * Maximum space available for extension headers.
195  */
196 #define RDS_HEADER_EXT_SPACE	16
197 
198 struct rds_header {
199 	__be64	h_sequence;
200 	__be64	h_ack;
201 	__be32	h_len;
202 	__be16	h_sport;
203 	__be16	h_dport;
204 	u8	h_flags;
205 	u8	h_credit;
206 	u8	h_padding[4];
207 	__sum16	h_csum;
208 
209 	u8	h_exthdr[RDS_HEADER_EXT_SPACE];
210 };
211 
212 /*
213  * Reserved - indicates end of extensions
214  */
215 #define RDS_EXTHDR_NONE		0
216 
217 /*
218  * This extension header is included in the very
219  * first message that is sent on a new connection,
220  * and identifies the protocol level. This will help
221  * rolling updates if a future change requires breaking
222  * the protocol.
223  * NB: This is no longer true for IB, where we do a version
224  * negotiation during the connection setup phase (protocol
225  * version information is included in the RDMA CM private data).
226  */
227 #define RDS_EXTHDR_VERSION	1
228 struct rds_ext_header_version {
229 	__be32			h_version;
230 };
231 
232 /*
233  * This extension header is included in the RDS message
234  * chasing an RDMA operation.
235  */
236 #define RDS_EXTHDR_RDMA		2
237 struct rds_ext_header_rdma {
238 	__be32			h_rdma_rkey;
239 };
240 
241 /*
242  * This extension header tells the peer about the
243  * destination <R_Key,offset> of the requested RDMA
244  * operation.
245  */
246 #define RDS_EXTHDR_RDMA_DEST	3
247 struct rds_ext_header_rdma_dest {
248 	__be32			h_rdma_rkey;
249 	__be32			h_rdma_offset;
250 };
251 
252 /* Extension header announcing number of paths.
253  * Implicit length = 2 bytes.
254  */
255 #define RDS_EXTHDR_NPATHS	5
256 #define RDS_EXTHDR_GEN_NUM	6
257 
258 #define __RDS_EXTHDR_MAX	16 /* for now */
259 #define RDS_RX_MAX_TRACES	(RDS_MSG_RX_DGRAM_TRACE_MAX + 1)
260 #define	RDS_MSG_RX_HDR		0
261 #define	RDS_MSG_RX_START	1
262 #define	RDS_MSG_RX_END		2
263 #define	RDS_MSG_RX_CMSG		3
264 
265 struct rds_incoming {
266 	refcount_t		i_refcount;
267 	struct list_head	i_item;
268 	struct rds_connection	*i_conn;
269 	struct rds_conn_path	*i_conn_path;
270 	struct rds_header	i_hdr;
271 	unsigned long		i_rx_jiffies;
272 	__be32			i_saddr;
273 
274 	rds_rdma_cookie_t	i_rdma_cookie;
275 	struct timeval		i_rx_tstamp;
276 	u64			i_rx_lat_trace[RDS_RX_MAX_TRACES];
277 };
278 
279 struct rds_mr {
280 	struct rb_node		r_rb_node;
281 	refcount_t		r_refcount;
282 	u32			r_key;
283 
284 	/* A copy of the creation flags */
285 	unsigned int		r_use_once:1;
286 	unsigned int		r_invalidate:1;
287 	unsigned int		r_write:1;
288 
289 	/* This is for RDS_MR_DEAD.
290 	 * It would be nice & consistent to make this part of the above
291 	 * bit field here, but we need to use test_and_set_bit.
292 	 */
293 	unsigned long		r_state;
294 	struct rds_sock		*r_sock; /* back pointer to the socket that owns us */
295 	struct rds_transport	*r_trans;
296 	void			*r_trans_private;
297 };
298 
299 /* Flags for mr->r_state */
300 #define RDS_MR_DEAD		0
301 
302 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
303 {
304 	return r_key | (((u64) offset) << 32);
305 }
306 
307 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
308 {
309 	return cookie;
310 }
311 
312 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
313 {
314 	return cookie >> 32;
315 }
316 
317 /* atomic operation types */
318 #define RDS_ATOMIC_TYPE_CSWP		0
319 #define RDS_ATOMIC_TYPE_FADD		1
320 
321 /*
322  * m_sock_item and m_conn_item are on lists that are serialized under
323  * conn->c_lock.  m_sock_item has additional meaning in that once it is empty
324  * the message will not be put back on the retransmit list after being sent.
325  * messages that are canceled while being sent rely on this.
326  *
327  * m_inc is used by loopback so that it can pass an incoming message straight
328  * back up into the rx path.  It embeds a wire header which is also used by
329  * the send path, which is kind of awkward.
330  *
331  * m_sock_item indicates the message's presence on a socket's send or receive
332  * queue.  m_rs will point to that socket.
333  *
334  * m_daddr is used by cancellation to prune messages to a given destination.
335  *
336  * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
337  * nesting.  As paths iterate over messages on a sock, or conn, they must
338  * also lock the conn, or sock, to remove the message from those lists too.
339  * Testing the flag to determine if the message is still on the lists lets
340  * us avoid testing the list_head directly.  That means each path can use
341  * the message's list_head to keep it on a local list while juggling locks
342  * without confusing the other path.
343  *
344  * m_ack_seq is an optional field set by transports who need a different
345  * sequence number range to invalidate.  They can use this in a callback
346  * that they pass to rds_send_drop_acked() to see if each message has been
347  * acked.  The HAS_ACK_SEQ flag can be used to detect messages which haven't
348  * had ack_seq set yet.
349  */
350 #define RDS_MSG_ON_SOCK		1
351 #define RDS_MSG_ON_CONN		2
352 #define RDS_MSG_HAS_ACK_SEQ	3
353 #define RDS_MSG_ACK_REQUIRED	4
354 #define RDS_MSG_RETRANSMITTED	5
355 #define RDS_MSG_MAPPED		6
356 #define RDS_MSG_PAGEVEC		7
357 #define RDS_MSG_FLUSH		8
358 
359 struct rds_message {
360 	refcount_t		m_refcount;
361 	struct list_head	m_sock_item;
362 	struct list_head	m_conn_item;
363 	struct rds_incoming	m_inc;
364 	u64			m_ack_seq;
365 	__be32			m_daddr;
366 	unsigned long		m_flags;
367 
368 	/* Never access m_rs without holding m_rs_lock.
369 	 * Lock nesting is
370 	 *  rm->m_rs_lock
371 	 *   -> rs->rs_lock
372 	 */
373 	spinlock_t		m_rs_lock;
374 	wait_queue_head_t	m_flush_wait;
375 
376 	struct rds_sock		*m_rs;
377 
378 	/* cookie to send to remote, in rds header */
379 	rds_rdma_cookie_t	m_rdma_cookie;
380 
381 	unsigned int		m_used_sgs;
382 	unsigned int		m_total_sgs;
383 
384 	void			*m_final_op;
385 
386 	struct {
387 		struct rm_atomic_op {
388 			int			op_type;
389 			union {
390 				struct {
391 					uint64_t	compare;
392 					uint64_t	swap;
393 					uint64_t	compare_mask;
394 					uint64_t	swap_mask;
395 				} op_m_cswp;
396 				struct {
397 					uint64_t	add;
398 					uint64_t	nocarry_mask;
399 				} op_m_fadd;
400 			};
401 
402 			u32			op_rkey;
403 			u64			op_remote_addr;
404 			unsigned int		op_notify:1;
405 			unsigned int		op_recverr:1;
406 			unsigned int		op_mapped:1;
407 			unsigned int		op_silent:1;
408 			unsigned int		op_active:1;
409 			struct scatterlist	*op_sg;
410 			struct rds_notifier	*op_notifier;
411 
412 			struct rds_mr		*op_rdma_mr;
413 		} atomic;
414 		struct rm_rdma_op {
415 			u32			op_rkey;
416 			u64			op_remote_addr;
417 			unsigned int		op_write:1;
418 			unsigned int		op_fence:1;
419 			unsigned int		op_notify:1;
420 			unsigned int		op_recverr:1;
421 			unsigned int		op_mapped:1;
422 			unsigned int		op_silent:1;
423 			unsigned int		op_active:1;
424 			unsigned int		op_bytes;
425 			unsigned int		op_nents;
426 			unsigned int		op_count;
427 			struct scatterlist	*op_sg;
428 			struct rds_notifier	*op_notifier;
429 
430 			struct rds_mr		*op_rdma_mr;
431 		} rdma;
432 		struct rm_data_op {
433 			unsigned int		op_active:1;
434 			unsigned int		op_notify:1;
435 			unsigned int		op_nents;
436 			unsigned int		op_count;
437 			unsigned int		op_dmasg;
438 			unsigned int		op_dmaoff;
439 			struct scatterlist	*op_sg;
440 		} data;
441 	};
442 };
443 
444 /*
445  * The RDS notifier is used (optionally) to tell the application about
446  * completed RDMA operations. Rather than keeping the whole rds message
447  * around on the queue, we allocate a small notifier that is put on the
448  * socket's notifier_list. Notifications are delivered to the application
449  * through control messages.
450  */
451 struct rds_notifier {
452 	struct list_head	n_list;
453 	uint64_t		n_user_token;
454 	int			n_status;
455 };
456 
457 /**
458  * struct rds_transport -  transport specific behavioural hooks
459  *
460  * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
461  *        part of a message.  The caller serializes on the send_sem so this
462  *        doesn't need to be reentrant for a given conn.  The header must be
463  *        sent before the data payload.  .xmit must be prepared to send a
464  *        message with no data payload.  .xmit should return the number of
465  *        bytes that were sent down the connection, including header bytes.
466  *        Returning 0 tells the caller that it doesn't need to perform any
467  *        additional work now.  This is usually the case when the transport has
468  *        filled the sending queue for its connection and will handle
469  *        triggering the rds thread to continue the send when space becomes
470  *        available.  Returning -EAGAIN tells the caller to retry the send
471  *        immediately.  Returning -ENOMEM tells the caller to retry the send at
472  *        some point in the future.
473  *
474  * @conn_shutdown: conn_shutdown stops traffic on the given connection.  Once
475  *                 it returns the connection can not call rds_recv_incoming().
476  *                 This will only be called once after conn_connect returns
477  *                 non-zero success and will The caller serializes this with
478  *                 the send and connecting paths (xmit_* and conn_*).  The
479  *                 transport is responsible for other serialization, including
480  *                 rds_recv_incoming().  This is called in process context but
481  *                 should try hard not to block.
482  */
483 
484 struct rds_transport {
485 	char			t_name[TRANSNAMSIZ];
486 	struct list_head	t_item;
487 	struct module		*t_owner;
488 	unsigned int		t_prefer_loopback:1,
489 				t_mp_capable:1;
490 	unsigned int		t_type;
491 
492 	int (*laddr_check)(struct net *net, __be32 addr);
493 	int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
494 	void (*conn_free)(void *data);
495 	int (*conn_path_connect)(struct rds_conn_path *cp);
496 	void (*conn_path_shutdown)(struct rds_conn_path *conn);
497 	void (*xmit_path_prepare)(struct rds_conn_path *cp);
498 	void (*xmit_path_complete)(struct rds_conn_path *cp);
499 	int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
500 		    unsigned int hdr_off, unsigned int sg, unsigned int off);
501 	int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
502 	int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
503 	int (*recv_path)(struct rds_conn_path *cp);
504 	int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
505 	void (*inc_free)(struct rds_incoming *inc);
506 
507 	int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
508 				 struct rdma_cm_event *event);
509 	int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
510 	void (*cm_connect_complete)(struct rds_connection *conn,
511 				    struct rdma_cm_event *event);
512 
513 	unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
514 					unsigned int avail);
515 	void (*exit)(void);
516 	void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
517 			struct rds_sock *rs, u32 *key_ret);
518 	void (*sync_mr)(void *trans_private, int direction);
519 	void (*free_mr)(void *trans_private, int invalidate);
520 	void (*flush_mrs)(void);
521 	bool (*t_unloading)(struct rds_connection *conn);
522 };
523 
524 struct rds_sock {
525 	struct sock		rs_sk;
526 
527 	u64			rs_user_addr;
528 	u64			rs_user_bytes;
529 
530 	/*
531 	 * bound_addr used for both incoming and outgoing, no INADDR_ANY
532 	 * support.
533 	 */
534 	struct rhash_head	rs_bound_node;
535 	u64			rs_bound_key;
536 	__be32			rs_bound_addr;
537 	__be32			rs_conn_addr;
538 	__be16			rs_bound_port;
539 	__be16			rs_conn_port;
540 	struct rds_transport    *rs_transport;
541 
542 	/*
543 	 * rds_sendmsg caches the conn it used the last time around.
544 	 * This helps avoid costly lookups.
545 	 */
546 	struct rds_connection	*rs_conn;
547 
548 	/* flag indicating we were congested or not */
549 	int			rs_congested;
550 	/* seen congestion (ENOBUFS) when sending? */
551 	int			rs_seen_congestion;
552 
553 	/* rs_lock protects all these adjacent members before the newline */
554 	spinlock_t		rs_lock;
555 	struct list_head	rs_send_queue;
556 	u32			rs_snd_bytes;
557 	int			rs_rcv_bytes;
558 	struct list_head	rs_notify_queue;	/* currently used for failed RDMAs */
559 
560 	/* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
561 	 * to decide whether the application should be woken up.
562 	 * If not set, we use rs_cong_track to find out whether a cong map
563 	 * update arrived.
564 	 */
565 	uint64_t		rs_cong_mask;
566 	uint64_t		rs_cong_notify;
567 	struct list_head	rs_cong_list;
568 	unsigned long		rs_cong_track;
569 
570 	/*
571 	 * rs_recv_lock protects the receive queue, and is
572 	 * used to serialize with rds_release.
573 	 */
574 	rwlock_t		rs_recv_lock;
575 	struct list_head	rs_recv_queue;
576 
577 	/* just for stats reporting */
578 	struct list_head	rs_item;
579 
580 	/* these have their own lock */
581 	spinlock_t		rs_rdma_lock;
582 	struct rb_root		rs_rdma_keys;
583 
584 	/* Socket options - in case there will be more */
585 	unsigned char		rs_recverr,
586 				rs_cong_monitor;
587 	u32			rs_hash_initval;
588 
589 	/* Socket receive path trace points*/
590 	u8			rs_rx_traces;
591 	u8			rs_rx_trace[RDS_MSG_RX_DGRAM_TRACE_MAX];
592 };
593 
594 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
595 {
596 	return container_of(sk, struct rds_sock, rs_sk);
597 }
598 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
599 {
600 	return &rs->rs_sk;
601 }
602 
603 /*
604  * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
605  * to account for overhead.  We don't account for overhead, we just apply
606  * the number of payload bytes to the specified value.
607  */
608 static inline int rds_sk_sndbuf(struct rds_sock *rs)
609 {
610 	return rds_rs_to_sk(rs)->sk_sndbuf / 2;
611 }
612 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
613 {
614 	return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
615 }
616 
617 struct rds_statistics {
618 	uint64_t	s_conn_reset;
619 	uint64_t	s_recv_drop_bad_checksum;
620 	uint64_t	s_recv_drop_old_seq;
621 	uint64_t	s_recv_drop_no_sock;
622 	uint64_t	s_recv_drop_dead_sock;
623 	uint64_t	s_recv_deliver_raced;
624 	uint64_t	s_recv_delivered;
625 	uint64_t	s_recv_queued;
626 	uint64_t	s_recv_immediate_retry;
627 	uint64_t	s_recv_delayed_retry;
628 	uint64_t	s_recv_ack_required;
629 	uint64_t	s_recv_rdma_bytes;
630 	uint64_t	s_recv_ping;
631 	uint64_t	s_send_queue_empty;
632 	uint64_t	s_send_queue_full;
633 	uint64_t	s_send_lock_contention;
634 	uint64_t	s_send_lock_queue_raced;
635 	uint64_t	s_send_immediate_retry;
636 	uint64_t	s_send_delayed_retry;
637 	uint64_t	s_send_drop_acked;
638 	uint64_t	s_send_ack_required;
639 	uint64_t	s_send_queued;
640 	uint64_t	s_send_rdma;
641 	uint64_t	s_send_rdma_bytes;
642 	uint64_t	s_send_pong;
643 	uint64_t	s_page_remainder_hit;
644 	uint64_t	s_page_remainder_miss;
645 	uint64_t	s_copy_to_user;
646 	uint64_t	s_copy_from_user;
647 	uint64_t	s_cong_update_queued;
648 	uint64_t	s_cong_update_received;
649 	uint64_t	s_cong_send_error;
650 	uint64_t	s_cong_send_blocked;
651 	uint64_t	s_recv_bytes_added_to_socket;
652 	uint64_t	s_recv_bytes_removed_from_socket;
653 
654 };
655 
656 /* af_rds.c */
657 void rds_sock_addref(struct rds_sock *rs);
658 void rds_sock_put(struct rds_sock *rs);
659 void rds_wake_sk_sleep(struct rds_sock *rs);
660 static inline void __rds_wake_sk_sleep(struct sock *sk)
661 {
662 	wait_queue_head_t *waitq = sk_sleep(sk);
663 
664 	if (!sock_flag(sk, SOCK_DEAD) && waitq)
665 		wake_up(waitq);
666 }
667 extern wait_queue_head_t rds_poll_waitq;
668 
669 
670 /* bind.c */
671 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
672 void rds_remove_bound(struct rds_sock *rs);
673 struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
674 int rds_bind_lock_init(void);
675 void rds_bind_lock_destroy(void);
676 
677 /* cong.c */
678 int rds_cong_get_maps(struct rds_connection *conn);
679 void rds_cong_add_conn(struct rds_connection *conn);
680 void rds_cong_remove_conn(struct rds_connection *conn);
681 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
682 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
683 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
684 void rds_cong_queue_updates(struct rds_cong_map *map);
685 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
686 int rds_cong_updated_since(unsigned long *recent);
687 void rds_cong_add_socket(struct rds_sock *);
688 void rds_cong_remove_socket(struct rds_sock *);
689 void rds_cong_exit(void);
690 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
691 
692 /* conn.c */
693 extern u32 rds_gen_num;
694 int rds_conn_init(void);
695 void rds_conn_exit(void);
696 struct rds_connection *rds_conn_create(struct net *net,
697 				       __be32 laddr, __be32 faddr,
698 				       struct rds_transport *trans, gfp_t gfp);
699 struct rds_connection *rds_conn_create_outgoing(struct net *net,
700 						__be32 laddr, __be32 faddr,
701 			       struct rds_transport *trans, gfp_t gfp);
702 void rds_conn_shutdown(struct rds_conn_path *cpath);
703 void rds_conn_destroy(struct rds_connection *conn);
704 void rds_conn_drop(struct rds_connection *conn);
705 void rds_conn_path_drop(struct rds_conn_path *cpath, bool destroy);
706 void rds_conn_connect_if_down(struct rds_connection *conn);
707 void rds_conn_path_connect_if_down(struct rds_conn_path *cp);
708 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
709 			  struct rds_info_iterator *iter,
710 			  struct rds_info_lengths *lens,
711 			  int (*visitor)(struct rds_connection *, void *),
712 			  size_t item_len);
713 
714 __printf(2, 3)
715 void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...);
716 #define rds_conn_path_error(cp, fmt...) \
717 	__rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt)
718 
719 static inline int
720 rds_conn_path_transition(struct rds_conn_path *cp, int old, int new)
721 {
722 	return atomic_cmpxchg(&cp->cp_state, old, new) == old;
723 }
724 
725 static inline int
726 rds_conn_transition(struct rds_connection *conn, int old, int new)
727 {
728 	WARN_ON(conn->c_trans->t_mp_capable);
729 	return rds_conn_path_transition(&conn->c_path[0], old, new);
730 }
731 
732 static inline int
733 rds_conn_path_state(struct rds_conn_path *cp)
734 {
735 	return atomic_read(&cp->cp_state);
736 }
737 
738 static inline int
739 rds_conn_state(struct rds_connection *conn)
740 {
741 	WARN_ON(conn->c_trans->t_mp_capable);
742 	return rds_conn_path_state(&conn->c_path[0]);
743 }
744 
745 static inline int
746 rds_conn_path_up(struct rds_conn_path *cp)
747 {
748 	return atomic_read(&cp->cp_state) == RDS_CONN_UP;
749 }
750 
751 static inline int
752 rds_conn_up(struct rds_connection *conn)
753 {
754 	WARN_ON(conn->c_trans->t_mp_capable);
755 	return rds_conn_path_up(&conn->c_path[0]);
756 }
757 
758 static inline int
759 rds_conn_path_connecting(struct rds_conn_path *cp)
760 {
761 	return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING;
762 }
763 
764 static inline int
765 rds_conn_connecting(struct rds_connection *conn)
766 {
767 	WARN_ON(conn->c_trans->t_mp_capable);
768 	return rds_conn_path_connecting(&conn->c_path[0]);
769 }
770 
771 /* message.c */
772 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
773 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
774 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from);
775 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
776 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
777 				 __be16 dport, u64 seq);
778 int rds_message_add_extension(struct rds_header *hdr,
779 			      unsigned int type, const void *data, unsigned int len);
780 int rds_message_next_extension(struct rds_header *hdr,
781 			       unsigned int *pos, void *buf, unsigned int *buflen);
782 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
783 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
784 void rds_message_inc_free(struct rds_incoming *inc);
785 void rds_message_addref(struct rds_message *rm);
786 void rds_message_put(struct rds_message *rm);
787 void rds_message_wait(struct rds_message *rm);
788 void rds_message_unmapped(struct rds_message *rm);
789 
790 static inline void rds_message_make_checksum(struct rds_header *hdr)
791 {
792 	hdr->h_csum = 0;
793 	hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
794 }
795 
796 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
797 {
798 	return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
799 }
800 
801 
802 /* page.c */
803 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
804 			     gfp_t gfp);
805 void rds_page_exit(void);
806 
807 /* recv.c */
808 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
809 		  __be32 saddr);
810 void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn,
811 		       __be32 saddr);
812 void rds_inc_put(struct rds_incoming *inc);
813 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
814 		       struct rds_incoming *inc, gfp_t gfp);
815 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
816 		int msg_flags);
817 void rds_clear_recv_queue(struct rds_sock *rs);
818 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
819 void rds_inc_info_copy(struct rds_incoming *inc,
820 		       struct rds_info_iterator *iter,
821 		       __be32 saddr, __be32 daddr, int flip);
822 
823 /* send.c */
824 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
825 void rds_send_path_reset(struct rds_conn_path *conn);
826 int rds_send_xmit(struct rds_conn_path *cp);
827 struct sockaddr_in;
828 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
829 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
830 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
831 			 is_acked_func is_acked);
832 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
833 			      is_acked_func is_acked);
834 void rds_send_ping(struct rds_connection *conn, int cp_index);
835 int rds_send_pong(struct rds_conn_path *cp, __be16 dport);
836 
837 /* rdma.c */
838 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
839 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
840 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
841 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
842 void rds_rdma_drop_keys(struct rds_sock *rs);
843 int rds_rdma_extra_size(struct rds_rdma_args *args);
844 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
845 			  struct cmsghdr *cmsg);
846 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
847 			  struct cmsghdr *cmsg);
848 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
849 			  struct cmsghdr *cmsg);
850 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
851 			  struct cmsghdr *cmsg);
852 void rds_rdma_free_op(struct rm_rdma_op *ro);
853 void rds_atomic_free_op(struct rm_atomic_op *ao);
854 void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
855 void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
856 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
857 		    struct cmsghdr *cmsg);
858 
859 void __rds_put_mr_final(struct rds_mr *mr);
860 static inline void rds_mr_put(struct rds_mr *mr)
861 {
862 	if (refcount_dec_and_test(&mr->r_refcount))
863 		__rds_put_mr_final(mr);
864 }
865 
866 static inline bool rds_destroy_pending(struct rds_connection *conn)
867 {
868 	return !check_net(rds_conn_net(conn)) ||
869 	       (conn->c_trans->t_unloading && conn->c_trans->t_unloading(conn));
870 }
871 
872 /* stats.c */
873 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
874 #define rds_stats_inc_which(which, member) do {		\
875 	per_cpu(which, get_cpu()).member++;		\
876 	put_cpu();					\
877 } while (0)
878 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
879 #define rds_stats_add_which(which, member, count) do {		\
880 	per_cpu(which, get_cpu()).member += count;	\
881 	put_cpu();					\
882 } while (0)
883 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
884 int rds_stats_init(void);
885 void rds_stats_exit(void);
886 void rds_stats_info_copy(struct rds_info_iterator *iter,
887 			 uint64_t *values, const char *const *names,
888 			 size_t nr);
889 
890 /* sysctl.c */
891 int rds_sysctl_init(void);
892 void rds_sysctl_exit(void);
893 extern unsigned long rds_sysctl_sndbuf_min;
894 extern unsigned long rds_sysctl_sndbuf_default;
895 extern unsigned long rds_sysctl_sndbuf_max;
896 extern unsigned long rds_sysctl_reconnect_min_jiffies;
897 extern unsigned long rds_sysctl_reconnect_max_jiffies;
898 extern unsigned int  rds_sysctl_max_unacked_packets;
899 extern unsigned int  rds_sysctl_max_unacked_bytes;
900 extern unsigned int  rds_sysctl_ping_enable;
901 extern unsigned long rds_sysctl_trace_flags;
902 extern unsigned int  rds_sysctl_trace_level;
903 
904 /* threads.c */
905 int rds_threads_init(void);
906 void rds_threads_exit(void);
907 extern struct workqueue_struct *rds_wq;
908 void rds_queue_reconnect(struct rds_conn_path *cp);
909 void rds_connect_worker(struct work_struct *);
910 void rds_shutdown_worker(struct work_struct *);
911 void rds_send_worker(struct work_struct *);
912 void rds_recv_worker(struct work_struct *);
913 void rds_connect_path_complete(struct rds_conn_path *conn, int curr);
914 void rds_connect_complete(struct rds_connection *conn);
915 
916 /* transport.c */
917 void rds_trans_register(struct rds_transport *trans);
918 void rds_trans_unregister(struct rds_transport *trans);
919 struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr);
920 void rds_trans_put(struct rds_transport *trans);
921 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
922 				       unsigned int avail);
923 struct rds_transport *rds_trans_get(int t_type);
924 int rds_trans_init(void);
925 void rds_trans_exit(void);
926 
927 #endif
928