xref: /openbmc/linux/net/rds/rds.h (revision 0c874100)
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 #include <linux/in6.h>
14 
15 #include "info.h"
16 
17 /*
18  * RDS Network protocol version
19  */
20 #define RDS_PROTOCOL_3_0	0x0300
21 #define RDS_PROTOCOL_3_1	0x0301
22 #define RDS_PROTOCOL_VERSION	RDS_PROTOCOL_3_1
23 #define RDS_PROTOCOL_MAJOR(v)	((v) >> 8)
24 #define RDS_PROTOCOL_MINOR(v)	((v) & 255)
25 #define RDS_PROTOCOL(maj, min)	(((maj) << 8) | min)
26 
27 /* The following ports, 16385, 18634, 18635, are registered with IANA as
28  * the ports to be used for RDS over TCP and UDP.  Currently, only RDS over
29  * TCP and RDS over IB/RDMA are implemented.  18634 is the historical value
30  * used for the RDMA_CM listener port.  RDS/TCP uses port 16385.  After
31  * IPv6 work, RDMA_CM also uses 16385 as the listener port.  18634 is kept
32  * to ensure compatibility with older RDS modules.  Those ports are defined
33  * in each transport's header file.
34  */
35 #define RDS_PORT	18634
36 
37 #ifdef ATOMIC64_INIT
38 #define KERNEL_HAS_ATOMIC64
39 #endif
40 
41 #ifdef RDS_DEBUG
42 #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
43 #else
44 /* sigh, pr_debug() causes unused variable warnings */
45 static inline __printf(1, 2)
46 void rdsdebug(char *fmt, ...)
47 {
48 }
49 #endif
50 
51 /* XXX is there one of these somewhere? */
52 #define ceil(x, y) \
53 	({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; })
54 
55 #define RDS_FRAG_SHIFT	12
56 #define RDS_FRAG_SIZE	((unsigned int)(1 << RDS_FRAG_SHIFT))
57 
58 /* Used to limit both RDMA and non-RDMA RDS message to 1MB */
59 #define RDS_MAX_MSG_SIZE	((unsigned int)(1 << 20))
60 
61 #define RDS_CONG_MAP_BYTES	(65536 / 8)
62 #define RDS_CONG_MAP_PAGES	(PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
63 #define RDS_CONG_MAP_PAGE_BITS	(PAGE_SIZE * 8)
64 
65 struct rds_cong_map {
66 	struct rb_node		m_rb_node;
67 	struct in6_addr		m_addr;
68 	wait_queue_head_t	m_waitq;
69 	struct list_head	m_conn_list;
70 	unsigned long		m_page_addrs[RDS_CONG_MAP_PAGES];
71 };
72 
73 
74 /*
75  * This is how we will track the connection state:
76  * A connection is always in one of the following
77  * states. Updates to the state are atomic and imply
78  * a memory barrier.
79  */
80 enum {
81 	RDS_CONN_DOWN = 0,
82 	RDS_CONN_CONNECTING,
83 	RDS_CONN_DISCONNECTING,
84 	RDS_CONN_UP,
85 	RDS_CONN_RESETTING,
86 	RDS_CONN_ERROR,
87 };
88 
89 /* Bits for c_flags */
90 #define RDS_LL_SEND_FULL	0
91 #define RDS_RECONNECT_PENDING	1
92 #define RDS_IN_XMIT		2
93 #define RDS_RECV_REFILL		3
94 #define	RDS_DESTROY_PENDING	4
95 
96 /* Max number of multipaths per RDS connection. Must be a power of 2 */
97 #define	RDS_MPATH_WORKERS	8
98 #define	RDS_MPATH_HASH(rs, n) (jhash_1word((rs)->rs_bound_port, \
99 			       (rs)->rs_hash_initval) & ((n) - 1))
100 
101 #define IS_CANONICAL(laddr, faddr) (htonl(laddr) < htonl(faddr))
102 
103 /* Per mpath connection state */
104 struct rds_conn_path {
105 	struct rds_connection	*cp_conn;
106 	struct rds_message	*cp_xmit_rm;
107 	unsigned long		cp_xmit_sg;
108 	unsigned int		cp_xmit_hdr_off;
109 	unsigned int		cp_xmit_data_off;
110 	unsigned int		cp_xmit_atomic_sent;
111 	unsigned int		cp_xmit_rdma_sent;
112 	unsigned int		cp_xmit_data_sent;
113 
114 	spinlock_t		cp_lock;		/* protect msg queues */
115 	u64			cp_next_tx_seq;
116 	struct list_head	cp_send_queue;
117 	struct list_head	cp_retrans;
118 
119 	u64			cp_next_rx_seq;
120 
121 	void			*cp_transport_data;
122 
123 	atomic_t		cp_state;
124 	unsigned long		cp_send_gen;
125 	unsigned long		cp_flags;
126 	unsigned long		cp_reconnect_jiffies;
127 	struct delayed_work	cp_send_w;
128 	struct delayed_work	cp_recv_w;
129 	struct delayed_work	cp_conn_w;
130 	struct work_struct	cp_down_w;
131 	struct mutex		cp_cm_lock;	/* protect cp_state & cm */
132 	wait_queue_head_t	cp_waitq;
133 
134 	unsigned int		cp_unacked_packets;
135 	unsigned int		cp_unacked_bytes;
136 	unsigned int		cp_index;
137 };
138 
139 /* One rds_connection per RDS address pair */
140 struct rds_connection {
141 	struct hlist_node	c_hash_node;
142 	struct in6_addr		c_laddr;
143 	struct in6_addr		c_faddr;
144 	int			c_dev_if; /* ifindex used for this conn */
145 	int			c_bound_if; /* ifindex of c_laddr */
146 	unsigned int		c_loopback:1,
147 				c_isv6:1,
148 				c_ping_triggered:1,
149 				c_pad_to_32:29;
150 	int			c_npaths;
151 	struct rds_connection	*c_passive;
152 	struct rds_transport	*c_trans;
153 
154 	struct rds_cong_map	*c_lcong;
155 	struct rds_cong_map	*c_fcong;
156 
157 	/* Protocol version */
158 	unsigned int		c_version;
159 	possible_net_t		c_net;
160 
161 	struct list_head	c_map_item;
162 	unsigned long		c_map_queued;
163 
164 	struct rds_conn_path	*c_path;
165 	wait_queue_head_t	c_hs_waitq; /* handshake waitq */
166 
167 	u32			c_my_gen_num;
168 	u32			c_peer_gen_num;
169 };
170 
171 static inline
172 struct net *rds_conn_net(struct rds_connection *conn)
173 {
174 	return read_pnet(&conn->c_net);
175 }
176 
177 static inline
178 void rds_conn_net_set(struct rds_connection *conn, struct net *net)
179 {
180 	write_pnet(&conn->c_net, net);
181 }
182 
183 #define RDS_FLAG_CONG_BITMAP	0x01
184 #define RDS_FLAG_ACK_REQUIRED	0x02
185 #define RDS_FLAG_RETRANSMITTED	0x04
186 #define RDS_MAX_ADV_CREDIT	255
187 
188 /* RDS_FLAG_PROBE_PORT is the reserved sport used for sending a ping
189  * probe to exchange control information before establishing a connection.
190  * Currently the control information that is exchanged is the number of
191  * supported paths. If the peer is a legacy (older kernel revision) peer,
192  * it would return a pong message without additional control information
193  * that would then alert the sender that the peer was an older rev.
194  */
195 #define RDS_FLAG_PROBE_PORT	1
196 #define	RDS_HS_PROBE(sport, dport) \
197 		((sport == RDS_FLAG_PROBE_PORT && dport == 0) || \
198 		 (sport == 0 && dport == RDS_FLAG_PROBE_PORT))
199 /*
200  * Maximum space available for extension headers.
201  */
202 #define RDS_HEADER_EXT_SPACE	16
203 
204 struct rds_header {
205 	__be64	h_sequence;
206 	__be64	h_ack;
207 	__be32	h_len;
208 	__be16	h_sport;
209 	__be16	h_dport;
210 	u8	h_flags;
211 	u8	h_credit;
212 	u8	h_padding[4];
213 	__sum16	h_csum;
214 
215 	u8	h_exthdr[RDS_HEADER_EXT_SPACE];
216 };
217 
218 /*
219  * Reserved - indicates end of extensions
220  */
221 #define RDS_EXTHDR_NONE		0
222 
223 /*
224  * This extension header is included in the very
225  * first message that is sent on a new connection,
226  * and identifies the protocol level. This will help
227  * rolling updates if a future change requires breaking
228  * the protocol.
229  * NB: This is no longer true for IB, where we do a version
230  * negotiation during the connection setup phase (protocol
231  * version information is included in the RDMA CM private data).
232  */
233 #define RDS_EXTHDR_VERSION	1
234 struct rds_ext_header_version {
235 	__be32			h_version;
236 };
237 
238 /*
239  * This extension header is included in the RDS message
240  * chasing an RDMA operation.
241  */
242 #define RDS_EXTHDR_RDMA		2
243 struct rds_ext_header_rdma {
244 	__be32			h_rdma_rkey;
245 };
246 
247 /*
248  * This extension header tells the peer about the
249  * destination <R_Key,offset> of the requested RDMA
250  * operation.
251  */
252 #define RDS_EXTHDR_RDMA_DEST	3
253 struct rds_ext_header_rdma_dest {
254 	__be32			h_rdma_rkey;
255 	__be32			h_rdma_offset;
256 };
257 
258 /* Extension header announcing number of paths.
259  * Implicit length = 2 bytes.
260  */
261 #define RDS_EXTHDR_NPATHS	5
262 #define RDS_EXTHDR_GEN_NUM	6
263 
264 #define __RDS_EXTHDR_MAX	16 /* for now */
265 #define RDS_RX_MAX_TRACES	(RDS_MSG_RX_DGRAM_TRACE_MAX + 1)
266 #define	RDS_MSG_RX_HDR		0
267 #define	RDS_MSG_RX_START	1
268 #define	RDS_MSG_RX_END		2
269 #define	RDS_MSG_RX_CMSG		3
270 
271 struct rds_incoming {
272 	refcount_t		i_refcount;
273 	struct list_head	i_item;
274 	struct rds_connection	*i_conn;
275 	struct rds_conn_path	*i_conn_path;
276 	struct rds_header	i_hdr;
277 	unsigned long		i_rx_jiffies;
278 	struct in6_addr		i_saddr;
279 
280 	rds_rdma_cookie_t	i_rdma_cookie;
281 	ktime_t			i_rx_tstamp;
282 	u64			i_rx_lat_trace[RDS_RX_MAX_TRACES];
283 };
284 
285 struct rds_mr {
286 	struct rb_node		r_rb_node;
287 	refcount_t		r_refcount;
288 	u32			r_key;
289 
290 	/* A copy of the creation flags */
291 	unsigned int		r_use_once:1;
292 	unsigned int		r_invalidate:1;
293 	unsigned int		r_write:1;
294 
295 	/* This is for RDS_MR_DEAD.
296 	 * It would be nice & consistent to make this part of the above
297 	 * bit field here, but we need to use test_and_set_bit.
298 	 */
299 	unsigned long		r_state;
300 	struct rds_sock		*r_sock; /* back pointer to the socket that owns us */
301 	struct rds_transport	*r_trans;
302 	void			*r_trans_private;
303 };
304 
305 /* Flags for mr->r_state */
306 #define RDS_MR_DEAD		0
307 
308 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
309 {
310 	return r_key | (((u64) offset) << 32);
311 }
312 
313 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
314 {
315 	return cookie;
316 }
317 
318 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
319 {
320 	return cookie >> 32;
321 }
322 
323 /* atomic operation types */
324 #define RDS_ATOMIC_TYPE_CSWP		0
325 #define RDS_ATOMIC_TYPE_FADD		1
326 
327 /*
328  * m_sock_item and m_conn_item are on lists that are serialized under
329  * conn->c_lock.  m_sock_item has additional meaning in that once it is empty
330  * the message will not be put back on the retransmit list after being sent.
331  * messages that are canceled while being sent rely on this.
332  *
333  * m_inc is used by loopback so that it can pass an incoming message straight
334  * back up into the rx path.  It embeds a wire header which is also used by
335  * the send path, which is kind of awkward.
336  *
337  * m_sock_item indicates the message's presence on a socket's send or receive
338  * queue.  m_rs will point to that socket.
339  *
340  * m_daddr is used by cancellation to prune messages to a given destination.
341  *
342  * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
343  * nesting.  As paths iterate over messages on a sock, or conn, they must
344  * also lock the conn, or sock, to remove the message from those lists too.
345  * Testing the flag to determine if the message is still on the lists lets
346  * us avoid testing the list_head directly.  That means each path can use
347  * the message's list_head to keep it on a local list while juggling locks
348  * without confusing the other path.
349  *
350  * m_ack_seq is an optional field set by transports who need a different
351  * sequence number range to invalidate.  They can use this in a callback
352  * that they pass to rds_send_drop_acked() to see if each message has been
353  * acked.  The HAS_ACK_SEQ flag can be used to detect messages which haven't
354  * had ack_seq set yet.
355  */
356 #define RDS_MSG_ON_SOCK		1
357 #define RDS_MSG_ON_CONN		2
358 #define RDS_MSG_HAS_ACK_SEQ	3
359 #define RDS_MSG_ACK_REQUIRED	4
360 #define RDS_MSG_RETRANSMITTED	5
361 #define RDS_MSG_MAPPED		6
362 #define RDS_MSG_PAGEVEC		7
363 #define RDS_MSG_FLUSH		8
364 
365 struct rds_znotifier {
366 	struct mmpin		z_mmp;
367 	u32			z_cookie;
368 };
369 
370 struct rds_msg_zcopy_info {
371 	struct list_head rs_zcookie_next;
372 	union {
373 		struct rds_znotifier znotif;
374 		struct rds_zcopy_cookies zcookies;
375 	};
376 };
377 
378 struct rds_msg_zcopy_queue {
379 	struct list_head zcookie_head;
380 	spinlock_t lock; /* protects zcookie_head queue */
381 };
382 
383 static inline void rds_message_zcopy_queue_init(struct rds_msg_zcopy_queue *q)
384 {
385 	spin_lock_init(&q->lock);
386 	INIT_LIST_HEAD(&q->zcookie_head);
387 }
388 
389 struct rds_message {
390 	refcount_t		m_refcount;
391 	struct list_head	m_sock_item;
392 	struct list_head	m_conn_item;
393 	struct rds_incoming	m_inc;
394 	u64			m_ack_seq;
395 	struct in6_addr		m_daddr;
396 	unsigned long		m_flags;
397 
398 	/* Never access m_rs without holding m_rs_lock.
399 	 * Lock nesting is
400 	 *  rm->m_rs_lock
401 	 *   -> rs->rs_lock
402 	 */
403 	spinlock_t		m_rs_lock;
404 	wait_queue_head_t	m_flush_wait;
405 
406 	struct rds_sock		*m_rs;
407 
408 	/* cookie to send to remote, in rds header */
409 	rds_rdma_cookie_t	m_rdma_cookie;
410 
411 	unsigned int		m_used_sgs;
412 	unsigned int		m_total_sgs;
413 
414 	void			*m_final_op;
415 
416 	struct {
417 		struct rm_atomic_op {
418 			int			op_type;
419 			union {
420 				struct {
421 					uint64_t	compare;
422 					uint64_t	swap;
423 					uint64_t	compare_mask;
424 					uint64_t	swap_mask;
425 				} op_m_cswp;
426 				struct {
427 					uint64_t	add;
428 					uint64_t	nocarry_mask;
429 				} op_m_fadd;
430 			};
431 
432 			u32			op_rkey;
433 			u64			op_remote_addr;
434 			unsigned int		op_notify:1;
435 			unsigned int		op_recverr:1;
436 			unsigned int		op_mapped:1;
437 			unsigned int		op_silent:1;
438 			unsigned int		op_active:1;
439 			struct scatterlist	*op_sg;
440 			struct rds_notifier	*op_notifier;
441 
442 			struct rds_mr		*op_rdma_mr;
443 		} atomic;
444 		struct rm_rdma_op {
445 			u32			op_rkey;
446 			u64			op_remote_addr;
447 			unsigned int		op_write:1;
448 			unsigned int		op_fence:1;
449 			unsigned int		op_notify:1;
450 			unsigned int		op_recverr:1;
451 			unsigned int		op_mapped:1;
452 			unsigned int		op_silent:1;
453 			unsigned int		op_active:1;
454 			unsigned int		op_bytes;
455 			unsigned int		op_nents;
456 			unsigned int		op_count;
457 			struct scatterlist	*op_sg;
458 			struct rds_notifier	*op_notifier;
459 
460 			struct rds_mr		*op_rdma_mr;
461 		} rdma;
462 		struct rm_data_op {
463 			unsigned int		op_active:1;
464 			unsigned int		op_notify:1;
465 			unsigned int		op_nents;
466 			unsigned int		op_count;
467 			unsigned int		op_dmasg;
468 			unsigned int		op_dmaoff;
469 			struct rds_znotifier	*op_mmp_znotifier;
470 			struct scatterlist	*op_sg;
471 		} data;
472 	};
473 
474 	struct rds_conn_path *m_conn_path;
475 };
476 
477 /*
478  * The RDS notifier is used (optionally) to tell the application about
479  * completed RDMA operations. Rather than keeping the whole rds message
480  * around on the queue, we allocate a small notifier that is put on the
481  * socket's notifier_list. Notifications are delivered to the application
482  * through control messages.
483  */
484 struct rds_notifier {
485 	struct list_head	n_list;
486 	uint64_t		n_user_token;
487 	int			n_status;
488 };
489 
490 /* Available as part of RDS core, so doesn't need to participate
491  * in get_preferred transport etc
492  */
493 #define	RDS_TRANS_LOOP	3
494 
495 /**
496  * struct rds_transport -  transport specific behavioural hooks
497  *
498  * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
499  *        part of a message.  The caller serializes on the send_sem so this
500  *        doesn't need to be reentrant for a given conn.  The header must be
501  *        sent before the data payload.  .xmit must be prepared to send a
502  *        message with no data payload.  .xmit should return the number of
503  *        bytes that were sent down the connection, including header bytes.
504  *        Returning 0 tells the caller that it doesn't need to perform any
505  *        additional work now.  This is usually the case when the transport has
506  *        filled the sending queue for its connection and will handle
507  *        triggering the rds thread to continue the send when space becomes
508  *        available.  Returning -EAGAIN tells the caller to retry the send
509  *        immediately.  Returning -ENOMEM tells the caller to retry the send at
510  *        some point in the future.
511  *
512  * @conn_shutdown: conn_shutdown stops traffic on the given connection.  Once
513  *                 it returns the connection can not call rds_recv_incoming().
514  *                 This will only be called once after conn_connect returns
515  *                 non-zero success and will The caller serializes this with
516  *                 the send and connecting paths (xmit_* and conn_*).  The
517  *                 transport is responsible for other serialization, including
518  *                 rds_recv_incoming().  This is called in process context but
519  *                 should try hard not to block.
520  */
521 
522 struct rds_transport {
523 	char			t_name[TRANSNAMSIZ];
524 	struct list_head	t_item;
525 	struct module		*t_owner;
526 	unsigned int		t_prefer_loopback:1,
527 				t_mp_capable:1;
528 	unsigned int		t_type;
529 
530 	int (*laddr_check)(struct net *net, const struct in6_addr *addr,
531 			   __u32 scope_id);
532 	int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
533 	void (*conn_free)(void *data);
534 	int (*conn_path_connect)(struct rds_conn_path *cp);
535 	void (*conn_path_shutdown)(struct rds_conn_path *conn);
536 	void (*xmit_path_prepare)(struct rds_conn_path *cp);
537 	void (*xmit_path_complete)(struct rds_conn_path *cp);
538 	int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
539 		    unsigned int hdr_off, unsigned int sg, unsigned int off);
540 	int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
541 	int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
542 	int (*recv_path)(struct rds_conn_path *cp);
543 	int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
544 	void (*inc_free)(struct rds_incoming *inc);
545 
546 	int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
547 				 struct rdma_cm_event *event, bool isv6);
548 	int (*cm_initiate_connect)(struct rdma_cm_id *cm_id, bool isv6);
549 	void (*cm_connect_complete)(struct rds_connection *conn,
550 				    struct rdma_cm_event *event);
551 
552 	unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
553 					unsigned int avail);
554 	void (*exit)(void);
555 	void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
556 			struct rds_sock *rs, u32 *key_ret,
557 			struct rds_connection *conn);
558 	void (*sync_mr)(void *trans_private, int direction);
559 	void (*free_mr)(void *trans_private, int invalidate);
560 	void (*flush_mrs)(void);
561 	bool (*t_unloading)(struct rds_connection *conn);
562 };
563 
564 /* Bind hash table key length.  It is the sum of the size of a struct
565  * in6_addr, a scope_id  and a port.
566  */
567 #define RDS_BOUND_KEY_LEN \
568 	(sizeof(struct in6_addr) + sizeof(__u32) + sizeof(__be16))
569 
570 struct rds_sock {
571 	struct sock		rs_sk;
572 
573 	u64			rs_user_addr;
574 	u64			rs_user_bytes;
575 
576 	/*
577 	 * bound_addr used for both incoming and outgoing, no INADDR_ANY
578 	 * support.
579 	 */
580 	struct rhash_head	rs_bound_node;
581 	u8			rs_bound_key[RDS_BOUND_KEY_LEN];
582 	struct sockaddr_in6	rs_bound_sin6;
583 #define rs_bound_addr		rs_bound_sin6.sin6_addr
584 #define rs_bound_addr_v4	rs_bound_sin6.sin6_addr.s6_addr32[3]
585 #define rs_bound_port		rs_bound_sin6.sin6_port
586 #define rs_bound_scope_id	rs_bound_sin6.sin6_scope_id
587 	struct in6_addr		rs_conn_addr;
588 #define rs_conn_addr_v4		rs_conn_addr.s6_addr32[3]
589 	__be16			rs_conn_port;
590 	struct rds_transport    *rs_transport;
591 
592 	/*
593 	 * rds_sendmsg caches the conn it used the last time around.
594 	 * This helps avoid costly lookups.
595 	 */
596 	struct rds_connection	*rs_conn;
597 
598 	/* flag indicating we were congested or not */
599 	int			rs_congested;
600 	/* seen congestion (ENOBUFS) when sending? */
601 	int			rs_seen_congestion;
602 
603 	/* rs_lock protects all these adjacent members before the newline */
604 	spinlock_t		rs_lock;
605 	struct list_head	rs_send_queue;
606 	u32			rs_snd_bytes;
607 	int			rs_rcv_bytes;
608 	struct list_head	rs_notify_queue;	/* currently used for failed RDMAs */
609 
610 	/* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
611 	 * to decide whether the application should be woken up.
612 	 * If not set, we use rs_cong_track to find out whether a cong map
613 	 * update arrived.
614 	 */
615 	uint64_t		rs_cong_mask;
616 	uint64_t		rs_cong_notify;
617 	struct list_head	rs_cong_list;
618 	unsigned long		rs_cong_track;
619 
620 	/*
621 	 * rs_recv_lock protects the receive queue, and is
622 	 * used to serialize with rds_release.
623 	 */
624 	rwlock_t		rs_recv_lock;
625 	struct list_head	rs_recv_queue;
626 
627 	/* just for stats reporting */
628 	struct list_head	rs_item;
629 
630 	/* these have their own lock */
631 	spinlock_t		rs_rdma_lock;
632 	struct rb_root		rs_rdma_keys;
633 
634 	/* Socket options - in case there will be more */
635 	unsigned char		rs_recverr,
636 				rs_cong_monitor;
637 	u32			rs_hash_initval;
638 
639 	/* Socket receive path trace points*/
640 	u8			rs_rx_traces;
641 	u8			rs_rx_trace[RDS_MSG_RX_DGRAM_TRACE_MAX];
642 	struct rds_msg_zcopy_queue rs_zcookie_queue;
643 };
644 
645 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
646 {
647 	return container_of(sk, struct rds_sock, rs_sk);
648 }
649 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
650 {
651 	return &rs->rs_sk;
652 }
653 
654 /*
655  * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
656  * to account for overhead.  We don't account for overhead, we just apply
657  * the number of payload bytes to the specified value.
658  */
659 static inline int rds_sk_sndbuf(struct rds_sock *rs)
660 {
661 	return rds_rs_to_sk(rs)->sk_sndbuf / 2;
662 }
663 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
664 {
665 	return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
666 }
667 
668 struct rds_statistics {
669 	uint64_t	s_conn_reset;
670 	uint64_t	s_recv_drop_bad_checksum;
671 	uint64_t	s_recv_drop_old_seq;
672 	uint64_t	s_recv_drop_no_sock;
673 	uint64_t	s_recv_drop_dead_sock;
674 	uint64_t	s_recv_deliver_raced;
675 	uint64_t	s_recv_delivered;
676 	uint64_t	s_recv_queued;
677 	uint64_t	s_recv_immediate_retry;
678 	uint64_t	s_recv_delayed_retry;
679 	uint64_t	s_recv_ack_required;
680 	uint64_t	s_recv_rdma_bytes;
681 	uint64_t	s_recv_ping;
682 	uint64_t	s_send_queue_empty;
683 	uint64_t	s_send_queue_full;
684 	uint64_t	s_send_lock_contention;
685 	uint64_t	s_send_lock_queue_raced;
686 	uint64_t	s_send_immediate_retry;
687 	uint64_t	s_send_delayed_retry;
688 	uint64_t	s_send_drop_acked;
689 	uint64_t	s_send_ack_required;
690 	uint64_t	s_send_queued;
691 	uint64_t	s_send_rdma;
692 	uint64_t	s_send_rdma_bytes;
693 	uint64_t	s_send_pong;
694 	uint64_t	s_page_remainder_hit;
695 	uint64_t	s_page_remainder_miss;
696 	uint64_t	s_copy_to_user;
697 	uint64_t	s_copy_from_user;
698 	uint64_t	s_cong_update_queued;
699 	uint64_t	s_cong_update_received;
700 	uint64_t	s_cong_send_error;
701 	uint64_t	s_cong_send_blocked;
702 	uint64_t	s_recv_bytes_added_to_socket;
703 	uint64_t	s_recv_bytes_removed_from_socket;
704 
705 };
706 
707 /* af_rds.c */
708 void rds_sock_addref(struct rds_sock *rs);
709 void rds_sock_put(struct rds_sock *rs);
710 void rds_wake_sk_sleep(struct rds_sock *rs);
711 static inline void __rds_wake_sk_sleep(struct sock *sk)
712 {
713 	wait_queue_head_t *waitq = sk_sleep(sk);
714 
715 	if (!sock_flag(sk, SOCK_DEAD) && waitq)
716 		wake_up(waitq);
717 }
718 extern wait_queue_head_t rds_poll_waitq;
719 
720 
721 /* bind.c */
722 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
723 void rds_remove_bound(struct rds_sock *rs);
724 struct rds_sock *rds_find_bound(const struct in6_addr *addr, __be16 port,
725 				__u32 scope_id);
726 int rds_bind_lock_init(void);
727 void rds_bind_lock_destroy(void);
728 
729 /* cong.c */
730 int rds_cong_get_maps(struct rds_connection *conn);
731 void rds_cong_add_conn(struct rds_connection *conn);
732 void rds_cong_remove_conn(struct rds_connection *conn);
733 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
734 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
735 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
736 void rds_cong_queue_updates(struct rds_cong_map *map);
737 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
738 int rds_cong_updated_since(unsigned long *recent);
739 void rds_cong_add_socket(struct rds_sock *);
740 void rds_cong_remove_socket(struct rds_sock *);
741 void rds_cong_exit(void);
742 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
743 
744 /* connection.c */
745 extern u32 rds_gen_num;
746 int rds_conn_init(void);
747 void rds_conn_exit(void);
748 struct rds_connection *rds_conn_create(struct net *net,
749 				       const struct in6_addr *laddr,
750 				       const struct in6_addr *faddr,
751 				       struct rds_transport *trans, gfp_t gfp,
752 				       int dev_if);
753 struct rds_connection *rds_conn_create_outgoing(struct net *net,
754 						const struct in6_addr *laddr,
755 						const struct in6_addr *faddr,
756 						struct rds_transport *trans,
757 						gfp_t gfp, int dev_if);
758 void rds_conn_shutdown(struct rds_conn_path *cpath);
759 void rds_conn_destroy(struct rds_connection *conn);
760 void rds_conn_drop(struct rds_connection *conn);
761 void rds_conn_path_drop(struct rds_conn_path *cpath, bool destroy);
762 void rds_conn_connect_if_down(struct rds_connection *conn);
763 void rds_conn_path_connect_if_down(struct rds_conn_path *cp);
764 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
765 			  struct rds_info_iterator *iter,
766 			  struct rds_info_lengths *lens,
767 			  int (*visitor)(struct rds_connection *, void *),
768 			  u64 *buffer,
769 			  size_t item_len);
770 
771 __printf(2, 3)
772 void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...);
773 #define rds_conn_path_error(cp, fmt...) \
774 	__rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt)
775 
776 static inline int
777 rds_conn_path_transition(struct rds_conn_path *cp, int old, int new)
778 {
779 	return atomic_cmpxchg(&cp->cp_state, old, new) == old;
780 }
781 
782 static inline int
783 rds_conn_transition(struct rds_connection *conn, int old, int new)
784 {
785 	WARN_ON(conn->c_trans->t_mp_capable);
786 	return rds_conn_path_transition(&conn->c_path[0], old, new);
787 }
788 
789 static inline int
790 rds_conn_path_state(struct rds_conn_path *cp)
791 {
792 	return atomic_read(&cp->cp_state);
793 }
794 
795 static inline int
796 rds_conn_state(struct rds_connection *conn)
797 {
798 	WARN_ON(conn->c_trans->t_mp_capable);
799 	return rds_conn_path_state(&conn->c_path[0]);
800 }
801 
802 static inline int
803 rds_conn_path_up(struct rds_conn_path *cp)
804 {
805 	return atomic_read(&cp->cp_state) == RDS_CONN_UP;
806 }
807 
808 static inline int
809 rds_conn_up(struct rds_connection *conn)
810 {
811 	WARN_ON(conn->c_trans->t_mp_capable);
812 	return rds_conn_path_up(&conn->c_path[0]);
813 }
814 
815 static inline int
816 rds_conn_path_connecting(struct rds_conn_path *cp)
817 {
818 	return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING;
819 }
820 
821 static inline int
822 rds_conn_connecting(struct rds_connection *conn)
823 {
824 	WARN_ON(conn->c_trans->t_mp_capable);
825 	return rds_conn_path_connecting(&conn->c_path[0]);
826 }
827 
828 /* message.c */
829 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
830 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
831 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from,
832 			       bool zcopy);
833 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
834 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
835 				 __be16 dport, u64 seq);
836 int rds_message_add_extension(struct rds_header *hdr,
837 			      unsigned int type, const void *data, unsigned int len);
838 int rds_message_next_extension(struct rds_header *hdr,
839 			       unsigned int *pos, void *buf, unsigned int *buflen);
840 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
841 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
842 void rds_message_inc_free(struct rds_incoming *inc);
843 void rds_message_addref(struct rds_message *rm);
844 void rds_message_put(struct rds_message *rm);
845 void rds_message_wait(struct rds_message *rm);
846 void rds_message_unmapped(struct rds_message *rm);
847 void rds_notify_msg_zcopy_purge(struct rds_msg_zcopy_queue *info);
848 
849 static inline void rds_message_make_checksum(struct rds_header *hdr)
850 {
851 	hdr->h_csum = 0;
852 	hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
853 }
854 
855 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
856 {
857 	return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
858 }
859 
860 
861 /* page.c */
862 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
863 			     gfp_t gfp);
864 void rds_page_exit(void);
865 
866 /* recv.c */
867 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
868 		  struct in6_addr *saddr);
869 void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn,
870 		       struct in6_addr *saddr);
871 void rds_inc_put(struct rds_incoming *inc);
872 void rds_recv_incoming(struct rds_connection *conn, struct in6_addr *saddr,
873 		       struct in6_addr *daddr,
874 		       struct rds_incoming *inc, gfp_t gfp);
875 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
876 		int msg_flags);
877 void rds_clear_recv_queue(struct rds_sock *rs);
878 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
879 void rds_inc_info_copy(struct rds_incoming *inc,
880 		       struct rds_info_iterator *iter,
881 		       __be32 saddr, __be32 daddr, int flip);
882 void rds6_inc_info_copy(struct rds_incoming *inc,
883 			struct rds_info_iterator *iter,
884 			struct in6_addr *saddr, struct in6_addr *daddr,
885 			int flip);
886 
887 /* send.c */
888 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
889 void rds_send_path_reset(struct rds_conn_path *conn);
890 int rds_send_xmit(struct rds_conn_path *cp);
891 struct sockaddr_in;
892 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in6 *dest);
893 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
894 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
895 			 is_acked_func is_acked);
896 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
897 			      is_acked_func is_acked);
898 void rds_send_ping(struct rds_connection *conn, int cp_index);
899 int rds_send_pong(struct rds_conn_path *cp, __be16 dport);
900 
901 /* rdma.c */
902 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
903 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
904 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
905 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
906 void rds_rdma_drop_keys(struct rds_sock *rs);
907 int rds_rdma_extra_size(struct rds_rdma_args *args);
908 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
909 			  struct cmsghdr *cmsg);
910 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
911 			  struct cmsghdr *cmsg);
912 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
913 			  struct cmsghdr *cmsg);
914 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
915 			  struct cmsghdr *cmsg);
916 void rds_rdma_free_op(struct rm_rdma_op *ro);
917 void rds_atomic_free_op(struct rm_atomic_op *ao);
918 void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
919 void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
920 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
921 		    struct cmsghdr *cmsg);
922 
923 void __rds_put_mr_final(struct rds_mr *mr);
924 static inline void rds_mr_put(struct rds_mr *mr)
925 {
926 	if (refcount_dec_and_test(&mr->r_refcount))
927 		__rds_put_mr_final(mr);
928 }
929 
930 static inline bool rds_destroy_pending(struct rds_connection *conn)
931 {
932 	return !check_net(rds_conn_net(conn)) ||
933 	       (conn->c_trans->t_unloading && conn->c_trans->t_unloading(conn));
934 }
935 
936 /* stats.c */
937 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
938 #define rds_stats_inc_which(which, member) do {		\
939 	per_cpu(which, get_cpu()).member++;		\
940 	put_cpu();					\
941 } while (0)
942 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
943 #define rds_stats_add_which(which, member, count) do {		\
944 	per_cpu(which, get_cpu()).member += count;	\
945 	put_cpu();					\
946 } while (0)
947 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
948 int rds_stats_init(void);
949 void rds_stats_exit(void);
950 void rds_stats_info_copy(struct rds_info_iterator *iter,
951 			 uint64_t *values, const char *const *names,
952 			 size_t nr);
953 
954 /* sysctl.c */
955 int rds_sysctl_init(void);
956 void rds_sysctl_exit(void);
957 extern unsigned long rds_sysctl_sndbuf_min;
958 extern unsigned long rds_sysctl_sndbuf_default;
959 extern unsigned long rds_sysctl_sndbuf_max;
960 extern unsigned long rds_sysctl_reconnect_min_jiffies;
961 extern unsigned long rds_sysctl_reconnect_max_jiffies;
962 extern unsigned int  rds_sysctl_max_unacked_packets;
963 extern unsigned int  rds_sysctl_max_unacked_bytes;
964 extern unsigned int  rds_sysctl_ping_enable;
965 extern unsigned long rds_sysctl_trace_flags;
966 extern unsigned int  rds_sysctl_trace_level;
967 
968 /* threads.c */
969 int rds_threads_init(void);
970 void rds_threads_exit(void);
971 extern struct workqueue_struct *rds_wq;
972 void rds_queue_reconnect(struct rds_conn_path *cp);
973 void rds_connect_worker(struct work_struct *);
974 void rds_shutdown_worker(struct work_struct *);
975 void rds_send_worker(struct work_struct *);
976 void rds_recv_worker(struct work_struct *);
977 void rds_connect_path_complete(struct rds_conn_path *conn, int curr);
978 void rds_connect_complete(struct rds_connection *conn);
979 int rds_addr_cmp(const struct in6_addr *a1, const struct in6_addr *a2);
980 
981 /* transport.c */
982 void rds_trans_register(struct rds_transport *trans);
983 void rds_trans_unregister(struct rds_transport *trans);
984 struct rds_transport *rds_trans_get_preferred(struct net *net,
985 					      const struct in6_addr *addr,
986 					      __u32 scope_id);
987 void rds_trans_put(struct rds_transport *trans);
988 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
989 				       unsigned int avail);
990 struct rds_transport *rds_trans_get(int t_type);
991 int rds_trans_init(void);
992 void rds_trans_exit(void);
993 
994 #endif
995