xref: /linux/net/rds/rds.h (revision 8f8d5745bb520c76b81abef4a2cb3023d0313bfd)
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_4_0	0x0400
23 #define RDS_PROTOCOL_4_1	0x0401
24 #define RDS_PROTOCOL_VERSION	RDS_PROTOCOL_3_1
25 #define RDS_PROTOCOL_MAJOR(v)	((v) >> 8)
26 #define RDS_PROTOCOL_MINOR(v)	((v) & 255)
27 #define RDS_PROTOCOL(maj, min)	(((maj) << 8) | min)
28 #define RDS_PROTOCOL_COMPAT_VERSION	RDS_PROTOCOL_3_1
29 
30 /* The following ports, 16385, 18634, 18635, are registered with IANA as
31  * the ports to be used for RDS over TCP and UDP.  Currently, only RDS over
32  * TCP and RDS over IB/RDMA are implemented.  18634 is the historical value
33  * used for the RDMA_CM listener port.  RDS/TCP uses port 16385.  After
34  * IPv6 work, RDMA_CM also uses 16385 as the listener port.  18634 is kept
35  * to ensure compatibility with older RDS modules.  Those ports are defined
36  * in each transport's header file.
37  */
38 #define RDS_PORT	18634
39 
40 #ifdef ATOMIC64_INIT
41 #define KERNEL_HAS_ATOMIC64
42 #endif
43 
44 #ifdef RDS_DEBUG
45 #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
46 #else
47 /* sigh, pr_debug() causes unused variable warnings */
48 static inline __printf(1, 2)
49 void rdsdebug(char *fmt, ...)
50 {
51 }
52 #endif
53 
54 #define RDS_FRAG_SHIFT	12
55 #define RDS_FRAG_SIZE	((unsigned int)(1 << RDS_FRAG_SHIFT))
56 
57 /* Used to limit both RDMA and non-RDMA RDS message to 1MB */
58 #define RDS_MAX_MSG_SIZE	((unsigned int)(1 << 20))
59 
60 #define RDS_CONG_MAP_BYTES	(65536 / 8)
61 #define RDS_CONG_MAP_PAGES	(PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
62 #define RDS_CONG_MAP_PAGE_BITS	(PAGE_SIZE * 8)
63 
64 struct rds_cong_map {
65 	struct rb_node		m_rb_node;
66 	struct in6_addr		m_addr;
67 	wait_queue_head_t	m_waitq;
68 	struct list_head	m_conn_list;
69 	unsigned long		m_page_addrs[RDS_CONG_MAP_PAGES];
70 };
71 
72 
73 /*
74  * This is how we will track the connection state:
75  * A connection is always in one of the following
76  * states. Updates to the state are atomic and imply
77  * a memory barrier.
78  */
79 enum {
80 	RDS_CONN_DOWN = 0,
81 	RDS_CONN_CONNECTING,
82 	RDS_CONN_DISCONNECTING,
83 	RDS_CONN_UP,
84 	RDS_CONN_RESETTING,
85 	RDS_CONN_ERROR,
86 };
87 
88 /* Bits for c_flags */
89 #define RDS_LL_SEND_FULL	0
90 #define RDS_RECONNECT_PENDING	1
91 #define RDS_IN_XMIT		2
92 #define RDS_RECV_REFILL		3
93 #define	RDS_DESTROY_PENDING	4
94 
95 /* Max number of multipaths per RDS connection. Must be a power of 2 */
96 #define	RDS_MPATH_WORKERS	8
97 #define	RDS_MPATH_HASH(rs, n) (jhash_1word((rs)->rs_bound_port, \
98 			       (rs)->rs_hash_initval) & ((n) - 1))
99 
100 #define IS_CANONICAL(laddr, faddr) (htonl(laddr) < htonl(faddr))
101 
102 /* Per mpath connection state */
103 struct rds_conn_path {
104 	struct rds_connection	*cp_conn;
105 	struct rds_message	*cp_xmit_rm;
106 	unsigned long		cp_xmit_sg;
107 	unsigned int		cp_xmit_hdr_off;
108 	unsigned int		cp_xmit_data_off;
109 	unsigned int		cp_xmit_atomic_sent;
110 	unsigned int		cp_xmit_rdma_sent;
111 	unsigned int		cp_xmit_data_sent;
112 
113 	spinlock_t		cp_lock;		/* protect msg queues */
114 	u64			cp_next_tx_seq;
115 	struct list_head	cp_send_queue;
116 	struct list_head	cp_retrans;
117 
118 	u64			cp_next_rx_seq;
119 
120 	void			*cp_transport_data;
121 
122 	atomic_t		cp_state;
123 	unsigned long		cp_send_gen;
124 	unsigned long		cp_flags;
125 	unsigned long		cp_reconnect_jiffies;
126 	struct delayed_work	cp_send_w;
127 	struct delayed_work	cp_recv_w;
128 	struct delayed_work	cp_conn_w;
129 	struct work_struct	cp_down_w;
130 	struct mutex		cp_cm_lock;	/* protect cp_state & cm */
131 	wait_queue_head_t	cp_waitq;
132 
133 	unsigned int		cp_unacked_packets;
134 	unsigned int		cp_unacked_bytes;
135 	unsigned int		cp_index;
136 };
137 
138 /* One rds_connection per RDS address pair */
139 struct rds_connection {
140 	struct hlist_node	c_hash_node;
141 	struct in6_addr		c_laddr;
142 	struct in6_addr		c_faddr;
143 	int			c_dev_if; /* ifindex used for this conn */
144 	int			c_bound_if; /* ifindex of c_laddr */
145 	unsigned int		c_loopback:1,
146 				c_isv6:1,
147 				c_ping_triggered:1,
148 				c_pad_to_32:29;
149 	int			c_npaths;
150 	struct rds_connection	*c_passive;
151 	struct rds_transport	*c_trans;
152 
153 	struct rds_cong_map	*c_lcong;
154 	struct rds_cong_map	*c_fcong;
155 
156 	/* Protocol version */
157 	unsigned int		c_proposed_version;
158 	unsigned int		c_version;
159 	possible_net_t		c_net;
160 
161 	/* TOS */
162 	u8			c_tos;
163 
164 	struct list_head	c_map_item;
165 	unsigned long		c_map_queued;
166 
167 	struct rds_conn_path	*c_path;
168 	wait_queue_head_t	c_hs_waitq; /* handshake waitq */
169 
170 	u32			c_my_gen_num;
171 	u32			c_peer_gen_num;
172 };
173 
174 static inline
175 struct net *rds_conn_net(struct rds_connection *conn)
176 {
177 	return read_pnet(&conn->c_net);
178 }
179 
180 static inline
181 void rds_conn_net_set(struct rds_connection *conn, struct net *net)
182 {
183 	write_pnet(&conn->c_net, net);
184 }
185 
186 #define RDS_FLAG_CONG_BITMAP	0x01
187 #define RDS_FLAG_ACK_REQUIRED	0x02
188 #define RDS_FLAG_RETRANSMITTED	0x04
189 #define RDS_MAX_ADV_CREDIT	255
190 
191 /* RDS_FLAG_PROBE_PORT is the reserved sport used for sending a ping
192  * probe to exchange control information before establishing a connection.
193  * Currently the control information that is exchanged is the number of
194  * supported paths. If the peer is a legacy (older kernel revision) peer,
195  * it would return a pong message without additional control information
196  * that would then alert the sender that the peer was an older rev.
197  */
198 #define RDS_FLAG_PROBE_PORT	1
199 #define	RDS_HS_PROBE(sport, dport) \
200 		((sport == RDS_FLAG_PROBE_PORT && dport == 0) || \
201 		 (sport == 0 && dport == RDS_FLAG_PROBE_PORT))
202 /*
203  * Maximum space available for extension headers.
204  */
205 #define RDS_HEADER_EXT_SPACE	16
206 
207 struct rds_header {
208 	__be64	h_sequence;
209 	__be64	h_ack;
210 	__be32	h_len;
211 	__be16	h_sport;
212 	__be16	h_dport;
213 	u8	h_flags;
214 	u8	h_credit;
215 	u8	h_padding[4];
216 	__sum16	h_csum;
217 
218 	u8	h_exthdr[RDS_HEADER_EXT_SPACE];
219 };
220 
221 /*
222  * Reserved - indicates end of extensions
223  */
224 #define RDS_EXTHDR_NONE		0
225 
226 /*
227  * This extension header is included in the very
228  * first message that is sent on a new connection,
229  * and identifies the protocol level. This will help
230  * rolling updates if a future change requires breaking
231  * the protocol.
232  * NB: This is no longer true for IB, where we do a version
233  * negotiation during the connection setup phase (protocol
234  * version information is included in the RDMA CM private data).
235  */
236 #define RDS_EXTHDR_VERSION	1
237 struct rds_ext_header_version {
238 	__be32			h_version;
239 };
240 
241 /*
242  * This extension header is included in the RDS message
243  * chasing an RDMA operation.
244  */
245 #define RDS_EXTHDR_RDMA		2
246 struct rds_ext_header_rdma {
247 	__be32			h_rdma_rkey;
248 };
249 
250 /*
251  * This extension header tells the peer about the
252  * destination <R_Key,offset> of the requested RDMA
253  * operation.
254  */
255 #define RDS_EXTHDR_RDMA_DEST	3
256 struct rds_ext_header_rdma_dest {
257 	__be32			h_rdma_rkey;
258 	__be32			h_rdma_offset;
259 };
260 
261 /* Extension header announcing number of paths.
262  * Implicit length = 2 bytes.
263  */
264 #define RDS_EXTHDR_NPATHS	5
265 #define RDS_EXTHDR_GEN_NUM	6
266 
267 #define __RDS_EXTHDR_MAX	16 /* for now */
268 #define RDS_RX_MAX_TRACES	(RDS_MSG_RX_DGRAM_TRACE_MAX + 1)
269 #define	RDS_MSG_RX_HDR		0
270 #define	RDS_MSG_RX_START	1
271 #define	RDS_MSG_RX_END		2
272 #define	RDS_MSG_RX_CMSG		3
273 
274 struct rds_incoming {
275 	refcount_t		i_refcount;
276 	struct list_head	i_item;
277 	struct rds_connection	*i_conn;
278 	struct rds_conn_path	*i_conn_path;
279 	struct rds_header	i_hdr;
280 	unsigned long		i_rx_jiffies;
281 	struct in6_addr		i_saddr;
282 
283 	rds_rdma_cookie_t	i_rdma_cookie;
284 	ktime_t			i_rx_tstamp;
285 	u64			i_rx_lat_trace[RDS_RX_MAX_TRACES];
286 };
287 
288 struct rds_mr {
289 	struct rb_node		r_rb_node;
290 	refcount_t		r_refcount;
291 	u32			r_key;
292 
293 	/* A copy of the creation flags */
294 	unsigned int		r_use_once:1;
295 	unsigned int		r_invalidate:1;
296 	unsigned int		r_write:1;
297 
298 	/* This is for RDS_MR_DEAD.
299 	 * It would be nice & consistent to make this part of the above
300 	 * bit field here, but we need to use test_and_set_bit.
301 	 */
302 	unsigned long		r_state;
303 	struct rds_sock		*r_sock; /* back pointer to the socket that owns us */
304 	struct rds_transport	*r_trans;
305 	void			*r_trans_private;
306 };
307 
308 /* Flags for mr->r_state */
309 #define RDS_MR_DEAD		0
310 
311 static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
312 {
313 	return r_key | (((u64) offset) << 32);
314 }
315 
316 static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
317 {
318 	return cookie;
319 }
320 
321 static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
322 {
323 	return cookie >> 32;
324 }
325 
326 /* atomic operation types */
327 #define RDS_ATOMIC_TYPE_CSWP		0
328 #define RDS_ATOMIC_TYPE_FADD		1
329 
330 /*
331  * m_sock_item and m_conn_item are on lists that are serialized under
332  * conn->c_lock.  m_sock_item has additional meaning in that once it is empty
333  * the message will not be put back on the retransmit list after being sent.
334  * messages that are canceled while being sent rely on this.
335  *
336  * m_inc is used by loopback so that it can pass an incoming message straight
337  * back up into the rx path.  It embeds a wire header which is also used by
338  * the send path, which is kind of awkward.
339  *
340  * m_sock_item indicates the message's presence on a socket's send or receive
341  * queue.  m_rs will point to that socket.
342  *
343  * m_daddr is used by cancellation to prune messages to a given destination.
344  *
345  * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
346  * nesting.  As paths iterate over messages on a sock, or conn, they must
347  * also lock the conn, or sock, to remove the message from those lists too.
348  * Testing the flag to determine if the message is still on the lists lets
349  * us avoid testing the list_head directly.  That means each path can use
350  * the message's list_head to keep it on a local list while juggling locks
351  * without confusing the other path.
352  *
353  * m_ack_seq is an optional field set by transports who need a different
354  * sequence number range to invalidate.  They can use this in a callback
355  * that they pass to rds_send_drop_acked() to see if each message has been
356  * acked.  The HAS_ACK_SEQ flag can be used to detect messages which haven't
357  * had ack_seq set yet.
358  */
359 #define RDS_MSG_ON_SOCK		1
360 #define RDS_MSG_ON_CONN		2
361 #define RDS_MSG_HAS_ACK_SEQ	3
362 #define RDS_MSG_ACK_REQUIRED	4
363 #define RDS_MSG_RETRANSMITTED	5
364 #define RDS_MSG_MAPPED		6
365 #define RDS_MSG_PAGEVEC		7
366 #define RDS_MSG_FLUSH		8
367 
368 struct rds_znotifier {
369 	struct mmpin		z_mmp;
370 	u32			z_cookie;
371 };
372 
373 struct rds_msg_zcopy_info {
374 	struct list_head rs_zcookie_next;
375 	union {
376 		struct rds_znotifier znotif;
377 		struct rds_zcopy_cookies zcookies;
378 	};
379 };
380 
381 struct rds_msg_zcopy_queue {
382 	struct list_head zcookie_head;
383 	spinlock_t lock; /* protects zcookie_head queue */
384 };
385 
386 static inline void rds_message_zcopy_queue_init(struct rds_msg_zcopy_queue *q)
387 {
388 	spin_lock_init(&q->lock);
389 	INIT_LIST_HEAD(&q->zcookie_head);
390 }
391 
392 struct rds_iov_vector {
393 	struct rds_iovec *iov;
394 	int               len;
395 };
396 
397 struct rds_iov_vector_arr {
398 	struct rds_iov_vector *vec;
399 	int                    len;
400 	int                    indx;
401 	int                    incr;
402 };
403 
404 struct rds_message {
405 	refcount_t		m_refcount;
406 	struct list_head	m_sock_item;
407 	struct list_head	m_conn_item;
408 	struct rds_incoming	m_inc;
409 	u64			m_ack_seq;
410 	struct in6_addr		m_daddr;
411 	unsigned long		m_flags;
412 
413 	/* Never access m_rs without holding m_rs_lock.
414 	 * Lock nesting is
415 	 *  rm->m_rs_lock
416 	 *   -> rs->rs_lock
417 	 */
418 	spinlock_t		m_rs_lock;
419 	wait_queue_head_t	m_flush_wait;
420 
421 	struct rds_sock		*m_rs;
422 
423 	/* cookie to send to remote, in rds header */
424 	rds_rdma_cookie_t	m_rdma_cookie;
425 
426 	unsigned int		m_used_sgs;
427 	unsigned int		m_total_sgs;
428 
429 	void			*m_final_op;
430 
431 	struct {
432 		struct rm_atomic_op {
433 			int			op_type;
434 			union {
435 				struct {
436 					uint64_t	compare;
437 					uint64_t	swap;
438 					uint64_t	compare_mask;
439 					uint64_t	swap_mask;
440 				} op_m_cswp;
441 				struct {
442 					uint64_t	add;
443 					uint64_t	nocarry_mask;
444 				} op_m_fadd;
445 			};
446 
447 			u32			op_rkey;
448 			u64			op_remote_addr;
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 			struct scatterlist	*op_sg;
455 			struct rds_notifier	*op_notifier;
456 
457 			struct rds_mr		*op_rdma_mr;
458 		} atomic;
459 		struct rm_rdma_op {
460 			u32			op_rkey;
461 			u64			op_remote_addr;
462 			unsigned int		op_write:1;
463 			unsigned int		op_fence:1;
464 			unsigned int		op_notify:1;
465 			unsigned int		op_recverr:1;
466 			unsigned int		op_mapped:1;
467 			unsigned int		op_silent:1;
468 			unsigned int		op_active:1;
469 			unsigned int		op_bytes;
470 			unsigned int		op_nents;
471 			unsigned int		op_count;
472 			struct scatterlist	*op_sg;
473 			struct rds_notifier	*op_notifier;
474 
475 			struct rds_mr		*op_rdma_mr;
476 		} rdma;
477 		struct rm_data_op {
478 			unsigned int		op_active:1;
479 			unsigned int		op_notify:1;
480 			unsigned int		op_nents;
481 			unsigned int		op_count;
482 			unsigned int		op_dmasg;
483 			unsigned int		op_dmaoff;
484 			struct rds_znotifier	*op_mmp_znotifier;
485 			struct scatterlist	*op_sg;
486 		} data;
487 	};
488 
489 	struct rds_conn_path *m_conn_path;
490 };
491 
492 /*
493  * The RDS notifier is used (optionally) to tell the application about
494  * completed RDMA operations. Rather than keeping the whole rds message
495  * around on the queue, we allocate a small notifier that is put on the
496  * socket's notifier_list. Notifications are delivered to the application
497  * through control messages.
498  */
499 struct rds_notifier {
500 	struct list_head	n_list;
501 	uint64_t		n_user_token;
502 	int			n_status;
503 };
504 
505 /* Available as part of RDS core, so doesn't need to participate
506  * in get_preferred transport etc
507  */
508 #define	RDS_TRANS_LOOP	3
509 
510 /**
511  * struct rds_transport -  transport specific behavioural hooks
512  *
513  * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
514  *        part of a message.  The caller serializes on the send_sem so this
515  *        doesn't need to be reentrant for a given conn.  The header must be
516  *        sent before the data payload.  .xmit must be prepared to send a
517  *        message with no data payload.  .xmit should return the number of
518  *        bytes that were sent down the connection, including header bytes.
519  *        Returning 0 tells the caller that it doesn't need to perform any
520  *        additional work now.  This is usually the case when the transport has
521  *        filled the sending queue for its connection and will handle
522  *        triggering the rds thread to continue the send when space becomes
523  *        available.  Returning -EAGAIN tells the caller to retry the send
524  *        immediately.  Returning -ENOMEM tells the caller to retry the send at
525  *        some point in the future.
526  *
527  * @conn_shutdown: conn_shutdown stops traffic on the given connection.  Once
528  *                 it returns the connection can not call rds_recv_incoming().
529  *                 This will only be called once after conn_connect returns
530  *                 non-zero success and will The caller serializes this with
531  *                 the send and connecting paths (xmit_* and conn_*).  The
532  *                 transport is responsible for other serialization, including
533  *                 rds_recv_incoming().  This is called in process context but
534  *                 should try hard not to block.
535  */
536 
537 struct rds_transport {
538 	char			t_name[TRANSNAMSIZ];
539 	struct list_head	t_item;
540 	struct module		*t_owner;
541 	unsigned int		t_prefer_loopback:1,
542 				t_mp_capable:1;
543 	unsigned int		t_type;
544 
545 	int (*laddr_check)(struct net *net, const struct in6_addr *addr,
546 			   __u32 scope_id);
547 	int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
548 	void (*conn_free)(void *data);
549 	int (*conn_path_connect)(struct rds_conn_path *cp);
550 	void (*conn_path_shutdown)(struct rds_conn_path *conn);
551 	void (*xmit_path_prepare)(struct rds_conn_path *cp);
552 	void (*xmit_path_complete)(struct rds_conn_path *cp);
553 	int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
554 		    unsigned int hdr_off, unsigned int sg, unsigned int off);
555 	int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
556 	int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
557 	int (*recv_path)(struct rds_conn_path *cp);
558 	int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
559 	void (*inc_free)(struct rds_incoming *inc);
560 
561 	int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
562 				 struct rdma_cm_event *event, bool isv6);
563 	int (*cm_initiate_connect)(struct rdma_cm_id *cm_id, bool isv6);
564 	void (*cm_connect_complete)(struct rds_connection *conn,
565 				    struct rdma_cm_event *event);
566 
567 	unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
568 					unsigned int avail);
569 	void (*exit)(void);
570 	void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
571 			struct rds_sock *rs, u32 *key_ret,
572 			struct rds_connection *conn);
573 	void (*sync_mr)(void *trans_private, int direction);
574 	void (*free_mr)(void *trans_private, int invalidate);
575 	void (*flush_mrs)(void);
576 	bool (*t_unloading)(struct rds_connection *conn);
577 	u8 (*get_tos_map)(u8 tos);
578 };
579 
580 /* Bind hash table key length.  It is the sum of the size of a struct
581  * in6_addr, a scope_id  and a port.
582  */
583 #define RDS_BOUND_KEY_LEN \
584 	(sizeof(struct in6_addr) + sizeof(__u32) + sizeof(__be16))
585 
586 struct rds_sock {
587 	struct sock		rs_sk;
588 
589 	u64			rs_user_addr;
590 	u64			rs_user_bytes;
591 
592 	/*
593 	 * bound_addr used for both incoming and outgoing, no INADDR_ANY
594 	 * support.
595 	 */
596 	struct rhash_head	rs_bound_node;
597 	u8			rs_bound_key[RDS_BOUND_KEY_LEN];
598 	struct sockaddr_in6	rs_bound_sin6;
599 #define rs_bound_addr		rs_bound_sin6.sin6_addr
600 #define rs_bound_addr_v4	rs_bound_sin6.sin6_addr.s6_addr32[3]
601 #define rs_bound_port		rs_bound_sin6.sin6_port
602 #define rs_bound_scope_id	rs_bound_sin6.sin6_scope_id
603 	struct in6_addr		rs_conn_addr;
604 #define rs_conn_addr_v4		rs_conn_addr.s6_addr32[3]
605 	__be16			rs_conn_port;
606 	struct rds_transport    *rs_transport;
607 
608 	/*
609 	 * rds_sendmsg caches the conn it used the last time around.
610 	 * This helps avoid costly lookups.
611 	 */
612 	struct rds_connection	*rs_conn;
613 
614 	/* flag indicating we were congested or not */
615 	int			rs_congested;
616 	/* seen congestion (ENOBUFS) when sending? */
617 	int			rs_seen_congestion;
618 
619 	/* rs_lock protects all these adjacent members before the newline */
620 	spinlock_t		rs_lock;
621 	struct list_head	rs_send_queue;
622 	u32			rs_snd_bytes;
623 	int			rs_rcv_bytes;
624 	struct list_head	rs_notify_queue;	/* currently used for failed RDMAs */
625 
626 	/* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
627 	 * to decide whether the application should be woken up.
628 	 * If not set, we use rs_cong_track to find out whether a cong map
629 	 * update arrived.
630 	 */
631 	uint64_t		rs_cong_mask;
632 	uint64_t		rs_cong_notify;
633 	struct list_head	rs_cong_list;
634 	unsigned long		rs_cong_track;
635 
636 	/*
637 	 * rs_recv_lock protects the receive queue, and is
638 	 * used to serialize with rds_release.
639 	 */
640 	rwlock_t		rs_recv_lock;
641 	struct list_head	rs_recv_queue;
642 
643 	/* just for stats reporting */
644 	struct list_head	rs_item;
645 
646 	/* these have their own lock */
647 	spinlock_t		rs_rdma_lock;
648 	struct rb_root		rs_rdma_keys;
649 
650 	/* Socket options - in case there will be more */
651 	unsigned char		rs_recverr,
652 				rs_cong_monitor;
653 	u32			rs_hash_initval;
654 
655 	/* Socket receive path trace points*/
656 	u8			rs_rx_traces;
657 	u8			rs_rx_trace[RDS_MSG_RX_DGRAM_TRACE_MAX];
658 	struct rds_msg_zcopy_queue rs_zcookie_queue;
659 	u8			rs_tos;
660 };
661 
662 static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
663 {
664 	return container_of(sk, struct rds_sock, rs_sk);
665 }
666 static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
667 {
668 	return &rs->rs_sk;
669 }
670 
671 /*
672  * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
673  * to account for overhead.  We don't account for overhead, we just apply
674  * the number of payload bytes to the specified value.
675  */
676 static inline int rds_sk_sndbuf(struct rds_sock *rs)
677 {
678 	return rds_rs_to_sk(rs)->sk_sndbuf / 2;
679 }
680 static inline int rds_sk_rcvbuf(struct rds_sock *rs)
681 {
682 	return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
683 }
684 
685 struct rds_statistics {
686 	uint64_t	s_conn_reset;
687 	uint64_t	s_recv_drop_bad_checksum;
688 	uint64_t	s_recv_drop_old_seq;
689 	uint64_t	s_recv_drop_no_sock;
690 	uint64_t	s_recv_drop_dead_sock;
691 	uint64_t	s_recv_deliver_raced;
692 	uint64_t	s_recv_delivered;
693 	uint64_t	s_recv_queued;
694 	uint64_t	s_recv_immediate_retry;
695 	uint64_t	s_recv_delayed_retry;
696 	uint64_t	s_recv_ack_required;
697 	uint64_t	s_recv_rdma_bytes;
698 	uint64_t	s_recv_ping;
699 	uint64_t	s_send_queue_empty;
700 	uint64_t	s_send_queue_full;
701 	uint64_t	s_send_lock_contention;
702 	uint64_t	s_send_lock_queue_raced;
703 	uint64_t	s_send_immediate_retry;
704 	uint64_t	s_send_delayed_retry;
705 	uint64_t	s_send_drop_acked;
706 	uint64_t	s_send_ack_required;
707 	uint64_t	s_send_queued;
708 	uint64_t	s_send_rdma;
709 	uint64_t	s_send_rdma_bytes;
710 	uint64_t	s_send_pong;
711 	uint64_t	s_page_remainder_hit;
712 	uint64_t	s_page_remainder_miss;
713 	uint64_t	s_copy_to_user;
714 	uint64_t	s_copy_from_user;
715 	uint64_t	s_cong_update_queued;
716 	uint64_t	s_cong_update_received;
717 	uint64_t	s_cong_send_error;
718 	uint64_t	s_cong_send_blocked;
719 	uint64_t	s_recv_bytes_added_to_socket;
720 	uint64_t	s_recv_bytes_removed_from_socket;
721 
722 };
723 
724 /* af_rds.c */
725 void rds_sock_addref(struct rds_sock *rs);
726 void rds_sock_put(struct rds_sock *rs);
727 void rds_wake_sk_sleep(struct rds_sock *rs);
728 static inline void __rds_wake_sk_sleep(struct sock *sk)
729 {
730 	wait_queue_head_t *waitq = sk_sleep(sk);
731 
732 	if (!sock_flag(sk, SOCK_DEAD) && waitq)
733 		wake_up(waitq);
734 }
735 extern wait_queue_head_t rds_poll_waitq;
736 
737 
738 /* bind.c */
739 int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
740 void rds_remove_bound(struct rds_sock *rs);
741 struct rds_sock *rds_find_bound(const struct in6_addr *addr, __be16 port,
742 				__u32 scope_id);
743 int rds_bind_lock_init(void);
744 void rds_bind_lock_destroy(void);
745 
746 /* cong.c */
747 int rds_cong_get_maps(struct rds_connection *conn);
748 void rds_cong_add_conn(struct rds_connection *conn);
749 void rds_cong_remove_conn(struct rds_connection *conn);
750 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
751 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
752 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
753 void rds_cong_queue_updates(struct rds_cong_map *map);
754 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
755 int rds_cong_updated_since(unsigned long *recent);
756 void rds_cong_add_socket(struct rds_sock *);
757 void rds_cong_remove_socket(struct rds_sock *);
758 void rds_cong_exit(void);
759 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
760 
761 /* connection.c */
762 extern u32 rds_gen_num;
763 int rds_conn_init(void);
764 void rds_conn_exit(void);
765 struct rds_connection *rds_conn_create(struct net *net,
766 				       const struct in6_addr *laddr,
767 				       const struct in6_addr *faddr,
768 				       struct rds_transport *trans,
769 				       u8 tos, gfp_t gfp,
770 				       int dev_if);
771 struct rds_connection *rds_conn_create_outgoing(struct net *net,
772 						const struct in6_addr *laddr,
773 						const struct in6_addr *faddr,
774 						struct rds_transport *trans,
775 						u8 tos, gfp_t gfp, int dev_if);
776 void rds_conn_shutdown(struct rds_conn_path *cpath);
777 void rds_conn_destroy(struct rds_connection *conn);
778 void rds_conn_drop(struct rds_connection *conn);
779 void rds_conn_path_drop(struct rds_conn_path *cpath, bool destroy);
780 void rds_conn_connect_if_down(struct rds_connection *conn);
781 void rds_conn_path_connect_if_down(struct rds_conn_path *cp);
782 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
783 			  struct rds_info_iterator *iter,
784 			  struct rds_info_lengths *lens,
785 			  int (*visitor)(struct rds_connection *, void *),
786 			  u64 *buffer,
787 			  size_t item_len);
788 
789 __printf(2, 3)
790 void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...);
791 #define rds_conn_path_error(cp, fmt...) \
792 	__rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt)
793 
794 static inline int
795 rds_conn_path_transition(struct rds_conn_path *cp, int old, int new)
796 {
797 	return atomic_cmpxchg(&cp->cp_state, old, new) == old;
798 }
799 
800 static inline int
801 rds_conn_transition(struct rds_connection *conn, int old, int new)
802 {
803 	WARN_ON(conn->c_trans->t_mp_capable);
804 	return rds_conn_path_transition(&conn->c_path[0], old, new);
805 }
806 
807 static inline int
808 rds_conn_path_state(struct rds_conn_path *cp)
809 {
810 	return atomic_read(&cp->cp_state);
811 }
812 
813 static inline int
814 rds_conn_state(struct rds_connection *conn)
815 {
816 	WARN_ON(conn->c_trans->t_mp_capable);
817 	return rds_conn_path_state(&conn->c_path[0]);
818 }
819 
820 static inline int
821 rds_conn_path_up(struct rds_conn_path *cp)
822 {
823 	return atomic_read(&cp->cp_state) == RDS_CONN_UP;
824 }
825 
826 static inline int
827 rds_conn_up(struct rds_connection *conn)
828 {
829 	WARN_ON(conn->c_trans->t_mp_capable);
830 	return rds_conn_path_up(&conn->c_path[0]);
831 }
832 
833 static inline int
834 rds_conn_path_connecting(struct rds_conn_path *cp)
835 {
836 	return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING;
837 }
838 
839 static inline int
840 rds_conn_connecting(struct rds_connection *conn)
841 {
842 	WARN_ON(conn->c_trans->t_mp_capable);
843 	return rds_conn_path_connecting(&conn->c_path[0]);
844 }
845 
846 /* message.c */
847 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
848 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents,
849 					  int *ret);
850 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from,
851 			       bool zcopy);
852 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
853 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
854 				 __be16 dport, u64 seq);
855 int rds_message_add_extension(struct rds_header *hdr,
856 			      unsigned int type, const void *data, unsigned int len);
857 int rds_message_next_extension(struct rds_header *hdr,
858 			       unsigned int *pos, void *buf, unsigned int *buflen);
859 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
860 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
861 void rds_message_inc_free(struct rds_incoming *inc);
862 void rds_message_addref(struct rds_message *rm);
863 void rds_message_put(struct rds_message *rm);
864 void rds_message_wait(struct rds_message *rm);
865 void rds_message_unmapped(struct rds_message *rm);
866 void rds_notify_msg_zcopy_purge(struct rds_msg_zcopy_queue *info);
867 
868 static inline void rds_message_make_checksum(struct rds_header *hdr)
869 {
870 	hdr->h_csum = 0;
871 	hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
872 }
873 
874 static inline int rds_message_verify_checksum(const struct rds_header *hdr)
875 {
876 	return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
877 }
878 
879 
880 /* page.c */
881 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
882 			     gfp_t gfp);
883 void rds_page_exit(void);
884 
885 /* recv.c */
886 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
887 		  struct in6_addr *saddr);
888 void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn,
889 		       struct in6_addr *saddr);
890 void rds_inc_put(struct rds_incoming *inc);
891 void rds_recv_incoming(struct rds_connection *conn, struct in6_addr *saddr,
892 		       struct in6_addr *daddr,
893 		       struct rds_incoming *inc, gfp_t gfp);
894 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
895 		int msg_flags);
896 void rds_clear_recv_queue(struct rds_sock *rs);
897 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
898 void rds_inc_info_copy(struct rds_incoming *inc,
899 		       struct rds_info_iterator *iter,
900 		       __be32 saddr, __be32 daddr, int flip);
901 void rds6_inc_info_copy(struct rds_incoming *inc,
902 			struct rds_info_iterator *iter,
903 			struct in6_addr *saddr, struct in6_addr *daddr,
904 			int flip);
905 
906 /* send.c */
907 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
908 void rds_send_path_reset(struct rds_conn_path *conn);
909 int rds_send_xmit(struct rds_conn_path *cp);
910 struct sockaddr_in;
911 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in6 *dest);
912 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
913 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
914 			 is_acked_func is_acked);
915 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
916 			      is_acked_func is_acked);
917 void rds_send_ping(struct rds_connection *conn, int cp_index);
918 int rds_send_pong(struct rds_conn_path *cp, __be16 dport);
919 
920 /* rdma.c */
921 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
922 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
923 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
924 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
925 void rds_rdma_drop_keys(struct rds_sock *rs);
926 int rds_rdma_extra_size(struct rds_rdma_args *args,
927 			struct rds_iov_vector *iov);
928 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
929 			  struct cmsghdr *cmsg);
930 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
931 			  struct cmsghdr *cmsg,
932 			  struct rds_iov_vector *vec);
933 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
934 			  struct cmsghdr *cmsg);
935 void rds_rdma_free_op(struct rm_rdma_op *ro);
936 void rds_atomic_free_op(struct rm_atomic_op *ao);
937 void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
938 void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
939 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
940 		    struct cmsghdr *cmsg);
941 
942 void __rds_put_mr_final(struct rds_mr *mr);
943 static inline void rds_mr_put(struct rds_mr *mr)
944 {
945 	if (refcount_dec_and_test(&mr->r_refcount))
946 		__rds_put_mr_final(mr);
947 }
948 
949 static inline bool rds_destroy_pending(struct rds_connection *conn)
950 {
951 	return !check_net(rds_conn_net(conn)) ||
952 	       (conn->c_trans->t_unloading && conn->c_trans->t_unloading(conn));
953 }
954 
955 /* stats.c */
956 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
957 #define rds_stats_inc_which(which, member) do {		\
958 	per_cpu(which, get_cpu()).member++;		\
959 	put_cpu();					\
960 } while (0)
961 #define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
962 #define rds_stats_add_which(which, member, count) do {		\
963 	per_cpu(which, get_cpu()).member += count;	\
964 	put_cpu();					\
965 } while (0)
966 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
967 int rds_stats_init(void);
968 void rds_stats_exit(void);
969 void rds_stats_info_copy(struct rds_info_iterator *iter,
970 			 uint64_t *values, const char *const *names,
971 			 size_t nr);
972 
973 /* sysctl.c */
974 int rds_sysctl_init(void);
975 void rds_sysctl_exit(void);
976 extern unsigned long rds_sysctl_sndbuf_min;
977 extern unsigned long rds_sysctl_sndbuf_default;
978 extern unsigned long rds_sysctl_sndbuf_max;
979 extern unsigned long rds_sysctl_reconnect_min_jiffies;
980 extern unsigned long rds_sysctl_reconnect_max_jiffies;
981 extern unsigned int  rds_sysctl_max_unacked_packets;
982 extern unsigned int  rds_sysctl_max_unacked_bytes;
983 extern unsigned int  rds_sysctl_ping_enable;
984 extern unsigned long rds_sysctl_trace_flags;
985 extern unsigned int  rds_sysctl_trace_level;
986 
987 /* threads.c */
988 int rds_threads_init(void);
989 void rds_threads_exit(void);
990 extern struct workqueue_struct *rds_wq;
991 void rds_queue_reconnect(struct rds_conn_path *cp);
992 void rds_connect_worker(struct work_struct *);
993 void rds_shutdown_worker(struct work_struct *);
994 void rds_send_worker(struct work_struct *);
995 void rds_recv_worker(struct work_struct *);
996 void rds_connect_path_complete(struct rds_conn_path *conn, int curr);
997 void rds_connect_complete(struct rds_connection *conn);
998 int rds_addr_cmp(const struct in6_addr *a1, const struct in6_addr *a2);
999 
1000 /* transport.c */
1001 void rds_trans_register(struct rds_transport *trans);
1002 void rds_trans_unregister(struct rds_transport *trans);
1003 struct rds_transport *rds_trans_get_preferred(struct net *net,
1004 					      const struct in6_addr *addr,
1005 					      __u32 scope_id);
1006 void rds_trans_put(struct rds_transport *trans);
1007 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
1008 				       unsigned int avail);
1009 struct rds_transport *rds_trans_get(int t_type);
1010 int rds_trans_init(void);
1011 void rds_trans_exit(void);
1012 
1013 #endif
1014