Loading...
1#ifndef _RDS_RDS_H
2#define _RDS_RDS_H
3
4#include <net/sock.h>
5#include <linux/scatterlist.h>
6#include <linux/highmem.h>
7#include <rdma/rdma_cm.h>
8#include <linux/mutex.h>
9#include <linux/rds.h>
10#include <linux/rhashtable.h>
11
12#include "info.h"
13
14/*
15 * RDS Network protocol version
16 */
17#define RDS_PROTOCOL_3_0 0x0300
18#define RDS_PROTOCOL_3_1 0x0301
19#define RDS_PROTOCOL_VERSION RDS_PROTOCOL_3_1
20#define RDS_PROTOCOL_MAJOR(v) ((v) >> 8)
21#define RDS_PROTOCOL_MINOR(v) ((v) & 255)
22#define RDS_PROTOCOL(maj, min) (((maj) << 8) | min)
23
24/*
25 * XXX randomly chosen, but at least seems to be unused:
26 * # 18464-18768 Unassigned
27 * We should do better. We want a reserved port to discourage unpriv'ed
28 * userspace from listening.
29 */
30#define RDS_PORT 18634
31
32#ifdef ATOMIC64_INIT
33#define KERNEL_HAS_ATOMIC64
34#endif
35
36#ifdef DEBUG
37#define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
38#else
39/* sigh, pr_debug() causes unused variable warnings */
40static inline __printf(1, 2)
41void rdsdebug(char *fmt, ...)
42{
43}
44#endif
45
46/* XXX is there one of these somewhere? */
47#define ceil(x, y) \
48 ({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; })
49
50#define RDS_FRAG_SHIFT 12
51#define RDS_FRAG_SIZE ((unsigned int)(1 << RDS_FRAG_SHIFT))
52
53#define RDS_CONG_MAP_BYTES (65536 / 8)
54#define RDS_CONG_MAP_PAGES (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
55#define RDS_CONG_MAP_PAGE_BITS (PAGE_SIZE * 8)
56
57struct rds_cong_map {
58 struct rb_node m_rb_node;
59 __be32 m_addr;
60 wait_queue_head_t m_waitq;
61 struct list_head m_conn_list;
62 unsigned long m_page_addrs[RDS_CONG_MAP_PAGES];
63};
64
65
66/*
67 * This is how we will track the connection state:
68 * A connection is always in one of the following
69 * states. Updates to the state are atomic and imply
70 * a memory barrier.
71 */
72enum {
73 RDS_CONN_DOWN = 0,
74 RDS_CONN_CONNECTING,
75 RDS_CONN_DISCONNECTING,
76 RDS_CONN_UP,
77 RDS_CONN_ERROR,
78};
79
80/* Bits for c_flags */
81#define RDS_LL_SEND_FULL 0
82#define RDS_RECONNECT_PENDING 1
83#define RDS_IN_XMIT 2
84#define RDS_RECV_REFILL 3
85
86struct rds_connection {
87 struct hlist_node c_hash_node;
88 __be32 c_laddr;
89 __be32 c_faddr;
90 unsigned int c_loopback:1,
91 c_outgoing:1,
92 c_pad_to_32:30;
93 struct rds_connection *c_passive;
94
95 struct rds_cong_map *c_lcong;
96 struct rds_cong_map *c_fcong;
97
98 struct rds_message *c_xmit_rm;
99 unsigned long c_xmit_sg;
100 unsigned int c_xmit_hdr_off;
101 unsigned int c_xmit_data_off;
102 unsigned int c_xmit_atomic_sent;
103 unsigned int c_xmit_rdma_sent;
104 unsigned int c_xmit_data_sent;
105
106 spinlock_t c_lock; /* protect msg queues */
107 u64 c_next_tx_seq;
108 struct list_head c_send_queue;
109 struct list_head c_retrans;
110
111 u64 c_next_rx_seq;
112
113 struct rds_transport *c_trans;
114 void *c_transport_data;
115
116 atomic_t c_state;
117 unsigned long c_send_gen;
118 unsigned long c_flags;
119 unsigned long c_reconnect_jiffies;
120 struct delayed_work c_send_w;
121 struct delayed_work c_recv_w;
122 struct delayed_work c_conn_w;
123 struct work_struct c_down_w;
124 struct mutex c_cm_lock; /* protect conn state & cm */
125 wait_queue_head_t c_waitq;
126
127 struct list_head c_map_item;
128 unsigned long c_map_queued;
129
130 unsigned int c_unacked_packets;
131 unsigned int c_unacked_bytes;
132
133 /* Protocol version */
134 unsigned int c_version;
135 possible_net_t c_net;
136};
137
138static inline
139struct net *rds_conn_net(struct rds_connection *conn)
140{
141 return read_pnet(&conn->c_net);
142}
143
144static inline
145void rds_conn_net_set(struct rds_connection *conn, struct net *net)
146{
147 write_pnet(&conn->c_net, net);
148}
149
150#define RDS_FLAG_CONG_BITMAP 0x01
151#define RDS_FLAG_ACK_REQUIRED 0x02
152#define RDS_FLAG_RETRANSMITTED 0x04
153#define RDS_MAX_ADV_CREDIT 255
154
155/*
156 * Maximum space available for extension headers.
157 */
158#define RDS_HEADER_EXT_SPACE 16
159
160struct rds_header {
161 __be64 h_sequence;
162 __be64 h_ack;
163 __be32 h_len;
164 __be16 h_sport;
165 __be16 h_dport;
166 u8 h_flags;
167 u8 h_credit;
168 u8 h_padding[4];
169 __sum16 h_csum;
170
171 u8 h_exthdr[RDS_HEADER_EXT_SPACE];
172};
173
174/*
175 * Reserved - indicates end of extensions
176 */
177#define RDS_EXTHDR_NONE 0
178
179/*
180 * This extension header is included in the very
181 * first message that is sent on a new connection,
182 * and identifies the protocol level. This will help
183 * rolling updates if a future change requires breaking
184 * the protocol.
185 * NB: This is no longer true for IB, where we do a version
186 * negotiation during the connection setup phase (protocol
187 * version information is included in the RDMA CM private data).
188 */
189#define RDS_EXTHDR_VERSION 1
190struct rds_ext_header_version {
191 __be32 h_version;
192};
193
194/*
195 * This extension header is included in the RDS message
196 * chasing an RDMA operation.
197 */
198#define RDS_EXTHDR_RDMA 2
199struct rds_ext_header_rdma {
200 __be32 h_rdma_rkey;
201};
202
203/*
204 * This extension header tells the peer about the
205 * destination <R_Key,offset> of the requested RDMA
206 * operation.
207 */
208#define RDS_EXTHDR_RDMA_DEST 3
209struct rds_ext_header_rdma_dest {
210 __be32 h_rdma_rkey;
211 __be32 h_rdma_offset;
212};
213
214#define __RDS_EXTHDR_MAX 16 /* for now */
215
216struct rds_incoming {
217 atomic_t i_refcount;
218 struct list_head i_item;
219 struct rds_connection *i_conn;
220 struct rds_header i_hdr;
221 unsigned long i_rx_jiffies;
222 __be32 i_saddr;
223
224 rds_rdma_cookie_t i_rdma_cookie;
225 struct timeval i_rx_tstamp;
226};
227
228struct rds_mr {
229 struct rb_node r_rb_node;
230 atomic_t r_refcount;
231 u32 r_key;
232
233 /* A copy of the creation flags */
234 unsigned int r_use_once:1;
235 unsigned int r_invalidate:1;
236 unsigned int r_write:1;
237
238 /* This is for RDS_MR_DEAD.
239 * It would be nice & consistent to make this part of the above
240 * bit field here, but we need to use test_and_set_bit.
241 */
242 unsigned long r_state;
243 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
244 struct rds_transport *r_trans;
245 void *r_trans_private;
246};
247
248/* Flags for mr->r_state */
249#define RDS_MR_DEAD 0
250
251static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
252{
253 return r_key | (((u64) offset) << 32);
254}
255
256static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
257{
258 return cookie;
259}
260
261static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
262{
263 return cookie >> 32;
264}
265
266/* atomic operation types */
267#define RDS_ATOMIC_TYPE_CSWP 0
268#define RDS_ATOMIC_TYPE_FADD 1
269
270/*
271 * m_sock_item and m_conn_item are on lists that are serialized under
272 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
273 * the message will not be put back on the retransmit list after being sent.
274 * messages that are canceled while being sent rely on this.
275 *
276 * m_inc is used by loopback so that it can pass an incoming message straight
277 * back up into the rx path. It embeds a wire header which is also used by
278 * the send path, which is kind of awkward.
279 *
280 * m_sock_item indicates the message's presence on a socket's send or receive
281 * queue. m_rs will point to that socket.
282 *
283 * m_daddr is used by cancellation to prune messages to a given destination.
284 *
285 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
286 * nesting. As paths iterate over messages on a sock, or conn, they must
287 * also lock the conn, or sock, to remove the message from those lists too.
288 * Testing the flag to determine if the message is still on the lists lets
289 * us avoid testing the list_head directly. That means each path can use
290 * the message's list_head to keep it on a local list while juggling locks
291 * without confusing the other path.
292 *
293 * m_ack_seq is an optional field set by transports who need a different
294 * sequence number range to invalidate. They can use this in a callback
295 * that they pass to rds_send_drop_acked() to see if each message has been
296 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
297 * had ack_seq set yet.
298 */
299#define RDS_MSG_ON_SOCK 1
300#define RDS_MSG_ON_CONN 2
301#define RDS_MSG_HAS_ACK_SEQ 3
302#define RDS_MSG_ACK_REQUIRED 4
303#define RDS_MSG_RETRANSMITTED 5
304#define RDS_MSG_MAPPED 6
305#define RDS_MSG_PAGEVEC 7
306
307struct rds_message {
308 atomic_t m_refcount;
309 struct list_head m_sock_item;
310 struct list_head m_conn_item;
311 struct rds_incoming m_inc;
312 u64 m_ack_seq;
313 __be32 m_daddr;
314 unsigned long m_flags;
315
316 /* Never access m_rs without holding m_rs_lock.
317 * Lock nesting is
318 * rm->m_rs_lock
319 * -> rs->rs_lock
320 */
321 spinlock_t m_rs_lock;
322 wait_queue_head_t m_flush_wait;
323
324 struct rds_sock *m_rs;
325
326 /* cookie to send to remote, in rds header */
327 rds_rdma_cookie_t m_rdma_cookie;
328
329 unsigned int m_used_sgs;
330 unsigned int m_total_sgs;
331
332 void *m_final_op;
333
334 struct {
335 struct rm_atomic_op {
336 int op_type;
337 union {
338 struct {
339 uint64_t compare;
340 uint64_t swap;
341 uint64_t compare_mask;
342 uint64_t swap_mask;
343 } op_m_cswp;
344 struct {
345 uint64_t add;
346 uint64_t nocarry_mask;
347 } op_m_fadd;
348 };
349
350 u32 op_rkey;
351 u64 op_remote_addr;
352 unsigned int op_notify:1;
353 unsigned int op_recverr:1;
354 unsigned int op_mapped:1;
355 unsigned int op_silent:1;
356 unsigned int op_active:1;
357 struct scatterlist *op_sg;
358 struct rds_notifier *op_notifier;
359
360 struct rds_mr *op_rdma_mr;
361 } atomic;
362 struct rm_rdma_op {
363 u32 op_rkey;
364 u64 op_remote_addr;
365 unsigned int op_write:1;
366 unsigned int op_fence:1;
367 unsigned int op_notify:1;
368 unsigned int op_recverr:1;
369 unsigned int op_mapped:1;
370 unsigned int op_silent:1;
371 unsigned int op_active:1;
372 unsigned int op_bytes;
373 unsigned int op_nents;
374 unsigned int op_count;
375 struct scatterlist *op_sg;
376 struct rds_notifier *op_notifier;
377
378 struct rds_mr *op_rdma_mr;
379 } rdma;
380 struct rm_data_op {
381 unsigned int op_active:1;
382 unsigned int op_nents;
383 unsigned int op_count;
384 unsigned int op_dmasg;
385 unsigned int op_dmaoff;
386 struct scatterlist *op_sg;
387 } data;
388 };
389};
390
391/*
392 * The RDS notifier is used (optionally) to tell the application about
393 * completed RDMA operations. Rather than keeping the whole rds message
394 * around on the queue, we allocate a small notifier that is put on the
395 * socket's notifier_list. Notifications are delivered to the application
396 * through control messages.
397 */
398struct rds_notifier {
399 struct list_head n_list;
400 uint64_t n_user_token;
401 int n_status;
402};
403
404/**
405 * struct rds_transport - transport specific behavioural hooks
406 *
407 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
408 * part of a message. The caller serializes on the send_sem so this
409 * doesn't need to be reentrant for a given conn. The header must be
410 * sent before the data payload. .xmit must be prepared to send a
411 * message with no data payload. .xmit should return the number of
412 * bytes that were sent down the connection, including header bytes.
413 * Returning 0 tells the caller that it doesn't need to perform any
414 * additional work now. This is usually the case when the transport has
415 * filled the sending queue for its connection and will handle
416 * triggering the rds thread to continue the send when space becomes
417 * available. Returning -EAGAIN tells the caller to retry the send
418 * immediately. Returning -ENOMEM tells the caller to retry the send at
419 * some point in the future.
420 *
421 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
422 * it returns the connection can not call rds_recv_incoming().
423 * This will only be called once after conn_connect returns
424 * non-zero success and will The caller serializes this with
425 * the send and connecting paths (xmit_* and conn_*). The
426 * transport is responsible for other serialization, including
427 * rds_recv_incoming(). This is called in process context but
428 * should try hard not to block.
429 */
430
431struct rds_transport {
432 char t_name[TRANSNAMSIZ];
433 struct list_head t_item;
434 struct module *t_owner;
435 unsigned int t_prefer_loopback:1;
436 unsigned int t_type;
437
438 int (*laddr_check)(struct net *net, __be32 addr);
439 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
440 void (*conn_free)(void *data);
441 int (*conn_connect)(struct rds_connection *conn);
442 void (*conn_shutdown)(struct rds_connection *conn);
443 void (*xmit_prepare)(struct rds_connection *conn);
444 void (*xmit_complete)(struct rds_connection *conn);
445 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
446 unsigned int hdr_off, unsigned int sg, unsigned int off);
447 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
448 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
449 int (*recv)(struct rds_connection *conn);
450 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
451 void (*inc_free)(struct rds_incoming *inc);
452
453 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
454 struct rdma_cm_event *event);
455 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
456 void (*cm_connect_complete)(struct rds_connection *conn,
457 struct rdma_cm_event *event);
458
459 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
460 unsigned int avail);
461 void (*exit)(void);
462 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
463 struct rds_sock *rs, u32 *key_ret);
464 void (*sync_mr)(void *trans_private, int direction);
465 void (*free_mr)(void *trans_private, int invalidate);
466 void (*flush_mrs)(void);
467};
468
469struct rds_sock {
470 struct sock rs_sk;
471
472 u64 rs_user_addr;
473 u64 rs_user_bytes;
474
475 /*
476 * bound_addr used for both incoming and outgoing, no INADDR_ANY
477 * support.
478 */
479 struct rhash_head rs_bound_node;
480 u64 rs_bound_key;
481 __be32 rs_bound_addr;
482 __be32 rs_conn_addr;
483 __be16 rs_bound_port;
484 __be16 rs_conn_port;
485 struct rds_transport *rs_transport;
486
487 /*
488 * rds_sendmsg caches the conn it used the last time around.
489 * This helps avoid costly lookups.
490 */
491 struct rds_connection *rs_conn;
492
493 /* flag indicating we were congested or not */
494 int rs_congested;
495 /* seen congestion (ENOBUFS) when sending? */
496 int rs_seen_congestion;
497
498 /* rs_lock protects all these adjacent members before the newline */
499 spinlock_t rs_lock;
500 struct list_head rs_send_queue;
501 u32 rs_snd_bytes;
502 int rs_rcv_bytes;
503 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
504
505 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
506 * to decide whether the application should be woken up.
507 * If not set, we use rs_cong_track to find out whether a cong map
508 * update arrived.
509 */
510 uint64_t rs_cong_mask;
511 uint64_t rs_cong_notify;
512 struct list_head rs_cong_list;
513 unsigned long rs_cong_track;
514
515 /*
516 * rs_recv_lock protects the receive queue, and is
517 * used to serialize with rds_release.
518 */
519 rwlock_t rs_recv_lock;
520 struct list_head rs_recv_queue;
521
522 /* just for stats reporting */
523 struct list_head rs_item;
524
525 /* these have their own lock */
526 spinlock_t rs_rdma_lock;
527 struct rb_root rs_rdma_keys;
528
529 /* Socket options - in case there will be more */
530 unsigned char rs_recverr,
531 rs_cong_monitor;
532};
533
534static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
535{
536 return container_of(sk, struct rds_sock, rs_sk);
537}
538static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
539{
540 return &rs->rs_sk;
541}
542
543/*
544 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
545 * to account for overhead. We don't account for overhead, we just apply
546 * the number of payload bytes to the specified value.
547 */
548static inline int rds_sk_sndbuf(struct rds_sock *rs)
549{
550 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
551}
552static inline int rds_sk_rcvbuf(struct rds_sock *rs)
553{
554 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
555}
556
557struct rds_statistics {
558 uint64_t s_conn_reset;
559 uint64_t s_recv_drop_bad_checksum;
560 uint64_t s_recv_drop_old_seq;
561 uint64_t s_recv_drop_no_sock;
562 uint64_t s_recv_drop_dead_sock;
563 uint64_t s_recv_deliver_raced;
564 uint64_t s_recv_delivered;
565 uint64_t s_recv_queued;
566 uint64_t s_recv_immediate_retry;
567 uint64_t s_recv_delayed_retry;
568 uint64_t s_recv_ack_required;
569 uint64_t s_recv_rdma_bytes;
570 uint64_t s_recv_ping;
571 uint64_t s_send_queue_empty;
572 uint64_t s_send_queue_full;
573 uint64_t s_send_lock_contention;
574 uint64_t s_send_lock_queue_raced;
575 uint64_t s_send_immediate_retry;
576 uint64_t s_send_delayed_retry;
577 uint64_t s_send_drop_acked;
578 uint64_t s_send_ack_required;
579 uint64_t s_send_queued;
580 uint64_t s_send_rdma;
581 uint64_t s_send_rdma_bytes;
582 uint64_t s_send_pong;
583 uint64_t s_page_remainder_hit;
584 uint64_t s_page_remainder_miss;
585 uint64_t s_copy_to_user;
586 uint64_t s_copy_from_user;
587 uint64_t s_cong_update_queued;
588 uint64_t s_cong_update_received;
589 uint64_t s_cong_send_error;
590 uint64_t s_cong_send_blocked;
591};
592
593/* af_rds.c */
594void rds_sock_addref(struct rds_sock *rs);
595void rds_sock_put(struct rds_sock *rs);
596void rds_wake_sk_sleep(struct rds_sock *rs);
597static inline void __rds_wake_sk_sleep(struct sock *sk)
598{
599 wait_queue_head_t *waitq = sk_sleep(sk);
600
601 if (!sock_flag(sk, SOCK_DEAD) && waitq)
602 wake_up(waitq);
603}
604extern wait_queue_head_t rds_poll_waitq;
605
606
607/* bind.c */
608int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
609void rds_remove_bound(struct rds_sock *rs);
610struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
611int rds_bind_lock_init(void);
612void rds_bind_lock_destroy(void);
613
614/* cong.c */
615int rds_cong_get_maps(struct rds_connection *conn);
616void rds_cong_add_conn(struct rds_connection *conn);
617void rds_cong_remove_conn(struct rds_connection *conn);
618void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
619void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
620int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
621void rds_cong_queue_updates(struct rds_cong_map *map);
622void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
623int rds_cong_updated_since(unsigned long *recent);
624void rds_cong_add_socket(struct rds_sock *);
625void rds_cong_remove_socket(struct rds_sock *);
626void rds_cong_exit(void);
627struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
628
629/* conn.c */
630int rds_conn_init(void);
631void rds_conn_exit(void);
632struct rds_connection *rds_conn_create(struct net *net,
633 __be32 laddr, __be32 faddr,
634 struct rds_transport *trans, gfp_t gfp);
635struct rds_connection *rds_conn_create_outgoing(struct net *net,
636 __be32 laddr, __be32 faddr,
637 struct rds_transport *trans, gfp_t gfp);
638void rds_conn_shutdown(struct rds_connection *conn);
639void rds_conn_destroy(struct rds_connection *conn);
640void rds_conn_drop(struct rds_connection *conn);
641void rds_conn_connect_if_down(struct rds_connection *conn);
642void rds_for_each_conn_info(struct socket *sock, unsigned int len,
643 struct rds_info_iterator *iter,
644 struct rds_info_lengths *lens,
645 int (*visitor)(struct rds_connection *, void *),
646 size_t item_len);
647__printf(2, 3)
648void __rds_conn_error(struct rds_connection *conn, const char *, ...);
649#define rds_conn_error(conn, fmt...) \
650 __rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
651
652static inline int
653rds_conn_transition(struct rds_connection *conn, int old, int new)
654{
655 return atomic_cmpxchg(&conn->c_state, old, new) == old;
656}
657
658static inline int
659rds_conn_state(struct rds_connection *conn)
660{
661 return atomic_read(&conn->c_state);
662}
663
664static inline int
665rds_conn_up(struct rds_connection *conn)
666{
667 return atomic_read(&conn->c_state) == RDS_CONN_UP;
668}
669
670static inline int
671rds_conn_connecting(struct rds_connection *conn)
672{
673 return atomic_read(&conn->c_state) == RDS_CONN_CONNECTING;
674}
675
676/* message.c */
677struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
678struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
679int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from);
680struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
681void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
682 __be16 dport, u64 seq);
683int rds_message_add_extension(struct rds_header *hdr,
684 unsigned int type, const void *data, unsigned int len);
685int rds_message_next_extension(struct rds_header *hdr,
686 unsigned int *pos, void *buf, unsigned int *buflen);
687int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
688int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
689void rds_message_inc_free(struct rds_incoming *inc);
690void rds_message_addref(struct rds_message *rm);
691void rds_message_put(struct rds_message *rm);
692void rds_message_wait(struct rds_message *rm);
693void rds_message_unmapped(struct rds_message *rm);
694
695static inline void rds_message_make_checksum(struct rds_header *hdr)
696{
697 hdr->h_csum = 0;
698 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
699}
700
701static inline int rds_message_verify_checksum(const struct rds_header *hdr)
702{
703 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
704}
705
706
707/* page.c */
708int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
709 gfp_t gfp);
710int rds_page_copy_user(struct page *page, unsigned long offset,
711 void __user *ptr, unsigned long bytes,
712 int to_user);
713#define rds_page_copy_to_user(page, offset, ptr, bytes) \
714 rds_page_copy_user(page, offset, ptr, bytes, 1)
715#define rds_page_copy_from_user(page, offset, ptr, bytes) \
716 rds_page_copy_user(page, offset, ptr, bytes, 0)
717void rds_page_exit(void);
718
719/* recv.c */
720void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
721 __be32 saddr);
722void rds_inc_put(struct rds_incoming *inc);
723void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
724 struct rds_incoming *inc, gfp_t gfp);
725int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
726 int msg_flags);
727void rds_clear_recv_queue(struct rds_sock *rs);
728int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
729void rds_inc_info_copy(struct rds_incoming *inc,
730 struct rds_info_iterator *iter,
731 __be32 saddr, __be32 daddr, int flip);
732
733/* send.c */
734int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
735void rds_send_reset(struct rds_connection *conn);
736int rds_send_xmit(struct rds_connection *conn);
737struct sockaddr_in;
738void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
739typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
740void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
741 is_acked_func is_acked);
742int rds_send_pong(struct rds_connection *conn, __be16 dport);
743struct rds_message *rds_send_get_message(struct rds_connection *,
744 struct rm_rdma_op *);
745
746/* rdma.c */
747void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
748int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
749int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
750int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
751void rds_rdma_drop_keys(struct rds_sock *rs);
752int rds_rdma_extra_size(struct rds_rdma_args *args);
753int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
754 struct cmsghdr *cmsg);
755int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
756 struct cmsghdr *cmsg);
757int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
758 struct cmsghdr *cmsg);
759int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
760 struct cmsghdr *cmsg);
761void rds_rdma_free_op(struct rm_rdma_op *ro);
762void rds_atomic_free_op(struct rm_atomic_op *ao);
763void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
764void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
765int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
766 struct cmsghdr *cmsg);
767
768void __rds_put_mr_final(struct rds_mr *mr);
769static inline void rds_mr_put(struct rds_mr *mr)
770{
771 if (atomic_dec_and_test(&mr->r_refcount))
772 __rds_put_mr_final(mr);
773}
774
775/* stats.c */
776DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
777#define rds_stats_inc_which(which, member) do { \
778 per_cpu(which, get_cpu()).member++; \
779 put_cpu(); \
780} while (0)
781#define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
782#define rds_stats_add_which(which, member, count) do { \
783 per_cpu(which, get_cpu()).member += count; \
784 put_cpu(); \
785} while (0)
786#define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
787int rds_stats_init(void);
788void rds_stats_exit(void);
789void rds_stats_info_copy(struct rds_info_iterator *iter,
790 uint64_t *values, const char *const *names,
791 size_t nr);
792
793/* sysctl.c */
794int rds_sysctl_init(void);
795void rds_sysctl_exit(void);
796extern unsigned long rds_sysctl_sndbuf_min;
797extern unsigned long rds_sysctl_sndbuf_default;
798extern unsigned long rds_sysctl_sndbuf_max;
799extern unsigned long rds_sysctl_reconnect_min_jiffies;
800extern unsigned long rds_sysctl_reconnect_max_jiffies;
801extern unsigned int rds_sysctl_max_unacked_packets;
802extern unsigned int rds_sysctl_max_unacked_bytes;
803extern unsigned int rds_sysctl_ping_enable;
804extern unsigned long rds_sysctl_trace_flags;
805extern unsigned int rds_sysctl_trace_level;
806
807/* threads.c */
808int rds_threads_init(void);
809void rds_threads_exit(void);
810extern struct workqueue_struct *rds_wq;
811void rds_queue_reconnect(struct rds_connection *conn);
812void rds_connect_worker(struct work_struct *);
813void rds_shutdown_worker(struct work_struct *);
814void rds_send_worker(struct work_struct *);
815void rds_recv_worker(struct work_struct *);
816void rds_connect_complete(struct rds_connection *conn);
817
818/* transport.c */
819int rds_trans_register(struct rds_transport *trans);
820void rds_trans_unregister(struct rds_transport *trans);
821struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr);
822void rds_trans_put(struct rds_transport *trans);
823unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
824 unsigned int avail);
825struct rds_transport *rds_trans_get(int t_type);
826int rds_trans_init(void);
827void rds_trans_exit(void);
828
829#endif
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 */
42static inline __printf(1, 2)
43void 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
62struct 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 */
77enum {
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 */
101struct 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 */
137struct 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
165static inline
166struct net *rds_conn_net(struct rds_connection *conn)
167{
168 return read_pnet(&conn->c_net);
169}
170
171static inline
172void 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
198struct 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
228struct 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
237struct 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
247struct 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
265struct 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
279struct 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
302static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
303{
304 return r_key | (((u64) offset) << 32);
305}
306
307static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
308{
309 return cookie;
310}
311
312static 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
359struct rds_znotifier {
360 struct mmpin z_mmp;
361 u32 z_cookie;
362};
363
364struct rds_msg_zcopy_info {
365 struct list_head rs_zcookie_next;
366 union {
367 struct rds_znotifier znotif;
368 struct rds_zcopy_cookies zcookies;
369 };
370};
371
372struct rds_msg_zcopy_queue {
373 struct list_head zcookie_head;
374 spinlock_t lock; /* protects zcookie_head queue */
375};
376
377static inline void rds_message_zcopy_queue_init(struct rds_msg_zcopy_queue *q)
378{
379 spin_lock_init(&q->lock);
380 INIT_LIST_HEAD(&q->zcookie_head);
381}
382
383struct rds_message {
384 refcount_t m_refcount;
385 struct list_head m_sock_item;
386 struct list_head m_conn_item;
387 struct rds_incoming m_inc;
388 u64 m_ack_seq;
389 __be32 m_daddr;
390 unsigned long m_flags;
391
392 /* Never access m_rs without holding m_rs_lock.
393 * Lock nesting is
394 * rm->m_rs_lock
395 * -> rs->rs_lock
396 */
397 spinlock_t m_rs_lock;
398 wait_queue_head_t m_flush_wait;
399
400 struct rds_sock *m_rs;
401
402 /* cookie to send to remote, in rds header */
403 rds_rdma_cookie_t m_rdma_cookie;
404
405 unsigned int m_used_sgs;
406 unsigned int m_total_sgs;
407
408 void *m_final_op;
409
410 struct {
411 struct rm_atomic_op {
412 int op_type;
413 union {
414 struct {
415 uint64_t compare;
416 uint64_t swap;
417 uint64_t compare_mask;
418 uint64_t swap_mask;
419 } op_m_cswp;
420 struct {
421 uint64_t add;
422 uint64_t nocarry_mask;
423 } op_m_fadd;
424 };
425
426 u32 op_rkey;
427 u64 op_remote_addr;
428 unsigned int op_notify:1;
429 unsigned int op_recverr:1;
430 unsigned int op_mapped:1;
431 unsigned int op_silent:1;
432 unsigned int op_active:1;
433 struct scatterlist *op_sg;
434 struct rds_notifier *op_notifier;
435
436 struct rds_mr *op_rdma_mr;
437 } atomic;
438 struct rm_rdma_op {
439 u32 op_rkey;
440 u64 op_remote_addr;
441 unsigned int op_write:1;
442 unsigned int op_fence:1;
443 unsigned int op_notify:1;
444 unsigned int op_recverr:1;
445 unsigned int op_mapped:1;
446 unsigned int op_silent:1;
447 unsigned int op_active:1;
448 unsigned int op_bytes;
449 unsigned int op_nents;
450 unsigned int op_count;
451 struct scatterlist *op_sg;
452 struct rds_notifier *op_notifier;
453
454 struct rds_mr *op_rdma_mr;
455 } rdma;
456 struct rm_data_op {
457 unsigned int op_active:1;
458 unsigned int op_notify:1;
459 unsigned int op_nents;
460 unsigned int op_count;
461 unsigned int op_dmasg;
462 unsigned int op_dmaoff;
463 struct rds_znotifier *op_mmp_znotifier;
464 struct scatterlist *op_sg;
465 } data;
466 };
467};
468
469/*
470 * The RDS notifier is used (optionally) to tell the application about
471 * completed RDMA operations. Rather than keeping the whole rds message
472 * around on the queue, we allocate a small notifier that is put on the
473 * socket's notifier_list. Notifications are delivered to the application
474 * through control messages.
475 */
476struct rds_notifier {
477 struct list_head n_list;
478 uint64_t n_user_token;
479 int n_status;
480};
481
482/**
483 * struct rds_transport - transport specific behavioural hooks
484 *
485 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
486 * part of a message. The caller serializes on the send_sem so this
487 * doesn't need to be reentrant for a given conn. The header must be
488 * sent before the data payload. .xmit must be prepared to send a
489 * message with no data payload. .xmit should return the number of
490 * bytes that were sent down the connection, including header bytes.
491 * Returning 0 tells the caller that it doesn't need to perform any
492 * additional work now. This is usually the case when the transport has
493 * filled the sending queue for its connection and will handle
494 * triggering the rds thread to continue the send when space becomes
495 * available. Returning -EAGAIN tells the caller to retry the send
496 * immediately. Returning -ENOMEM tells the caller to retry the send at
497 * some point in the future.
498 *
499 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
500 * it returns the connection can not call rds_recv_incoming().
501 * This will only be called once after conn_connect returns
502 * non-zero success and will The caller serializes this with
503 * the send and connecting paths (xmit_* and conn_*). The
504 * transport is responsible for other serialization, including
505 * rds_recv_incoming(). This is called in process context but
506 * should try hard not to block.
507 */
508
509struct rds_transport {
510 char t_name[TRANSNAMSIZ];
511 struct list_head t_item;
512 struct module *t_owner;
513 unsigned int t_prefer_loopback:1,
514 t_mp_capable:1;
515 unsigned int t_type;
516
517 int (*laddr_check)(struct net *net, __be32 addr);
518 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
519 void (*conn_free)(void *data);
520 int (*conn_path_connect)(struct rds_conn_path *cp);
521 void (*conn_path_shutdown)(struct rds_conn_path *conn);
522 void (*xmit_path_prepare)(struct rds_conn_path *cp);
523 void (*xmit_path_complete)(struct rds_conn_path *cp);
524 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
525 unsigned int hdr_off, unsigned int sg, unsigned int off);
526 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
527 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
528 int (*recv_path)(struct rds_conn_path *cp);
529 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
530 void (*inc_free)(struct rds_incoming *inc);
531
532 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
533 struct rdma_cm_event *event);
534 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
535 void (*cm_connect_complete)(struct rds_connection *conn,
536 struct rdma_cm_event *event);
537
538 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
539 unsigned int avail);
540 void (*exit)(void);
541 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
542 struct rds_sock *rs, u32 *key_ret);
543 void (*sync_mr)(void *trans_private, int direction);
544 void (*free_mr)(void *trans_private, int invalidate);
545 void (*flush_mrs)(void);
546 bool (*t_unloading)(struct rds_connection *conn);
547};
548
549struct rds_sock {
550 struct sock rs_sk;
551
552 u64 rs_user_addr;
553 u64 rs_user_bytes;
554
555 /*
556 * bound_addr used for both incoming and outgoing, no INADDR_ANY
557 * support.
558 */
559 struct rhash_head rs_bound_node;
560 u64 rs_bound_key;
561 __be32 rs_bound_addr;
562 __be32 rs_conn_addr;
563 __be16 rs_bound_port;
564 __be16 rs_conn_port;
565 struct rds_transport *rs_transport;
566
567 /*
568 * rds_sendmsg caches the conn it used the last time around.
569 * This helps avoid costly lookups.
570 */
571 struct rds_connection *rs_conn;
572
573 /* flag indicating we were congested or not */
574 int rs_congested;
575 /* seen congestion (ENOBUFS) when sending? */
576 int rs_seen_congestion;
577
578 /* rs_lock protects all these adjacent members before the newline */
579 spinlock_t rs_lock;
580 struct list_head rs_send_queue;
581 u32 rs_snd_bytes;
582 int rs_rcv_bytes;
583 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
584
585 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
586 * to decide whether the application should be woken up.
587 * If not set, we use rs_cong_track to find out whether a cong map
588 * update arrived.
589 */
590 uint64_t rs_cong_mask;
591 uint64_t rs_cong_notify;
592 struct list_head rs_cong_list;
593 unsigned long rs_cong_track;
594
595 /*
596 * rs_recv_lock protects the receive queue, and is
597 * used to serialize with rds_release.
598 */
599 rwlock_t rs_recv_lock;
600 struct list_head rs_recv_queue;
601
602 /* just for stats reporting */
603 struct list_head rs_item;
604
605 /* these have their own lock */
606 spinlock_t rs_rdma_lock;
607 struct rb_root rs_rdma_keys;
608
609 /* Socket options - in case there will be more */
610 unsigned char rs_recverr,
611 rs_cong_monitor;
612 u32 rs_hash_initval;
613
614 /* Socket receive path trace points*/
615 u8 rs_rx_traces;
616 u8 rs_rx_trace[RDS_MSG_RX_DGRAM_TRACE_MAX];
617 struct rds_msg_zcopy_queue rs_zcookie_queue;
618};
619
620static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
621{
622 return container_of(sk, struct rds_sock, rs_sk);
623}
624static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
625{
626 return &rs->rs_sk;
627}
628
629/*
630 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
631 * to account for overhead. We don't account for overhead, we just apply
632 * the number of payload bytes to the specified value.
633 */
634static inline int rds_sk_sndbuf(struct rds_sock *rs)
635{
636 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
637}
638static inline int rds_sk_rcvbuf(struct rds_sock *rs)
639{
640 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
641}
642
643struct rds_statistics {
644 uint64_t s_conn_reset;
645 uint64_t s_recv_drop_bad_checksum;
646 uint64_t s_recv_drop_old_seq;
647 uint64_t s_recv_drop_no_sock;
648 uint64_t s_recv_drop_dead_sock;
649 uint64_t s_recv_deliver_raced;
650 uint64_t s_recv_delivered;
651 uint64_t s_recv_queued;
652 uint64_t s_recv_immediate_retry;
653 uint64_t s_recv_delayed_retry;
654 uint64_t s_recv_ack_required;
655 uint64_t s_recv_rdma_bytes;
656 uint64_t s_recv_ping;
657 uint64_t s_send_queue_empty;
658 uint64_t s_send_queue_full;
659 uint64_t s_send_lock_contention;
660 uint64_t s_send_lock_queue_raced;
661 uint64_t s_send_immediate_retry;
662 uint64_t s_send_delayed_retry;
663 uint64_t s_send_drop_acked;
664 uint64_t s_send_ack_required;
665 uint64_t s_send_queued;
666 uint64_t s_send_rdma;
667 uint64_t s_send_rdma_bytes;
668 uint64_t s_send_pong;
669 uint64_t s_page_remainder_hit;
670 uint64_t s_page_remainder_miss;
671 uint64_t s_copy_to_user;
672 uint64_t s_copy_from_user;
673 uint64_t s_cong_update_queued;
674 uint64_t s_cong_update_received;
675 uint64_t s_cong_send_error;
676 uint64_t s_cong_send_blocked;
677 uint64_t s_recv_bytes_added_to_socket;
678 uint64_t s_recv_bytes_removed_from_socket;
679
680};
681
682/* af_rds.c */
683void rds_sock_addref(struct rds_sock *rs);
684void rds_sock_put(struct rds_sock *rs);
685void rds_wake_sk_sleep(struct rds_sock *rs);
686static inline void __rds_wake_sk_sleep(struct sock *sk)
687{
688 wait_queue_head_t *waitq = sk_sleep(sk);
689
690 if (!sock_flag(sk, SOCK_DEAD) && waitq)
691 wake_up(waitq);
692}
693extern wait_queue_head_t rds_poll_waitq;
694
695
696/* bind.c */
697int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
698void rds_remove_bound(struct rds_sock *rs);
699struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
700int rds_bind_lock_init(void);
701void rds_bind_lock_destroy(void);
702
703/* cong.c */
704int rds_cong_get_maps(struct rds_connection *conn);
705void rds_cong_add_conn(struct rds_connection *conn);
706void rds_cong_remove_conn(struct rds_connection *conn);
707void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
708void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
709int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
710void rds_cong_queue_updates(struct rds_cong_map *map);
711void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
712int rds_cong_updated_since(unsigned long *recent);
713void rds_cong_add_socket(struct rds_sock *);
714void rds_cong_remove_socket(struct rds_sock *);
715void rds_cong_exit(void);
716struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
717
718/* conn.c */
719extern u32 rds_gen_num;
720int rds_conn_init(void);
721void rds_conn_exit(void);
722struct rds_connection *rds_conn_create(struct net *net,
723 __be32 laddr, __be32 faddr,
724 struct rds_transport *trans, gfp_t gfp);
725struct rds_connection *rds_conn_create_outgoing(struct net *net,
726 __be32 laddr, __be32 faddr,
727 struct rds_transport *trans, gfp_t gfp);
728void rds_conn_shutdown(struct rds_conn_path *cpath);
729void rds_conn_destroy(struct rds_connection *conn);
730void rds_conn_drop(struct rds_connection *conn);
731void rds_conn_path_drop(struct rds_conn_path *cpath, bool destroy);
732void rds_conn_connect_if_down(struct rds_connection *conn);
733void rds_conn_path_connect_if_down(struct rds_conn_path *cp);
734void rds_for_each_conn_info(struct socket *sock, unsigned int len,
735 struct rds_info_iterator *iter,
736 struct rds_info_lengths *lens,
737 int (*visitor)(struct rds_connection *, void *),
738 u64 *buffer,
739 size_t item_len);
740
741__printf(2, 3)
742void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...);
743#define rds_conn_path_error(cp, fmt...) \
744 __rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt)
745
746static inline int
747rds_conn_path_transition(struct rds_conn_path *cp, int old, int new)
748{
749 return atomic_cmpxchg(&cp->cp_state, old, new) == old;
750}
751
752static inline int
753rds_conn_transition(struct rds_connection *conn, int old, int new)
754{
755 WARN_ON(conn->c_trans->t_mp_capable);
756 return rds_conn_path_transition(&conn->c_path[0], old, new);
757}
758
759static inline int
760rds_conn_path_state(struct rds_conn_path *cp)
761{
762 return atomic_read(&cp->cp_state);
763}
764
765static inline int
766rds_conn_state(struct rds_connection *conn)
767{
768 WARN_ON(conn->c_trans->t_mp_capable);
769 return rds_conn_path_state(&conn->c_path[0]);
770}
771
772static inline int
773rds_conn_path_up(struct rds_conn_path *cp)
774{
775 return atomic_read(&cp->cp_state) == RDS_CONN_UP;
776}
777
778static inline int
779rds_conn_up(struct rds_connection *conn)
780{
781 WARN_ON(conn->c_trans->t_mp_capable);
782 return rds_conn_path_up(&conn->c_path[0]);
783}
784
785static inline int
786rds_conn_path_connecting(struct rds_conn_path *cp)
787{
788 return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING;
789}
790
791static inline int
792rds_conn_connecting(struct rds_connection *conn)
793{
794 WARN_ON(conn->c_trans->t_mp_capable);
795 return rds_conn_path_connecting(&conn->c_path[0]);
796}
797
798/* message.c */
799struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
800struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
801int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from,
802 bool zcopy);
803struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
804void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
805 __be16 dport, u64 seq);
806int rds_message_add_extension(struct rds_header *hdr,
807 unsigned int type, const void *data, unsigned int len);
808int rds_message_next_extension(struct rds_header *hdr,
809 unsigned int *pos, void *buf, unsigned int *buflen);
810int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
811int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
812void rds_message_inc_free(struct rds_incoming *inc);
813void rds_message_addref(struct rds_message *rm);
814void rds_message_put(struct rds_message *rm);
815void rds_message_wait(struct rds_message *rm);
816void rds_message_unmapped(struct rds_message *rm);
817void rds_notify_msg_zcopy_purge(struct rds_msg_zcopy_queue *info);
818
819static inline void rds_message_make_checksum(struct rds_header *hdr)
820{
821 hdr->h_csum = 0;
822 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
823}
824
825static inline int rds_message_verify_checksum(const struct rds_header *hdr)
826{
827 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
828}
829
830
831/* page.c */
832int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
833 gfp_t gfp);
834void rds_page_exit(void);
835
836/* recv.c */
837void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
838 __be32 saddr);
839void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn,
840 __be32 saddr);
841void rds_inc_put(struct rds_incoming *inc);
842void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
843 struct rds_incoming *inc, gfp_t gfp);
844int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
845 int msg_flags);
846void rds_clear_recv_queue(struct rds_sock *rs);
847int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
848void rds_inc_info_copy(struct rds_incoming *inc,
849 struct rds_info_iterator *iter,
850 __be32 saddr, __be32 daddr, int flip);
851
852/* send.c */
853int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
854void rds_send_path_reset(struct rds_conn_path *conn);
855int rds_send_xmit(struct rds_conn_path *cp);
856struct sockaddr_in;
857void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
858typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
859void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
860 is_acked_func is_acked);
861void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
862 is_acked_func is_acked);
863void rds_send_ping(struct rds_connection *conn, int cp_index);
864int rds_send_pong(struct rds_conn_path *cp, __be16 dport);
865
866/* rdma.c */
867void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
868int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
869int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
870int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
871void rds_rdma_drop_keys(struct rds_sock *rs);
872int rds_rdma_extra_size(struct rds_rdma_args *args);
873int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
874 struct cmsghdr *cmsg);
875int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
876 struct cmsghdr *cmsg);
877int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
878 struct cmsghdr *cmsg);
879int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
880 struct cmsghdr *cmsg);
881void rds_rdma_free_op(struct rm_rdma_op *ro);
882void rds_atomic_free_op(struct rm_atomic_op *ao);
883void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
884void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
885int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
886 struct cmsghdr *cmsg);
887
888void __rds_put_mr_final(struct rds_mr *mr);
889static inline void rds_mr_put(struct rds_mr *mr)
890{
891 if (refcount_dec_and_test(&mr->r_refcount))
892 __rds_put_mr_final(mr);
893}
894
895static inline bool rds_destroy_pending(struct rds_connection *conn)
896{
897 return !check_net(rds_conn_net(conn)) ||
898 (conn->c_trans->t_unloading && conn->c_trans->t_unloading(conn));
899}
900
901/* stats.c */
902DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
903#define rds_stats_inc_which(which, member) do { \
904 per_cpu(which, get_cpu()).member++; \
905 put_cpu(); \
906} while (0)
907#define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
908#define rds_stats_add_which(which, member, count) do { \
909 per_cpu(which, get_cpu()).member += count; \
910 put_cpu(); \
911} while (0)
912#define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
913int rds_stats_init(void);
914void rds_stats_exit(void);
915void rds_stats_info_copy(struct rds_info_iterator *iter,
916 uint64_t *values, const char *const *names,
917 size_t nr);
918
919/* sysctl.c */
920int rds_sysctl_init(void);
921void rds_sysctl_exit(void);
922extern unsigned long rds_sysctl_sndbuf_min;
923extern unsigned long rds_sysctl_sndbuf_default;
924extern unsigned long rds_sysctl_sndbuf_max;
925extern unsigned long rds_sysctl_reconnect_min_jiffies;
926extern unsigned long rds_sysctl_reconnect_max_jiffies;
927extern unsigned int rds_sysctl_max_unacked_packets;
928extern unsigned int rds_sysctl_max_unacked_bytes;
929extern unsigned int rds_sysctl_ping_enable;
930extern unsigned long rds_sysctl_trace_flags;
931extern unsigned int rds_sysctl_trace_level;
932
933/* threads.c */
934int rds_threads_init(void);
935void rds_threads_exit(void);
936extern struct workqueue_struct *rds_wq;
937void rds_queue_reconnect(struct rds_conn_path *cp);
938void rds_connect_worker(struct work_struct *);
939void rds_shutdown_worker(struct work_struct *);
940void rds_send_worker(struct work_struct *);
941void rds_recv_worker(struct work_struct *);
942void rds_connect_path_complete(struct rds_conn_path *conn, int curr);
943void rds_connect_complete(struct rds_connection *conn);
944
945/* transport.c */
946void rds_trans_register(struct rds_transport *trans);
947void rds_trans_unregister(struct rds_transport *trans);
948struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr);
949void rds_trans_put(struct rds_transport *trans);
950unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
951 unsigned int avail);
952struct rds_transport *rds_trans_get(int t_type);
953int rds_trans_init(void);
954void rds_trans_exit(void);
955
956#endif