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
11#include "info.h"
12
13/*
14 * RDS Network protocol version
15 */
16#define RDS_PROTOCOL_3_0 0x0300
17#define RDS_PROTOCOL_3_1 0x0301
18#define RDS_PROTOCOL_VERSION RDS_PROTOCOL_3_1
19#define RDS_PROTOCOL_MAJOR(v) ((v) >> 8)
20#define RDS_PROTOCOL_MINOR(v) ((v) & 255)
21#define RDS_PROTOCOL(maj, min) (((maj) << 8) | min)
22
23/*
24 * XXX randomly chosen, but at least seems to be unused:
25 * # 18464-18768 Unassigned
26 * We should do better. We want a reserved port to discourage unpriv'ed
27 * userspace from listening.
28 */
29#define RDS_PORT 18634
30
31#ifdef ATOMIC64_INIT
32#define KERNEL_HAS_ATOMIC64
33#endif
34
35#ifdef DEBUG
36#define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
37#else
38/* sigh, pr_debug() causes unused variable warnings */
39static inline void __attribute__ ((format (printf, 1, 2)))
40rdsdebug(char *fmt, ...)
41{
42}
43#endif
44
45/* XXX is there one of these somewhere? */
46#define ceil(x, y) \
47 ({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; })
48
49#define RDS_FRAG_SHIFT 12
50#define RDS_FRAG_SIZE ((unsigned int)(1 << RDS_FRAG_SHIFT))
51
52#define RDS_CONG_MAP_BYTES (65536 / 8)
53#define RDS_CONG_MAP_PAGES (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
54#define RDS_CONG_MAP_PAGE_BITS (PAGE_SIZE * 8)
55
56struct rds_cong_map {
57 struct rb_node m_rb_node;
58 __be32 m_addr;
59 wait_queue_head_t m_waitq;
60 struct list_head m_conn_list;
61 unsigned long m_page_addrs[RDS_CONG_MAP_PAGES];
62};
63
64
65/*
66 * This is how we will track the connection state:
67 * A connection is always in one of the following
68 * states. Updates to the state are atomic and imply
69 * a memory barrier.
70 */
71enum {
72 RDS_CONN_DOWN = 0,
73 RDS_CONN_CONNECTING,
74 RDS_CONN_DISCONNECTING,
75 RDS_CONN_UP,
76 RDS_CONN_ERROR,
77};
78
79/* Bits for c_flags */
80#define RDS_LL_SEND_FULL 0
81#define RDS_RECONNECT_PENDING 1
82#define RDS_IN_XMIT 2
83
84struct rds_connection {
85 struct hlist_node c_hash_node;
86 __be32 c_laddr;
87 __be32 c_faddr;
88 unsigned int c_loopback:1;
89 struct rds_connection *c_passive;
90
91 struct rds_cong_map *c_lcong;
92 struct rds_cong_map *c_fcong;
93
94 struct rds_message *c_xmit_rm;
95 unsigned long c_xmit_sg;
96 unsigned int c_xmit_hdr_off;
97 unsigned int c_xmit_data_off;
98 unsigned int c_xmit_atomic_sent;
99 unsigned int c_xmit_rdma_sent;
100 unsigned int c_xmit_data_sent;
101
102 spinlock_t c_lock; /* protect msg queues */
103 u64 c_next_tx_seq;
104 struct list_head c_send_queue;
105 struct list_head c_retrans;
106
107 u64 c_next_rx_seq;
108
109 struct rds_transport *c_trans;
110 void *c_transport_data;
111
112 atomic_t c_state;
113 unsigned long c_flags;
114 unsigned long c_reconnect_jiffies;
115 struct delayed_work c_send_w;
116 struct delayed_work c_recv_w;
117 struct delayed_work c_conn_w;
118 struct work_struct c_down_w;
119 struct mutex c_cm_lock; /* protect conn state & cm */
120 wait_queue_head_t c_waitq;
121
122 struct list_head c_map_item;
123 unsigned long c_map_queued;
124
125 unsigned int c_unacked_packets;
126 unsigned int c_unacked_bytes;
127
128 /* Protocol version */
129 unsigned int c_version;
130};
131
132#define RDS_FLAG_CONG_BITMAP 0x01
133#define RDS_FLAG_ACK_REQUIRED 0x02
134#define RDS_FLAG_RETRANSMITTED 0x04
135#define RDS_MAX_ADV_CREDIT 255
136
137/*
138 * Maximum space available for extension headers.
139 */
140#define RDS_HEADER_EXT_SPACE 16
141
142struct rds_header {
143 __be64 h_sequence;
144 __be64 h_ack;
145 __be32 h_len;
146 __be16 h_sport;
147 __be16 h_dport;
148 u8 h_flags;
149 u8 h_credit;
150 u8 h_padding[4];
151 __sum16 h_csum;
152
153 u8 h_exthdr[RDS_HEADER_EXT_SPACE];
154};
155
156/*
157 * Reserved - indicates end of extensions
158 */
159#define RDS_EXTHDR_NONE 0
160
161/*
162 * This extension header is included in the very
163 * first message that is sent on a new connection,
164 * and identifies the protocol level. This will help
165 * rolling updates if a future change requires breaking
166 * the protocol.
167 * NB: This is no longer true for IB, where we do a version
168 * negotiation during the connection setup phase (protocol
169 * version information is included in the RDMA CM private data).
170 */
171#define RDS_EXTHDR_VERSION 1
172struct rds_ext_header_version {
173 __be32 h_version;
174};
175
176/*
177 * This extension header is included in the RDS message
178 * chasing an RDMA operation.
179 */
180#define RDS_EXTHDR_RDMA 2
181struct rds_ext_header_rdma {
182 __be32 h_rdma_rkey;
183};
184
185/*
186 * This extension header tells the peer about the
187 * destination <R_Key,offset> of the requested RDMA
188 * operation.
189 */
190#define RDS_EXTHDR_RDMA_DEST 3
191struct rds_ext_header_rdma_dest {
192 __be32 h_rdma_rkey;
193 __be32 h_rdma_offset;
194};
195
196#define __RDS_EXTHDR_MAX 16 /* for now */
197
198struct rds_incoming {
199 atomic_t i_refcount;
200 struct list_head i_item;
201 struct rds_connection *i_conn;
202 struct rds_header i_hdr;
203 unsigned long i_rx_jiffies;
204 __be32 i_saddr;
205
206 rds_rdma_cookie_t i_rdma_cookie;
207};
208
209struct rds_mr {
210 struct rb_node r_rb_node;
211 atomic_t r_refcount;
212 u32 r_key;
213
214 /* A copy of the creation flags */
215 unsigned int r_use_once:1;
216 unsigned int r_invalidate:1;
217 unsigned int r_write:1;
218
219 /* This is for RDS_MR_DEAD.
220 * It would be nice & consistent to make this part of the above
221 * bit field here, but we need to use test_and_set_bit.
222 */
223 unsigned long r_state;
224 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
225 struct rds_transport *r_trans;
226 void *r_trans_private;
227};
228
229/* Flags for mr->r_state */
230#define RDS_MR_DEAD 0
231
232static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
233{
234 return r_key | (((u64) offset) << 32);
235}
236
237static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
238{
239 return cookie;
240}
241
242static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
243{
244 return cookie >> 32;
245}
246
247/* atomic operation types */
248#define RDS_ATOMIC_TYPE_CSWP 0
249#define RDS_ATOMIC_TYPE_FADD 1
250
251/*
252 * m_sock_item and m_conn_item are on lists that are serialized under
253 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
254 * the message will not be put back on the retransmit list after being sent.
255 * messages that are canceled while being sent rely on this.
256 *
257 * m_inc is used by loopback so that it can pass an incoming message straight
258 * back up into the rx path. It embeds a wire header which is also used by
259 * the send path, which is kind of awkward.
260 *
261 * m_sock_item indicates the message's presence on a socket's send or receive
262 * queue. m_rs will point to that socket.
263 *
264 * m_daddr is used by cancellation to prune messages to a given destination.
265 *
266 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
267 * nesting. As paths iterate over messages on a sock, or conn, they must
268 * also lock the conn, or sock, to remove the message from those lists too.
269 * Testing the flag to determine if the message is still on the lists lets
270 * us avoid testing the list_head directly. That means each path can use
271 * the message's list_head to keep it on a local list while juggling locks
272 * without confusing the other path.
273 *
274 * m_ack_seq is an optional field set by transports who need a different
275 * sequence number range to invalidate. They can use this in a callback
276 * that they pass to rds_send_drop_acked() to see if each message has been
277 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
278 * had ack_seq set yet.
279 */
280#define RDS_MSG_ON_SOCK 1
281#define RDS_MSG_ON_CONN 2
282#define RDS_MSG_HAS_ACK_SEQ 3
283#define RDS_MSG_ACK_REQUIRED 4
284#define RDS_MSG_RETRANSMITTED 5
285#define RDS_MSG_MAPPED 6
286#define RDS_MSG_PAGEVEC 7
287
288struct rds_message {
289 atomic_t m_refcount;
290 struct list_head m_sock_item;
291 struct list_head m_conn_item;
292 struct rds_incoming m_inc;
293 u64 m_ack_seq;
294 __be32 m_daddr;
295 unsigned long m_flags;
296
297 /* Never access m_rs without holding m_rs_lock.
298 * Lock nesting is
299 * rm->m_rs_lock
300 * -> rs->rs_lock
301 */
302 spinlock_t m_rs_lock;
303 wait_queue_head_t m_flush_wait;
304
305 struct rds_sock *m_rs;
306
307 /* cookie to send to remote, in rds header */
308 rds_rdma_cookie_t m_rdma_cookie;
309
310 unsigned int m_used_sgs;
311 unsigned int m_total_sgs;
312
313 void *m_final_op;
314
315 struct {
316 struct rm_atomic_op {
317 int op_type;
318 union {
319 struct {
320 uint64_t compare;
321 uint64_t swap;
322 uint64_t compare_mask;
323 uint64_t swap_mask;
324 } op_m_cswp;
325 struct {
326 uint64_t add;
327 uint64_t nocarry_mask;
328 } op_m_fadd;
329 };
330
331 u32 op_rkey;
332 u64 op_remote_addr;
333 unsigned int op_notify:1;
334 unsigned int op_recverr:1;
335 unsigned int op_mapped:1;
336 unsigned int op_silent:1;
337 unsigned int op_active:1;
338 struct scatterlist *op_sg;
339 struct rds_notifier *op_notifier;
340
341 struct rds_mr *op_rdma_mr;
342 } atomic;
343 struct rm_rdma_op {
344 u32 op_rkey;
345 u64 op_remote_addr;
346 unsigned int op_write:1;
347 unsigned int op_fence:1;
348 unsigned int op_notify:1;
349 unsigned int op_recverr:1;
350 unsigned int op_mapped:1;
351 unsigned int op_silent:1;
352 unsigned int op_active:1;
353 unsigned int op_bytes;
354 unsigned int op_nents;
355 unsigned int op_count;
356 struct scatterlist *op_sg;
357 struct rds_notifier *op_notifier;
358
359 struct rds_mr *op_rdma_mr;
360 } rdma;
361 struct rm_data_op {
362 unsigned int op_active:1;
363 unsigned int op_nents;
364 unsigned int op_count;
365 struct scatterlist *op_sg;
366 } data;
367 };
368};
369
370/*
371 * The RDS notifier is used (optionally) to tell the application about
372 * completed RDMA operations. Rather than keeping the whole rds message
373 * around on the queue, we allocate a small notifier that is put on the
374 * socket's notifier_list. Notifications are delivered to the application
375 * through control messages.
376 */
377struct rds_notifier {
378 struct list_head n_list;
379 uint64_t n_user_token;
380 int n_status;
381};
382
383/**
384 * struct rds_transport - transport specific behavioural hooks
385 *
386 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
387 * part of a message. The caller serializes on the send_sem so this
388 * doesn't need to be reentrant for a given conn. The header must be
389 * sent before the data payload. .xmit must be prepared to send a
390 * message with no data payload. .xmit should return the number of
391 * bytes that were sent down the connection, including header bytes.
392 * Returning 0 tells the caller that it doesn't need to perform any
393 * additional work now. This is usually the case when the transport has
394 * filled the sending queue for its connection and will handle
395 * triggering the rds thread to continue the send when space becomes
396 * available. Returning -EAGAIN tells the caller to retry the send
397 * immediately. Returning -ENOMEM tells the caller to retry the send at
398 * some point in the future.
399 *
400 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
401 * it returns the connection can not call rds_recv_incoming().
402 * This will only be called once after conn_connect returns
403 * non-zero success and will The caller serializes this with
404 * the send and connecting paths (xmit_* and conn_*). The
405 * transport is responsible for other serialization, including
406 * rds_recv_incoming(). This is called in process context but
407 * should try hard not to block.
408 */
409
410#define RDS_TRANS_IB 0
411#define RDS_TRANS_IWARP 1
412#define RDS_TRANS_TCP 2
413#define RDS_TRANS_COUNT 3
414
415struct rds_transport {
416 char t_name[TRANSNAMSIZ];
417 struct list_head t_item;
418 struct module *t_owner;
419 unsigned int t_prefer_loopback:1;
420 unsigned int t_type;
421
422 int (*laddr_check)(__be32 addr);
423 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
424 void (*conn_free)(void *data);
425 int (*conn_connect)(struct rds_connection *conn);
426 void (*conn_shutdown)(struct rds_connection *conn);
427 void (*xmit_prepare)(struct rds_connection *conn);
428 void (*xmit_complete)(struct rds_connection *conn);
429 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
430 unsigned int hdr_off, unsigned int sg, unsigned int off);
431 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
432 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
433 int (*recv)(struct rds_connection *conn);
434 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iovec *iov,
435 size_t size);
436 void (*inc_free)(struct rds_incoming *inc);
437
438 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
439 struct rdma_cm_event *event);
440 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
441 void (*cm_connect_complete)(struct rds_connection *conn,
442 struct rdma_cm_event *event);
443
444 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
445 unsigned int avail);
446 void (*exit)(void);
447 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
448 struct rds_sock *rs, u32 *key_ret);
449 void (*sync_mr)(void *trans_private, int direction);
450 void (*free_mr)(void *trans_private, int invalidate);
451 void (*flush_mrs)(void);
452};
453
454struct rds_sock {
455 struct sock rs_sk;
456
457 u64 rs_user_addr;
458 u64 rs_user_bytes;
459
460 /*
461 * bound_addr used for both incoming and outgoing, no INADDR_ANY
462 * support.
463 */
464 struct hlist_node rs_bound_node;
465 __be32 rs_bound_addr;
466 __be32 rs_conn_addr;
467 __be16 rs_bound_port;
468 __be16 rs_conn_port;
469 struct rds_transport *rs_transport;
470
471 /*
472 * rds_sendmsg caches the conn it used the last time around.
473 * This helps avoid costly lookups.
474 */
475 struct rds_connection *rs_conn;
476
477 /* flag indicating we were congested or not */
478 int rs_congested;
479 /* seen congestion (ENOBUFS) when sending? */
480 int rs_seen_congestion;
481
482 /* rs_lock protects all these adjacent members before the newline */
483 spinlock_t rs_lock;
484 struct list_head rs_send_queue;
485 u32 rs_snd_bytes;
486 int rs_rcv_bytes;
487 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
488
489 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
490 * to decide whether the application should be woken up.
491 * If not set, we use rs_cong_track to find out whether a cong map
492 * update arrived.
493 */
494 uint64_t rs_cong_mask;
495 uint64_t rs_cong_notify;
496 struct list_head rs_cong_list;
497 unsigned long rs_cong_track;
498
499 /*
500 * rs_recv_lock protects the receive queue, and is
501 * used to serialize with rds_release.
502 */
503 rwlock_t rs_recv_lock;
504 struct list_head rs_recv_queue;
505
506 /* just for stats reporting */
507 struct list_head rs_item;
508
509 /* these have their own lock */
510 spinlock_t rs_rdma_lock;
511 struct rb_root rs_rdma_keys;
512
513 /* Socket options - in case there will be more */
514 unsigned char rs_recverr,
515 rs_cong_monitor;
516};
517
518static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
519{
520 return container_of(sk, struct rds_sock, rs_sk);
521}
522static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
523{
524 return &rs->rs_sk;
525}
526
527/*
528 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
529 * to account for overhead. We don't account for overhead, we just apply
530 * the number of payload bytes to the specified value.
531 */
532static inline int rds_sk_sndbuf(struct rds_sock *rs)
533{
534 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
535}
536static inline int rds_sk_rcvbuf(struct rds_sock *rs)
537{
538 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
539}
540
541struct rds_statistics {
542 uint64_t s_conn_reset;
543 uint64_t s_recv_drop_bad_checksum;
544 uint64_t s_recv_drop_old_seq;
545 uint64_t s_recv_drop_no_sock;
546 uint64_t s_recv_drop_dead_sock;
547 uint64_t s_recv_deliver_raced;
548 uint64_t s_recv_delivered;
549 uint64_t s_recv_queued;
550 uint64_t s_recv_immediate_retry;
551 uint64_t s_recv_delayed_retry;
552 uint64_t s_recv_ack_required;
553 uint64_t s_recv_rdma_bytes;
554 uint64_t s_recv_ping;
555 uint64_t s_send_queue_empty;
556 uint64_t s_send_queue_full;
557 uint64_t s_send_lock_contention;
558 uint64_t s_send_lock_queue_raced;
559 uint64_t s_send_immediate_retry;
560 uint64_t s_send_delayed_retry;
561 uint64_t s_send_drop_acked;
562 uint64_t s_send_ack_required;
563 uint64_t s_send_queued;
564 uint64_t s_send_rdma;
565 uint64_t s_send_rdma_bytes;
566 uint64_t s_send_pong;
567 uint64_t s_page_remainder_hit;
568 uint64_t s_page_remainder_miss;
569 uint64_t s_copy_to_user;
570 uint64_t s_copy_from_user;
571 uint64_t s_cong_update_queued;
572 uint64_t s_cong_update_received;
573 uint64_t s_cong_send_error;
574 uint64_t s_cong_send_blocked;
575};
576
577/* af_rds.c */
578char *rds_str_array(char **array, size_t elements, size_t index);
579void rds_sock_addref(struct rds_sock *rs);
580void rds_sock_put(struct rds_sock *rs);
581void rds_wake_sk_sleep(struct rds_sock *rs);
582static inline void __rds_wake_sk_sleep(struct sock *sk)
583{
584 wait_queue_head_t *waitq = sk_sleep(sk);
585
586 if (!sock_flag(sk, SOCK_DEAD) && waitq)
587 wake_up(waitq);
588}
589extern wait_queue_head_t rds_poll_waitq;
590
591
592/* bind.c */
593int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
594void rds_remove_bound(struct rds_sock *rs);
595struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
596
597/* cong.c */
598int rds_cong_get_maps(struct rds_connection *conn);
599void rds_cong_add_conn(struct rds_connection *conn);
600void rds_cong_remove_conn(struct rds_connection *conn);
601void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
602void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
603int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
604void rds_cong_queue_updates(struct rds_cong_map *map);
605void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
606int rds_cong_updated_since(unsigned long *recent);
607void rds_cong_add_socket(struct rds_sock *);
608void rds_cong_remove_socket(struct rds_sock *);
609void rds_cong_exit(void);
610struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
611
612/* conn.c */
613int rds_conn_init(void);
614void rds_conn_exit(void);
615struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
616 struct rds_transport *trans, gfp_t gfp);
617struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
618 struct rds_transport *trans, gfp_t gfp);
619void rds_conn_shutdown(struct rds_connection *conn);
620void rds_conn_destroy(struct rds_connection *conn);
621void rds_conn_drop(struct rds_connection *conn);
622void rds_conn_connect_if_down(struct rds_connection *conn);
623void rds_for_each_conn_info(struct socket *sock, unsigned int len,
624 struct rds_info_iterator *iter,
625 struct rds_info_lengths *lens,
626 int (*visitor)(struct rds_connection *, void *),
627 size_t item_len);
628void __rds_conn_error(struct rds_connection *conn, const char *, ...)
629 __attribute__ ((format (printf, 2, 3)));
630#define rds_conn_error(conn, fmt...) \
631 __rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
632
633static inline int
634rds_conn_transition(struct rds_connection *conn, int old, int new)
635{
636 return atomic_cmpxchg(&conn->c_state, old, new) == old;
637}
638
639static inline int
640rds_conn_state(struct rds_connection *conn)
641{
642 return atomic_read(&conn->c_state);
643}
644
645static inline int
646rds_conn_up(struct rds_connection *conn)
647{
648 return atomic_read(&conn->c_state) == RDS_CONN_UP;
649}
650
651static inline int
652rds_conn_connecting(struct rds_connection *conn)
653{
654 return atomic_read(&conn->c_state) == RDS_CONN_CONNECTING;
655}
656
657/* message.c */
658struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
659struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
660int rds_message_copy_from_user(struct rds_message *rm, struct iovec *first_iov,
661 size_t total_len);
662struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
663void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
664 __be16 dport, u64 seq);
665int rds_message_add_extension(struct rds_header *hdr,
666 unsigned int type, const void *data, unsigned int len);
667int rds_message_next_extension(struct rds_header *hdr,
668 unsigned int *pos, void *buf, unsigned int *buflen);
669int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
670int rds_message_inc_copy_to_user(struct rds_incoming *inc,
671 struct iovec *first_iov, size_t size);
672void rds_message_inc_free(struct rds_incoming *inc);
673void rds_message_addref(struct rds_message *rm);
674void rds_message_put(struct rds_message *rm);
675void rds_message_wait(struct rds_message *rm);
676void rds_message_unmapped(struct rds_message *rm);
677
678static inline void rds_message_make_checksum(struct rds_header *hdr)
679{
680 hdr->h_csum = 0;
681 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
682}
683
684static inline int rds_message_verify_checksum(const struct rds_header *hdr)
685{
686 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
687}
688
689
690/* page.c */
691int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
692 gfp_t gfp);
693int rds_page_copy_user(struct page *page, unsigned long offset,
694 void __user *ptr, unsigned long bytes,
695 int to_user);
696#define rds_page_copy_to_user(page, offset, ptr, bytes) \
697 rds_page_copy_user(page, offset, ptr, bytes, 1)
698#define rds_page_copy_from_user(page, offset, ptr, bytes) \
699 rds_page_copy_user(page, offset, ptr, bytes, 0)
700void rds_page_exit(void);
701
702/* recv.c */
703void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
704 __be32 saddr);
705void rds_inc_put(struct rds_incoming *inc);
706void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
707 struct rds_incoming *inc, gfp_t gfp, enum km_type km);
708int rds_recvmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
709 size_t size, int msg_flags);
710void rds_clear_recv_queue(struct rds_sock *rs);
711int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
712void rds_inc_info_copy(struct rds_incoming *inc,
713 struct rds_info_iterator *iter,
714 __be32 saddr, __be32 daddr, int flip);
715
716/* send.c */
717int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
718 size_t payload_len);
719void rds_send_reset(struct rds_connection *conn);
720int rds_send_xmit(struct rds_connection *conn);
721struct sockaddr_in;
722void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
723typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
724void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
725 is_acked_func is_acked);
726int rds_send_pong(struct rds_connection *conn, __be16 dport);
727struct rds_message *rds_send_get_message(struct rds_connection *,
728 struct rm_rdma_op *);
729
730/* rdma.c */
731void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
732int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
733int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
734int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
735void rds_rdma_drop_keys(struct rds_sock *rs);
736int rds_rdma_extra_size(struct rds_rdma_args *args);
737int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
738 struct cmsghdr *cmsg);
739int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
740 struct cmsghdr *cmsg);
741int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
742 struct cmsghdr *cmsg);
743int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
744 struct cmsghdr *cmsg);
745void rds_rdma_free_op(struct rm_rdma_op *ro);
746void rds_atomic_free_op(struct rm_atomic_op *ao);
747void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
748void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
749int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
750 struct cmsghdr *cmsg);
751
752extern void __rds_put_mr_final(struct rds_mr *mr);
753static inline void rds_mr_put(struct rds_mr *mr)
754{
755 if (atomic_dec_and_test(&mr->r_refcount))
756 __rds_put_mr_final(mr);
757}
758
759/* stats.c */
760DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
761#define rds_stats_inc_which(which, member) do { \
762 per_cpu(which, get_cpu()).member++; \
763 put_cpu(); \
764} while (0)
765#define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
766#define rds_stats_add_which(which, member, count) do { \
767 per_cpu(which, get_cpu()).member += count; \
768 put_cpu(); \
769} while (0)
770#define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
771int rds_stats_init(void);
772void rds_stats_exit(void);
773void rds_stats_info_copy(struct rds_info_iterator *iter,
774 uint64_t *values, const char *const *names,
775 size_t nr);
776
777/* sysctl.c */
778int rds_sysctl_init(void);
779void rds_sysctl_exit(void);
780extern unsigned long rds_sysctl_sndbuf_min;
781extern unsigned long rds_sysctl_sndbuf_default;
782extern unsigned long rds_sysctl_sndbuf_max;
783extern unsigned long rds_sysctl_reconnect_min_jiffies;
784extern unsigned long rds_sysctl_reconnect_max_jiffies;
785extern unsigned int rds_sysctl_max_unacked_packets;
786extern unsigned int rds_sysctl_max_unacked_bytes;
787extern unsigned int rds_sysctl_ping_enable;
788extern unsigned long rds_sysctl_trace_flags;
789extern unsigned int rds_sysctl_trace_level;
790
791/* threads.c */
792int rds_threads_init(void);
793void rds_threads_exit(void);
794extern struct workqueue_struct *rds_wq;
795void rds_queue_reconnect(struct rds_connection *conn);
796void rds_connect_worker(struct work_struct *);
797void rds_shutdown_worker(struct work_struct *);
798void rds_send_worker(struct work_struct *);
799void rds_recv_worker(struct work_struct *);
800void rds_connect_complete(struct rds_connection *conn);
801
802/* transport.c */
803int rds_trans_register(struct rds_transport *trans);
804void rds_trans_unregister(struct rds_transport *trans);
805struct rds_transport *rds_trans_get_preferred(__be32 addr);
806void rds_trans_put(struct rds_transport *trans);
807unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
808 unsigned int avail);
809int rds_trans_init(void);
810void rds_trans_exit(void);
811
812#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#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 */
48static inline __printf(1, 2)
49void 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
64struct 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 */
79enum {
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 */
103struct 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 */
139struct 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
174static inline
175struct net *rds_conn_net(struct rds_connection *conn)
176{
177 return read_pnet(&conn->c_net);
178}
179
180static inline
181void 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
207struct 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
237struct 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
246struct 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
256struct 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/* The following values are whitelisted for usercopy */
275struct rds_inc_usercopy {
276 rds_rdma_cookie_t rdma_cookie;
277 ktime_t rx_tstamp;
278};
279
280struct rds_incoming {
281 refcount_t i_refcount;
282 struct list_head i_item;
283 struct rds_connection *i_conn;
284 struct rds_conn_path *i_conn_path;
285 struct rds_header i_hdr;
286 unsigned long i_rx_jiffies;
287 struct in6_addr i_saddr;
288
289 struct rds_inc_usercopy i_usercopy;
290 u64 i_rx_lat_trace[RDS_RX_MAX_TRACES];
291};
292
293struct rds_mr {
294 struct rb_node r_rb_node;
295 refcount_t r_refcount;
296 u32 r_key;
297
298 /* A copy of the creation flags */
299 unsigned int r_use_once:1;
300 unsigned int r_invalidate:1;
301 unsigned int r_write:1;
302
303 /* This is for RDS_MR_DEAD.
304 * It would be nice & consistent to make this part of the above
305 * bit field here, but we need to use test_and_set_bit.
306 */
307 unsigned long r_state;
308 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
309 struct rds_transport *r_trans;
310 void *r_trans_private;
311};
312
313/* Flags for mr->r_state */
314#define RDS_MR_DEAD 0
315
316static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
317{
318 return r_key | (((u64) offset) << 32);
319}
320
321static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
322{
323 return cookie;
324}
325
326static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
327{
328 return cookie >> 32;
329}
330
331/* atomic operation types */
332#define RDS_ATOMIC_TYPE_CSWP 0
333#define RDS_ATOMIC_TYPE_FADD 1
334
335/*
336 * m_sock_item and m_conn_item are on lists that are serialized under
337 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
338 * the message will not be put back on the retransmit list after being sent.
339 * messages that are canceled while being sent rely on this.
340 *
341 * m_inc is used by loopback so that it can pass an incoming message straight
342 * back up into the rx path. It embeds a wire header which is also used by
343 * the send path, which is kind of awkward.
344 *
345 * m_sock_item indicates the message's presence on a socket's send or receive
346 * queue. m_rs will point to that socket.
347 *
348 * m_daddr is used by cancellation to prune messages to a given destination.
349 *
350 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
351 * nesting. As paths iterate over messages on a sock, or conn, they must
352 * also lock the conn, or sock, to remove the message from those lists too.
353 * Testing the flag to determine if the message is still on the lists lets
354 * us avoid testing the list_head directly. That means each path can use
355 * the message's list_head to keep it on a local list while juggling locks
356 * without confusing the other path.
357 *
358 * m_ack_seq is an optional field set by transports who need a different
359 * sequence number range to invalidate. They can use this in a callback
360 * that they pass to rds_send_drop_acked() to see if each message has been
361 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
362 * had ack_seq set yet.
363 */
364#define RDS_MSG_ON_SOCK 1
365#define RDS_MSG_ON_CONN 2
366#define RDS_MSG_HAS_ACK_SEQ 3
367#define RDS_MSG_ACK_REQUIRED 4
368#define RDS_MSG_RETRANSMITTED 5
369#define RDS_MSG_MAPPED 6
370#define RDS_MSG_PAGEVEC 7
371#define RDS_MSG_FLUSH 8
372
373struct rds_znotifier {
374 struct mmpin z_mmp;
375 u32 z_cookie;
376};
377
378struct rds_msg_zcopy_info {
379 struct list_head rs_zcookie_next;
380 union {
381 struct rds_znotifier znotif;
382 struct rds_zcopy_cookies zcookies;
383 };
384};
385
386struct rds_msg_zcopy_queue {
387 struct list_head zcookie_head;
388 spinlock_t lock; /* protects zcookie_head queue */
389};
390
391static inline void rds_message_zcopy_queue_init(struct rds_msg_zcopy_queue *q)
392{
393 spin_lock_init(&q->lock);
394 INIT_LIST_HEAD(&q->zcookie_head);
395}
396
397struct rds_iov_vector {
398 struct rds_iovec *iov;
399 int len;
400};
401
402struct rds_iov_vector_arr {
403 struct rds_iov_vector *vec;
404 int len;
405 int indx;
406 int incr;
407};
408
409struct rds_message {
410 refcount_t m_refcount;
411 struct list_head m_sock_item;
412 struct list_head m_conn_item;
413 struct rds_incoming m_inc;
414 u64 m_ack_seq;
415 struct in6_addr m_daddr;
416 unsigned long m_flags;
417
418 /* Never access m_rs without holding m_rs_lock.
419 * Lock nesting is
420 * rm->m_rs_lock
421 * -> rs->rs_lock
422 */
423 spinlock_t m_rs_lock;
424 wait_queue_head_t m_flush_wait;
425
426 struct rds_sock *m_rs;
427
428 /* cookie to send to remote, in rds header */
429 rds_rdma_cookie_t m_rdma_cookie;
430
431 unsigned int m_used_sgs;
432 unsigned int m_total_sgs;
433
434 void *m_final_op;
435
436 struct {
437 struct rm_atomic_op {
438 int op_type;
439 union {
440 struct {
441 uint64_t compare;
442 uint64_t swap;
443 uint64_t compare_mask;
444 uint64_t swap_mask;
445 } op_m_cswp;
446 struct {
447 uint64_t add;
448 uint64_t nocarry_mask;
449 } op_m_fadd;
450 };
451
452 u32 op_rkey;
453 u64 op_remote_addr;
454 unsigned int op_notify:1;
455 unsigned int op_recverr:1;
456 unsigned int op_mapped:1;
457 unsigned int op_silent:1;
458 unsigned int op_active:1;
459 struct scatterlist *op_sg;
460 struct rds_notifier *op_notifier;
461
462 struct rds_mr *op_rdma_mr;
463 } atomic;
464 struct rm_rdma_op {
465 u32 op_rkey;
466 u64 op_remote_addr;
467 unsigned int op_write:1;
468 unsigned int op_fence:1;
469 unsigned int op_notify:1;
470 unsigned int op_recverr:1;
471 unsigned int op_mapped:1;
472 unsigned int op_silent:1;
473 unsigned int op_active:1;
474 unsigned int op_bytes;
475 unsigned int op_nents;
476 unsigned int op_count;
477 struct scatterlist *op_sg;
478 struct rds_notifier *op_notifier;
479
480 struct rds_mr *op_rdma_mr;
481 } rdma;
482 struct rm_data_op {
483 unsigned int op_active:1;
484 unsigned int op_nents;
485 unsigned int op_count;
486 unsigned int op_dmasg;
487 unsigned int op_dmaoff;
488 struct rds_znotifier *op_mmp_znotifier;
489 struct scatterlist *op_sg;
490 } data;
491 };
492
493 struct rds_conn_path *m_conn_path;
494};
495
496/*
497 * The RDS notifier is used (optionally) to tell the application about
498 * completed RDMA operations. Rather than keeping the whole rds message
499 * around on the queue, we allocate a small notifier that is put on the
500 * socket's notifier_list. Notifications are delivered to the application
501 * through control messages.
502 */
503struct rds_notifier {
504 struct list_head n_list;
505 uint64_t n_user_token;
506 int n_status;
507};
508
509/* Available as part of RDS core, so doesn't need to participate
510 * in get_preferred transport etc
511 */
512#define RDS_TRANS_LOOP 3
513
514/**
515 * struct rds_transport - transport specific behavioural hooks
516 *
517 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
518 * part of a message. The caller serializes on the send_sem so this
519 * doesn't need to be reentrant for a given conn. The header must be
520 * sent before the data payload. .xmit must be prepared to send a
521 * message with no data payload. .xmit should return the number of
522 * bytes that were sent down the connection, including header bytes.
523 * Returning 0 tells the caller that it doesn't need to perform any
524 * additional work now. This is usually the case when the transport has
525 * filled the sending queue for its connection and will handle
526 * triggering the rds thread to continue the send when space becomes
527 * available. Returning -EAGAIN tells the caller to retry the send
528 * immediately. Returning -ENOMEM tells the caller to retry the send at
529 * some point in the future.
530 *
531 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
532 * it returns the connection can not call rds_recv_incoming().
533 * This will only be called once after conn_connect returns
534 * non-zero success and will The caller serializes this with
535 * the send and connecting paths (xmit_* and conn_*). The
536 * transport is responsible for other serialization, including
537 * rds_recv_incoming(). This is called in process context but
538 * should try hard not to block.
539 */
540
541struct rds_transport {
542 char t_name[TRANSNAMSIZ];
543 struct list_head t_item;
544 struct module *t_owner;
545 unsigned int t_prefer_loopback:1,
546 t_mp_capable:1;
547 unsigned int t_type;
548
549 int (*laddr_check)(struct net *net, const struct in6_addr *addr,
550 __u32 scope_id);
551 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
552 void (*conn_free)(void *data);
553 int (*conn_path_connect)(struct rds_conn_path *cp);
554 void (*conn_path_shutdown)(struct rds_conn_path *conn);
555 void (*xmit_path_prepare)(struct rds_conn_path *cp);
556 void (*xmit_path_complete)(struct rds_conn_path *cp);
557 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
558 unsigned int hdr_off, unsigned int sg, unsigned int off);
559 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
560 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
561 int (*recv_path)(struct rds_conn_path *cp);
562 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
563 void (*inc_free)(struct rds_incoming *inc);
564
565 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
566 struct rdma_cm_event *event, bool isv6);
567 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id, bool isv6);
568 void (*cm_connect_complete)(struct rds_connection *conn,
569 struct rdma_cm_event *event);
570
571 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
572 unsigned int avail);
573 void (*exit)(void);
574 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
575 struct rds_sock *rs, u32 *key_ret,
576 struct rds_connection *conn);
577 void (*sync_mr)(void *trans_private, int direction);
578 void (*free_mr)(void *trans_private, int invalidate);
579 void (*flush_mrs)(void);
580 bool (*t_unloading)(struct rds_connection *conn);
581 u8 (*get_tos_map)(u8 tos);
582};
583
584/* Bind hash table key length. It is the sum of the size of a struct
585 * in6_addr, a scope_id and a port.
586 */
587#define RDS_BOUND_KEY_LEN \
588 (sizeof(struct in6_addr) + sizeof(__u32) + sizeof(__be16))
589
590struct rds_sock {
591 struct sock rs_sk;
592
593 u64 rs_user_addr;
594 u64 rs_user_bytes;
595
596 /*
597 * bound_addr used for both incoming and outgoing, no INADDR_ANY
598 * support.
599 */
600 struct rhash_head rs_bound_node;
601 u8 rs_bound_key[RDS_BOUND_KEY_LEN];
602 struct sockaddr_in6 rs_bound_sin6;
603#define rs_bound_addr rs_bound_sin6.sin6_addr
604#define rs_bound_addr_v4 rs_bound_sin6.sin6_addr.s6_addr32[3]
605#define rs_bound_port rs_bound_sin6.sin6_port
606#define rs_bound_scope_id rs_bound_sin6.sin6_scope_id
607 struct in6_addr rs_conn_addr;
608#define rs_conn_addr_v4 rs_conn_addr.s6_addr32[3]
609 __be16 rs_conn_port;
610 struct rds_transport *rs_transport;
611
612 /*
613 * rds_sendmsg caches the conn it used the last time around.
614 * This helps avoid costly lookups.
615 */
616 struct rds_connection *rs_conn;
617
618 /* flag indicating we were congested or not */
619 int rs_congested;
620 /* seen congestion (ENOBUFS) when sending? */
621 int rs_seen_congestion;
622
623 /* rs_lock protects all these adjacent members before the newline */
624 spinlock_t rs_lock;
625 struct list_head rs_send_queue;
626 u32 rs_snd_bytes;
627 int rs_rcv_bytes;
628 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
629
630 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
631 * to decide whether the application should be woken up.
632 * If not set, we use rs_cong_track to find out whether a cong map
633 * update arrived.
634 */
635 uint64_t rs_cong_mask;
636 uint64_t rs_cong_notify;
637 struct list_head rs_cong_list;
638 unsigned long rs_cong_track;
639
640 /*
641 * rs_recv_lock protects the receive queue, and is
642 * used to serialize with rds_release.
643 */
644 rwlock_t rs_recv_lock;
645 struct list_head rs_recv_queue;
646
647 /* just for stats reporting */
648 struct list_head rs_item;
649
650 /* these have their own lock */
651 spinlock_t rs_rdma_lock;
652 struct rb_root rs_rdma_keys;
653
654 /* Socket options - in case there will be more */
655 unsigned char rs_recverr,
656 rs_cong_monitor;
657 u32 rs_hash_initval;
658
659 /* Socket receive path trace points*/
660 u8 rs_rx_traces;
661 u8 rs_rx_trace[RDS_MSG_RX_DGRAM_TRACE_MAX];
662 struct rds_msg_zcopy_queue rs_zcookie_queue;
663 u8 rs_tos;
664};
665
666static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
667{
668 return container_of(sk, struct rds_sock, rs_sk);
669}
670static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
671{
672 return &rs->rs_sk;
673}
674
675/*
676 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
677 * to account for overhead. We don't account for overhead, we just apply
678 * the number of payload bytes to the specified value.
679 */
680static inline int rds_sk_sndbuf(struct rds_sock *rs)
681{
682 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
683}
684static inline int rds_sk_rcvbuf(struct rds_sock *rs)
685{
686 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
687}
688
689struct rds_statistics {
690 uint64_t s_conn_reset;
691 uint64_t s_recv_drop_bad_checksum;
692 uint64_t s_recv_drop_old_seq;
693 uint64_t s_recv_drop_no_sock;
694 uint64_t s_recv_drop_dead_sock;
695 uint64_t s_recv_deliver_raced;
696 uint64_t s_recv_delivered;
697 uint64_t s_recv_queued;
698 uint64_t s_recv_immediate_retry;
699 uint64_t s_recv_delayed_retry;
700 uint64_t s_recv_ack_required;
701 uint64_t s_recv_rdma_bytes;
702 uint64_t s_recv_ping;
703 uint64_t s_send_queue_empty;
704 uint64_t s_send_queue_full;
705 uint64_t s_send_lock_contention;
706 uint64_t s_send_lock_queue_raced;
707 uint64_t s_send_immediate_retry;
708 uint64_t s_send_delayed_retry;
709 uint64_t s_send_drop_acked;
710 uint64_t s_send_ack_required;
711 uint64_t s_send_queued;
712 uint64_t s_send_rdma;
713 uint64_t s_send_rdma_bytes;
714 uint64_t s_send_pong;
715 uint64_t s_page_remainder_hit;
716 uint64_t s_page_remainder_miss;
717 uint64_t s_copy_to_user;
718 uint64_t s_copy_from_user;
719 uint64_t s_cong_update_queued;
720 uint64_t s_cong_update_received;
721 uint64_t s_cong_send_error;
722 uint64_t s_cong_send_blocked;
723 uint64_t s_recv_bytes_added_to_socket;
724 uint64_t s_recv_bytes_removed_from_socket;
725 uint64_t s_send_stuck_rm;
726};
727
728/* af_rds.c */
729void rds_sock_addref(struct rds_sock *rs);
730void rds_sock_put(struct rds_sock *rs);
731void rds_wake_sk_sleep(struct rds_sock *rs);
732static inline void __rds_wake_sk_sleep(struct sock *sk)
733{
734 wait_queue_head_t *waitq = sk_sleep(sk);
735
736 if (!sock_flag(sk, SOCK_DEAD) && waitq)
737 wake_up(waitq);
738}
739extern wait_queue_head_t rds_poll_waitq;
740
741
742/* bind.c */
743int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
744void rds_remove_bound(struct rds_sock *rs);
745struct rds_sock *rds_find_bound(const struct in6_addr *addr, __be16 port,
746 __u32 scope_id);
747int rds_bind_lock_init(void);
748void rds_bind_lock_destroy(void);
749
750/* cong.c */
751int rds_cong_get_maps(struct rds_connection *conn);
752void rds_cong_add_conn(struct rds_connection *conn);
753void rds_cong_remove_conn(struct rds_connection *conn);
754void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
755void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
756int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
757void rds_cong_queue_updates(struct rds_cong_map *map);
758void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
759int rds_cong_updated_since(unsigned long *recent);
760void rds_cong_add_socket(struct rds_sock *);
761void rds_cong_remove_socket(struct rds_sock *);
762void rds_cong_exit(void);
763struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
764
765/* connection.c */
766extern u32 rds_gen_num;
767int rds_conn_init(void);
768void rds_conn_exit(void);
769struct rds_connection *rds_conn_create(struct net *net,
770 const struct in6_addr *laddr,
771 const struct in6_addr *faddr,
772 struct rds_transport *trans,
773 u8 tos, gfp_t gfp,
774 int dev_if);
775struct rds_connection *rds_conn_create_outgoing(struct net *net,
776 const struct in6_addr *laddr,
777 const struct in6_addr *faddr,
778 struct rds_transport *trans,
779 u8 tos, gfp_t gfp, int dev_if);
780void rds_conn_shutdown(struct rds_conn_path *cpath);
781void rds_conn_destroy(struct rds_connection *conn);
782void rds_conn_drop(struct rds_connection *conn);
783void rds_conn_path_drop(struct rds_conn_path *cpath, bool destroy);
784void rds_conn_connect_if_down(struct rds_connection *conn);
785void rds_conn_path_connect_if_down(struct rds_conn_path *cp);
786void rds_for_each_conn_info(struct socket *sock, unsigned int len,
787 struct rds_info_iterator *iter,
788 struct rds_info_lengths *lens,
789 int (*visitor)(struct rds_connection *, void *),
790 u64 *buffer,
791 size_t item_len);
792
793__printf(2, 3)
794void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...);
795#define rds_conn_path_error(cp, fmt...) \
796 __rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt)
797
798static inline int
799rds_conn_path_transition(struct rds_conn_path *cp, int old, int new)
800{
801 return atomic_cmpxchg(&cp->cp_state, old, new) == old;
802}
803
804static inline int
805rds_conn_transition(struct rds_connection *conn, int old, int new)
806{
807 WARN_ON(conn->c_trans->t_mp_capable);
808 return rds_conn_path_transition(&conn->c_path[0], old, new);
809}
810
811static inline int
812rds_conn_path_state(struct rds_conn_path *cp)
813{
814 return atomic_read(&cp->cp_state);
815}
816
817static inline int
818rds_conn_state(struct rds_connection *conn)
819{
820 WARN_ON(conn->c_trans->t_mp_capable);
821 return rds_conn_path_state(&conn->c_path[0]);
822}
823
824static inline int
825rds_conn_path_up(struct rds_conn_path *cp)
826{
827 return atomic_read(&cp->cp_state) == RDS_CONN_UP;
828}
829
830static inline int
831rds_conn_up(struct rds_connection *conn)
832{
833 WARN_ON(conn->c_trans->t_mp_capable);
834 return rds_conn_path_up(&conn->c_path[0]);
835}
836
837static inline int
838rds_conn_path_connecting(struct rds_conn_path *cp)
839{
840 return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING;
841}
842
843static inline int
844rds_conn_connecting(struct rds_connection *conn)
845{
846 WARN_ON(conn->c_trans->t_mp_capable);
847 return rds_conn_path_connecting(&conn->c_path[0]);
848}
849
850/* message.c */
851struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
852struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents,
853 int *ret);
854int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from,
855 bool zcopy);
856struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
857void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
858 __be16 dport, u64 seq);
859int rds_message_add_extension(struct rds_header *hdr,
860 unsigned int type, const void *data, unsigned int len);
861int rds_message_next_extension(struct rds_header *hdr,
862 unsigned int *pos, void *buf, unsigned int *buflen);
863int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
864int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
865void rds_message_inc_free(struct rds_incoming *inc);
866void rds_message_addref(struct rds_message *rm);
867void rds_message_put(struct rds_message *rm);
868void rds_message_wait(struct rds_message *rm);
869void rds_message_unmapped(struct rds_message *rm);
870void rds_notify_msg_zcopy_purge(struct rds_msg_zcopy_queue *info);
871
872static inline void rds_message_make_checksum(struct rds_header *hdr)
873{
874 hdr->h_csum = 0;
875 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
876}
877
878static inline int rds_message_verify_checksum(const struct rds_header *hdr)
879{
880 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
881}
882
883
884/* page.c */
885int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
886 gfp_t gfp);
887void rds_page_exit(void);
888
889/* recv.c */
890void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
891 struct in6_addr *saddr);
892void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn,
893 struct in6_addr *saddr);
894void rds_inc_put(struct rds_incoming *inc);
895void rds_recv_incoming(struct rds_connection *conn, struct in6_addr *saddr,
896 struct in6_addr *daddr,
897 struct rds_incoming *inc, gfp_t gfp);
898int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
899 int msg_flags);
900void rds_clear_recv_queue(struct rds_sock *rs);
901int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
902void rds_inc_info_copy(struct rds_incoming *inc,
903 struct rds_info_iterator *iter,
904 __be32 saddr, __be32 daddr, int flip);
905void rds6_inc_info_copy(struct rds_incoming *inc,
906 struct rds_info_iterator *iter,
907 struct in6_addr *saddr, struct in6_addr *daddr,
908 int flip);
909
910/* send.c */
911int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
912void rds_send_path_reset(struct rds_conn_path *conn);
913int rds_send_xmit(struct rds_conn_path *cp);
914struct sockaddr_in;
915void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in6 *dest);
916typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
917void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
918 is_acked_func is_acked);
919void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
920 is_acked_func is_acked);
921void rds_send_ping(struct rds_connection *conn, int cp_index);
922int rds_send_pong(struct rds_conn_path *cp, __be16 dport);
923
924/* rdma.c */
925void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
926int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
927int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
928int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
929void rds_rdma_drop_keys(struct rds_sock *rs);
930int rds_rdma_extra_size(struct rds_rdma_args *args,
931 struct rds_iov_vector *iov);
932int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
933 struct cmsghdr *cmsg);
934int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
935 struct cmsghdr *cmsg,
936 struct rds_iov_vector *vec);
937int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
938 struct cmsghdr *cmsg);
939void rds_rdma_free_op(struct rm_rdma_op *ro);
940void rds_atomic_free_op(struct rm_atomic_op *ao);
941void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
942void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
943int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
944 struct cmsghdr *cmsg);
945
946void __rds_put_mr_final(struct rds_mr *mr);
947static inline void rds_mr_put(struct rds_mr *mr)
948{
949 if (refcount_dec_and_test(&mr->r_refcount))
950 __rds_put_mr_final(mr);
951}
952
953static inline bool rds_destroy_pending(struct rds_connection *conn)
954{
955 return !check_net(rds_conn_net(conn)) ||
956 (conn->c_trans->t_unloading && conn->c_trans->t_unloading(conn));
957}
958
959/* stats.c */
960DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
961#define rds_stats_inc_which(which, member) do { \
962 per_cpu(which, get_cpu()).member++; \
963 put_cpu(); \
964} while (0)
965#define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
966#define rds_stats_add_which(which, member, count) do { \
967 per_cpu(which, get_cpu()).member += count; \
968 put_cpu(); \
969} while (0)
970#define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
971int rds_stats_init(void);
972void rds_stats_exit(void);
973void rds_stats_info_copy(struct rds_info_iterator *iter,
974 uint64_t *values, const char *const *names,
975 size_t nr);
976
977/* sysctl.c */
978int rds_sysctl_init(void);
979void rds_sysctl_exit(void);
980extern unsigned long rds_sysctl_sndbuf_min;
981extern unsigned long rds_sysctl_sndbuf_default;
982extern unsigned long rds_sysctl_sndbuf_max;
983extern unsigned long rds_sysctl_reconnect_min_jiffies;
984extern unsigned long rds_sysctl_reconnect_max_jiffies;
985extern unsigned int rds_sysctl_max_unacked_packets;
986extern unsigned int rds_sysctl_max_unacked_bytes;
987extern unsigned int rds_sysctl_ping_enable;
988extern unsigned long rds_sysctl_trace_flags;
989extern unsigned int rds_sysctl_trace_level;
990
991/* threads.c */
992int rds_threads_init(void);
993void rds_threads_exit(void);
994extern struct workqueue_struct *rds_wq;
995void rds_queue_reconnect(struct rds_conn_path *cp);
996void rds_connect_worker(struct work_struct *);
997void rds_shutdown_worker(struct work_struct *);
998void rds_send_worker(struct work_struct *);
999void rds_recv_worker(struct work_struct *);
1000void rds_connect_path_complete(struct rds_conn_path *conn, int curr);
1001void rds_connect_complete(struct rds_connection *conn);
1002int rds_addr_cmp(const struct in6_addr *a1, const struct in6_addr *a2);
1003
1004/* transport.c */
1005void rds_trans_register(struct rds_transport *trans);
1006void rds_trans_unregister(struct rds_transport *trans);
1007struct rds_transport *rds_trans_get_preferred(struct net *net,
1008 const struct in6_addr *addr,
1009 __u32 scope_id);
1010void rds_trans_put(struct rds_transport *trans);
1011unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
1012 unsigned int avail);
1013struct rds_transport *rds_trans_get(int t_type);
1014int rds_trans_init(void);
1015void rds_trans_exit(void);
1016
1017#endif