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#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#ifdef RDS_DEBUG
44#define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
45#else
46/* sigh, pr_debug() causes unused variable warnings */
47static inline __printf(1, 2)
48void rdsdebug(char *fmt, ...)
49{
50}
51#endif
52
53#define RDS_FRAG_SHIFT 12
54#define RDS_FRAG_SIZE ((unsigned int)(1 << RDS_FRAG_SHIFT))
55
56/* Used to limit both RDMA and non-RDMA RDS message to 1MB */
57#define RDS_MAX_MSG_SIZE ((unsigned int)(1 << 20))
58
59#define RDS_CONG_MAP_BYTES (65536 / 8)
60#define RDS_CONG_MAP_PAGES (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
61#define RDS_CONG_MAP_PAGE_BITS (PAGE_SIZE * 8)
62
63struct rds_cong_map {
64 struct rb_node m_rb_node;
65 struct in6_addr m_addr;
66 wait_queue_head_t m_waitq;
67 struct list_head m_conn_list;
68 unsigned long m_page_addrs[RDS_CONG_MAP_PAGES];
69};
70
71
72/*
73 * This is how we will track the connection state:
74 * A connection is always in one of the following
75 * states. Updates to the state are atomic and imply
76 * a memory barrier.
77 */
78enum {
79 RDS_CONN_DOWN = 0,
80 RDS_CONN_CONNECTING,
81 RDS_CONN_DISCONNECTING,
82 RDS_CONN_UP,
83 RDS_CONN_RESETTING,
84 RDS_CONN_ERROR,
85};
86
87/* Bits for c_flags */
88#define RDS_LL_SEND_FULL 0
89#define RDS_RECONNECT_PENDING 1
90#define RDS_IN_XMIT 2
91#define RDS_RECV_REFILL 3
92#define RDS_DESTROY_PENDING 4
93
94/* Max number of multipaths per RDS connection. Must be a power of 2 */
95#define RDS_MPATH_WORKERS 8
96#define RDS_MPATH_HASH(rs, n) (jhash_1word((rs)->rs_bound_port, \
97 (rs)->rs_hash_initval) & ((n) - 1))
98
99#define IS_CANONICAL(laddr, faddr) (htonl(laddr) < htonl(faddr))
100
101/* Per mpath connection state */
102struct rds_conn_path {
103 struct rds_connection *cp_conn;
104 struct rds_message *cp_xmit_rm;
105 unsigned long cp_xmit_sg;
106 unsigned int cp_xmit_hdr_off;
107 unsigned int cp_xmit_data_off;
108 unsigned int cp_xmit_atomic_sent;
109 unsigned int cp_xmit_rdma_sent;
110 unsigned int cp_xmit_data_sent;
111
112 spinlock_t cp_lock; /* protect msg queues */
113 u64 cp_next_tx_seq;
114 struct list_head cp_send_queue;
115 struct list_head cp_retrans;
116
117 u64 cp_next_rx_seq;
118
119 void *cp_transport_data;
120
121 atomic_t cp_state;
122 unsigned long cp_send_gen;
123 unsigned long cp_flags;
124 unsigned long cp_reconnect_jiffies;
125 struct delayed_work cp_send_w;
126 struct delayed_work cp_recv_w;
127 struct delayed_work cp_conn_w;
128 struct work_struct cp_down_w;
129 struct mutex cp_cm_lock; /* protect cp_state & cm */
130 wait_queue_head_t cp_waitq;
131
132 unsigned int cp_unacked_packets;
133 unsigned int cp_unacked_bytes;
134 unsigned int cp_index;
135};
136
137/* One rds_connection per RDS address pair */
138struct rds_connection {
139 struct hlist_node c_hash_node;
140 struct in6_addr c_laddr;
141 struct in6_addr c_faddr;
142 int c_dev_if; /* ifindex used for this conn */
143 int c_bound_if; /* ifindex of c_laddr */
144 unsigned int c_loopback:1,
145 c_isv6:1,
146 c_ping_triggered:1,
147 c_pad_to_32:29;
148 int c_npaths;
149 struct rds_connection *c_passive;
150 struct rds_transport *c_trans;
151
152 struct rds_cong_map *c_lcong;
153 struct rds_cong_map *c_fcong;
154
155 /* Protocol version */
156 unsigned int c_proposed_version;
157 unsigned int c_version;
158 possible_net_t c_net;
159
160 /* TOS */
161 u8 c_tos;
162
163 struct list_head c_map_item;
164 unsigned long c_map_queued;
165
166 struct rds_conn_path *c_path;
167 wait_queue_head_t c_hs_waitq; /* handshake waitq */
168
169 u32 c_my_gen_num;
170 u32 c_peer_gen_num;
171};
172
173static inline
174struct net *rds_conn_net(struct rds_connection *conn)
175{
176 return read_pnet(&conn->c_net);
177}
178
179static inline
180void rds_conn_net_set(struct rds_connection *conn, struct net *net)
181{
182 write_pnet(&conn->c_net, net);
183}
184
185#define RDS_FLAG_CONG_BITMAP 0x01
186#define RDS_FLAG_ACK_REQUIRED 0x02
187#define RDS_FLAG_RETRANSMITTED 0x04
188#define RDS_MAX_ADV_CREDIT 255
189
190/* RDS_FLAG_PROBE_PORT is the reserved sport used for sending a ping
191 * probe to exchange control information before establishing a connection.
192 * Currently the control information that is exchanged is the number of
193 * supported paths. If the peer is a legacy (older kernel revision) peer,
194 * it would return a pong message without additional control information
195 * that would then alert the sender that the peer was an older rev.
196 */
197#define RDS_FLAG_PROBE_PORT 1
198#define RDS_HS_PROBE(sport, dport) \
199 ((sport == RDS_FLAG_PROBE_PORT && dport == 0) || \
200 (sport == 0 && dport == RDS_FLAG_PROBE_PORT))
201/*
202 * Maximum space available for extension headers.
203 */
204#define RDS_HEADER_EXT_SPACE 16
205
206struct rds_header {
207 __be64 h_sequence;
208 __be64 h_ack;
209 __be32 h_len;
210 __be16 h_sport;
211 __be16 h_dport;
212 u8 h_flags;
213 u8 h_credit;
214 u8 h_padding[4];
215 __sum16 h_csum;
216
217 u8 h_exthdr[RDS_HEADER_EXT_SPACE];
218};
219
220/*
221 * Reserved - indicates end of extensions
222 */
223#define RDS_EXTHDR_NONE 0
224
225/*
226 * This extension header is included in the very
227 * first message that is sent on a new connection,
228 * and identifies the protocol level. This will help
229 * rolling updates if a future change requires breaking
230 * the protocol.
231 * NB: This is no longer true for IB, where we do a version
232 * negotiation during the connection setup phase (protocol
233 * version information is included in the RDMA CM private data).
234 */
235#define RDS_EXTHDR_VERSION 1
236struct rds_ext_header_version {
237 __be32 h_version;
238};
239
240/*
241 * This extension header is included in the RDS message
242 * chasing an RDMA operation.
243 */
244#define RDS_EXTHDR_RDMA 2
245struct rds_ext_header_rdma {
246 __be32 h_rdma_rkey;
247};
248
249/*
250 * This extension header tells the peer about the
251 * destination <R_Key,offset> of the requested RDMA
252 * operation.
253 */
254#define RDS_EXTHDR_RDMA_DEST 3
255struct rds_ext_header_rdma_dest {
256 __be32 h_rdma_rkey;
257 __be32 h_rdma_offset;
258};
259
260/* Extension header announcing number of paths.
261 * Implicit length = 2 bytes.
262 */
263#define RDS_EXTHDR_NPATHS 5
264#define RDS_EXTHDR_GEN_NUM 6
265
266#define __RDS_EXTHDR_MAX 16 /* for now */
267#define RDS_RX_MAX_TRACES (RDS_MSG_RX_DGRAM_TRACE_MAX + 1)
268#define RDS_MSG_RX_HDR 0
269#define RDS_MSG_RX_START 1
270#define RDS_MSG_RX_END 2
271#define RDS_MSG_RX_CMSG 3
272
273/* The following values are whitelisted for usercopy */
274struct rds_inc_usercopy {
275 rds_rdma_cookie_t rdma_cookie;
276 ktime_t rx_tstamp;
277};
278
279struct rds_incoming {
280 refcount_t i_refcount;
281 struct list_head i_item;
282 struct rds_connection *i_conn;
283 struct rds_conn_path *i_conn_path;
284 struct rds_header i_hdr;
285 unsigned long i_rx_jiffies;
286 struct in6_addr i_saddr;
287
288 struct rds_inc_usercopy i_usercopy;
289 u64 i_rx_lat_trace[RDS_RX_MAX_TRACES];
290};
291
292struct rds_mr {
293 struct rb_node r_rb_node;
294 struct kref r_kref;
295 u32 r_key;
296
297 /* A copy of the creation flags */
298 unsigned int r_use_once:1;
299 unsigned int r_invalidate:1;
300 unsigned int r_write:1;
301
302 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
303 struct rds_transport *r_trans;
304 void *r_trans_private;
305};
306
307static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
308{
309 return r_key | (((u64) offset) << 32);
310}
311
312static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
313{
314 return cookie;
315}
316
317static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
318{
319 return cookie >> 32;
320}
321
322/* atomic operation types */
323#define RDS_ATOMIC_TYPE_CSWP 0
324#define RDS_ATOMIC_TYPE_FADD 1
325
326/*
327 * m_sock_item and m_conn_item are on lists that are serialized under
328 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
329 * the message will not be put back on the retransmit list after being sent.
330 * messages that are canceled while being sent rely on this.
331 *
332 * m_inc is used by loopback so that it can pass an incoming message straight
333 * back up into the rx path. It embeds a wire header which is also used by
334 * the send path, which is kind of awkward.
335 *
336 * m_sock_item indicates the message's presence on a socket's send or receive
337 * queue. m_rs will point to that socket.
338 *
339 * m_daddr is used by cancellation to prune messages to a given destination.
340 *
341 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
342 * nesting. As paths iterate over messages on a sock, or conn, they must
343 * also lock the conn, or sock, to remove the message from those lists too.
344 * Testing the flag to determine if the message is still on the lists lets
345 * us avoid testing the list_head directly. That means each path can use
346 * the message's list_head to keep it on a local list while juggling locks
347 * without confusing the other path.
348 *
349 * m_ack_seq is an optional field set by transports who need a different
350 * sequence number range to invalidate. They can use this in a callback
351 * that they pass to rds_send_drop_acked() to see if each message has been
352 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
353 * had ack_seq set yet.
354 */
355#define RDS_MSG_ON_SOCK 1
356#define RDS_MSG_ON_CONN 2
357#define RDS_MSG_HAS_ACK_SEQ 3
358#define RDS_MSG_ACK_REQUIRED 4
359#define RDS_MSG_RETRANSMITTED 5
360#define RDS_MSG_MAPPED 6
361#define RDS_MSG_PAGEVEC 7
362#define RDS_MSG_FLUSH 8
363
364struct rds_znotifier {
365 struct mmpin z_mmp;
366 u32 z_cookie;
367};
368
369struct rds_msg_zcopy_info {
370 struct list_head rs_zcookie_next;
371 union {
372 struct rds_znotifier znotif;
373 struct rds_zcopy_cookies zcookies;
374 };
375};
376
377struct rds_msg_zcopy_queue {
378 struct list_head zcookie_head;
379 spinlock_t lock; /* protects zcookie_head queue */
380};
381
382static inline void rds_message_zcopy_queue_init(struct rds_msg_zcopy_queue *q)
383{
384 spin_lock_init(&q->lock);
385 INIT_LIST_HEAD(&q->zcookie_head);
386}
387
388struct rds_iov_vector {
389 struct rds_iovec *iov;
390 int len;
391};
392
393struct rds_iov_vector_arr {
394 struct rds_iov_vector *vec;
395 int len;
396 int indx;
397 int incr;
398};
399
400struct rds_message {
401 refcount_t m_refcount;
402 struct list_head m_sock_item;
403 struct list_head m_conn_item;
404 struct rds_incoming m_inc;
405 u64 m_ack_seq;
406 struct in6_addr m_daddr;
407 unsigned long m_flags;
408
409 /* Never access m_rs without holding m_rs_lock.
410 * Lock nesting is
411 * rm->m_rs_lock
412 * -> rs->rs_lock
413 */
414 spinlock_t m_rs_lock;
415 wait_queue_head_t m_flush_wait;
416
417 struct rds_sock *m_rs;
418
419 /* cookie to send to remote, in rds header */
420 rds_rdma_cookie_t m_rdma_cookie;
421
422 unsigned int m_used_sgs;
423 unsigned int m_total_sgs;
424
425 void *m_final_op;
426
427 struct {
428 struct rm_atomic_op {
429 int op_type;
430 union {
431 struct {
432 uint64_t compare;
433 uint64_t swap;
434 uint64_t compare_mask;
435 uint64_t swap_mask;
436 } op_m_cswp;
437 struct {
438 uint64_t add;
439 uint64_t nocarry_mask;
440 } op_m_fadd;
441 };
442
443 u32 op_rkey;
444 u64 op_remote_addr;
445 unsigned int op_notify:1;
446 unsigned int op_recverr:1;
447 unsigned int op_mapped:1;
448 unsigned int op_silent:1;
449 unsigned int op_active:1;
450 struct scatterlist *op_sg;
451 struct rds_notifier *op_notifier;
452
453 struct rds_mr *op_rdma_mr;
454 } atomic;
455 struct rm_rdma_op {
456 u32 op_rkey;
457 u64 op_remote_addr;
458 unsigned int op_write:1;
459 unsigned int op_fence:1;
460 unsigned int op_notify:1;
461 unsigned int op_recverr:1;
462 unsigned int op_mapped:1;
463 unsigned int op_silent:1;
464 unsigned int op_active:1;
465 unsigned int op_bytes;
466 unsigned int op_nents;
467 unsigned int op_count;
468 struct scatterlist *op_sg;
469 struct rds_notifier *op_notifier;
470
471 struct rds_mr *op_rdma_mr;
472
473 u64 op_odp_addr;
474 struct rds_mr *op_odp_mr;
475 } rdma;
476 struct rm_data_op {
477 unsigned int op_active:1;
478 unsigned int op_nents;
479 unsigned int op_count;
480 unsigned int op_dmasg;
481 unsigned int op_dmaoff;
482 struct rds_znotifier *op_mmp_znotifier;
483 struct scatterlist *op_sg;
484 } data;
485 };
486
487 struct rds_conn_path *m_conn_path;
488};
489
490/*
491 * The RDS notifier is used (optionally) to tell the application about
492 * completed RDMA operations. Rather than keeping the whole rds message
493 * around on the queue, we allocate a small notifier that is put on the
494 * socket's notifier_list. Notifications are delivered to the application
495 * through control messages.
496 */
497struct rds_notifier {
498 struct list_head n_list;
499 uint64_t n_user_token;
500 int n_status;
501};
502
503/* Available as part of RDS core, so doesn't need to participate
504 * in get_preferred transport etc
505 */
506#define RDS_TRANS_LOOP 3
507
508/**
509 * struct rds_transport - transport specific behavioural hooks
510 *
511 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
512 * part of a message. The caller serializes on the send_sem so this
513 * doesn't need to be reentrant for a given conn. The header must be
514 * sent before the data payload. .xmit must be prepared to send a
515 * message with no data payload. .xmit should return the number of
516 * bytes that were sent down the connection, including header bytes.
517 * Returning 0 tells the caller that it doesn't need to perform any
518 * additional work now. This is usually the case when the transport has
519 * filled the sending queue for its connection and will handle
520 * triggering the rds thread to continue the send when space becomes
521 * available. Returning -EAGAIN tells the caller to retry the send
522 * immediately. Returning -ENOMEM tells the caller to retry the send at
523 * some point in the future.
524 *
525 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
526 * it returns the connection can not call rds_recv_incoming().
527 * This will only be called once after conn_connect returns
528 * non-zero success and will The caller serializes this with
529 * the send and connecting paths (xmit_* and conn_*). The
530 * transport is responsible for other serialization, including
531 * rds_recv_incoming(). This is called in process context but
532 * should try hard not to block.
533 */
534
535struct rds_transport {
536 char t_name[TRANSNAMSIZ];
537 struct list_head t_item;
538 struct module *t_owner;
539 unsigned int t_prefer_loopback:1,
540 t_mp_capable:1;
541 unsigned int t_type;
542
543 int (*laddr_check)(struct net *net, const struct in6_addr *addr,
544 __u32 scope_id);
545 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
546 void (*conn_free)(void *data);
547 int (*conn_path_connect)(struct rds_conn_path *cp);
548 void (*conn_path_shutdown)(struct rds_conn_path *conn);
549 void (*xmit_path_prepare)(struct rds_conn_path *cp);
550 void (*xmit_path_complete)(struct rds_conn_path *cp);
551 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
552 unsigned int hdr_off, unsigned int sg, unsigned int off);
553 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
554 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
555 int (*recv_path)(struct rds_conn_path *cp);
556 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
557 void (*inc_free)(struct rds_incoming *inc);
558
559 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
560 struct rdma_cm_event *event, bool isv6);
561 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id, bool isv6);
562 void (*cm_connect_complete)(struct rds_connection *conn,
563 struct rdma_cm_event *event);
564
565 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
566 unsigned int avail);
567 void (*exit)(void);
568 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
569 struct rds_sock *rs, u32 *key_ret,
570 struct rds_connection *conn,
571 u64 start, u64 length, int need_odp);
572 void (*sync_mr)(void *trans_private, int direction);
573 void (*free_mr)(void *trans_private, int invalidate);
574 void (*flush_mrs)(void);
575 bool (*t_unloading)(struct rds_connection *conn);
576 u8 (*get_tos_map)(u8 tos);
577};
578
579/* Bind hash table key length. It is the sum of the size of a struct
580 * in6_addr, a scope_id and a port.
581 */
582#define RDS_BOUND_KEY_LEN \
583 (sizeof(struct in6_addr) + sizeof(__u32) + sizeof(__be16))
584
585struct rds_sock {
586 struct sock rs_sk;
587
588 u64 rs_user_addr;
589 u64 rs_user_bytes;
590
591 /*
592 * bound_addr used for both incoming and outgoing, no INADDR_ANY
593 * support.
594 */
595 struct rhash_head rs_bound_node;
596 u8 rs_bound_key[RDS_BOUND_KEY_LEN];
597 struct sockaddr_in6 rs_bound_sin6;
598#define rs_bound_addr rs_bound_sin6.sin6_addr
599#define rs_bound_addr_v4 rs_bound_sin6.sin6_addr.s6_addr32[3]
600#define rs_bound_port rs_bound_sin6.sin6_port
601#define rs_bound_scope_id rs_bound_sin6.sin6_scope_id
602 struct in6_addr rs_conn_addr;
603#define rs_conn_addr_v4 rs_conn_addr.s6_addr32[3]
604 __be16 rs_conn_port;
605 struct rds_transport *rs_transport;
606
607 /*
608 * rds_sendmsg caches the conn it used the last time around.
609 * This helps avoid costly lookups.
610 */
611 struct rds_connection *rs_conn;
612
613 /* flag indicating we were congested or not */
614 int rs_congested;
615 /* seen congestion (ENOBUFS) when sending? */
616 int rs_seen_congestion;
617
618 /* rs_lock protects all these adjacent members before the newline */
619 spinlock_t rs_lock;
620 struct list_head rs_send_queue;
621 u32 rs_snd_bytes;
622 int rs_rcv_bytes;
623 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
624
625 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
626 * to decide whether the application should be woken up.
627 * If not set, we use rs_cong_track to find out whether a cong map
628 * update arrived.
629 */
630 uint64_t rs_cong_mask;
631 uint64_t rs_cong_notify;
632 struct list_head rs_cong_list;
633 unsigned long rs_cong_track;
634
635 /*
636 * rs_recv_lock protects the receive queue, and is
637 * used to serialize with rds_release.
638 */
639 rwlock_t rs_recv_lock;
640 struct list_head rs_recv_queue;
641
642 /* just for stats reporting */
643 struct list_head rs_item;
644
645 /* these have their own lock */
646 spinlock_t rs_rdma_lock;
647 struct rb_root rs_rdma_keys;
648
649 /* Socket options - in case there will be more */
650 unsigned char rs_recverr,
651 rs_cong_monitor;
652 u32 rs_hash_initval;
653
654 /* Socket receive path trace points*/
655 u8 rs_rx_traces;
656 u8 rs_rx_trace[RDS_MSG_RX_DGRAM_TRACE_MAX];
657 struct rds_msg_zcopy_queue rs_zcookie_queue;
658 u8 rs_tos;
659};
660
661static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
662{
663 return container_of(sk, struct rds_sock, rs_sk);
664}
665static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
666{
667 return &rs->rs_sk;
668}
669
670/*
671 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
672 * to account for overhead. We don't account for overhead, we just apply
673 * the number of payload bytes to the specified value.
674 */
675static inline int rds_sk_sndbuf(struct rds_sock *rs)
676{
677 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
678}
679static inline int rds_sk_rcvbuf(struct rds_sock *rs)
680{
681 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
682}
683
684struct rds_statistics {
685 uint64_t s_conn_reset;
686 uint64_t s_recv_drop_bad_checksum;
687 uint64_t s_recv_drop_old_seq;
688 uint64_t s_recv_drop_no_sock;
689 uint64_t s_recv_drop_dead_sock;
690 uint64_t s_recv_deliver_raced;
691 uint64_t s_recv_delivered;
692 uint64_t s_recv_queued;
693 uint64_t s_recv_immediate_retry;
694 uint64_t s_recv_delayed_retry;
695 uint64_t s_recv_ack_required;
696 uint64_t s_recv_rdma_bytes;
697 uint64_t s_recv_ping;
698 uint64_t s_send_queue_empty;
699 uint64_t s_send_queue_full;
700 uint64_t s_send_lock_contention;
701 uint64_t s_send_lock_queue_raced;
702 uint64_t s_send_immediate_retry;
703 uint64_t s_send_delayed_retry;
704 uint64_t s_send_drop_acked;
705 uint64_t s_send_ack_required;
706 uint64_t s_send_queued;
707 uint64_t s_send_rdma;
708 uint64_t s_send_rdma_bytes;
709 uint64_t s_send_pong;
710 uint64_t s_page_remainder_hit;
711 uint64_t s_page_remainder_miss;
712 uint64_t s_copy_to_user;
713 uint64_t s_copy_from_user;
714 uint64_t s_cong_update_queued;
715 uint64_t s_cong_update_received;
716 uint64_t s_cong_send_error;
717 uint64_t s_cong_send_blocked;
718 uint64_t s_recv_bytes_added_to_socket;
719 uint64_t s_recv_bytes_removed_from_socket;
720 uint64_t s_send_stuck_rm;
721};
722
723/* af_rds.c */
724void rds_sock_addref(struct rds_sock *rs);
725void rds_sock_put(struct rds_sock *rs);
726void rds_wake_sk_sleep(struct rds_sock *rs);
727static inline void __rds_wake_sk_sleep(struct sock *sk)
728{
729 wait_queue_head_t *waitq = sk_sleep(sk);
730
731 if (!sock_flag(sk, SOCK_DEAD) && waitq)
732 wake_up(waitq);
733}
734extern wait_queue_head_t rds_poll_waitq;
735
736
737/* bind.c */
738int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
739void rds_remove_bound(struct rds_sock *rs);
740struct rds_sock *rds_find_bound(const struct in6_addr *addr, __be16 port,
741 __u32 scope_id);
742int rds_bind_lock_init(void);
743void rds_bind_lock_destroy(void);
744
745/* cong.c */
746int rds_cong_get_maps(struct rds_connection *conn);
747void rds_cong_add_conn(struct rds_connection *conn);
748void rds_cong_remove_conn(struct rds_connection *conn);
749void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
750void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
751int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
752void rds_cong_queue_updates(struct rds_cong_map *map);
753void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
754int rds_cong_updated_since(unsigned long *recent);
755void rds_cong_add_socket(struct rds_sock *);
756void rds_cong_remove_socket(struct rds_sock *);
757void rds_cong_exit(void);
758struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
759
760/* connection.c */
761extern u32 rds_gen_num;
762int rds_conn_init(void);
763void rds_conn_exit(void);
764struct rds_connection *rds_conn_create(struct net *net,
765 const struct in6_addr *laddr,
766 const struct in6_addr *faddr,
767 struct rds_transport *trans,
768 u8 tos, gfp_t gfp,
769 int dev_if);
770struct rds_connection *rds_conn_create_outgoing(struct net *net,
771 const struct in6_addr *laddr,
772 const struct in6_addr *faddr,
773 struct rds_transport *trans,
774 u8 tos, gfp_t gfp, int dev_if);
775void rds_conn_shutdown(struct rds_conn_path *cpath);
776void rds_conn_destroy(struct rds_connection *conn);
777void rds_conn_drop(struct rds_connection *conn);
778void rds_conn_path_drop(struct rds_conn_path *cpath, bool destroy);
779void rds_conn_connect_if_down(struct rds_connection *conn);
780void rds_conn_path_connect_if_down(struct rds_conn_path *cp);
781void rds_check_all_paths(struct rds_connection *conn);
782void rds_for_each_conn_info(struct socket *sock, unsigned int len,
783 struct rds_info_iterator *iter,
784 struct rds_info_lengths *lens,
785 int (*visitor)(struct rds_connection *, void *),
786 u64 *buffer,
787 size_t item_len);
788
789__printf(2, 3)
790void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...);
791#define rds_conn_path_error(cp, fmt...) \
792 __rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt)
793
794static inline int
795rds_conn_path_transition(struct rds_conn_path *cp, int old, int new)
796{
797 return atomic_cmpxchg(&cp->cp_state, old, new) == old;
798}
799
800static inline int
801rds_conn_transition(struct rds_connection *conn, int old, int new)
802{
803 WARN_ON(conn->c_trans->t_mp_capable);
804 return rds_conn_path_transition(&conn->c_path[0], old, new);
805}
806
807static inline int
808rds_conn_path_state(struct rds_conn_path *cp)
809{
810 return atomic_read(&cp->cp_state);
811}
812
813static inline int
814rds_conn_state(struct rds_connection *conn)
815{
816 WARN_ON(conn->c_trans->t_mp_capable);
817 return rds_conn_path_state(&conn->c_path[0]);
818}
819
820static inline int
821rds_conn_path_up(struct rds_conn_path *cp)
822{
823 return atomic_read(&cp->cp_state) == RDS_CONN_UP;
824}
825
826static inline int
827rds_conn_path_down(struct rds_conn_path *cp)
828{
829 return atomic_read(&cp->cp_state) == RDS_CONN_DOWN;
830}
831
832static inline int
833rds_conn_up(struct rds_connection *conn)
834{
835 WARN_ON(conn->c_trans->t_mp_capable);
836 return rds_conn_path_up(&conn->c_path[0]);
837}
838
839static inline int
840rds_conn_path_connecting(struct rds_conn_path *cp)
841{
842 return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING;
843}
844
845static inline int
846rds_conn_connecting(struct rds_connection *conn)
847{
848 WARN_ON(conn->c_trans->t_mp_capable);
849 return rds_conn_path_connecting(&conn->c_path[0]);
850}
851
852/* message.c */
853struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
854struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
855int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from,
856 bool zcopy);
857struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
858void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
859 __be16 dport, u64 seq);
860int rds_message_add_extension(struct rds_header *hdr,
861 unsigned int type, const void *data, unsigned int len);
862int rds_message_next_extension(struct rds_header *hdr,
863 unsigned int *pos, void *buf, unsigned int *buflen);
864int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
865int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
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, sockptr_t optval, int optlen);
927int rds_get_mr_for_dest(struct rds_sock *rs, sockptr_t optval, int optlen);
928int rds_free_mr(struct rds_sock *rs, sockptr_t 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 kref *kref);
947
948static inline bool rds_destroy_pending(struct rds_connection *conn)
949{
950 return !check_net(rds_conn_net(conn)) ||
951 (conn->c_trans->t_unloading && conn->c_trans->t_unloading(conn));
952}
953
954enum {
955 ODP_NOT_NEEDED,
956 ODP_ZEROBASED,
957 ODP_VIRTUAL
958};
959
960/* stats.c */
961DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
962#define rds_stats_inc_which(which, member) do { \
963 per_cpu(which, get_cpu()).member++; \
964 put_cpu(); \
965} while (0)
966#define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
967#define rds_stats_add_which(which, member, count) do { \
968 per_cpu(which, get_cpu()).member += count; \
969 put_cpu(); \
970} while (0)
971#define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
972int rds_stats_init(void);
973void rds_stats_exit(void);
974void rds_stats_info_copy(struct rds_info_iterator *iter,
975 uint64_t *values, const char *const *names,
976 size_t nr);
977
978/* sysctl.c */
979int rds_sysctl_init(void);
980void rds_sysctl_exit(void);
981extern unsigned long rds_sysctl_sndbuf_min;
982extern unsigned long rds_sysctl_sndbuf_default;
983extern unsigned long rds_sysctl_sndbuf_max;
984extern unsigned long rds_sysctl_reconnect_min_jiffies;
985extern unsigned long rds_sysctl_reconnect_max_jiffies;
986extern unsigned int rds_sysctl_max_unacked_packets;
987extern unsigned int rds_sysctl_max_unacked_bytes;
988extern unsigned int rds_sysctl_ping_enable;
989extern unsigned long rds_sysctl_trace_flags;
990extern unsigned int rds_sysctl_trace_level;
991
992/* threads.c */
993int rds_threads_init(void);
994void rds_threads_exit(void);
995extern struct workqueue_struct *rds_wq;
996void rds_queue_reconnect(struct rds_conn_path *cp);
997void rds_connect_worker(struct work_struct *);
998void rds_shutdown_worker(struct work_struct *);
999void rds_send_worker(struct work_struct *);
1000void rds_recv_worker(struct work_struct *);
1001void rds_connect_path_complete(struct rds_conn_path *conn, int curr);
1002void rds_connect_complete(struct rds_connection *conn);
1003int rds_addr_cmp(const struct in6_addr *a1, const struct in6_addr *a2);
1004
1005/* transport.c */
1006void rds_trans_register(struct rds_transport *trans);
1007void rds_trans_unregister(struct rds_transport *trans);
1008struct rds_transport *rds_trans_get_preferred(struct net *net,
1009 const struct in6_addr *addr,
1010 __u32 scope_id);
1011void rds_trans_put(struct rds_transport *trans);
1012unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
1013 unsigned int avail);
1014struct rds_transport *rds_trans_get(int t_type);
1015
1016#endif