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#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 RDS_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_RESETTING,
78 RDS_CONN_ERROR,
79};
80
81/* Bits for c_flags */
82#define RDS_LL_SEND_FULL 0
83#define RDS_RECONNECT_PENDING 1
84#define RDS_IN_XMIT 2
85#define RDS_RECV_REFILL 3
86
87/* Max number of multipaths per RDS connection. Must be a power of 2 */
88#define RDS_MPATH_WORKERS 8
89#define RDS_MPATH_HASH(rs, n) (jhash_1word((rs)->rs_bound_port, \
90 (rs)->rs_hash_initval) & ((n) - 1))
91
92/* Per mpath connection state */
93struct rds_conn_path {
94 struct rds_connection *cp_conn;
95 struct rds_message *cp_xmit_rm;
96 unsigned long cp_xmit_sg;
97 unsigned int cp_xmit_hdr_off;
98 unsigned int cp_xmit_data_off;
99 unsigned int cp_xmit_atomic_sent;
100 unsigned int cp_xmit_rdma_sent;
101 unsigned int cp_xmit_data_sent;
102
103 spinlock_t cp_lock; /* protect msg queues */
104 u64 cp_next_tx_seq;
105 struct list_head cp_send_queue;
106 struct list_head cp_retrans;
107
108 u64 cp_next_rx_seq;
109
110 void *cp_transport_data;
111
112 atomic_t cp_state;
113 unsigned long cp_send_gen;
114 unsigned long cp_flags;
115 unsigned long cp_reconnect_jiffies;
116 struct delayed_work cp_send_w;
117 struct delayed_work cp_recv_w;
118 struct delayed_work cp_conn_w;
119 struct work_struct cp_down_w;
120 struct mutex cp_cm_lock; /* protect cp_state & cm */
121 wait_queue_head_t cp_waitq;
122
123 unsigned int cp_unacked_packets;
124 unsigned int cp_unacked_bytes;
125 unsigned int cp_outgoing:1,
126 cp_pad_to_32:31;
127 unsigned int cp_index;
128};
129
130/* One rds_connection per RDS address pair */
131struct rds_connection {
132 struct hlist_node c_hash_node;
133 __be32 c_laddr;
134 __be32 c_faddr;
135 unsigned int c_loopback:1,
136 c_ping_triggered:1,
137 c_pad_to_32:30;
138 int c_npaths;
139 struct rds_connection *c_passive;
140 struct rds_transport *c_trans;
141
142 struct rds_cong_map *c_lcong;
143 struct rds_cong_map *c_fcong;
144
145 /* Protocol version */
146 unsigned int c_version;
147 possible_net_t c_net;
148
149 struct list_head c_map_item;
150 unsigned long c_map_queued;
151
152 struct rds_conn_path c_path[RDS_MPATH_WORKERS];
153 wait_queue_head_t c_hs_waitq; /* handshake waitq */
154
155 u32 c_my_gen_num;
156 u32 c_peer_gen_num;
157};
158
159static inline
160struct net *rds_conn_net(struct rds_connection *conn)
161{
162 return read_pnet(&conn->c_net);
163}
164
165static inline
166void rds_conn_net_set(struct rds_connection *conn, struct net *net)
167{
168 write_pnet(&conn->c_net, net);
169}
170
171#define RDS_FLAG_CONG_BITMAP 0x01
172#define RDS_FLAG_ACK_REQUIRED 0x02
173#define RDS_FLAG_RETRANSMITTED 0x04
174#define RDS_MAX_ADV_CREDIT 255
175
176/* RDS_FLAG_PROBE_PORT is the reserved sport used for sending a ping
177 * probe to exchange control information before establishing a connection.
178 * Currently the control information that is exchanged is the number of
179 * supported paths. If the peer is a legacy (older kernel revision) peer,
180 * it would return a pong message without additional control information
181 * that would then alert the sender that the peer was an older rev.
182 */
183#define RDS_FLAG_PROBE_PORT 1
184#define RDS_HS_PROBE(sport, dport) \
185 ((sport == RDS_FLAG_PROBE_PORT && dport == 0) || \
186 (sport == 0 && dport == RDS_FLAG_PROBE_PORT))
187/*
188 * Maximum space available for extension headers.
189 */
190#define RDS_HEADER_EXT_SPACE 16
191
192struct rds_header {
193 __be64 h_sequence;
194 __be64 h_ack;
195 __be32 h_len;
196 __be16 h_sport;
197 __be16 h_dport;
198 u8 h_flags;
199 u8 h_credit;
200 u8 h_padding[4];
201 __sum16 h_csum;
202
203 u8 h_exthdr[RDS_HEADER_EXT_SPACE];
204};
205
206/*
207 * Reserved - indicates end of extensions
208 */
209#define RDS_EXTHDR_NONE 0
210
211/*
212 * This extension header is included in the very
213 * first message that is sent on a new connection,
214 * and identifies the protocol level. This will help
215 * rolling updates if a future change requires breaking
216 * the protocol.
217 * NB: This is no longer true for IB, where we do a version
218 * negotiation during the connection setup phase (protocol
219 * version information is included in the RDMA CM private data).
220 */
221#define RDS_EXTHDR_VERSION 1
222struct rds_ext_header_version {
223 __be32 h_version;
224};
225
226/*
227 * This extension header is included in the RDS message
228 * chasing an RDMA operation.
229 */
230#define RDS_EXTHDR_RDMA 2
231struct rds_ext_header_rdma {
232 __be32 h_rdma_rkey;
233};
234
235/*
236 * This extension header tells the peer about the
237 * destination <R_Key,offset> of the requested RDMA
238 * operation.
239 */
240#define RDS_EXTHDR_RDMA_DEST 3
241struct rds_ext_header_rdma_dest {
242 __be32 h_rdma_rkey;
243 __be32 h_rdma_offset;
244};
245
246/* Extension header announcing number of paths.
247 * Implicit length = 2 bytes.
248 */
249#define RDS_EXTHDR_NPATHS 5
250#define RDS_EXTHDR_GEN_NUM 6
251
252#define __RDS_EXTHDR_MAX 16 /* for now */
253
254struct rds_incoming {
255 atomic_t i_refcount;
256 struct list_head i_item;
257 struct rds_connection *i_conn;
258 struct rds_conn_path *i_conn_path;
259 struct rds_header i_hdr;
260 unsigned long i_rx_jiffies;
261 __be32 i_saddr;
262
263 rds_rdma_cookie_t i_rdma_cookie;
264 struct timeval i_rx_tstamp;
265};
266
267struct rds_mr {
268 struct rb_node r_rb_node;
269 atomic_t r_refcount;
270 u32 r_key;
271
272 /* A copy of the creation flags */
273 unsigned int r_use_once:1;
274 unsigned int r_invalidate:1;
275 unsigned int r_write:1;
276
277 /* This is for RDS_MR_DEAD.
278 * It would be nice & consistent to make this part of the above
279 * bit field here, but we need to use test_and_set_bit.
280 */
281 unsigned long r_state;
282 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
283 struct rds_transport *r_trans;
284 void *r_trans_private;
285};
286
287/* Flags for mr->r_state */
288#define RDS_MR_DEAD 0
289
290static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
291{
292 return r_key | (((u64) offset) << 32);
293}
294
295static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
296{
297 return cookie;
298}
299
300static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
301{
302 return cookie >> 32;
303}
304
305/* atomic operation types */
306#define RDS_ATOMIC_TYPE_CSWP 0
307#define RDS_ATOMIC_TYPE_FADD 1
308
309/*
310 * m_sock_item and m_conn_item are on lists that are serialized under
311 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
312 * the message will not be put back on the retransmit list after being sent.
313 * messages that are canceled while being sent rely on this.
314 *
315 * m_inc is used by loopback so that it can pass an incoming message straight
316 * back up into the rx path. It embeds a wire header which is also used by
317 * the send path, which is kind of awkward.
318 *
319 * m_sock_item indicates the message's presence on a socket's send or receive
320 * queue. m_rs will point to that socket.
321 *
322 * m_daddr is used by cancellation to prune messages to a given destination.
323 *
324 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
325 * nesting. As paths iterate over messages on a sock, or conn, they must
326 * also lock the conn, or sock, to remove the message from those lists too.
327 * Testing the flag to determine if the message is still on the lists lets
328 * us avoid testing the list_head directly. That means each path can use
329 * the message's list_head to keep it on a local list while juggling locks
330 * without confusing the other path.
331 *
332 * m_ack_seq is an optional field set by transports who need a different
333 * sequence number range to invalidate. They can use this in a callback
334 * that they pass to rds_send_drop_acked() to see if each message has been
335 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
336 * had ack_seq set yet.
337 */
338#define RDS_MSG_ON_SOCK 1
339#define RDS_MSG_ON_CONN 2
340#define RDS_MSG_HAS_ACK_SEQ 3
341#define RDS_MSG_ACK_REQUIRED 4
342#define RDS_MSG_RETRANSMITTED 5
343#define RDS_MSG_MAPPED 6
344#define RDS_MSG_PAGEVEC 7
345#define RDS_MSG_FLUSH 8
346
347struct rds_message {
348 atomic_t m_refcount;
349 struct list_head m_sock_item;
350 struct list_head m_conn_item;
351 struct rds_incoming m_inc;
352 u64 m_ack_seq;
353 __be32 m_daddr;
354 unsigned long m_flags;
355
356 /* Never access m_rs without holding m_rs_lock.
357 * Lock nesting is
358 * rm->m_rs_lock
359 * -> rs->rs_lock
360 */
361 spinlock_t m_rs_lock;
362 wait_queue_head_t m_flush_wait;
363
364 struct rds_sock *m_rs;
365
366 /* cookie to send to remote, in rds header */
367 rds_rdma_cookie_t m_rdma_cookie;
368
369 unsigned int m_used_sgs;
370 unsigned int m_total_sgs;
371
372 void *m_final_op;
373
374 struct {
375 struct rm_atomic_op {
376 int op_type;
377 union {
378 struct {
379 uint64_t compare;
380 uint64_t swap;
381 uint64_t compare_mask;
382 uint64_t swap_mask;
383 } op_m_cswp;
384 struct {
385 uint64_t add;
386 uint64_t nocarry_mask;
387 } op_m_fadd;
388 };
389
390 u32 op_rkey;
391 u64 op_remote_addr;
392 unsigned int op_notify:1;
393 unsigned int op_recverr:1;
394 unsigned int op_mapped:1;
395 unsigned int op_silent:1;
396 unsigned int op_active:1;
397 struct scatterlist *op_sg;
398 struct rds_notifier *op_notifier;
399
400 struct rds_mr *op_rdma_mr;
401 } atomic;
402 struct rm_rdma_op {
403 u32 op_rkey;
404 u64 op_remote_addr;
405 unsigned int op_write:1;
406 unsigned int op_fence:1;
407 unsigned int op_notify:1;
408 unsigned int op_recverr:1;
409 unsigned int op_mapped:1;
410 unsigned int op_silent:1;
411 unsigned int op_active:1;
412 unsigned int op_bytes;
413 unsigned int op_nents;
414 unsigned int op_count;
415 struct scatterlist *op_sg;
416 struct rds_notifier *op_notifier;
417
418 struct rds_mr *op_rdma_mr;
419 } rdma;
420 struct rm_data_op {
421 unsigned int op_active:1;
422 unsigned int op_nents;
423 unsigned int op_count;
424 unsigned int op_dmasg;
425 unsigned int op_dmaoff;
426 struct scatterlist *op_sg;
427 } data;
428 };
429};
430
431/*
432 * The RDS notifier is used (optionally) to tell the application about
433 * completed RDMA operations. Rather than keeping the whole rds message
434 * around on the queue, we allocate a small notifier that is put on the
435 * socket's notifier_list. Notifications are delivered to the application
436 * through control messages.
437 */
438struct rds_notifier {
439 struct list_head n_list;
440 uint64_t n_user_token;
441 int n_status;
442};
443
444/**
445 * struct rds_transport - transport specific behavioural hooks
446 *
447 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
448 * part of a message. The caller serializes on the send_sem so this
449 * doesn't need to be reentrant for a given conn. The header must be
450 * sent before the data payload. .xmit must be prepared to send a
451 * message with no data payload. .xmit should return the number of
452 * bytes that were sent down the connection, including header bytes.
453 * Returning 0 tells the caller that it doesn't need to perform any
454 * additional work now. This is usually the case when the transport has
455 * filled the sending queue for its connection and will handle
456 * triggering the rds thread to continue the send when space becomes
457 * available. Returning -EAGAIN tells the caller to retry the send
458 * immediately. Returning -ENOMEM tells the caller to retry the send at
459 * some point in the future.
460 *
461 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
462 * it returns the connection can not call rds_recv_incoming().
463 * This will only be called once after conn_connect returns
464 * non-zero success and will The caller serializes this with
465 * the send and connecting paths (xmit_* and conn_*). The
466 * transport is responsible for other serialization, including
467 * rds_recv_incoming(). This is called in process context but
468 * should try hard not to block.
469 */
470
471struct rds_transport {
472 char t_name[TRANSNAMSIZ];
473 struct list_head t_item;
474 struct module *t_owner;
475 unsigned int t_prefer_loopback:1,
476 t_mp_capable:1;
477 unsigned int t_type;
478
479 int (*laddr_check)(struct net *net, __be32 addr);
480 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
481 void (*conn_free)(void *data);
482 int (*conn_path_connect)(struct rds_conn_path *cp);
483 void (*conn_path_shutdown)(struct rds_conn_path *conn);
484 void (*xmit_path_prepare)(struct rds_conn_path *cp);
485 void (*xmit_path_complete)(struct rds_conn_path *cp);
486 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
487 unsigned int hdr_off, unsigned int sg, unsigned int off);
488 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
489 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
490 int (*recv_path)(struct rds_conn_path *cp);
491 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
492 void (*inc_free)(struct rds_incoming *inc);
493
494 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
495 struct rdma_cm_event *event);
496 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
497 void (*cm_connect_complete)(struct rds_connection *conn,
498 struct rdma_cm_event *event);
499
500 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
501 unsigned int avail);
502 void (*exit)(void);
503 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
504 struct rds_sock *rs, u32 *key_ret);
505 void (*sync_mr)(void *trans_private, int direction);
506 void (*free_mr)(void *trans_private, int invalidate);
507 void (*flush_mrs)(void);
508};
509
510struct rds_sock {
511 struct sock rs_sk;
512
513 u64 rs_user_addr;
514 u64 rs_user_bytes;
515
516 /*
517 * bound_addr used for both incoming and outgoing, no INADDR_ANY
518 * support.
519 */
520 struct rhash_head rs_bound_node;
521 u64 rs_bound_key;
522 __be32 rs_bound_addr;
523 __be32 rs_conn_addr;
524 __be16 rs_bound_port;
525 __be16 rs_conn_port;
526 struct rds_transport *rs_transport;
527
528 /*
529 * rds_sendmsg caches the conn it used the last time around.
530 * This helps avoid costly lookups.
531 */
532 struct rds_connection *rs_conn;
533
534 /* flag indicating we were congested or not */
535 int rs_congested;
536 /* seen congestion (ENOBUFS) when sending? */
537 int rs_seen_congestion;
538
539 /* rs_lock protects all these adjacent members before the newline */
540 spinlock_t rs_lock;
541 struct list_head rs_send_queue;
542 u32 rs_snd_bytes;
543 int rs_rcv_bytes;
544 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
545
546 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
547 * to decide whether the application should be woken up.
548 * If not set, we use rs_cong_track to find out whether a cong map
549 * update arrived.
550 */
551 uint64_t rs_cong_mask;
552 uint64_t rs_cong_notify;
553 struct list_head rs_cong_list;
554 unsigned long rs_cong_track;
555
556 /*
557 * rs_recv_lock protects the receive queue, and is
558 * used to serialize with rds_release.
559 */
560 rwlock_t rs_recv_lock;
561 struct list_head rs_recv_queue;
562
563 /* just for stats reporting */
564 struct list_head rs_item;
565
566 /* these have their own lock */
567 spinlock_t rs_rdma_lock;
568 struct rb_root rs_rdma_keys;
569
570 /* Socket options - in case there will be more */
571 unsigned char rs_recverr,
572 rs_cong_monitor;
573 u32 rs_hash_initval;
574};
575
576static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
577{
578 return container_of(sk, struct rds_sock, rs_sk);
579}
580static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
581{
582 return &rs->rs_sk;
583}
584
585/*
586 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
587 * to account for overhead. We don't account for overhead, we just apply
588 * the number of payload bytes to the specified value.
589 */
590static inline int rds_sk_sndbuf(struct rds_sock *rs)
591{
592 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
593}
594static inline int rds_sk_rcvbuf(struct rds_sock *rs)
595{
596 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
597}
598
599struct rds_statistics {
600 uint64_t s_conn_reset;
601 uint64_t s_recv_drop_bad_checksum;
602 uint64_t s_recv_drop_old_seq;
603 uint64_t s_recv_drop_no_sock;
604 uint64_t s_recv_drop_dead_sock;
605 uint64_t s_recv_deliver_raced;
606 uint64_t s_recv_delivered;
607 uint64_t s_recv_queued;
608 uint64_t s_recv_immediate_retry;
609 uint64_t s_recv_delayed_retry;
610 uint64_t s_recv_ack_required;
611 uint64_t s_recv_rdma_bytes;
612 uint64_t s_recv_ping;
613 uint64_t s_send_queue_empty;
614 uint64_t s_send_queue_full;
615 uint64_t s_send_lock_contention;
616 uint64_t s_send_lock_queue_raced;
617 uint64_t s_send_immediate_retry;
618 uint64_t s_send_delayed_retry;
619 uint64_t s_send_drop_acked;
620 uint64_t s_send_ack_required;
621 uint64_t s_send_queued;
622 uint64_t s_send_rdma;
623 uint64_t s_send_rdma_bytes;
624 uint64_t s_send_pong;
625 uint64_t s_page_remainder_hit;
626 uint64_t s_page_remainder_miss;
627 uint64_t s_copy_to_user;
628 uint64_t s_copy_from_user;
629 uint64_t s_cong_update_queued;
630 uint64_t s_cong_update_received;
631 uint64_t s_cong_send_error;
632 uint64_t s_cong_send_blocked;
633};
634
635/* af_rds.c */
636void rds_sock_addref(struct rds_sock *rs);
637void rds_sock_put(struct rds_sock *rs);
638void rds_wake_sk_sleep(struct rds_sock *rs);
639static inline void __rds_wake_sk_sleep(struct sock *sk)
640{
641 wait_queue_head_t *waitq = sk_sleep(sk);
642
643 if (!sock_flag(sk, SOCK_DEAD) && waitq)
644 wake_up(waitq);
645}
646extern wait_queue_head_t rds_poll_waitq;
647
648
649/* bind.c */
650int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
651void rds_remove_bound(struct rds_sock *rs);
652struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
653int rds_bind_lock_init(void);
654void rds_bind_lock_destroy(void);
655
656/* cong.c */
657int rds_cong_get_maps(struct rds_connection *conn);
658void rds_cong_add_conn(struct rds_connection *conn);
659void rds_cong_remove_conn(struct rds_connection *conn);
660void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
661void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
662int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
663void rds_cong_queue_updates(struct rds_cong_map *map);
664void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
665int rds_cong_updated_since(unsigned long *recent);
666void rds_cong_add_socket(struct rds_sock *);
667void rds_cong_remove_socket(struct rds_sock *);
668void rds_cong_exit(void);
669struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
670
671/* conn.c */
672extern u32 rds_gen_num;
673int rds_conn_init(void);
674void rds_conn_exit(void);
675struct rds_connection *rds_conn_create(struct net *net,
676 __be32 laddr, __be32 faddr,
677 struct rds_transport *trans, gfp_t gfp);
678struct rds_connection *rds_conn_create_outgoing(struct net *net,
679 __be32 laddr, __be32 faddr,
680 struct rds_transport *trans, gfp_t gfp);
681void rds_conn_shutdown(struct rds_conn_path *cpath);
682void rds_conn_destroy(struct rds_connection *conn);
683void rds_conn_drop(struct rds_connection *conn);
684void rds_conn_path_drop(struct rds_conn_path *cpath);
685void rds_conn_connect_if_down(struct rds_connection *conn);
686void rds_conn_path_connect_if_down(struct rds_conn_path *cp);
687void rds_for_each_conn_info(struct socket *sock, unsigned int len,
688 struct rds_info_iterator *iter,
689 struct rds_info_lengths *lens,
690 int (*visitor)(struct rds_connection *, void *),
691 size_t item_len);
692
693__printf(2, 3)
694void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...);
695#define rds_conn_path_error(cp, fmt...) \
696 __rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt)
697
698static inline int
699rds_conn_path_transition(struct rds_conn_path *cp, int old, int new)
700{
701 return atomic_cmpxchg(&cp->cp_state, old, new) == old;
702}
703
704static inline int
705rds_conn_transition(struct rds_connection *conn, int old, int new)
706{
707 WARN_ON(conn->c_trans->t_mp_capable);
708 return rds_conn_path_transition(&conn->c_path[0], old, new);
709}
710
711static inline int
712rds_conn_path_state(struct rds_conn_path *cp)
713{
714 return atomic_read(&cp->cp_state);
715}
716
717static inline int
718rds_conn_state(struct rds_connection *conn)
719{
720 WARN_ON(conn->c_trans->t_mp_capable);
721 return rds_conn_path_state(&conn->c_path[0]);
722}
723
724static inline int
725rds_conn_path_up(struct rds_conn_path *cp)
726{
727 return atomic_read(&cp->cp_state) == RDS_CONN_UP;
728}
729
730static inline int
731rds_conn_up(struct rds_connection *conn)
732{
733 WARN_ON(conn->c_trans->t_mp_capable);
734 return rds_conn_path_up(&conn->c_path[0]);
735}
736
737static inline int
738rds_conn_path_connecting(struct rds_conn_path *cp)
739{
740 return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING;
741}
742
743static inline int
744rds_conn_connecting(struct rds_connection *conn)
745{
746 WARN_ON(conn->c_trans->t_mp_capable);
747 return rds_conn_path_connecting(&conn->c_path[0]);
748}
749
750/* message.c */
751struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
752struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
753int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from);
754struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
755void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
756 __be16 dport, u64 seq);
757int rds_message_add_extension(struct rds_header *hdr,
758 unsigned int type, const void *data, unsigned int len);
759int rds_message_next_extension(struct rds_header *hdr,
760 unsigned int *pos, void *buf, unsigned int *buflen);
761int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
762int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
763void rds_message_inc_free(struct rds_incoming *inc);
764void rds_message_addref(struct rds_message *rm);
765void rds_message_put(struct rds_message *rm);
766void rds_message_wait(struct rds_message *rm);
767void rds_message_unmapped(struct rds_message *rm);
768
769static inline void rds_message_make_checksum(struct rds_header *hdr)
770{
771 hdr->h_csum = 0;
772 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
773}
774
775static inline int rds_message_verify_checksum(const struct rds_header *hdr)
776{
777 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
778}
779
780
781/* page.c */
782int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
783 gfp_t gfp);
784int rds_page_copy_user(struct page *page, unsigned long offset,
785 void __user *ptr, unsigned long bytes,
786 int to_user);
787#define rds_page_copy_to_user(page, offset, ptr, bytes) \
788 rds_page_copy_user(page, offset, ptr, bytes, 1)
789#define rds_page_copy_from_user(page, offset, ptr, bytes) \
790 rds_page_copy_user(page, offset, ptr, bytes, 0)
791void rds_page_exit(void);
792
793/* recv.c */
794void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
795 __be32 saddr);
796void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn,
797 __be32 saddr);
798void rds_inc_put(struct rds_incoming *inc);
799void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
800 struct rds_incoming *inc, gfp_t gfp);
801int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
802 int msg_flags);
803void rds_clear_recv_queue(struct rds_sock *rs);
804int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
805void rds_inc_info_copy(struct rds_incoming *inc,
806 struct rds_info_iterator *iter,
807 __be32 saddr, __be32 daddr, int flip);
808
809/* send.c */
810int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
811void rds_send_path_reset(struct rds_conn_path *conn);
812int rds_send_xmit(struct rds_conn_path *cp);
813struct sockaddr_in;
814void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
815typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
816void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
817 is_acked_func is_acked);
818void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
819 is_acked_func is_acked);
820int rds_send_pong(struct rds_conn_path *cp, __be16 dport);
821
822/* rdma.c */
823void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
824int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
825int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
826int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
827void rds_rdma_drop_keys(struct rds_sock *rs);
828int rds_rdma_extra_size(struct rds_rdma_args *args);
829int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
830 struct cmsghdr *cmsg);
831int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
832 struct cmsghdr *cmsg);
833int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
834 struct cmsghdr *cmsg);
835int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
836 struct cmsghdr *cmsg);
837void rds_rdma_free_op(struct rm_rdma_op *ro);
838void rds_atomic_free_op(struct rm_atomic_op *ao);
839void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
840void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
841int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
842 struct cmsghdr *cmsg);
843
844void __rds_put_mr_final(struct rds_mr *mr);
845static inline void rds_mr_put(struct rds_mr *mr)
846{
847 if (atomic_dec_and_test(&mr->r_refcount))
848 __rds_put_mr_final(mr);
849}
850
851/* stats.c */
852DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
853#define rds_stats_inc_which(which, member) do { \
854 per_cpu(which, get_cpu()).member++; \
855 put_cpu(); \
856} while (0)
857#define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
858#define rds_stats_add_which(which, member, count) do { \
859 per_cpu(which, get_cpu()).member += count; \
860 put_cpu(); \
861} while (0)
862#define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
863int rds_stats_init(void);
864void rds_stats_exit(void);
865void rds_stats_info_copy(struct rds_info_iterator *iter,
866 uint64_t *values, const char *const *names,
867 size_t nr);
868
869/* sysctl.c */
870int rds_sysctl_init(void);
871void rds_sysctl_exit(void);
872extern unsigned long rds_sysctl_sndbuf_min;
873extern unsigned long rds_sysctl_sndbuf_default;
874extern unsigned long rds_sysctl_sndbuf_max;
875extern unsigned long rds_sysctl_reconnect_min_jiffies;
876extern unsigned long rds_sysctl_reconnect_max_jiffies;
877extern unsigned int rds_sysctl_max_unacked_packets;
878extern unsigned int rds_sysctl_max_unacked_bytes;
879extern unsigned int rds_sysctl_ping_enable;
880extern unsigned long rds_sysctl_trace_flags;
881extern unsigned int rds_sysctl_trace_level;
882
883/* threads.c */
884int rds_threads_init(void);
885void rds_threads_exit(void);
886extern struct workqueue_struct *rds_wq;
887void rds_queue_reconnect(struct rds_conn_path *cp);
888void rds_connect_worker(struct work_struct *);
889void rds_shutdown_worker(struct work_struct *);
890void rds_send_worker(struct work_struct *);
891void rds_recv_worker(struct work_struct *);
892void rds_connect_path_complete(struct rds_conn_path *conn, int curr);
893void rds_connect_complete(struct rds_connection *conn);
894
895/* transport.c */
896int rds_trans_register(struct rds_transport *trans);
897void rds_trans_unregister(struct rds_transport *trans);
898struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr);
899void rds_trans_put(struct rds_transport *trans);
900unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
901 unsigned int avail);
902struct rds_transport *rds_trans_get(int t_type);
903int rds_trans_init(void);
904void rds_trans_exit(void);
905
906#endif