blob: 7c4adbe8c2848ad4733155fb3ce18b3492ef0aa4 [file] [log] [blame]
Andy Grover39de8282009-02-24 15:30:19 +00001#ifndef _RDS_RDS_H
2#define _RDS_RDS_H
3
4#include <net/sock.h>
5#include <linux/scatterlist.h>
6#include <linux/highmem.h>
7#include <rdma/rdma_cm.h>
8#include <linux/mutex.h>
9#include <linux/rds.h>
10
11#include "info.h"
12
13/*
14 * RDS Network protocol version
15 */
16#define RDS_PROTOCOL_3_0 0x0300
17#define RDS_PROTOCOL_3_1 0x0301
18#define RDS_PROTOCOL_VERSION RDS_PROTOCOL_3_1
19#define RDS_PROTOCOL_MAJOR(v) ((v) >> 8)
20#define RDS_PROTOCOL_MINOR(v) ((v) & 255)
21#define RDS_PROTOCOL(maj, min) (((maj) << 8) | min)
22
23/*
24 * XXX randomly chosen, but at least seems to be unused:
25 * # 18464-18768 Unassigned
26 * We should do better. We want a reserved port to discourage unpriv'ed
27 * userspace from listening.
28 */
29#define RDS_PORT 18634
30
Andy Grover8cbd9602009-04-01 08:20:20 +000031#ifdef ATOMIC64_INIT
32#define KERNEL_HAS_ATOMIC64
33#endif
34
Andy Grover39de8282009-02-24 15:30:19 +000035#ifdef DEBUG
36#define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
37#else
38/* sigh, pr_debug() causes unused variable warnings */
39static inline void __attribute__ ((format (printf, 1, 2)))
40rdsdebug(char *fmt, ...)
41{
42}
43#endif
44
45/* XXX is there one of these somewhere? */
46#define ceil(x, y) \
47 ({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; })
48
49#define RDS_FRAG_SHIFT 12
50#define RDS_FRAG_SIZE ((unsigned int)(1 << RDS_FRAG_SHIFT))
51
52#define RDS_CONG_MAP_BYTES (65536 / 8)
53#define RDS_CONG_MAP_LONGS (RDS_CONG_MAP_BYTES / sizeof(unsigned long))
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
84struct rds_connection {
85 struct hlist_node c_hash_node;
86 __be32 c_laddr;
87 __be32 c_faddr;
88 unsigned int c_loopback:1;
89 struct rds_connection *c_passive;
90
91 struct rds_cong_map *c_lcong;
92 struct rds_cong_map *c_fcong;
93
94 struct mutex c_send_lock; /* protect send ring */
95 struct rds_message *c_xmit_rm;
96 unsigned long c_xmit_sg;
97 unsigned int c_xmit_hdr_off;
98 unsigned int c_xmit_data_off;
99 unsigned int c_xmit_rdma_sent;
100
101 spinlock_t c_lock; /* protect msg queues */
102 u64 c_next_tx_seq;
103 struct list_head c_send_queue;
104 struct list_head c_retrans;
105
106 u64 c_next_rx_seq;
107
108 struct rds_transport *c_trans;
109 void *c_transport_data;
110
111 atomic_t c_state;
112 unsigned long c_flags;
113 unsigned long c_reconnect_jiffies;
114 struct delayed_work c_send_w;
115 struct delayed_work c_recv_w;
116 struct delayed_work c_conn_w;
117 struct work_struct c_down_w;
118 struct mutex c_cm_lock; /* protect conn state & cm */
119
120 struct list_head c_map_item;
121 unsigned long c_map_queued;
122 unsigned long c_map_offset;
123 unsigned long c_map_bytes;
124
125 unsigned int c_unacked_packets;
126 unsigned int c_unacked_bytes;
127
128 /* Protocol version */
129 unsigned int c_version;
130};
131
132#define RDS_FLAG_CONG_BITMAP 0x01
133#define RDS_FLAG_ACK_REQUIRED 0x02
134#define RDS_FLAG_RETRANSMITTED 0x04
Steve Wise7b70d032009-04-09 14:09:39 +0000135#define RDS_MAX_ADV_CREDIT 255
Andy Grover39de8282009-02-24 15:30:19 +0000136
137/*
138 * Maximum space available for extension headers.
139 */
140#define RDS_HEADER_EXT_SPACE 16
141
142struct rds_header {
143 __be64 h_sequence;
144 __be64 h_ack;
145 __be32 h_len;
146 __be16 h_sport;
147 __be16 h_dport;
148 u8 h_flags;
149 u8 h_credit;
150 u8 h_padding[4];
151 __sum16 h_csum;
152
153 u8 h_exthdr[RDS_HEADER_EXT_SPACE];
154};
155
156/*
157 * Reserved - indicates end of extensions
158 */
159#define RDS_EXTHDR_NONE 0
160
161/*
162 * This extension header is included in the very
163 * first message that is sent on a new connection,
164 * and identifies the protocol level. This will help
165 * rolling updates if a future change requires breaking
166 * the protocol.
167 * NB: This is no longer true for IB, where we do a version
168 * negotiation during the connection setup phase (protocol
169 * version information is included in the RDMA CM private data).
170 */
171#define RDS_EXTHDR_VERSION 1
172struct rds_ext_header_version {
173 __be32 h_version;
174};
175
176/*
177 * This extension header is included in the RDS message
178 * chasing an RDMA operation.
179 */
180#define RDS_EXTHDR_RDMA 2
181struct rds_ext_header_rdma {
182 __be32 h_rdma_rkey;
183};
184
185/*
186 * This extension header tells the peer about the
187 * destination <R_Key,offset> of the requested RDMA
188 * operation.
189 */
190#define RDS_EXTHDR_RDMA_DEST 3
191struct rds_ext_header_rdma_dest {
192 __be32 h_rdma_rkey;
193 __be32 h_rdma_offset;
194};
195
196#define __RDS_EXTHDR_MAX 16 /* for now */
197
198struct rds_incoming {
199 atomic_t i_refcount;
200 struct list_head i_item;
201 struct rds_connection *i_conn;
202 struct rds_header i_hdr;
203 unsigned long i_rx_jiffies;
204 __be32 i_saddr;
205
206 rds_rdma_cookie_t i_rdma_cookie;
207};
208
Andy Grover21f79af2010-01-12 12:57:27 -0800209struct rds_mr {
210 struct rb_node r_rb_node;
211 atomic_t r_refcount;
212 u32 r_key;
213
214 /* A copy of the creation flags */
215 unsigned int r_use_once:1;
216 unsigned int r_invalidate:1;
217 unsigned int r_write:1;
218
219 /* This is for RDS_MR_DEAD.
220 * It would be nice & consistent to make this part of the above
221 * bit field here, but we need to use test_and_set_bit.
222 */
223 unsigned long r_state;
224 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
225 struct rds_transport *r_trans;
226 void *r_trans_private;
227};
228
229/* Flags for mr->r_state */
230#define RDS_MR_DEAD 0
231
232struct rds_rdma_op {
233 u32 r_key;
234 u64 r_remote_addr;
235 unsigned int r_write:1;
236 unsigned int r_fence:1;
237 unsigned int r_notify:1;
238 unsigned int r_recverr:1;
239 unsigned int r_mapped:1;
240 unsigned int r_active:1;
241 struct rds_notifier *r_notifier;
242 unsigned int r_bytes;
243 unsigned int r_nents;
244 unsigned int r_count;
245 struct scatterlist *r_sg;
246};
247
248static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
249{
250 return r_key | (((u64) offset) << 32);
251}
252
253static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
254{
255 return cookie;
256}
257
258static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
259{
260 return cookie >> 32;
261}
262
Andy Grover39de8282009-02-24 15:30:19 +0000263/*
264 * m_sock_item and m_conn_item are on lists that are serialized under
265 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
266 * the message will not be put back on the retransmit list after being sent.
267 * messages that are canceled while being sent rely on this.
268 *
269 * m_inc is used by loopback so that it can pass an incoming message straight
270 * back up into the rx path. It embeds a wire header which is also used by
271 * the send path, which is kind of awkward.
272 *
273 * m_sock_item indicates the message's presence on a socket's send or receive
274 * queue. m_rs will point to that socket.
275 *
276 * m_daddr is used by cancellation to prune messages to a given destination.
277 *
278 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
279 * nesting. As paths iterate over messages on a sock, or conn, they must
280 * also lock the conn, or sock, to remove the message from those lists too.
281 * Testing the flag to determine if the message is still on the lists lets
282 * us avoid testing the list_head directly. That means each path can use
283 * the message's list_head to keep it on a local list while juggling locks
284 * without confusing the other path.
285 *
286 * m_ack_seq is an optional field set by transports who need a different
287 * sequence number range to invalidate. They can use this in a callback
288 * that they pass to rds_send_drop_acked() to see if each message has been
289 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
290 * had ack_seq set yet.
291 */
292#define RDS_MSG_ON_SOCK 1
293#define RDS_MSG_ON_CONN 2
294#define RDS_MSG_HAS_ACK_SEQ 3
295#define RDS_MSG_ACK_REQUIRED 4
296#define RDS_MSG_RETRANSMITTED 5
297#define RDS_MSG_MAPPED 6
298#define RDS_MSG_PAGEVEC 7
299
300struct rds_message {
301 atomic_t m_refcount;
302 struct list_head m_sock_item;
303 struct list_head m_conn_item;
304 struct rds_incoming m_inc;
305 u64 m_ack_seq;
306 __be32 m_daddr;
307 unsigned long m_flags;
308
309 /* Never access m_rs without holding m_rs_lock.
310 * Lock nesting is
311 * rm->m_rs_lock
312 * -> rs->rs_lock
313 */
314 spinlock_t m_rs_lock;
315 struct rds_sock *m_rs;
Andy Grover39de8282009-02-24 15:30:19 +0000316 rds_rdma_cookie_t m_rdma_cookie;
Andy Grovere7791372010-01-12 12:15:02 -0800317 struct {
318 struct {
319 struct rds_rdma_op *m_rdma_op;
320 struct rds_mr *m_rdma_mr;
321 } rdma;
322 struct {
323 unsigned int m_nents;
324 unsigned int m_count;
Andy Groverfc445082010-01-12 12:56:06 -0800325 struct scatterlist *m_sg;
Andy Grovere7791372010-01-12 12:15:02 -0800326 } data;
327 };
Andy Groverfc445082010-01-12 12:56:06 -0800328 unsigned int m_used_sgs;
329 unsigned int m_total_sgs;
Andy Grover39de8282009-02-24 15:30:19 +0000330};
331
332/*
333 * The RDS notifier is used (optionally) to tell the application about
334 * completed RDMA operations. Rather than keeping the whole rds message
335 * around on the queue, we allocate a small notifier that is put on the
336 * socket's notifier_list. Notifications are delivered to the application
337 * through control messages.
338 */
339struct rds_notifier {
340 struct list_head n_list;
341 uint64_t n_user_token;
342 int n_status;
343};
344
345/**
346 * struct rds_transport - transport specific behavioural hooks
347 *
348 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
349 * part of a message. The caller serializes on the send_sem so this
350 * doesn't need to be reentrant for a given conn. The header must be
351 * sent before the data payload. .xmit must be prepared to send a
352 * message with no data payload. .xmit should return the number of
353 * bytes that were sent down the connection, including header bytes.
354 * Returning 0 tells the caller that it doesn't need to perform any
355 * additional work now. This is usually the case when the transport has
356 * filled the sending queue for its connection and will handle
357 * triggering the rds thread to continue the send when space becomes
358 * available. Returning -EAGAIN tells the caller to retry the send
359 * immediately. Returning -ENOMEM tells the caller to retry the send at
360 * some point in the future.
361 *
362 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
363 * it returns the connection can not call rds_recv_incoming().
364 * This will only be called once after conn_connect returns
365 * non-zero success and will The caller serializes this with
366 * the send and connecting paths (xmit_* and conn_*). The
367 * transport is responsible for other serialization, including
368 * rds_recv_incoming(). This is called in process context but
369 * should try hard not to block.
370 *
371 * @xmit_cong_map: This asks the transport to send the local bitmap down the
372 * given connection. XXX get a better story about the bitmap
373 * flag and header.
374 */
375
Andy Grover335776b2009-08-21 12:28:34 +0000376#define RDS_TRANS_IB 0
377#define RDS_TRANS_IWARP 1
378#define RDS_TRANS_TCP 2
379#define RDS_TRANS_COUNT 3
380
Andy Grover39de8282009-02-24 15:30:19 +0000381struct rds_transport {
382 char t_name[TRANSNAMSIZ];
383 struct list_head t_item;
384 struct module *t_owner;
385 unsigned int t_prefer_loopback:1;
Andy Grover335776b2009-08-21 12:28:34 +0000386 unsigned int t_type;
Andy Grover39de8282009-02-24 15:30:19 +0000387
388 int (*laddr_check)(__be32 addr);
389 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
390 void (*conn_free)(void *data);
391 int (*conn_connect)(struct rds_connection *conn);
392 void (*conn_shutdown)(struct rds_connection *conn);
393 void (*xmit_prepare)(struct rds_connection *conn);
394 void (*xmit_complete)(struct rds_connection *conn);
395 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
396 unsigned int hdr_off, unsigned int sg, unsigned int off);
397 int (*xmit_cong_map)(struct rds_connection *conn,
398 struct rds_cong_map *map, unsigned long offset);
399 int (*xmit_rdma)(struct rds_connection *conn, struct rds_rdma_op *op);
400 int (*recv)(struct rds_connection *conn);
401 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iovec *iov,
402 size_t size);
403 void (*inc_purge)(struct rds_incoming *inc);
404 void (*inc_free)(struct rds_incoming *inc);
405
406 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
407 struct rdma_cm_event *event);
408 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
409 void (*cm_connect_complete)(struct rds_connection *conn,
410 struct rdma_cm_event *event);
411
412 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
413 unsigned int avail);
414 void (*exit)(void);
415 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
416 struct rds_sock *rs, u32 *key_ret);
417 void (*sync_mr)(void *trans_private, int direction);
418 void (*free_mr)(void *trans_private, int invalidate);
419 void (*flush_mrs)(void);
420};
421
422struct rds_sock {
423 struct sock rs_sk;
424
425 u64 rs_user_addr;
426 u64 rs_user_bytes;
427
428 /*
429 * bound_addr used for both incoming and outgoing, no INADDR_ANY
430 * support.
431 */
432 struct rb_node rs_bound_node;
433 __be32 rs_bound_addr;
434 __be32 rs_conn_addr;
435 __be16 rs_bound_port;
436 __be16 rs_conn_port;
437
438 /*
439 * This is only used to communicate the transport between bind and
440 * initiating connections. All other trans use is referenced through
441 * the connection.
442 */
443 struct rds_transport *rs_transport;
444
445 /*
446 * rds_sendmsg caches the conn it used the last time around.
447 * This helps avoid costly lookups.
448 */
449 struct rds_connection *rs_conn;
450
451 /* flag indicating we were congested or not */
452 int rs_congested;
Andy Groverb98ba522010-03-11 13:50:04 +0000453 /* seen congestion (ENOBUFS) when sending? */
454 int rs_seen_congestion;
Andy Grover39de8282009-02-24 15:30:19 +0000455
456 /* rs_lock protects all these adjacent members before the newline */
457 spinlock_t rs_lock;
458 struct list_head rs_send_queue;
459 u32 rs_snd_bytes;
460 int rs_rcv_bytes;
461 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
462
463 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
464 * to decide whether the application should be woken up.
465 * If not set, we use rs_cong_track to find out whether a cong map
466 * update arrived.
467 */
468 uint64_t rs_cong_mask;
469 uint64_t rs_cong_notify;
470 struct list_head rs_cong_list;
471 unsigned long rs_cong_track;
472
473 /*
474 * rs_recv_lock protects the receive queue, and is
475 * used to serialize with rds_release.
476 */
477 rwlock_t rs_recv_lock;
478 struct list_head rs_recv_queue;
479
480 /* just for stats reporting */
481 struct list_head rs_item;
482
483 /* these have their own lock */
484 spinlock_t rs_rdma_lock;
485 struct rb_root rs_rdma_keys;
486
487 /* Socket options - in case there will be more */
488 unsigned char rs_recverr,
489 rs_cong_monitor;
490};
491
492static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
493{
494 return container_of(sk, struct rds_sock, rs_sk);
495}
496static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
497{
498 return &rs->rs_sk;
499}
500
501/*
502 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
503 * to account for overhead. We don't account for overhead, we just apply
504 * the number of payload bytes to the specified value.
505 */
506static inline int rds_sk_sndbuf(struct rds_sock *rs)
507{
508 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
509}
510static inline int rds_sk_rcvbuf(struct rds_sock *rs)
511{
512 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
513}
514
515struct rds_statistics {
516 uint64_t s_conn_reset;
517 uint64_t s_recv_drop_bad_checksum;
518 uint64_t s_recv_drop_old_seq;
519 uint64_t s_recv_drop_no_sock;
520 uint64_t s_recv_drop_dead_sock;
521 uint64_t s_recv_deliver_raced;
522 uint64_t s_recv_delivered;
523 uint64_t s_recv_queued;
524 uint64_t s_recv_immediate_retry;
525 uint64_t s_recv_delayed_retry;
526 uint64_t s_recv_ack_required;
527 uint64_t s_recv_rdma_bytes;
528 uint64_t s_recv_ping;
529 uint64_t s_send_queue_empty;
530 uint64_t s_send_queue_full;
531 uint64_t s_send_sem_contention;
532 uint64_t s_send_sem_queue_raced;
533 uint64_t s_send_immediate_retry;
534 uint64_t s_send_delayed_retry;
535 uint64_t s_send_drop_acked;
536 uint64_t s_send_ack_required;
537 uint64_t s_send_queued;
538 uint64_t s_send_rdma;
539 uint64_t s_send_rdma_bytes;
540 uint64_t s_send_pong;
541 uint64_t s_page_remainder_hit;
542 uint64_t s_page_remainder_miss;
543 uint64_t s_copy_to_user;
544 uint64_t s_copy_from_user;
545 uint64_t s_cong_update_queued;
546 uint64_t s_cong_update_received;
547 uint64_t s_cong_send_error;
548 uint64_t s_cong_send_blocked;
549};
550
551/* af_rds.c */
552void rds_sock_addref(struct rds_sock *rs);
553void rds_sock_put(struct rds_sock *rs);
554void rds_wake_sk_sleep(struct rds_sock *rs);
555static inline void __rds_wake_sk_sleep(struct sock *sk)
556{
Eric Dumazetaa395142010-04-20 13:03:51 +0000557 wait_queue_head_t *waitq = sk_sleep(sk);
Andy Grover39de8282009-02-24 15:30:19 +0000558
559 if (!sock_flag(sk, SOCK_DEAD) && waitq)
560 wake_up(waitq);
561}
562extern wait_queue_head_t rds_poll_waitq;
563
564
565/* bind.c */
566int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
567void rds_remove_bound(struct rds_sock *rs);
568struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
569
570/* cong.c */
571int rds_cong_get_maps(struct rds_connection *conn);
572void rds_cong_add_conn(struct rds_connection *conn);
573void rds_cong_remove_conn(struct rds_connection *conn);
574void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
575void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
576int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
577void rds_cong_queue_updates(struct rds_cong_map *map);
578void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
579int rds_cong_updated_since(unsigned long *recent);
580void rds_cong_add_socket(struct rds_sock *);
581void rds_cong_remove_socket(struct rds_sock *);
582void rds_cong_exit(void);
583struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
584
585/* conn.c */
586int __init rds_conn_init(void);
587void rds_conn_exit(void);
588struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
589 struct rds_transport *trans, gfp_t gfp);
590struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
591 struct rds_transport *trans, gfp_t gfp);
Andy Grover2dc39352010-06-11 13:49:13 -0700592void rds_conn_shutdown(struct rds_connection *conn);
Andy Grover39de8282009-02-24 15:30:19 +0000593void rds_conn_destroy(struct rds_connection *conn);
594void rds_conn_reset(struct rds_connection *conn);
595void rds_conn_drop(struct rds_connection *conn);
596void rds_for_each_conn_info(struct socket *sock, unsigned int len,
597 struct rds_info_iterator *iter,
598 struct rds_info_lengths *lens,
599 int (*visitor)(struct rds_connection *, void *),
600 size_t item_len);
601void __rds_conn_error(struct rds_connection *conn, const char *, ...)
602 __attribute__ ((format (printf, 2, 3)));
603#define rds_conn_error(conn, fmt...) \
604 __rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
605
606static inline int
607rds_conn_transition(struct rds_connection *conn, int old, int new)
608{
609 return atomic_cmpxchg(&conn->c_state, old, new) == old;
610}
611
612static inline int
613rds_conn_state(struct rds_connection *conn)
614{
615 return atomic_read(&conn->c_state);
616}
617
618static inline int
619rds_conn_up(struct rds_connection *conn)
620{
621 return atomic_read(&conn->c_state) == RDS_CONN_UP;
622}
623
624static inline int
625rds_conn_connecting(struct rds_connection *conn)
626{
627 return atomic_read(&conn->c_state) == RDS_CONN_CONNECTING;
628}
629
630/* message.c */
631struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
Andy Groverfc445082010-01-12 12:56:06 -0800632struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
633int rds_message_copy_from_user(struct rds_message *rm, struct iovec *first_iov,
Andy Grover39de8282009-02-24 15:30:19 +0000634 size_t total_len);
635struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
636void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
637 __be16 dport, u64 seq);
638int rds_message_add_extension(struct rds_header *hdr,
639 unsigned int type, const void *data, unsigned int len);
640int rds_message_next_extension(struct rds_header *hdr,
641 unsigned int *pos, void *buf, unsigned int *buflen);
642int rds_message_add_version_extension(struct rds_header *hdr, unsigned int version);
643int rds_message_get_version_extension(struct rds_header *hdr, unsigned int *version);
644int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
645int rds_message_inc_copy_to_user(struct rds_incoming *inc,
646 struct iovec *first_iov, size_t size);
647void rds_message_inc_purge(struct rds_incoming *inc);
648void rds_message_inc_free(struct rds_incoming *inc);
649void rds_message_addref(struct rds_message *rm);
650void rds_message_put(struct rds_message *rm);
651void rds_message_wait(struct rds_message *rm);
652void rds_message_unmapped(struct rds_message *rm);
653
654static inline void rds_message_make_checksum(struct rds_header *hdr)
655{
656 hdr->h_csum = 0;
657 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
658}
659
660static inline int rds_message_verify_checksum(const struct rds_header *hdr)
661{
662 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
663}
664
665
666/* page.c */
667int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
668 gfp_t gfp);
669int rds_page_copy_user(struct page *page, unsigned long offset,
670 void __user *ptr, unsigned long bytes,
671 int to_user);
672#define rds_page_copy_to_user(page, offset, ptr, bytes) \
673 rds_page_copy_user(page, offset, ptr, bytes, 1)
674#define rds_page_copy_from_user(page, offset, ptr, bytes) \
675 rds_page_copy_user(page, offset, ptr, bytes, 0)
676void rds_page_exit(void);
677
678/* recv.c */
679void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
680 __be32 saddr);
681void rds_inc_addref(struct rds_incoming *inc);
682void rds_inc_put(struct rds_incoming *inc);
683void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
684 struct rds_incoming *inc, gfp_t gfp, enum km_type km);
685int rds_recvmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
686 size_t size, int msg_flags);
687void rds_clear_recv_queue(struct rds_sock *rs);
688int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
689void rds_inc_info_copy(struct rds_incoming *inc,
690 struct rds_info_iterator *iter,
691 __be32 saddr, __be32 daddr, int flip);
692
693/* send.c */
694int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
695 size_t payload_len);
696void rds_send_reset(struct rds_connection *conn);
697int rds_send_xmit(struct rds_connection *conn);
698struct sockaddr_in;
699void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
700typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
701void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
702 is_acked_func is_acked);
703int rds_send_acked_before(struct rds_connection *conn, u64 seq);
704void rds_send_remove_from_sock(struct list_head *messages, int status);
705int rds_send_pong(struct rds_connection *conn, __be16 dport);
706struct rds_message *rds_send_get_message(struct rds_connection *,
707 struct rds_rdma_op *);
708
709/* rdma.c */
710void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
Andy Grover21f79af2010-01-12 12:57:27 -0800711int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
712int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
713int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
714void rds_rdma_drop_keys(struct rds_sock *rs);
715int rds_rdma_extra_size(struct rds_rdma_args *args);
716int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
717 struct cmsghdr *cmsg);
718int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
719 struct cmsghdr *cmsg);
720int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
721 struct cmsghdr *cmsg);
722int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
723 struct cmsghdr *cmsg);
724void rds_rdma_free_op(struct rds_rdma_op *ro);
725void rds_rdma_send_complete(struct rds_message *rm, int);
726
727extern void __rds_put_mr_final(struct rds_mr *mr);
728static inline void rds_mr_put(struct rds_mr *mr)
729{
730 if (atomic_dec_and_test(&mr->r_refcount))
731 __rds_put_mr_final(mr);
732}
Andy Grover39de8282009-02-24 15:30:19 +0000733
734/* stats.c */
David Howells9b8de742009-04-21 23:00:24 +0100735DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
Andy Grover39de8282009-02-24 15:30:19 +0000736#define rds_stats_inc_which(which, member) do { \
737 per_cpu(which, get_cpu()).member++; \
738 put_cpu(); \
739} while (0)
740#define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
741#define rds_stats_add_which(which, member, count) do { \
742 per_cpu(which, get_cpu()).member += count; \
743 put_cpu(); \
744} while (0)
745#define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
746int __init rds_stats_init(void);
747void rds_stats_exit(void);
748void rds_stats_info_copy(struct rds_info_iterator *iter,
Jan Engelhardt36cbd3d2009-08-05 10:42:58 -0700749 uint64_t *values, const char *const *names,
750 size_t nr);
Andy Grover39de8282009-02-24 15:30:19 +0000751
752/* sysctl.c */
753int __init rds_sysctl_init(void);
754void rds_sysctl_exit(void);
755extern unsigned long rds_sysctl_sndbuf_min;
756extern unsigned long rds_sysctl_sndbuf_default;
757extern unsigned long rds_sysctl_sndbuf_max;
758extern unsigned long rds_sysctl_reconnect_min_jiffies;
759extern unsigned long rds_sysctl_reconnect_max_jiffies;
760extern unsigned int rds_sysctl_max_unacked_packets;
761extern unsigned int rds_sysctl_max_unacked_bytes;
762extern unsigned int rds_sysctl_ping_enable;
763extern unsigned long rds_sysctl_trace_flags;
764extern unsigned int rds_sysctl_trace_level;
765
766/* threads.c */
767int __init rds_threads_init(void);
768void rds_threads_exit(void);
769extern struct workqueue_struct *rds_wq;
Andy Grover2dc39352010-06-11 13:49:13 -0700770void rds_queue_reconnect(struct rds_connection *conn);
Andy Grover39de8282009-02-24 15:30:19 +0000771void rds_connect_worker(struct work_struct *);
772void rds_shutdown_worker(struct work_struct *);
773void rds_send_worker(struct work_struct *);
774void rds_recv_worker(struct work_struct *);
775void rds_connect_complete(struct rds_connection *conn);
776
777/* transport.c */
778int rds_trans_register(struct rds_transport *trans);
779void rds_trans_unregister(struct rds_transport *trans);
780struct rds_transport *rds_trans_get_preferred(__be32 addr);
781unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
782 unsigned int avail);
783int __init rds_trans_init(void);
784void rds_trans_exit(void);
785
786#endif