blob: 9542449c0720864af4697dd73cfb0685fd821a55 [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
Zach Brown0f4b1c72010-06-04 14:41:41 -070083#define RDS_IN_XMIT 2
Andy Grover39de8282009-02-24 15:30:19 +000084
85struct rds_connection {
86 struct hlist_node c_hash_node;
87 __be32 c_laddr;
88 __be32 c_faddr;
89 unsigned int c_loopback:1;
90 struct rds_connection *c_passive;
91
92 struct rds_cong_map *c_lcong;
93 struct rds_cong_map *c_fcong;
94
Andy Grover39de8282009-02-24 15:30:19 +000095 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;
Andy Grover15133f62010-01-12 14:33:38 -080099 unsigned int c_xmit_atomic_sent;
Andy Grover5b2366b2010-02-03 19:36:44 -0800100 unsigned int c_xmit_rdma_sent;
101 unsigned int c_xmit_data_sent;
Andy Grover39de8282009-02-24 15:30:19 +0000102
103 spinlock_t c_lock; /* protect msg queues */
104 u64 c_next_tx_seq;
105 struct list_head c_send_queue;
106 struct list_head c_retrans;
107
108 u64 c_next_rx_seq;
109
110 struct rds_transport *c_trans;
111 void *c_transport_data;
112
113 atomic_t c_state;
114 unsigned long c_flags;
115 unsigned long c_reconnect_jiffies;
116 struct delayed_work c_send_w;
117 struct delayed_work c_recv_w;
118 struct delayed_work c_conn_w;
119 struct work_struct c_down_w;
120 struct mutex c_cm_lock; /* protect conn state & cm */
Zach Brown0f4b1c72010-06-04 14:41:41 -0700121 wait_queue_head_t c_waitq;
Andy Grover39de8282009-02-24 15:30:19 +0000122
123 struct list_head c_map_item;
124 unsigned long c_map_queued;
Andy Grover39de8282009-02-24 15:30:19 +0000125
126 unsigned int c_unacked_packets;
127 unsigned int c_unacked_bytes;
128
129 /* Protocol version */
130 unsigned int c_version;
131};
132
133#define RDS_FLAG_CONG_BITMAP 0x01
134#define RDS_FLAG_ACK_REQUIRED 0x02
135#define RDS_FLAG_RETRANSMITTED 0x04
Steve Wise7b70d032009-04-09 14:09:39 +0000136#define RDS_MAX_ADV_CREDIT 255
Andy Grover39de8282009-02-24 15:30:19 +0000137
138/*
139 * Maximum space available for extension headers.
140 */
141#define RDS_HEADER_EXT_SPACE 16
142
143struct rds_header {
144 __be64 h_sequence;
145 __be64 h_ack;
146 __be32 h_len;
147 __be16 h_sport;
148 __be16 h_dport;
149 u8 h_flags;
150 u8 h_credit;
151 u8 h_padding[4];
152 __sum16 h_csum;
153
154 u8 h_exthdr[RDS_HEADER_EXT_SPACE];
155};
156
157/*
158 * Reserved - indicates end of extensions
159 */
160#define RDS_EXTHDR_NONE 0
161
162/*
163 * This extension header is included in the very
164 * first message that is sent on a new connection,
165 * and identifies the protocol level. This will help
166 * rolling updates if a future change requires breaking
167 * the protocol.
168 * NB: This is no longer true for IB, where we do a version
169 * negotiation during the connection setup phase (protocol
170 * version information is included in the RDMA CM private data).
171 */
172#define RDS_EXTHDR_VERSION 1
173struct rds_ext_header_version {
174 __be32 h_version;
175};
176
177/*
178 * This extension header is included in the RDS message
179 * chasing an RDMA operation.
180 */
181#define RDS_EXTHDR_RDMA 2
182struct rds_ext_header_rdma {
183 __be32 h_rdma_rkey;
184};
185
186/*
187 * This extension header tells the peer about the
188 * destination <R_Key,offset> of the requested RDMA
189 * operation.
190 */
191#define RDS_EXTHDR_RDMA_DEST 3
192struct rds_ext_header_rdma_dest {
193 __be32 h_rdma_rkey;
194 __be32 h_rdma_offset;
195};
196
197#define __RDS_EXTHDR_MAX 16 /* for now */
198
199struct rds_incoming {
200 atomic_t i_refcount;
201 struct list_head i_item;
202 struct rds_connection *i_conn;
203 struct rds_header i_hdr;
204 unsigned long i_rx_jiffies;
205 __be32 i_saddr;
206
207 rds_rdma_cookie_t i_rdma_cookie;
208};
209
Andy Grover21f79af2010-01-12 12:57:27 -0800210struct rds_mr {
211 struct rb_node r_rb_node;
212 atomic_t r_refcount;
213 u32 r_key;
214
215 /* A copy of the creation flags */
216 unsigned int r_use_once:1;
217 unsigned int r_invalidate:1;
218 unsigned int r_write:1;
219
220 /* This is for RDS_MR_DEAD.
221 * It would be nice & consistent to make this part of the above
222 * bit field here, but we need to use test_and_set_bit.
223 */
224 unsigned long r_state;
225 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
226 struct rds_transport *r_trans;
227 void *r_trans_private;
228};
229
230/* Flags for mr->r_state */
231#define RDS_MR_DEAD 0
232
Andy Grover21f79af2010-01-12 12:57:27 -0800233static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
234{
235 return r_key | (((u64) offset) << 32);
236}
237
238static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
239{
240 return cookie;
241}
242
243static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
244{
245 return cookie >> 32;
246}
247
Andy Grover15133f62010-01-12 14:33:38 -0800248/* atomic operation types */
249#define RDS_ATOMIC_TYPE_CSWP 0
250#define RDS_ATOMIC_TYPE_FADD 1
251
Andy Grover39de8282009-02-24 15:30:19 +0000252/*
253 * m_sock_item and m_conn_item are on lists that are serialized under
254 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
255 * the message will not be put back on the retransmit list after being sent.
256 * messages that are canceled while being sent rely on this.
257 *
258 * m_inc is used by loopback so that it can pass an incoming message straight
259 * back up into the rx path. It embeds a wire header which is also used by
260 * the send path, which is kind of awkward.
261 *
262 * m_sock_item indicates the message's presence on a socket's send or receive
263 * queue. m_rs will point to that socket.
264 *
265 * m_daddr is used by cancellation to prune messages to a given destination.
266 *
267 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
268 * nesting. As paths iterate over messages on a sock, or conn, they must
269 * also lock the conn, or sock, to remove the message from those lists too.
270 * Testing the flag to determine if the message is still on the lists lets
271 * us avoid testing the list_head directly. That means each path can use
272 * the message's list_head to keep it on a local list while juggling locks
273 * without confusing the other path.
274 *
275 * m_ack_seq is an optional field set by transports who need a different
276 * sequence number range to invalidate. They can use this in a callback
277 * that they pass to rds_send_drop_acked() to see if each message has been
278 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
279 * had ack_seq set yet.
280 */
281#define RDS_MSG_ON_SOCK 1
282#define RDS_MSG_ON_CONN 2
283#define RDS_MSG_HAS_ACK_SEQ 3
284#define RDS_MSG_ACK_REQUIRED 4
285#define RDS_MSG_RETRANSMITTED 5
286#define RDS_MSG_MAPPED 6
287#define RDS_MSG_PAGEVEC 7
288
289struct rds_message {
290 atomic_t m_refcount;
291 struct list_head m_sock_item;
292 struct list_head m_conn_item;
293 struct rds_incoming m_inc;
294 u64 m_ack_seq;
295 __be32 m_daddr;
296 unsigned long m_flags;
297
298 /* Never access m_rs without holding m_rs_lock.
299 * Lock nesting is
300 * rm->m_rs_lock
301 * -> rs->rs_lock
302 */
303 spinlock_t m_rs_lock;
Chris Masonc83188d2010-04-21 13:09:28 -0700304 wait_queue_head_t m_flush_wait;
305
Andy Grover39de8282009-02-24 15:30:19 +0000306 struct rds_sock *m_rs;
Andy Grover7e3bd652010-03-01 16:04:59 -0800307
308 /* cookie to send to remote, in rds header */
Andy Grover39de8282009-02-24 15:30:19 +0000309 rds_rdma_cookie_t m_rdma_cookie;
Andy Grover7e3bd652010-03-01 16:04:59 -0800310
311 unsigned int m_used_sgs;
312 unsigned int m_total_sgs;
313
Andy Groverff3d7d32010-03-01 14:03:09 -0800314 void *m_final_op;
315
Andy Grovere7791372010-01-12 12:15:02 -0800316 struct {
Andy Grover15133f62010-01-12 14:33:38 -0800317 struct rm_atomic_op {
318 int op_type;
Andy Grover20c72bd2010-08-25 05:51:28 -0700319 union {
320 struct {
321 uint64_t compare;
322 uint64_t swap;
323 uint64_t compare_mask;
324 uint64_t swap_mask;
325 } op_m_cswp;
326 struct {
327 uint64_t add;
328 uint64_t nocarry_mask;
329 } op_m_fadd;
330 };
Andy Grover15133f62010-01-12 14:33:38 -0800331
332 u32 op_rkey;
333 u64 op_remote_addr;
334 unsigned int op_notify:1;
335 unsigned int op_recverr:1;
336 unsigned int op_mapped:1;
Andy Grover2c3a5f92010-03-01 16:10:40 -0800337 unsigned int op_silent:1;
Andy Grover15133f62010-01-12 14:33:38 -0800338 unsigned int op_active:1;
Andy Grover15133f62010-01-12 14:33:38 -0800339 struct scatterlist *op_sg;
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800340 struct rds_notifier *op_notifier;
Andy Grover15133f62010-01-12 14:33:38 -0800341
342 struct rds_mr *op_rdma_mr;
343 } atomic;
344 struct rm_rdma_op {
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800345 u32 op_rkey;
346 u64 op_remote_addr;
347 unsigned int op_write:1;
348 unsigned int op_fence:1;
349 unsigned int op_notify:1;
350 unsigned int op_recverr:1;
351 unsigned int op_mapped:1;
Andy Grover2c3a5f92010-03-01 16:10:40 -0800352 unsigned int op_silent:1;
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800353 unsigned int op_active:1;
354 unsigned int op_bytes;
355 unsigned int op_nents;
356 unsigned int op_count;
357 struct scatterlist *op_sg;
358 struct rds_notifier *op_notifier;
359
360 struct rds_mr *op_rdma_mr;
Andy Grovere7791372010-01-12 12:15:02 -0800361 } rdma;
Andy Grover15133f62010-01-12 14:33:38 -0800362 struct rm_data_op {
Andy Grover241eef32010-01-19 21:25:26 -0800363 unsigned int op_active:1;
Andy Grover6c7cc6e2010-01-27 18:04:18 -0800364 unsigned int op_nents;
365 unsigned int op_count;
366 struct scatterlist *op_sg;
Andy Grovere7791372010-01-12 12:15:02 -0800367 } data;
368 };
Andy Grover39de8282009-02-24 15:30:19 +0000369};
370
371/*
372 * The RDS notifier is used (optionally) to tell the application about
373 * completed RDMA operations. Rather than keeping the whole rds message
374 * around on the queue, we allocate a small notifier that is put on the
375 * socket's notifier_list. Notifications are delivered to the application
376 * through control messages.
377 */
378struct rds_notifier {
379 struct list_head n_list;
380 uint64_t n_user_token;
381 int n_status;
382};
383
384/**
385 * struct rds_transport - transport specific behavioural hooks
386 *
387 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
388 * part of a message. The caller serializes on the send_sem so this
389 * doesn't need to be reentrant for a given conn. The header must be
390 * sent before the data payload. .xmit must be prepared to send a
391 * message with no data payload. .xmit should return the number of
392 * bytes that were sent down the connection, including header bytes.
393 * Returning 0 tells the caller that it doesn't need to perform any
394 * additional work now. This is usually the case when the transport has
395 * filled the sending queue for its connection and will handle
396 * triggering the rds thread to continue the send when space becomes
397 * available. Returning -EAGAIN tells the caller to retry the send
398 * immediately. Returning -ENOMEM tells the caller to retry the send at
399 * some point in the future.
400 *
401 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
402 * it returns the connection can not call rds_recv_incoming().
403 * This will only be called once after conn_connect returns
404 * non-zero success and will The caller serializes this with
405 * the send and connecting paths (xmit_* and conn_*). The
406 * transport is responsible for other serialization, including
407 * rds_recv_incoming(). This is called in process context but
408 * should try hard not to block.
Andy Grover39de8282009-02-24 15:30:19 +0000409 */
410
Andy Grover335776b2009-08-21 12:28:34 +0000411#define RDS_TRANS_IB 0
412#define RDS_TRANS_IWARP 1
413#define RDS_TRANS_TCP 2
414#define RDS_TRANS_COUNT 3
415
Andy Grover39de8282009-02-24 15:30:19 +0000416struct rds_transport {
417 char t_name[TRANSNAMSIZ];
418 struct list_head t_item;
419 struct module *t_owner;
420 unsigned int t_prefer_loopback:1;
Andy Grover335776b2009-08-21 12:28:34 +0000421 unsigned int t_type;
Andy Grover39de8282009-02-24 15:30:19 +0000422
423 int (*laddr_check)(__be32 addr);
424 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
425 void (*conn_free)(void *data);
426 int (*conn_connect)(struct rds_connection *conn);
427 void (*conn_shutdown)(struct rds_connection *conn);
428 void (*xmit_prepare)(struct rds_connection *conn);
429 void (*xmit_complete)(struct rds_connection *conn);
430 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
431 unsigned int hdr_off, unsigned int sg, unsigned int off);
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800432 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
Andy Groverff3d7d32010-03-01 14:03:09 -0800433 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
Andy Grover39de8282009-02-24 15:30:19 +0000434 int (*recv)(struct rds_connection *conn);
435 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iovec *iov,
436 size_t size);
Andy Grover39de8282009-02-24 15:30:19 +0000437 void (*inc_free)(struct rds_incoming *inc);
438
439 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
440 struct rdma_cm_event *event);
441 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
442 void (*cm_connect_complete)(struct rds_connection *conn,
443 struct rdma_cm_event *event);
444
445 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
446 unsigned int avail);
447 void (*exit)(void);
448 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
449 struct rds_sock *rs, u32 *key_ret);
450 void (*sync_mr)(void *trans_private, int direction);
451 void (*free_mr)(void *trans_private, int invalidate);
452 void (*flush_mrs)(void);
453};
454
455struct rds_sock {
456 struct sock rs_sk;
457
458 u64 rs_user_addr;
459 u64 rs_user_bytes;
460
461 /*
462 * bound_addr used for both incoming and outgoing, no INADDR_ANY
463 * support.
464 */
Chris Mason38a4e5e2010-05-11 15:09:45 -0700465 struct hlist_node rs_bound_node;
Andy Grover39de8282009-02-24 15:30:19 +0000466 __be32 rs_bound_addr;
467 __be32 rs_conn_addr;
468 __be16 rs_bound_port;
469 __be16 rs_conn_port;
Andy Grover39de8282009-02-24 15:30:19 +0000470 struct rds_transport *rs_transport;
471
472 /*
473 * rds_sendmsg caches the conn it used the last time around.
474 * This helps avoid costly lookups.
475 */
476 struct rds_connection *rs_conn;
477
478 /* flag indicating we were congested or not */
479 int rs_congested;
Andy Groverb98ba522010-03-11 13:50:04 +0000480 /* seen congestion (ENOBUFS) when sending? */
481 int rs_seen_congestion;
Andy Grover39de8282009-02-24 15:30:19 +0000482
483 /* rs_lock protects all these adjacent members before the newline */
484 spinlock_t rs_lock;
485 struct list_head rs_send_queue;
486 u32 rs_snd_bytes;
487 int rs_rcv_bytes;
488 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
489
490 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
491 * to decide whether the application should be woken up.
492 * If not set, we use rs_cong_track to find out whether a cong map
493 * update arrived.
494 */
495 uint64_t rs_cong_mask;
496 uint64_t rs_cong_notify;
497 struct list_head rs_cong_list;
498 unsigned long rs_cong_track;
499
500 /*
501 * rs_recv_lock protects the receive queue, and is
502 * used to serialize with rds_release.
503 */
504 rwlock_t rs_recv_lock;
505 struct list_head rs_recv_queue;
506
507 /* just for stats reporting */
508 struct list_head rs_item;
509
510 /* these have their own lock */
511 spinlock_t rs_rdma_lock;
512 struct rb_root rs_rdma_keys;
513
514 /* Socket options - in case there will be more */
515 unsigned char rs_recverr,
516 rs_cong_monitor;
517};
518
519static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
520{
521 return container_of(sk, struct rds_sock, rs_sk);
522}
523static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
524{
525 return &rs->rs_sk;
526}
527
528/*
529 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
530 * to account for overhead. We don't account for overhead, we just apply
531 * the number of payload bytes to the specified value.
532 */
533static inline int rds_sk_sndbuf(struct rds_sock *rs)
534{
535 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
536}
537static inline int rds_sk_rcvbuf(struct rds_sock *rs)
538{
539 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
540}
541
542struct rds_statistics {
543 uint64_t s_conn_reset;
544 uint64_t s_recv_drop_bad_checksum;
545 uint64_t s_recv_drop_old_seq;
546 uint64_t s_recv_drop_no_sock;
547 uint64_t s_recv_drop_dead_sock;
548 uint64_t s_recv_deliver_raced;
549 uint64_t s_recv_delivered;
550 uint64_t s_recv_queued;
551 uint64_t s_recv_immediate_retry;
552 uint64_t s_recv_delayed_retry;
553 uint64_t s_recv_ack_required;
554 uint64_t s_recv_rdma_bytes;
555 uint64_t s_recv_ping;
556 uint64_t s_send_queue_empty;
557 uint64_t s_send_queue_full;
Andy Grover049ee3f2010-03-23 17:39:07 -0700558 uint64_t s_send_lock_contention;
559 uint64_t s_send_lock_queue_raced;
Andy Grover39de8282009-02-24 15:30:19 +0000560 uint64_t s_send_immediate_retry;
561 uint64_t s_send_delayed_retry;
562 uint64_t s_send_drop_acked;
563 uint64_t s_send_ack_required;
564 uint64_t s_send_queued;
565 uint64_t s_send_rdma;
566 uint64_t s_send_rdma_bytes;
567 uint64_t s_send_pong;
568 uint64_t s_page_remainder_hit;
569 uint64_t s_page_remainder_miss;
570 uint64_t s_copy_to_user;
571 uint64_t s_copy_from_user;
572 uint64_t s_cong_update_queued;
573 uint64_t s_cong_update_received;
574 uint64_t s_cong_send_error;
575 uint64_t s_cong_send_blocked;
576};
577
578/* af_rds.c */
Zach Brown59f740a2010-08-03 13:52:47 -0700579char *rds_str_array(char **array, size_t elements, size_t index);
Andy Grover39de8282009-02-24 15:30:19 +0000580void rds_sock_addref(struct rds_sock *rs);
581void rds_sock_put(struct rds_sock *rs);
582void rds_wake_sk_sleep(struct rds_sock *rs);
583static inline void __rds_wake_sk_sleep(struct sock *sk)
584{
Eric Dumazetaa395142010-04-20 13:03:51 +0000585 wait_queue_head_t *waitq = sk_sleep(sk);
Andy Grover39de8282009-02-24 15:30:19 +0000586
587 if (!sock_flag(sk, SOCK_DEAD) && waitq)
588 wake_up(waitq);
589}
590extern wait_queue_head_t rds_poll_waitq;
591
592
593/* bind.c */
594int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
595void rds_remove_bound(struct rds_sock *rs);
596struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
597
598/* cong.c */
599int rds_cong_get_maps(struct rds_connection *conn);
600void rds_cong_add_conn(struct rds_connection *conn);
601void rds_cong_remove_conn(struct rds_connection *conn);
602void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
603void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
604int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
605void rds_cong_queue_updates(struct rds_cong_map *map);
606void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
607int rds_cong_updated_since(unsigned long *recent);
608void rds_cong_add_socket(struct rds_sock *);
609void rds_cong_remove_socket(struct rds_sock *);
610void rds_cong_exit(void);
611struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
612
613/* conn.c */
Zach Brownef87b7e2010-07-09 12:26:20 -0700614int rds_conn_init(void);
Andy Grover39de8282009-02-24 15:30:19 +0000615void rds_conn_exit(void);
616struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
617 struct rds_transport *trans, gfp_t gfp);
618struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
619 struct rds_transport *trans, gfp_t gfp);
Andy Grover2dc39352010-06-11 13:49:13 -0700620void rds_conn_shutdown(struct rds_connection *conn);
Andy Grover39de8282009-02-24 15:30:19 +0000621void rds_conn_destroy(struct rds_connection *conn);
Andy Grover39de8282009-02-24 15:30:19 +0000622void rds_conn_drop(struct rds_connection *conn);
Zach Brownf3c68082010-05-24 13:14:36 -0700623void rds_conn_connect_if_down(struct rds_connection *conn);
Andy Grover39de8282009-02-24 15:30:19 +0000624void rds_for_each_conn_info(struct socket *sock, unsigned int len,
625 struct rds_info_iterator *iter,
626 struct rds_info_lengths *lens,
627 int (*visitor)(struct rds_connection *, void *),
628 size_t item_len);
629void __rds_conn_error(struct rds_connection *conn, const char *, ...)
630 __attribute__ ((format (printf, 2, 3)));
631#define rds_conn_error(conn, fmt...) \
632 __rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
633
634static inline int
635rds_conn_transition(struct rds_connection *conn, int old, int new)
636{
637 return atomic_cmpxchg(&conn->c_state, old, new) == old;
638}
639
640static inline int
641rds_conn_state(struct rds_connection *conn)
642{
643 return atomic_read(&conn->c_state);
644}
645
646static inline int
647rds_conn_up(struct rds_connection *conn)
648{
649 return atomic_read(&conn->c_state) == RDS_CONN_UP;
650}
651
652static inline int
653rds_conn_connecting(struct rds_connection *conn)
654{
655 return atomic_read(&conn->c_state) == RDS_CONN_CONNECTING;
656}
657
658/* message.c */
659struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
Andy Groverfc445082010-01-12 12:56:06 -0800660struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
661int rds_message_copy_from_user(struct rds_message *rm, struct iovec *first_iov,
Andy Grover39de8282009-02-24 15:30:19 +0000662 size_t total_len);
663struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
664void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
665 __be16 dport, u64 seq);
666int rds_message_add_extension(struct rds_header *hdr,
667 unsigned int type, const void *data, unsigned int len);
668int rds_message_next_extension(struct rds_header *hdr,
669 unsigned int *pos, void *buf, unsigned int *buflen);
Andy Grover39de8282009-02-24 15:30:19 +0000670int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
671int rds_message_inc_copy_to_user(struct rds_incoming *inc,
672 struct iovec *first_iov, size_t size);
Andy Grover39de8282009-02-24 15:30:19 +0000673void rds_message_inc_free(struct rds_incoming *inc);
674void rds_message_addref(struct rds_message *rm);
675void rds_message_put(struct rds_message *rm);
676void rds_message_wait(struct rds_message *rm);
677void rds_message_unmapped(struct rds_message *rm);
678
679static inline void rds_message_make_checksum(struct rds_header *hdr)
680{
681 hdr->h_csum = 0;
682 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
683}
684
685static inline int rds_message_verify_checksum(const struct rds_header *hdr)
686{
687 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
688}
689
690
691/* page.c */
692int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
693 gfp_t gfp);
694int rds_page_copy_user(struct page *page, unsigned long offset,
695 void __user *ptr, unsigned long bytes,
696 int to_user);
697#define rds_page_copy_to_user(page, offset, ptr, bytes) \
698 rds_page_copy_user(page, offset, ptr, bytes, 1)
699#define rds_page_copy_from_user(page, offset, ptr, bytes) \
700 rds_page_copy_user(page, offset, ptr, bytes, 0)
701void rds_page_exit(void);
702
703/* recv.c */
704void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
705 __be32 saddr);
Andy Grover39de8282009-02-24 15:30:19 +0000706void rds_inc_put(struct rds_incoming *inc);
707void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
708 struct rds_incoming *inc, gfp_t gfp, enum km_type km);
709int rds_recvmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
710 size_t size, int msg_flags);
711void rds_clear_recv_queue(struct rds_sock *rs);
712int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
713void rds_inc_info_copy(struct rds_incoming *inc,
714 struct rds_info_iterator *iter,
715 __be32 saddr, __be32 daddr, int flip);
716
717/* send.c */
718int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
719 size_t payload_len);
720void rds_send_reset(struct rds_connection *conn);
721int rds_send_xmit(struct rds_connection *conn);
722struct sockaddr_in;
723void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
724typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
725void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
726 is_acked_func is_acked);
Andy Grover39de8282009-02-24 15:30:19 +0000727int rds_send_pong(struct rds_connection *conn, __be16 dport);
728struct rds_message *rds_send_get_message(struct rds_connection *,
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800729 struct rm_rdma_op *);
Andy Grover39de8282009-02-24 15:30:19 +0000730
731/* rdma.c */
732void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
Andy Grover21f79af2010-01-12 12:57:27 -0800733int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
734int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
735int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
736void rds_rdma_drop_keys(struct rds_sock *rs);
737int rds_rdma_extra_size(struct rds_rdma_args *args);
738int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
739 struct cmsghdr *cmsg);
740int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
741 struct cmsghdr *cmsg);
742int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
743 struct cmsghdr *cmsg);
744int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
745 struct cmsghdr *cmsg);
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800746void rds_rdma_free_op(struct rm_rdma_op *ro);
Andy Groverd0ab25a2010-01-27 16:15:48 -0800747void rds_atomic_free_op(struct rm_atomic_op *ao);
Andy Grover15133f62010-01-12 14:33:38 -0800748void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
749void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
750int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
751 struct cmsghdr *cmsg);
Andy Grover21f79af2010-01-12 12:57:27 -0800752
753extern void __rds_put_mr_final(struct rds_mr *mr);
754static inline void rds_mr_put(struct rds_mr *mr)
755{
756 if (atomic_dec_and_test(&mr->r_refcount))
757 __rds_put_mr_final(mr);
758}
Andy Grover39de8282009-02-24 15:30:19 +0000759
760/* stats.c */
David Howells9b8de742009-04-21 23:00:24 +0100761DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
Andy Grover39de8282009-02-24 15:30:19 +0000762#define rds_stats_inc_which(which, member) do { \
763 per_cpu(which, get_cpu()).member++; \
764 put_cpu(); \
765} while (0)
766#define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
767#define rds_stats_add_which(which, member, count) do { \
768 per_cpu(which, get_cpu()).member += count; \
769 put_cpu(); \
770} while (0)
771#define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
Zach Brownef87b7e2010-07-09 12:26:20 -0700772int rds_stats_init(void);
Andy Grover39de8282009-02-24 15:30:19 +0000773void rds_stats_exit(void);
774void rds_stats_info_copy(struct rds_info_iterator *iter,
Jan Engelhardt36cbd3d2009-08-05 10:42:58 -0700775 uint64_t *values, const char *const *names,
776 size_t nr);
Andy Grover39de8282009-02-24 15:30:19 +0000777
778/* sysctl.c */
Zach Brownef87b7e2010-07-09 12:26:20 -0700779int rds_sysctl_init(void);
Andy Grover39de8282009-02-24 15:30:19 +0000780void rds_sysctl_exit(void);
781extern unsigned long rds_sysctl_sndbuf_min;
782extern unsigned long rds_sysctl_sndbuf_default;
783extern unsigned long rds_sysctl_sndbuf_max;
784extern unsigned long rds_sysctl_reconnect_min_jiffies;
785extern unsigned long rds_sysctl_reconnect_max_jiffies;
786extern unsigned int rds_sysctl_max_unacked_packets;
787extern unsigned int rds_sysctl_max_unacked_bytes;
788extern unsigned int rds_sysctl_ping_enable;
789extern unsigned long rds_sysctl_trace_flags;
790extern unsigned int rds_sysctl_trace_level;
791
792/* threads.c */
Zach Brownef87b7e2010-07-09 12:26:20 -0700793int rds_threads_init(void);
Andy Grover39de8282009-02-24 15:30:19 +0000794void rds_threads_exit(void);
795extern struct workqueue_struct *rds_wq;
Andy Grover2dc39352010-06-11 13:49:13 -0700796void rds_queue_reconnect(struct rds_connection *conn);
Andy Grover39de8282009-02-24 15:30:19 +0000797void rds_connect_worker(struct work_struct *);
798void rds_shutdown_worker(struct work_struct *);
799void rds_send_worker(struct work_struct *);
800void rds_recv_worker(struct work_struct *);
801void rds_connect_complete(struct rds_connection *conn);
802
803/* transport.c */
804int rds_trans_register(struct rds_transport *trans);
805void rds_trans_unregister(struct rds_transport *trans);
806struct rds_transport *rds_trans_get_preferred(__be32 addr);
Zach Brown5adb5bc2010-07-23 10:32:31 -0700807void rds_trans_put(struct rds_transport *trans);
Andy Grover39de8282009-02-24 15:30:19 +0000808unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
809 unsigned int avail);
Zach Brownef87b7e2010-07-09 12:26:20 -0700810int rds_trans_init(void);
Andy Grover39de8282009-02-24 15:30:19 +0000811void rds_trans_exit(void);
812
813#endif