blob: afb4048d0cfd2b530e8885cc9f2f28f96aac63f0 [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 */
Joe Perchesb9075fa2011-10-31 17:11:33 -070039static inline __printf(1, 2)
40void rdsdebug(char *fmt, ...)
Andy Grover39de8282009-02-24 15:30:19 +000041{
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)
Andy Grover39de8282009-02-24 15:30:19 +000053#define RDS_CONG_MAP_PAGES (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE)
54#define RDS_CONG_MAP_PAGE_BITS (PAGE_SIZE * 8)
55
56struct rds_cong_map {
57 struct rb_node m_rb_node;
58 __be32 m_addr;
59 wait_queue_head_t m_waitq;
60 struct list_head m_conn_list;
61 unsigned long m_page_addrs[RDS_CONG_MAP_PAGES];
62};
63
64
65/*
66 * This is how we will track the connection state:
67 * A connection is always in one of the following
68 * states. Updates to the state are atomic and imply
69 * a memory barrier.
70 */
71enum {
72 RDS_CONN_DOWN = 0,
73 RDS_CONN_CONNECTING,
74 RDS_CONN_DISCONNECTING,
75 RDS_CONN_UP,
76 RDS_CONN_ERROR,
77};
78
79/* Bits for c_flags */
80#define RDS_LL_SEND_FULL 0
81#define RDS_RECONNECT_PENDING 1
Zach Brown0f4b1c72010-06-04 14:41:41 -070082#define RDS_IN_XMIT 2
santosh.shilimkar@oracle.com73ce4312015-08-22 15:45:26 -070083#define RDS_RECV_REFILL 3
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;
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400114 unsigned long c_send_gen;
Andy Grover39de8282009-02-24 15:30:19 +0000115 unsigned long c_flags;
116 unsigned long c_reconnect_jiffies;
117 struct delayed_work c_send_w;
118 struct delayed_work c_recv_w;
119 struct delayed_work c_conn_w;
120 struct work_struct c_down_w;
121 struct mutex c_cm_lock; /* protect conn state & cm */
Zach Brown0f4b1c72010-06-04 14:41:41 -0700122 wait_queue_head_t c_waitq;
Andy Grover39de8282009-02-24 15:30:19 +0000123
124 struct list_head c_map_item;
125 unsigned long c_map_queued;
Andy Grover39de8282009-02-24 15:30:19 +0000126
127 unsigned int c_unacked_packets;
128 unsigned int c_unacked_bytes;
129
130 /* Protocol version */
131 unsigned int c_version;
Sowmini Varadhand5a8ac22015-08-05 01:43:25 -0400132 possible_net_t c_net;
Andy Grover39de8282009-02-24 15:30:19 +0000133};
134
Sowmini Varadhand5a8ac22015-08-05 01:43:25 -0400135static inline
136struct net *rds_conn_net(struct rds_connection *conn)
137{
138 return read_pnet(&conn->c_net);
139}
140
141static inline
142void rds_conn_net_set(struct rds_connection *conn, struct net *net)
143{
144 write_pnet(&conn->c_net, net);
145}
146
Andy Grover39de8282009-02-24 15:30:19 +0000147#define RDS_FLAG_CONG_BITMAP 0x01
148#define RDS_FLAG_ACK_REQUIRED 0x02
149#define RDS_FLAG_RETRANSMITTED 0x04
Steve Wise7b70d032009-04-09 14:09:39 +0000150#define RDS_MAX_ADV_CREDIT 255
Andy Grover39de8282009-02-24 15:30:19 +0000151
152/*
153 * Maximum space available for extension headers.
154 */
155#define RDS_HEADER_EXT_SPACE 16
156
157struct rds_header {
158 __be64 h_sequence;
159 __be64 h_ack;
160 __be32 h_len;
161 __be16 h_sport;
162 __be16 h_dport;
163 u8 h_flags;
164 u8 h_credit;
165 u8 h_padding[4];
166 __sum16 h_csum;
167
168 u8 h_exthdr[RDS_HEADER_EXT_SPACE];
169};
170
171/*
172 * Reserved - indicates end of extensions
173 */
174#define RDS_EXTHDR_NONE 0
175
176/*
177 * This extension header is included in the very
178 * first message that is sent on a new connection,
179 * and identifies the protocol level. This will help
180 * rolling updates if a future change requires breaking
181 * the protocol.
182 * NB: This is no longer true for IB, where we do a version
183 * negotiation during the connection setup phase (protocol
184 * version information is included in the RDMA CM private data).
185 */
186#define RDS_EXTHDR_VERSION 1
187struct rds_ext_header_version {
188 __be32 h_version;
189};
190
191/*
192 * This extension header is included in the RDS message
193 * chasing an RDMA operation.
194 */
195#define RDS_EXTHDR_RDMA 2
196struct rds_ext_header_rdma {
197 __be32 h_rdma_rkey;
198};
199
200/*
201 * This extension header tells the peer about the
202 * destination <R_Key,offset> of the requested RDMA
203 * operation.
204 */
205#define RDS_EXTHDR_RDMA_DEST 3
206struct rds_ext_header_rdma_dest {
207 __be32 h_rdma_rkey;
208 __be32 h_rdma_offset;
209};
210
211#define __RDS_EXTHDR_MAX 16 /* for now */
212
213struct rds_incoming {
214 atomic_t i_refcount;
215 struct list_head i_item;
216 struct rds_connection *i_conn;
217 struct rds_header i_hdr;
218 unsigned long i_rx_jiffies;
219 __be32 i_saddr;
220
221 rds_rdma_cookie_t i_rdma_cookie;
222};
223
Andy Grover21f79af2010-01-12 12:57:27 -0800224struct rds_mr {
225 struct rb_node r_rb_node;
226 atomic_t r_refcount;
227 u32 r_key;
228
229 /* A copy of the creation flags */
230 unsigned int r_use_once:1;
231 unsigned int r_invalidate:1;
232 unsigned int r_write:1;
233
234 /* This is for RDS_MR_DEAD.
235 * It would be nice & consistent to make this part of the above
236 * bit field here, but we need to use test_and_set_bit.
237 */
238 unsigned long r_state;
239 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
240 struct rds_transport *r_trans;
241 void *r_trans_private;
242};
243
244/* Flags for mr->r_state */
245#define RDS_MR_DEAD 0
246
Andy Grover21f79af2010-01-12 12:57:27 -0800247static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
248{
249 return r_key | (((u64) offset) << 32);
250}
251
252static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
253{
254 return cookie;
255}
256
257static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
258{
259 return cookie >> 32;
260}
261
Andy Grover15133f62010-01-12 14:33:38 -0800262/* atomic operation types */
263#define RDS_ATOMIC_TYPE_CSWP 0
264#define RDS_ATOMIC_TYPE_FADD 1
265
Andy Grover39de8282009-02-24 15:30:19 +0000266/*
267 * m_sock_item and m_conn_item are on lists that are serialized under
268 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
269 * the message will not be put back on the retransmit list after being sent.
270 * messages that are canceled while being sent rely on this.
271 *
272 * m_inc is used by loopback so that it can pass an incoming message straight
273 * back up into the rx path. It embeds a wire header which is also used by
274 * the send path, which is kind of awkward.
275 *
276 * m_sock_item indicates the message's presence on a socket's send or receive
277 * queue. m_rs will point to that socket.
278 *
279 * m_daddr is used by cancellation to prune messages to a given destination.
280 *
281 * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock
282 * nesting. As paths iterate over messages on a sock, or conn, they must
283 * also lock the conn, or sock, to remove the message from those lists too.
284 * Testing the flag to determine if the message is still on the lists lets
285 * us avoid testing the list_head directly. That means each path can use
286 * the message's list_head to keep it on a local list while juggling locks
287 * without confusing the other path.
288 *
289 * m_ack_seq is an optional field set by transports who need a different
290 * sequence number range to invalidate. They can use this in a callback
291 * that they pass to rds_send_drop_acked() to see if each message has been
292 * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't
293 * had ack_seq set yet.
294 */
295#define RDS_MSG_ON_SOCK 1
296#define RDS_MSG_ON_CONN 2
297#define RDS_MSG_HAS_ACK_SEQ 3
298#define RDS_MSG_ACK_REQUIRED 4
299#define RDS_MSG_RETRANSMITTED 5
300#define RDS_MSG_MAPPED 6
301#define RDS_MSG_PAGEVEC 7
302
303struct rds_message {
304 atomic_t m_refcount;
305 struct list_head m_sock_item;
306 struct list_head m_conn_item;
307 struct rds_incoming m_inc;
308 u64 m_ack_seq;
309 __be32 m_daddr;
310 unsigned long m_flags;
311
312 /* Never access m_rs without holding m_rs_lock.
313 * Lock nesting is
314 * rm->m_rs_lock
315 * -> rs->rs_lock
316 */
317 spinlock_t m_rs_lock;
Chris Masonc83188d2010-04-21 13:09:28 -0700318 wait_queue_head_t m_flush_wait;
319
Andy Grover39de8282009-02-24 15:30:19 +0000320 struct rds_sock *m_rs;
Andy Grover7e3bd652010-03-01 16:04:59 -0800321
322 /* cookie to send to remote, in rds header */
Andy Grover39de8282009-02-24 15:30:19 +0000323 rds_rdma_cookie_t m_rdma_cookie;
Andy Grover7e3bd652010-03-01 16:04:59 -0800324
325 unsigned int m_used_sgs;
326 unsigned int m_total_sgs;
327
Andy Groverff3d7d32010-03-01 14:03:09 -0800328 void *m_final_op;
329
Andy Grovere7791372010-01-12 12:15:02 -0800330 struct {
Andy Grover15133f62010-01-12 14:33:38 -0800331 struct rm_atomic_op {
332 int op_type;
Andy Grover20c72bd2010-08-25 05:51:28 -0700333 union {
334 struct {
335 uint64_t compare;
336 uint64_t swap;
337 uint64_t compare_mask;
338 uint64_t swap_mask;
339 } op_m_cswp;
340 struct {
341 uint64_t add;
342 uint64_t nocarry_mask;
343 } op_m_fadd;
344 };
Andy Grover15133f62010-01-12 14:33:38 -0800345
346 u32 op_rkey;
347 u64 op_remote_addr;
348 unsigned int op_notify:1;
349 unsigned int op_recverr:1;
350 unsigned int op_mapped:1;
Andy Grover2c3a5f92010-03-01 16:10:40 -0800351 unsigned int op_silent:1;
Andy Grover15133f62010-01-12 14:33:38 -0800352 unsigned int op_active:1;
Andy Grover15133f62010-01-12 14:33:38 -0800353 struct scatterlist *op_sg;
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800354 struct rds_notifier *op_notifier;
Andy Grover15133f62010-01-12 14:33:38 -0800355
356 struct rds_mr *op_rdma_mr;
357 } atomic;
358 struct rm_rdma_op {
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800359 u32 op_rkey;
360 u64 op_remote_addr;
361 unsigned int op_write:1;
362 unsigned int op_fence:1;
363 unsigned int op_notify:1;
364 unsigned int op_recverr:1;
365 unsigned int op_mapped:1;
Andy Grover2c3a5f92010-03-01 16:10:40 -0800366 unsigned int op_silent:1;
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800367 unsigned int op_active:1;
368 unsigned int op_bytes;
369 unsigned int op_nents;
370 unsigned int op_count;
371 struct scatterlist *op_sg;
372 struct rds_notifier *op_notifier;
373
374 struct rds_mr *op_rdma_mr;
Andy Grovere7791372010-01-12 12:15:02 -0800375 } rdma;
Andy Grover15133f62010-01-12 14:33:38 -0800376 struct rm_data_op {
Andy Grover241eef32010-01-19 21:25:26 -0800377 unsigned int op_active:1;
Andy Grover6c7cc6e2010-01-27 18:04:18 -0800378 unsigned int op_nents;
379 unsigned int op_count;
Wengang Wangd655a9f2015-05-21 13:11:40 +0800380 unsigned int op_dmasg;
381 unsigned int op_dmaoff;
Andy Grover6c7cc6e2010-01-27 18:04:18 -0800382 struct scatterlist *op_sg;
Andy Grovere7791372010-01-12 12:15:02 -0800383 } data;
384 };
Andy Grover39de8282009-02-24 15:30:19 +0000385};
386
387/*
388 * The RDS notifier is used (optionally) to tell the application about
389 * completed RDMA operations. Rather than keeping the whole rds message
390 * around on the queue, we allocate a small notifier that is put on the
391 * socket's notifier_list. Notifications are delivered to the application
392 * through control messages.
393 */
394struct rds_notifier {
395 struct list_head n_list;
396 uint64_t n_user_token;
397 int n_status;
398};
399
400/**
401 * struct rds_transport - transport specific behavioural hooks
402 *
403 * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send
404 * part of a message. The caller serializes on the send_sem so this
405 * doesn't need to be reentrant for a given conn. The header must be
406 * sent before the data payload. .xmit must be prepared to send a
407 * message with no data payload. .xmit should return the number of
408 * bytes that were sent down the connection, including header bytes.
409 * Returning 0 tells the caller that it doesn't need to perform any
410 * additional work now. This is usually the case when the transport has
411 * filled the sending queue for its connection and will handle
412 * triggering the rds thread to continue the send when space becomes
413 * available. Returning -EAGAIN tells the caller to retry the send
414 * immediately. Returning -ENOMEM tells the caller to retry the send at
415 * some point in the future.
416 *
417 * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once
418 * it returns the connection can not call rds_recv_incoming().
419 * This will only be called once after conn_connect returns
420 * non-zero success and will The caller serializes this with
421 * the send and connecting paths (xmit_* and conn_*). The
422 * transport is responsible for other serialization, including
423 * rds_recv_incoming(). This is called in process context but
424 * should try hard not to block.
Andy Grover39de8282009-02-24 15:30:19 +0000425 */
426
427struct rds_transport {
428 char t_name[TRANSNAMSIZ];
429 struct list_head t_item;
430 struct module *t_owner;
431 unsigned int t_prefer_loopback:1;
Andy Grover335776b2009-08-21 12:28:34 +0000432 unsigned int t_type;
Andy Grover39de8282009-02-24 15:30:19 +0000433
Sowmini Varadhand5a8ac22015-08-05 01:43:25 -0400434 int (*laddr_check)(struct net *net, __be32 addr);
Andy Grover39de8282009-02-24 15:30:19 +0000435 int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp);
436 void (*conn_free)(void *data);
437 int (*conn_connect)(struct rds_connection *conn);
438 void (*conn_shutdown)(struct rds_connection *conn);
439 void (*xmit_prepare)(struct rds_connection *conn);
440 void (*xmit_complete)(struct rds_connection *conn);
441 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
442 unsigned int hdr_off, unsigned int sg, unsigned int off);
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800443 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
Andy Groverff3d7d32010-03-01 14:03:09 -0800444 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
Andy Grover39de8282009-02-24 15:30:19 +0000445 int (*recv)(struct rds_connection *conn);
Al Viroc310e722014-11-20 09:21:14 -0500446 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to);
Andy Grover39de8282009-02-24 15:30:19 +0000447 void (*inc_free)(struct rds_incoming *inc);
448
449 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
450 struct rdma_cm_event *event);
451 int (*cm_initiate_connect)(struct rdma_cm_id *cm_id);
452 void (*cm_connect_complete)(struct rds_connection *conn,
453 struct rdma_cm_event *event);
454
455 unsigned int (*stats_info_copy)(struct rds_info_iterator *iter,
456 unsigned int avail);
457 void (*exit)(void);
458 void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
459 struct rds_sock *rs, u32 *key_ret);
460 void (*sync_mr)(void *trans_private, int direction);
461 void (*free_mr)(void *trans_private, int invalidate);
462 void (*flush_mrs)(void);
463};
464
465struct rds_sock {
466 struct sock rs_sk;
467
468 u64 rs_user_addr;
469 u64 rs_user_bytes;
470
471 /*
472 * bound_addr used for both incoming and outgoing, no INADDR_ANY
473 * support.
474 */
Chris Mason38a4e5e2010-05-11 15:09:45 -0700475 struct hlist_node rs_bound_node;
Andy Grover39de8282009-02-24 15:30:19 +0000476 __be32 rs_bound_addr;
477 __be32 rs_conn_addr;
478 __be16 rs_bound_port;
479 __be16 rs_conn_port;
Andy Grover39de8282009-02-24 15:30:19 +0000480 struct rds_transport *rs_transport;
481
482 /*
483 * rds_sendmsg caches the conn it used the last time around.
484 * This helps avoid costly lookups.
485 */
486 struct rds_connection *rs_conn;
487
488 /* flag indicating we were congested or not */
489 int rs_congested;
Andy Groverb98ba522010-03-11 13:50:04 +0000490 /* seen congestion (ENOBUFS) when sending? */
491 int rs_seen_congestion;
Andy Grover39de8282009-02-24 15:30:19 +0000492
493 /* rs_lock protects all these adjacent members before the newline */
494 spinlock_t rs_lock;
495 struct list_head rs_send_queue;
496 u32 rs_snd_bytes;
497 int rs_rcv_bytes;
498 struct list_head rs_notify_queue; /* currently used for failed RDMAs */
499
500 /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask
501 * to decide whether the application should be woken up.
502 * If not set, we use rs_cong_track to find out whether a cong map
503 * update arrived.
504 */
505 uint64_t rs_cong_mask;
506 uint64_t rs_cong_notify;
507 struct list_head rs_cong_list;
508 unsigned long rs_cong_track;
509
510 /*
511 * rs_recv_lock protects the receive queue, and is
512 * used to serialize with rds_release.
513 */
514 rwlock_t rs_recv_lock;
515 struct list_head rs_recv_queue;
516
517 /* just for stats reporting */
518 struct list_head rs_item;
519
520 /* these have their own lock */
521 spinlock_t rs_rdma_lock;
522 struct rb_root rs_rdma_keys;
523
524 /* Socket options - in case there will be more */
525 unsigned char rs_recverr,
526 rs_cong_monitor;
527};
528
529static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk)
530{
531 return container_of(sk, struct rds_sock, rs_sk);
532}
533static inline struct sock *rds_rs_to_sk(struct rds_sock *rs)
534{
535 return &rs->rs_sk;
536}
537
538/*
539 * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value
540 * to account for overhead. We don't account for overhead, we just apply
541 * the number of payload bytes to the specified value.
542 */
543static inline int rds_sk_sndbuf(struct rds_sock *rs)
544{
545 return rds_rs_to_sk(rs)->sk_sndbuf / 2;
546}
547static inline int rds_sk_rcvbuf(struct rds_sock *rs)
548{
549 return rds_rs_to_sk(rs)->sk_rcvbuf / 2;
550}
551
552struct rds_statistics {
553 uint64_t s_conn_reset;
554 uint64_t s_recv_drop_bad_checksum;
555 uint64_t s_recv_drop_old_seq;
556 uint64_t s_recv_drop_no_sock;
557 uint64_t s_recv_drop_dead_sock;
558 uint64_t s_recv_deliver_raced;
559 uint64_t s_recv_delivered;
560 uint64_t s_recv_queued;
561 uint64_t s_recv_immediate_retry;
562 uint64_t s_recv_delayed_retry;
563 uint64_t s_recv_ack_required;
564 uint64_t s_recv_rdma_bytes;
565 uint64_t s_recv_ping;
566 uint64_t s_send_queue_empty;
567 uint64_t s_send_queue_full;
Andy Grover049ee3f2010-03-23 17:39:07 -0700568 uint64_t s_send_lock_contention;
569 uint64_t s_send_lock_queue_raced;
Andy Grover39de8282009-02-24 15:30:19 +0000570 uint64_t s_send_immediate_retry;
571 uint64_t s_send_delayed_retry;
572 uint64_t s_send_drop_acked;
573 uint64_t s_send_ack_required;
574 uint64_t s_send_queued;
575 uint64_t s_send_rdma;
576 uint64_t s_send_rdma_bytes;
577 uint64_t s_send_pong;
578 uint64_t s_page_remainder_hit;
579 uint64_t s_page_remainder_miss;
580 uint64_t s_copy_to_user;
581 uint64_t s_copy_from_user;
582 uint64_t s_cong_update_queued;
583 uint64_t s_cong_update_received;
584 uint64_t s_cong_send_error;
585 uint64_t s_cong_send_blocked;
586};
587
588/* af_rds.c */
589void rds_sock_addref(struct rds_sock *rs);
590void rds_sock_put(struct rds_sock *rs);
591void rds_wake_sk_sleep(struct rds_sock *rs);
592static inline void __rds_wake_sk_sleep(struct sock *sk)
593{
Eric Dumazetaa395142010-04-20 13:03:51 +0000594 wait_queue_head_t *waitq = sk_sleep(sk);
Andy Grover39de8282009-02-24 15:30:19 +0000595
596 if (!sock_flag(sk, SOCK_DEAD) && waitq)
597 wake_up(waitq);
598}
599extern wait_queue_head_t rds_poll_waitq;
600
601
602/* bind.c */
603int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
604void rds_remove_bound(struct rds_sock *rs);
605struct rds_sock *rds_find_bound(__be32 addr, __be16 port);
606
607/* cong.c */
608int rds_cong_get_maps(struct rds_connection *conn);
609void rds_cong_add_conn(struct rds_connection *conn);
610void rds_cong_remove_conn(struct rds_connection *conn);
611void rds_cong_set_bit(struct rds_cong_map *map, __be16 port);
612void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port);
613int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs);
614void rds_cong_queue_updates(struct rds_cong_map *map);
615void rds_cong_map_updated(struct rds_cong_map *map, uint64_t);
616int rds_cong_updated_since(unsigned long *recent);
617void rds_cong_add_socket(struct rds_sock *);
618void rds_cong_remove_socket(struct rds_sock *);
619void rds_cong_exit(void);
620struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
621
622/* conn.c */
Zach Brownef87b7e2010-07-09 12:26:20 -0700623int rds_conn_init(void);
Andy Grover39de8282009-02-24 15:30:19 +0000624void rds_conn_exit(void);
Sowmini Varadhand5a8ac22015-08-05 01:43:25 -0400625struct rds_connection *rds_conn_create(struct net *net,
626 __be32 laddr, __be32 faddr,
Andy Grover39de8282009-02-24 15:30:19 +0000627 struct rds_transport *trans, gfp_t gfp);
Sowmini Varadhand5a8ac22015-08-05 01:43:25 -0400628struct rds_connection *rds_conn_create_outgoing(struct net *net,
629 __be32 laddr, __be32 faddr,
Andy Grover39de8282009-02-24 15:30:19 +0000630 struct rds_transport *trans, gfp_t gfp);
Andy Grover2dc39352010-06-11 13:49:13 -0700631void rds_conn_shutdown(struct rds_connection *conn);
Andy Grover39de8282009-02-24 15:30:19 +0000632void rds_conn_destroy(struct rds_connection *conn);
Andy Grover39de8282009-02-24 15:30:19 +0000633void rds_conn_drop(struct rds_connection *conn);
Zach Brownf3c68082010-05-24 13:14:36 -0700634void rds_conn_connect_if_down(struct rds_connection *conn);
Andy Grover39de8282009-02-24 15:30:19 +0000635void rds_for_each_conn_info(struct socket *sock, unsigned int len,
636 struct rds_info_iterator *iter,
637 struct rds_info_lengths *lens,
638 int (*visitor)(struct rds_connection *, void *),
639 size_t item_len);
Joe Perchesb9075fa2011-10-31 17:11:33 -0700640__printf(2, 3)
641void __rds_conn_error(struct rds_connection *conn, const char *, ...);
Andy Grover39de8282009-02-24 15:30:19 +0000642#define rds_conn_error(conn, fmt...) \
643 __rds_conn_error(conn, KERN_WARNING "RDS: " fmt)
644
645static inline int
646rds_conn_transition(struct rds_connection *conn, int old, int new)
647{
648 return atomic_cmpxchg(&conn->c_state, old, new) == old;
649}
650
651static inline int
652rds_conn_state(struct rds_connection *conn)
653{
654 return atomic_read(&conn->c_state);
655}
656
657static inline int
658rds_conn_up(struct rds_connection *conn)
659{
660 return atomic_read(&conn->c_state) == RDS_CONN_UP;
661}
662
663static inline int
664rds_conn_connecting(struct rds_connection *conn)
665{
666 return atomic_read(&conn->c_state) == RDS_CONN_CONNECTING;
667}
668
669/* message.c */
670struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
Andy Groverfc445082010-01-12 12:56:06 -0800671struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
Al Viro083735f2014-11-20 09:31:08 -0500672int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from);
Andy Grover39de8282009-02-24 15:30:19 +0000673struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
674void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
675 __be16 dport, u64 seq);
676int rds_message_add_extension(struct rds_header *hdr,
677 unsigned int type, const void *data, unsigned int len);
678int rds_message_next_extension(struct rds_header *hdr,
679 unsigned int *pos, void *buf, unsigned int *buflen);
Andy Grover39de8282009-02-24 15:30:19 +0000680int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
Al Viroc310e722014-11-20 09:21:14 -0500681int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
Andy Grover39de8282009-02-24 15:30:19 +0000682void rds_message_inc_free(struct rds_incoming *inc);
683void rds_message_addref(struct rds_message *rm);
684void rds_message_put(struct rds_message *rm);
685void rds_message_wait(struct rds_message *rm);
686void rds_message_unmapped(struct rds_message *rm);
687
688static inline void rds_message_make_checksum(struct rds_header *hdr)
689{
690 hdr->h_csum = 0;
691 hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2);
692}
693
694static inline int rds_message_verify_checksum(const struct rds_header *hdr)
695{
696 return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0;
697}
698
699
700/* page.c */
701int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
702 gfp_t gfp);
703int rds_page_copy_user(struct page *page, unsigned long offset,
704 void __user *ptr, unsigned long bytes,
705 int to_user);
706#define rds_page_copy_to_user(page, offset, ptr, bytes) \
707 rds_page_copy_user(page, offset, ptr, bytes, 1)
708#define rds_page_copy_from_user(page, offset, ptr, bytes) \
709 rds_page_copy_user(page, offset, ptr, bytes, 0)
710void rds_page_exit(void);
711
712/* recv.c */
713void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
714 __be32 saddr);
Andy Grover39de8282009-02-24 15:30:19 +0000715void rds_inc_put(struct rds_incoming *inc);
716void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
Cong Wang6114eab2011-11-25 23:14:40 +0800717 struct rds_incoming *inc, gfp_t gfp);
Ying Xue1b784142015-03-02 15:37:48 +0800718int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
719 int msg_flags);
Andy Grover39de8282009-02-24 15:30:19 +0000720void rds_clear_recv_queue(struct rds_sock *rs);
721int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg);
722void rds_inc_info_copy(struct rds_incoming *inc,
723 struct rds_info_iterator *iter,
724 __be32 saddr, __be32 daddr, int flip);
725
726/* send.c */
Ying Xue1b784142015-03-02 15:37:48 +0800727int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len);
Andy Grover39de8282009-02-24 15:30:19 +0000728void rds_send_reset(struct rds_connection *conn);
729int rds_send_xmit(struct rds_connection *conn);
730struct sockaddr_in;
731void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
732typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
733void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
734 is_acked_func is_acked);
Andy Grover39de8282009-02-24 15:30:19 +0000735int rds_send_pong(struct rds_connection *conn, __be16 dport);
736struct rds_message *rds_send_get_message(struct rds_connection *,
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800737 struct rm_rdma_op *);
Andy Grover39de8282009-02-24 15:30:19 +0000738
739/* rdma.c */
740void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
Andy Grover21f79af2010-01-12 12:57:27 -0800741int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
742int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
743int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
744void rds_rdma_drop_keys(struct rds_sock *rs);
745int rds_rdma_extra_size(struct rds_rdma_args *args);
746int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
747 struct cmsghdr *cmsg);
748int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
749 struct cmsghdr *cmsg);
750int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
751 struct cmsghdr *cmsg);
752int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
753 struct cmsghdr *cmsg);
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800754void rds_rdma_free_op(struct rm_rdma_op *ro);
Andy Groverd0ab25a2010-01-27 16:15:48 -0800755void rds_atomic_free_op(struct rm_atomic_op *ao);
Andy Grover15133f62010-01-12 14:33:38 -0800756void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
757void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
758int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
759 struct cmsghdr *cmsg);
Andy Grover21f79af2010-01-12 12:57:27 -0800760
Joe Perchesc1b12032013-10-18 13:48:25 -0700761void __rds_put_mr_final(struct rds_mr *mr);
Andy Grover21f79af2010-01-12 12:57:27 -0800762static inline void rds_mr_put(struct rds_mr *mr)
763{
764 if (atomic_dec_and_test(&mr->r_refcount))
765 __rds_put_mr_final(mr);
766}
Andy Grover39de8282009-02-24 15:30:19 +0000767
768/* stats.c */
David Howells9b8de742009-04-21 23:00:24 +0100769DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
Andy Grover39de8282009-02-24 15:30:19 +0000770#define rds_stats_inc_which(which, member) do { \
771 per_cpu(which, get_cpu()).member++; \
772 put_cpu(); \
773} while (0)
774#define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member)
775#define rds_stats_add_which(which, member, count) do { \
776 per_cpu(which, get_cpu()).member += count; \
777 put_cpu(); \
778} while (0)
779#define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
Zach Brownef87b7e2010-07-09 12:26:20 -0700780int rds_stats_init(void);
Andy Grover39de8282009-02-24 15:30:19 +0000781void rds_stats_exit(void);
782void rds_stats_info_copy(struct rds_info_iterator *iter,
Jan Engelhardt36cbd3d2009-08-05 10:42:58 -0700783 uint64_t *values, const char *const *names,
784 size_t nr);
Andy Grover39de8282009-02-24 15:30:19 +0000785
786/* sysctl.c */
Zach Brownef87b7e2010-07-09 12:26:20 -0700787int rds_sysctl_init(void);
Andy Grover39de8282009-02-24 15:30:19 +0000788void rds_sysctl_exit(void);
789extern unsigned long rds_sysctl_sndbuf_min;
790extern unsigned long rds_sysctl_sndbuf_default;
791extern unsigned long rds_sysctl_sndbuf_max;
792extern unsigned long rds_sysctl_reconnect_min_jiffies;
793extern unsigned long rds_sysctl_reconnect_max_jiffies;
794extern unsigned int rds_sysctl_max_unacked_packets;
795extern unsigned int rds_sysctl_max_unacked_bytes;
796extern unsigned int rds_sysctl_ping_enable;
797extern unsigned long rds_sysctl_trace_flags;
798extern unsigned int rds_sysctl_trace_level;
799
800/* threads.c */
Zach Brownef87b7e2010-07-09 12:26:20 -0700801int rds_threads_init(void);
Andy Grover39de8282009-02-24 15:30:19 +0000802void rds_threads_exit(void);
803extern struct workqueue_struct *rds_wq;
Andy Grover2dc39352010-06-11 13:49:13 -0700804void rds_queue_reconnect(struct rds_connection *conn);
Andy Grover39de8282009-02-24 15:30:19 +0000805void rds_connect_worker(struct work_struct *);
806void rds_shutdown_worker(struct work_struct *);
807void rds_send_worker(struct work_struct *);
808void rds_recv_worker(struct work_struct *);
809void rds_connect_complete(struct rds_connection *conn);
810
811/* transport.c */
812int rds_trans_register(struct rds_transport *trans);
813void rds_trans_unregister(struct rds_transport *trans);
Sowmini Varadhand5a8ac22015-08-05 01:43:25 -0400814struct rds_transport *rds_trans_get_preferred(struct net *net, __be32 addr);
Zach Brown5adb5bc2010-07-23 10:32:31 -0700815void rds_trans_put(struct rds_transport *trans);
Andy Grover39de8282009-02-24 15:30:19 +0000816unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
817 unsigned int avail);
Sowmini Varadhand97dac52015-05-29 17:28:08 -0400818struct rds_transport *rds_trans_get(int t_type);
Zach Brownef87b7e2010-07-09 12:26:20 -0700819int rds_trans_init(void);
Andy Grover39de8282009-02-24 15:30:19 +0000820void rds_trans_exit(void);
821
822#endif